Artificial Intelligence in Capsule Endoscopy: A Gamechanger for a Groundbreaking Technique 0323996477, 9780323996471

Citation preview

Artificial Intelligence in Capsule Endoscopy A Gamechanger for a Groundbreaking Technique

This page intentionally left blank

Artificial Intelligence in Capsule Endoscopy A Gamechanger for a Groundbreaking Technique Edited by

Miguel Mascarenhas Faculty of Medicine, University of Porto, Porto, Portugal; Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal

He´lder Cardoso Faculty of Medicine, University of Porto, Porto, Portugal; Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal; World Gastroenterology Organization Porto Training Center, Porto, Portugal

Guilherme Macedo Faculty of Medicine, University of Porto, Porto, Portugal; Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal; World Gastroenterology Organization Porto Training Center, Porto, Portugal

Academic Press is an imprint of Elsevier 125 London Wall, London EC2Y 5AS, United Kingdom 525 B Street, Suite 1650, San Diego, CA 92101, United States 50 Hampshire Street, 5th Floor, Cambridge, MA 02139, United States The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, United Kingdom Copyright © 2023 Elsevier Inc. All rights reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, recording, or any information storage and retrieval system, without permission in writing from the publisher. Details on how to seek permission, further information about the Publisher’s permissions policies and our arrangements with organizations such as the Copyright Clearance Center and the Copyright Licensing Agency, can be found at our website: www.elsevier.com/permissions. This book and the individual contributions contained in it are protected under copyright by the Publisher (other than as may be noted herein). Notices Knowledge and best practice in this field are constantly changing. As new research and experience broaden our understanding, changes in research methods, professional practices, or medical treatment may become necessary. Practitioners and researchers must always rely on their own experience and knowledge in evaluating and using any information, methods, compounds, or experiments described herein. In using such information or methods they should be mindful of their own safety and the safety of others, including parties for whom they have a professional responsibility. To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors, assume any liability for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions, or ideas contained in the material herein. ISBN: 978-0-323-99647-1 For Information on all Academic Press publications visit our website at https://www.elsevier.com/books-and-journals

Publisher: Stacy Masucci Editorial Project Manager: Timothy J. Bennett Production Project Manager: Sajana Devasi P K Cover Designer: Greg Harris Typeset by MPS Limited, Chennai, India

Contents List of contributors ................................................................................................ xiii Preface .................................................................................................................. xvii Acknowledgments ..................................................................................................xix

CHAPTER 1 Artificial intelligence: machine learning, deep learning, and applications in gastrointestinal endoscopy ..................................................................... 1 Joa˜o Afonso, Miguel Martins, Joa˜o Ferreira and Miguel Mascarenhas Definition of artificial intelligence................................................ 1 Machine learning versus deep learning......................................... 2 Machine learning .......................................................................2 Deep learning.............................................................................2 Examples of artificial intelligence applicability ........................... 3 Online experience .......................................................................... 4 Robotics ......................................................................................... 4 Vehicles ......................................................................................... 5 Fake news detection and cybersecurity......................................... 5 Artificial intelligence in healthcare as a facilitating technology...................................................................................... 5 Artificial intelligence in medicine ................................................ 5 Capsule endoscopy: a brief introduction....................................... 7 References...................................................................................... 8

CHAPTER 2 Wireless capsule endoscopy: concept and modalities.................................................................... 11 Pablo Cortegoso Valdivia and Marco Pennazio Background .................................................................................. 11 Types of capsules ........................................................................ 11 Indications.................................................................................... 12 Suspected small bowel bleeding .............................................12 Small bowel tumors.................................................................14 Hereditary polyposis syndromes .............................................14 Celiac disease ..........................................................................15 Crohn’s disease........................................................................15 Colon examination...................................................................15 Future perspectives ...................................................................... 16 References.................................................................................... 16

v

vi

Contents

CHAPTER 3 Capsule endoscopy: wide clinical scope .................. 21 Pilar Esteban Delgado, Renato Medas, Eunice Trindade and Enrique Pe´rez-Cuadrado Martı´nez Body............................................................................................. 21 Indications in capsule endoscopy................................................ 22 Suspected middle digestive hemorrhage: obscure gastrointestinal bleeding ..........................................................22 Crohn’s disease........................................................................25 Evaluation of refractory celiac disease ...................................30 Screening of polyposis syndromes: familial adenomatous polyposis and Peutz Jeghers syndrome...........31 Suspected small intestine tumors ............................................33 Graft versus host disease .........................................................33 Capsule endoscopy clinical scope in pediatrics.......................... 36 Indications................................................................................37 Occult gastrointestinal bleeding ..............................................37 Inflammatory bowel disease....................................................38 Polyposis syndromes ...............................................................39 Other indications......................................................................40 Limitations of endoscopic capsule .............................................. 40 Challenges in pediatrics...........................................................40 Bowel cleansing.......................................................................41 Capsule retention .....................................................................42 Conclusions.................................................................................. 46 References.................................................................................... 46

CHAPTER 4 The role of capsule endoscopy in diagnosis and clinical management of obscure gastrointestinal bleeding....................................................................... 53 Nayantara Coelho-Prabhu, Shabana F. Pasha and Jonathan Leighton Introduction.................................................................................. 53 Suspected small bowel bleeding ................................................. 54 Timing of capsule endoscopy...................................................... 58 Contraindications and complications of capsule endoscopy .................................................................................... 59 Advanced technologies in capsules............................................. 61 Artificial intelligence in capsule endoscopy............................... 61 References.................................................................................... 62

Contents

CHAPTER 5 The role of capsule endoscopy in diagnosis and clinical management of inflammatory bowel disease ............................................................. 69 Isabel Garrido, Patrı´cia Andrade, Susana Lopes and Guilherme Macedo Introduction.................................................................................. 69 Crohn’s disease............................................................................ 70 Ulcerative colitis.......................................................................... 70 Capsule endoscopy in suspected Crohn’s disease ...................... 72 Capsule endoscopy in patients with established Crohn’s disease............................................................................ 75 Assessment of postoperative recurrence ..................................... 76 Role of capsule endoscopy in reclassification of inflammatory bowel disease ............................................................................... 77 Colon capsule endoscopy ............................................................ 78 Colon capsule endoscopy in Crohn’s disease ............................. 79 Colon capsule endoscopy in ulcerative colitis............................ 81 Cost-effectiveness of colon capsule endoscopy in inflammatory bowel disease........................................................ 83 Complications of capsule endoscopy .......................................... 84 New research areas for future ..................................................... 85 Conclusion ................................................................................... 86 References.................................................................................... 86

CHAPTER 6 Artificial intelligence for automatic detection of blood and hematic residues ....................................... 91 Gerardo Blanco Sr, Oscar Mondragon and Omar Solo´rzano Artificial intelligence................................................................... 92 Support vector machines ............................................................. 94 Artificial neural network ............................................................. 94 Convolutional neural network ..................................................... 94 ESNavi .....................................................................................96 SSD 1 ResNet50......................................................................96 Inception-Resnet-V2................................................................96 Recent outcomes of artificial intelligence in detecting active bleeding and hematic residues.......................................... 97 Acknowledgments ....................................................................... 97 References.................................................................................... 97

vii

viii

Contents

CHAPTER 7 Artificial intelligence in capsule endoscopy for detection of ulcers and erosions ............................. 101 Shabana F. Pasha and Jean-Christophe Saurin Introduction................................................................................ 101 Capsule endoscopes and current challenges ............................. 102 Capsule endoscopes (currently available and in process of development)........................................................102 Current challenges in capsule endoscopy .............................103 Capsule endoscopy scoring systems for small bowel inflammation.............................................................................. 104 Lewis score ............................................................................105 Capsule endoscopy Crohn’s disease activity index ..............106 Capsule endoscopy software enhancements to improve detection of inflammatory lesions............................................. 106 Image enhanced endoscopy...................................................106 Artificial intelligence and its application in capsule endoscopy .................................................................................. 108 Artificial intelligence for detection of small bowel ulcerations and erosions ............................................................ 109 Automatic detection of ulcers and erosions.............................. 110 Grading of ulcers and erosions severity.................................... 114 Artificial intelligence in next-generation capsule endoscopes ................................................................................. 114 Conclusions................................................................................ 115 References.................................................................................. 115 Further reading .......................................................................... 119

CHAPTER 8 Artificial intelligence for protruding lesions ........... 121 Xavier Dray, Aymeric Histace, Alexander Robertson and Santi Segui Introduction................................................................................ 121 State-of-the-art technological aspects ....................................... 123 State-of-the-art clinical aspects ................................................. 123 Esophagus ..............................................................................123 Stomach .................................................................................125 Small bowel ...........................................................................130 Colon......................................................................................141 Perspectives on challenges and developments.......................... 142 Conclusion ................................................................................. 143 Conflict of interest..................................................................... 143 References.................................................................................. 144

Contents

CHAPTER 9 Artificial intelligence for vascular lesions.............. 149 Pere Gilabert, Pablo Laiz and Santi Seguı´ Introduction................................................................................ 149 Wireless capsule endoscopy and artificial intelligence ............ 149 Vascular lesions in gastrointestinal tract...............................150 Datasets ...................................................................................... 151 KIDs dataset...........................................................................152 Red lesion endoscopy dataset................................................152 CAD CAP 2020 ...................................................................152 Kvasir Capsule.....................................................................154 Artificial intelligence methods for vascular lesions ................. 154 Conclusions................................................................................ 159 References.................................................................................. 160

CHAPTER 10 Artificial intelligence for luminal content analysis and miscellaneous findings....................... 163 Nuno Almeida and Pedro Figueiredo Introduction................................................................................ 163 Small bowel preparation and luminal content .......................... 164 Lymphangiectasia and other miscellaneous findings ............... 170 Hookworms and foreign bodies ................................................ 172 Discussion and conclusions....................................................... 175 Acknowledgments ..................................................................... 175 Disclosures/transparency declaration ........................................ 176 References.................................................................................. 176

CHAPTER 11 Small bowel and colon cleansing in capsule endoscopy ................................................................. 181 Vı´tor Macedo Silva, Bruno Rosa, Francisco Mendes, Miguel Mascarenhas, Miguel Mascarenhas Saraiva and Jose´ Cotter Introduction................................................................................ 181 Small bowel capsule endoscopy preparation ............................ 181 Diet and fasting .....................................................................182 Oral purgatives ......................................................................182 Prokinetic drugs .....................................................................184 Antifoaming agents................................................................184 Water ingestion......................................................................185 Colon capsule endoscopy preparation....................................... 185 Diet and fasting .....................................................................185 Oral purgatives ......................................................................186

ix

x

Contents

Boosters..................................................................................186 Prokinetic drugs .....................................................................187 Small bowel capsule endoscopy cleansing quality evaluation................................................................................... 187 Automated scores ..................................................................187 Operator-dependent scores ....................................................188 Colon capsule endoscopy cleansing quality evaluation ........... 190 Final remarks ............................................................................. 193 References.................................................................................. 194

CHAPTER 12 Introducing blockchain technology in data storage to foster big data and artificial intelligence applications in healthcare systems ......... 199 Miguel Mascarenhas, Andre´ Santos and Guilherme Macedo Introduction................................................................................ 199 A brief picture of present-day medical challenges ................... 200 Emergence of blockchain in healthcare .................................... 202 Blockchain and its utility for big data and artificial intelligence in healthcare.......................................................203 Blockchain and use of big data and artificial intelligence in imaging ..........................................................205 Growing field of artificial intelligence applied to capsule endoscopy ..................................................................... 206 Limitations and challenges to applications of blockchain in healthcare............................................................ 208 Advantages of using blockchain in capsule endoscopy: how it can be enhanced with artificial intelligence tools......... 210 Concluding remarks................................................................... 210 Acknowledgments ..................................................................... 211 Conflicts of interest ................................................................... 211 References.................................................................................. 211

CHAPTER 13 Magnetic capsule endoscopy: concept and application of artificial intelligence ........................ 217 Chen He, Qiwen Wang, Xi Jiang, Bin Jiang, Yang-Yang Qian, Jun Pan and Zhuan Liao Types of magnetic capsule endoscopy and their feasibility.......... 217 Hand-held magnetic capsule endoscopy ...............................217 Magnetic resonance imaging-based magnetic capsule endoscopy ..............................................................................218 Robotic magnetic capsule endoscopy ...................................220

Contents

Operation procedure, indications, and contradictions of magnetic capsule endoscopy ..................................................... 223 Operation procedure of gastric examination.........................226 Indications and contradictions...............................................227 Overview of artificial intelligence and its integration into gastrointestinal practice ............................................................. 228 Artificial intelligence: Definition and role in technology enhancement .......................................................230 Development and validation of artificial intelligence systems in gastrointestinal practice.......................................230 Current artificial intelligence applications in magnetic capsule endoscopy ..................................................................... 231 Artificial intelligence-assisted magnetic capsule endoscopy localization strategy ............................................231 Artificial intelligence-assisted magnetic capsule endoscopy diagnostic procedure ...........................................233 Prospects of artificial intelligence in magnetic capsule endoscopy .................................................................................. 237 References.................................................................................. 238

CHAPTER 14 Nonwhite light endoscopy in capsule endoscopy: Fujinon Intelligent Chromo Endoscopy and blue mode......................................................................... 243 Catarina Gomes, Emanuel Dias and Rolando Pinho Background ................................................................................ 243 White light ................................................................................. 243 Virtual chromoendoscopy in capsule endoscopy...................... 244 Fujinon Intelligent Chromo Endoscopy system....................244 Blue mode..............................................................................246 Narrow band imaging ............................................................246 Evidence of virtual chromoendoscopy in capsule endoscopy .................................................................................. 247 Fujinon Intelligent Chromo Endoscopy ................................247 Blue mode..............................................................................248 Fujinon Intelligent Chromo Endoscopy and blue mode ..............................................................................248 Narrow band imaging ............................................................248 Other virtual chromoendoscopy methods .............................249 Conclusion ................................................................................. 251 References.................................................................................. 251

xi

xii

Contents

CHAPTER 15 Colon capsule endoscopy and artificial intelligence: a perfect match for panendoscopy ........................................................... 255 Tiago Ribeiro, Ignacio Ferna´ndez-Urien and He´lder Cardoso Introduction................................................................................ 255 Principles of colon capsule endoscopy .................................255 Indications for colon capsule endoscopy/panendoscopy................ 256 Colorectal cancer screening ..................................................256 Inflammatory bowel disease..................................................259 Gastrointestinal bleeding and anemia ...................................261 Limitations of colon capsule endoscopy................................... 261 Impact of artificial intelligence ................................................. 262 Future directions ........................................................................ 264 Conclusion ................................................................................. 265 References.................................................................................. 265 Index ......................................................................................................................271

List of contributors Joa˜o Afonso Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal Nuno Almeida Gastroenterology Department, Coimbra Hospital and Universitary Center, Coimbra, Portugal; Faculty of Medicine, University of Coimbra, Coimbra, Portugal Patrı´cia Andrade Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal; World Gastroenterology Organization Porto Training Center, Porto, Portugal; Faculty of Medicine, University of Porto, Porto, Portugal Gerardo Blanco Sr Department of Endoscopy, Hospital de Especialidades, Centro Me´dico Nacional Siglo XXI, Instituto Mexicano del Seguro Social, Mexico City, Mexico He´lder Cardoso Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal; Faculty of Medicine, University of Porto, Porto, Portugal; World Gastroenterology Organization Porto Training Center, Porto, Portugal Nayantara Coelho-Prabhu Mayo Clinic, Rochester, MN, United States Pablo Cortegoso Valdivia Gastroenterology and Endoscopy Unit, University Hospital of Parma, University of Parma, Parma, Italy Jose´ Cotter Gastroenterology Department, Hospital da Senhora da Oliveira, Guimara˜es, Portugal; Life and Health Sciences Research Institute (ICVS), School of Medicine, University of Minho, Braga, Portugal; ICVS/3B’s, PT Government Associate Laboratory, Braga/Guimara˜es, Portugal Pilar Esteban Delgado Small Bowel Unit, Endoscopy Department, University Hospital Morales Meseguer, Murcia, Spain Emanuel Dias Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal Xavier Dray Center for Digestive Endoscopy, Saint Antoine Hospital, APHP, Sorbonne University, Paris, France; ENSEA, CNRS, ETIS UMR 8051, CY Cergy Paris University, Cergy, France Ignacio Ferna´ndez-Urien Navarra Hospital Complex, Pamplona, Spain

xiii

xiv

List of contributors

Joa˜o Ferreira Mechanical Engineering Department, Faculty of Engineering of the University of Porto, Porto, Portugal; Institute of Science and Innovation in Mechanical and Industrial Engineering (INEGI), Porto, Portugal Pedro Figueiredo Gastroenterology Department, Coimbra Hospital and Universitary Center, Coimbra, Portugal; Faculty of Medicine, University of Coimbra, Coimbra, Portugal Isabel Garrido Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal; World Gastroenterology Organization Porto Training Center, Porto, Portugal Pere Gilabert Department of Mathematics and Computer Science, University of Barcelona, Barcelona, Spain Catarina Gomes Gastroenterology Department, Centro Hospitalar Vila Nova de Gaia/Espinho, Vila Nova de Gaia, Porto, Portugal Chen He Department of Gastroenterology, Changhai Hospital, Naval Medical University, Shanghai, P.R. China Aymeric Histace ENSEA, CNRS, ETIS UMR 8051, CY Cergy Paris University, Cergy, France Bin Jiang Department of Gastroenterology, The First Naval Hospital of Southern Theater Command, Guangdong, P.R. China Xi Jiang Department of Gastroenterology, The First Naval Hospital of Southern Theater Command, Guangdong, P.R. China Pablo Laiz Department of Mathematics and Computer Science, University of Barcelona, Barcelona, Spain Jonathan Leighton Division of Gastroenterology and Hepatology, Mayo Clinic, Scottsdale, AZ, United States Zhuan Liao Department of Gastroenterology, Changhai Hospital, Naval Medical University, Shanghai, P.R. China Susana Lopes Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal; World Gastroenterology Organization Porto Training Center, Porto, Portugal; Faculty of Medicine, University of Porto, Porto, Portugal

List of contributors

Guilherme Macedo Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal; World Gastroenterology Organization Porto Training Center, Porto, Portugal; Faculty of Medicine, University of Porto, Porto, Portugal Enrique Pe´rez-Cuadrado Martı´nez Small Bowel Unit, Endoscopy Department, University Hospital Morales Meseguer, Murcia, Spain Miguel Martins Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal Miguel Mascarenhas Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal; Faculty of Medicine, University of Porto, Porto, Portugal Renato Medas Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal; Faculty of Medicine, University of Porto, Porto, Portugal Francisco Mendes Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal Oscar Mondragon Department of Endoscopy, Hospital de Especialidades, Centro Me´dico Nacional Siglo XXI, Instituto Mexicano del Seguro Social, Mexico City, Mexico Jun Pan Department of Endoscopy, Hospital de Especialidades, Centro Me´dico Nacional Siglo XXI, Instituto Mexicano del Seguro Social, Mexico City, Mexico Shabana F. Pasha Division of Gastroenterology and Hepatology, Mayo Clinic, Scottsdale, AZ, United States Marco Pennazio University Division of Gastroenterology, City of Health and Science University Hospital, University of Turin, Turin, Italy Rolando Pinho Gastroenterology Department, Centro Hospitalar Vila Nova de Gaia/Espinho, Vila Nova de Gaia, Porto, Portugal Yang-Yang Qian Department of Gastroenterology, Changhai Hospital, Naval Medical University, Shanghai, P.R. China Tiago Ribeiro Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal Alexander Robertson Leicester General Hospital, Leicester, United Kingdom

xv

xvi

List of contributors

Bruno Rosa Gastroenterology Department, Hospital da Senhora da Oliveira, Guimara˜es, Portugal; Life and Health Sciences Research Institute (ICVS), School of Medicine, University of Minho, Braga, Portugal; ICVS/3B’s, PT Government Associate Laboratory, Braga/Guimara˜es, Portugal Andre´ Santos Centro Hospitalar do Baixo Vouga, Aveiro, Portugal Miguel Mascarenhas Saraiva Manoph Gastroenterology Clinic, Porto, Portugal Jean-Christophe Saurin Department of Gastroenterology, E. Herriot Hospital, Claude Bernard University, Lyon, France Santi Segui Department of Mathematics and Computer Science, University of Barcelona, Barcelona, Spain Vı´tor Macedo Silva Gastroenterology Department, Hospital da Senhora da Oliveira, Guimara˜es, Portugal; Life and Health Sciences Research Institute (ICVS), School of Medicine, University of Minho, Braga, Portugal; ICVS/3B’s, PT Government Associate Laboratory, Braga/Guimara˜es, Portugal Omar Solo´rzano Department of Endoscopy, Hospital de Especialidades, Centro Me´dico Nacional Siglo XXI, Instituto Mexicano del Seguro Social, Mexico City, Mexico Eunice Trindade Unit of Pediatric Gastroenterology and Nutrition, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal Qiwen Wang Department of Gastroenterology, Changhai Hospital, Naval Medical University, Shanghai, P.R. China

Preface Artificial intelligence (AI) has been subtly and progressively taking part in our daily life. The applicability of heuristic algorithms has revolutionized the way we manage large databases and AI, providing an adequate framework for systematic analysis, which promptly revealed its enormous potential in Gastroenterology and Endoscopy. The excitement of the innovation of a plausible application of advanced systems in our practice generated an overwhelming enthusiasm in our community, and clinicians rapidly dove into a new lexicon, such as convolutional neural network models, deep learning methods, training machines, and computer-aided detection systems. Soon, we all realized that the cross-pollination research with biomedical engineers, informaticians, and clinicians was more than a circumstantial drift of our mindset but an indispensable move toward a new advancing frontier. The exponential development of the usefulness of AI in capsule endoscopy requires consideration of its medium- and long-term impacts on clinical practice. Indeed, the advent of deep learning in gastrointestinal endoscopy, with its evolutionary character, could lead to a paradigm shift in clinical activity in this setting. In this book, we aim to showcase the state of the art of AI in the field of capsule endoscopy, with examples of cutting-edge research being carried out in this field. Measurable improvement in clinical benefits is the ultimate goal of developing this disruptive technology. It seems to us, the three pillars in our Gastroenterology practice—prevention, diagnosis, and treatment of digestive diseases—would ultimately be profoundly affected by the judicious use of AI, touching many of our basic concerns: accuracy, quality assessment, time management, and access to medical care. If AI is still the road not taken by many clinicians, then this is the right time to make all the difference. Guilherme Macedo

xvii

This page intentionally left blank

Acknowledgments Thanks to my family for their unconditional support. I am eternally grateful to Professor Guilherme Macedo, my mentor and catalyst throughout my journey. To Dr. He´lder Cardoso and Dr. Patrı´cia Andrade for their crucial role in my specific training in gastroenterology and their constant friendship. To Professor Joa˜o Ferreira for his friendship, trust, and complicity in the multiple challenges to overcome. To Professor Renato Natal for his enthusiasm and constant support, which were decisive in the start and success of the project. To those, I taught, but from whom I learned much more. Thanks to everyone, in particular, Joa˜o and Tiago, for their constant belief, dynamism, and strength in the most difficult moments. To my patients, the main motivating factor of my entire journey. To the adversities, disagreements, and injustices, which made me more resilient and tenacious in outlining the achieved path. Miguel Mascarenhas Thanks to my beautiful and loving family and to all those who made this book possible. He´lder Cardoso

xix

This page intentionally left blank

CHAPTER

Artificial intelligence: machine learning, deep learning, and applications in gastrointestinal endoscopy

1

Joa˜o Afonso1, Miguel Martins1, Joa˜o Ferreira2,3 and Miguel Mascarenhas1,4 1

Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal 2 Mechanical Engineering Department, Faculty of Engineering of the University of Porto, Porto, Portugal 3 Institute of Science and Innovation in Mechanical and Industrial Engineering (INEGI), Porto, Portugal 4 Faculty of Medicine, University of Porto, Porto, Portugal

The history of our world is marked by constant and challenging changes that revolutionize how we relate to each other, how we work, and how we live. Currently, it is almost impossible not to be confronted with artificial intelligence (AI) technologies in practically all spheres of our life. New features and new technological advances are being revealed daily in the development of this type of technology. We live with them every day without even realizing it in self-driving cars, websites, robots, and even in the palm of our hands on our smartphones. With the improvement of hardware capable of performing progressively more complex tasks, the potential of AI systems has skyrocketed and, with it, the interest associated in developing these technologies.

Definition of artificial intelligence AI is defined as the use of computers and technology to simulate intelligent behavior and critical thinking comparable to that of a human being. Intelligence consists of all the intellectual characteristics of an individual, which include the ability to know, understand, reason, think, and interpret. It brings together several competencies and is subdivided into different types, with the computer’s ability to emulate each type being different. For example, the computer’s capacity to perform tasks dependent on logical-mathematical intelligence is far superior to the ability of creative intelligence, where the need to create new patterns is imperative [1]. One of the problems associated with the application of AI results from Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00003-4 © 2023 Elsevier Inc. All rights reserved.

1

2

CHAPTER 1 Artificial intelligence: machine learning, deep learning

the unrealistic expectations instilled by popular culture, for example, in movies and books, where the anthropomorphization of machines leads the viewer to create a stereotype of robotic consciousness, which is not achievable in practice. Since the beginning of computerization, humanity has tried to mimic the workings of the human mind and transfer the inherent abilities to machines. However, the inability to scrutinize the complete and complex functioning of the human mind prevents its adequate transfer to automated mechanisms. Thus, still far from the complete cloning of the human mind, new and provocative mechanisms have emerged capable of performing tasks close to sentience with remarkable efficiency.

Machine learning versus deep learning Machine learning Machine learning (ML) is a subset of AI and is defined as the ability of a computer to learn new tasks through data analysis, either supervised or unsupervised [2]. In a regular computer program, which is not dependent on ML algorithms, a series of rules and conditions are generated, and the outputs depend on the analysis of these conditions, usually of the if/then type [3]. ML-based algorithms, capable of generating conclusions based on the data entered, do not require the manual creation of these rules [4]. A classic example of how ML works is the ability to recognize digits or letters written manually. Traditionally, to achieve recognition of a digit, all possible combinations of pixels to generate that digit would have to be accounted for, which proves to be an unfeasible task. To generate an AI engine capable of this objective, we start by exposing the algorithm to thousands of possible examples of the digit and images not containing the digit in question. With the increasing amount of data the algorithm is able to recognize the object in question, improving its accuracy with continued exposure to new data [5].

Deep learning Deep learning is a subtype of ML in which the structure of the neural network presents a hierarchy inspired by the functioning of the human brain [6]. Deep neural networks use the compositional hierarchy of signals, in which higher-level features are obtained by combining lower-level ones. This structure allows the accomplishment of complex tasks [6]. The relationship between these different concepts can be seen in Fig. 1.1. A paradigmatic example of the applicability of deep learning technologies is the development of convolutional neural networks (CNNs), inspired by the functioning of the human visual cortex and developed specifically for image analysis [7]. In these networks, neurons are organized to respond to specific areas of the

Examples of artificial intelligence applicability

FIGURE 1.1 Machine learning and Deep learning brief description.

image, similar to what happens in the human visual field. A schematic representation of the functioning of CNN can be seen in Fig. 1.2. The neurons, when activated, propagate the information obtained to the subsequent layers, and the analysis of activation patterns allows for obtaining the final output [8]. CNNs require less preprocessing and are less dependent on prior knowledge and human effort [8]. CNNs perform better in image detection and recognition than other deep learning technologies [7].

Examples of artificial intelligence applicability The applications of AI technologies in our society are immense and sometimes surprising. We use speech recognition algorithms every time we talk to our smartphone, and it responds accordingly. When our car automatically corrects the route, helping the driver to avoid crashes, computerized technologies are applied. Even in our homes, AI technologies, for example, can automatically manage the temperature based on our personal preferences and time spent indoors and link this data with weather updates [9]. The development of these technologies is guided by structural rules, which have as their fundamental driving force the need to perform tasks more efficiently with reduced time and cost associated with it.

3

4

CHAPTER 1 Artificial intelligence: machine learning, deep learning

FIGURE 1.2 Neural Networks Architecture emulating neurobiological processes.

Online experience Our online experience is shaped by various AI algorithms, often without our awareness. When we type in a search engine, it suggests several options to complete the search, which serves a number of purposes. It makes the user experience easier as you only need to type a few characters and also saves your time. Furthermore, it reduces the possibilities of unnecessary searches via typos, making the search engine’s work easier. This intricate connection between the user and the server is facilitated by computerized methods that adapt the search possibilities of the user according to his previous research and interests [10]. The same rationale is applied in online shopping and in the advertisements shown to us on various web pages, which are modified according to the clicks or searches we perform in the browser. Data analysis generates a symbiosis between service providers, web pages, and users, benefitting everyone. Vendors are able to advertise more effectively to a targeted audience who are more likely to be interested in the products advertised; websites are able to offer a better service and increase their profit; and users are not annoyed with ads they are not interested in [11].

Robotics Robotics is the branch of computer science and engineering responsible for developing machines that perform programmed tasks without human intervention. Despite being two completely different entities, robotics and AI may have an integrated role to play [12]. An example of robotics is a robot vacuum cleaner that maps the house and plans its cleaning; the robot vacuum cleaner may be enabled to detect obstacles automatically by incorporating AI algorithms reducing human intervention and giving an enhanced experience [13].

Artificial intelligence in medicine

Vehicles The era of self-driving cars is clearly under progress, and although this is not yet a reality, several technologies dependent on AI that can assist drivers are already in practice. Automatic car parking, driver drowsiness detection, and automatic route corrections are already available computer-fueled driving aids [14].

Fake news detection and cybersecurity AI algorithms are useful in detecting and preventing the spread of fake news and cybersecurity. The development of social networks and other means of disseminating news has revolutionized the way we acquire information. However, the ease of access to social media content also makes the dissemination of fake news easy, given the speed with which a text or article is shared. The information transmitted through social networks are so massive that it is almost impossible to analyze and confirm its veracity in a short span of time. Automated methods can be useful in detecting fake news, particularly through detection methods like clickbait, spam, and phishing [15].

Artificial intelligence in healthcare as a facilitating technology With so many and vast examples of applications of AI in different areas of our society, it is not surprising that healthcare is also an area with a growing interest in the applicability of AI technologies. The use of AI in healthcare involves ethical and legal issues, so careful and exhaustive validation of these technologies in the field is a must. However, if used appropriately and as an adjunct to clinical practice, AI in healthcare can play an important role in reducing errors and saving time and resources.

Artificial intelligence in medicine The application of AI in the medical field is possibly the greatest breakthrough of the century. Adopting AI technologies in the field of medicine can revolutionize the way medical data is collected, screening and monitoring diseases, and the discovery of new drugs, leading to more effective and accurate healthcare services. Currently, there have been promising advances in image-based medical specialties such as radiology, dermatology, and ophthalmology. The development of

5

6

CHAPTER 1 Artificial intelligence: machine learning, deep learning

CNN has revolutionized image pattern recognition, enabling greater data analysis performance and growth in the number of implemented projects with clinical impact [16]. Implementation of AI algorithms during imagiologic monitorization of oncologic patients can assist radiologists in assessing the progression of the disease quantitatively, rather than just qualitatively [17]. In a general practitioner’s office, the diagnosis of skin cancer can be improved by applying AI screening software to evaluate skin lesions, and only referring cases with moderate-to-high probability of malignancy to dermatologists [18]. The evaluation of retinopathies in elderly patients, especially in those with multiple cardiovascular risk factors and diabetes, can be facilitated and become more accessible by using AI technology in assessing the ocular fundus [19]. Gastroenterology has always been a highly innovative field using cutting-edge technologies to provide better care for patients. It is therefore not surprising that the application of AI is exponentially growing, not only in the field of radiology, dermatology, ophthalmology and gastrointestinal (GI) endoscopy, but also in the field of hepatology, inflammatory bowel disease, and digestive pathology. There have been major advances in colonoscopy with the introduction of realtime automatic quality assurance systems. AI can identify landmarks, quantify withdrawal time, evaluate intestinal preparation quality, and detect and classify polyp type, size, and characteristics [20]. Evidence shows that using real-time AI feedback during colonoscopy is associated with a higher adenoma detection rate, not only by young trainees but also by expert gastroenterologists [21]. AI algorithms can also be used in assessing the upper GI tract to detect and evaluate premalignant and malignant lesions. Even with high-definition and advanced endoscopic imaging methods like digital chromoendoscopy, finding dysplasia in Barrett’s esophagus can be an intricate task. The missing rates are still not insignificant, and there can be false negatives if a biopsy is not taken at the ideal location. AI algorithms have shown promising results in identifying the location with a higher probability of dysplasia within Barrett’s metaplastic tissue, improving its detection and treatment significantly [22]. Upper and lower endoscopy, especially in unsedated patients, require efficiency, fluidity of movement, and a high level of concentration. One of the biggest concerns of any gastroenterologist is missing out on an important lesion that could be biopsied or treated, dramatically changing the natural course of the disease. AI technology can be used to alert endoscopists of blind spots during the procedure, contributing to a higher quality endoscopy [23]. AI can be useful in other areas of expertise in gastroenterology, such as in assessing endoscopic and histologic disease activity in inflammatory bowel disease, predicting clinical responses in biologic or oncologic therapies, making endoscopic capsule evaluation less time-consuming, and assisting in the diagnosis of biliary and pancreatic diseases [23,24]. The training curriculum of young gastroenterologists might need to be updated soon with the inclusion of the applications of AI technologies.

Capsule endoscopy: a brief introduction

Although AI is getting wide acceptance, some challenges still need to be considered [23]. First, we need to consider the technical complexity behind deep learning in terms of computing power, data acquisition, and standardization dealing with inherent selection bias, validation in randomized multicenter trials, privacy, and security. Second, we need to bear in mind that physicians and patients might not easily accept information given by AI black box algorithms that they cannot fully understand. Third, there is still an ethical and legal gap regarding AI application in the medical field, which can be difficult to solve: Who will be held responsible for untoward events if it was not a human decision? The medical community is witnessing a structural paradigm shift with booming health data science. AI is becoming an increasingly hot topic in every field of medicine since it has the potential for major improvements in clinical decisions and possibly still be cost-effective. Physicians, software engineers, and researchers need to work together to enhance the accuracy and external validity of the models while addressing the possible constraints of this technology.

Capsule endoscopy: a brief introduction Among the different areas of gastroenterology, capsule endoscopy is one that may benefit most from its use. Capsule endoscopy plays a key role in the investigation and management of small bowel diseases [25]. However, viewing and analyzing a capsule endoscopy video is a time-consuming and error-prone process, as a single video can produce up to 50,000 images and take up to 120 min to analyze [26]. This modality, therefore, becomes a fertile ground for the creation and application of computer-aided diagnostic tools, either because of the high number of data it produces, facilitating the creation of robust datasets and thus allowing the creation of more effective AI networks, or for the pertinence of its application, helping doctors in the analysis of these videos by reducing the time needed to read them, thereby reducing the errors associated with it. Several pilot studies have been developed within the scope of capsule endoscopy for the automatic detection of different lesions, such as the detection of digestive bleeding, through the detection of blood and hematic residues in the intestinal lumen, or the detection of potentially bleeding vascular lesions [27 30]. AI algorithms have also been developed for the detection of protruding lesions, evaluation of celiac disease activity, detection of parasites, and assessment of inflammatory bowel disease activity through automatic detection of ulcers and erosions [31 35]. These AI tools can also be applied to detect small bowel lesions in device-assisted enteroscopy in real time [36]. Undoubtedly, among the different fields of medicine, gastroenterology is one of the areas that could benefit most from the implementation of AI technologies, with capsule endoscopy being the paradigmatic example of the applicability of AI algorithms in clinical practice.

7

8

CHAPTER 1 Artificial intelligence: machine learning, deep learning

References [1] Kumar S. Advantages and disadvantages of artificial intelligence. Towar Data Sci [Internet]. 2019;(October). Available from: https://towardsdatascience.com/advantages-and-disadvantages-of-artificial-intelligence-182a5ef6588c. [2] Le Berre C, Sandborn WJ, Aridhi S, Devignes M-D, Fournier L, Smaı¨l-Tabbone M, et al. Application of artificial intelligence to gastroenterology and hepatology. Gastroenterology. 2020;158(1):76 94 e2. [3] Litch M, Karofsky A. Artificial intelligence. Philos Film 2020;(1):102 29. [4] Handelman GS, Kok HK, Chandra RV, Razavi AH, Lee MJ, Asadi H. eDoctor: machine learning and the future of medicine. J Intern Med 2018;284(6):603 19. [5] El-Sawy A, El-Bakry H, Loey M. CNN for handwritten arabic digits recognition based on LeNet-5. Int Conf Adv Intell Syst Inform. Springer; 2016. p. 566 75. [6] Amisha Malik P, Pathania M, Rathaur V. Overview of artificial intelligence in medicine. J Fam Med Prim Care 2019;. [7] Li N, Zhao X, Yang Y, Zou X. Objects classification by learning-based visual saliency model and convolutional neural network. Comput Intell Neurosci [Internet] 2016;2016:7942501. Available from: https://pubmed.ncbi.nlm.nih.gov/27803711. [8] Kim J, Kim J, Jang G-J, Lee M. Fast learning method for convolutional neural networks using extreme learning machine and its application to lane detection. Neural Netw J Int Neural Netw Soc 2017;87:109 21. [9] Parker D, Sutherland K, Chasar D. Evaluation of the space heating and cooling energy savings of smart thermostats in a hot-humid climate using long-term data. ACEEE Summer Study Energy Effic Build. 2016;(Nevius 2000):1 15. [10] Collins C, Dennehy D, Conboy K, Mikalef P. Artificial intelligence in information systems research: a systematic literature review and research agenda. Int J Inf Manage [Internet] 2021;60:102383. Available from: https://www.sciencedirect.com/ science/article/pii/S0268401221000761. [11] Ullal M, Hawaldar I, Soni R, Nadeem M. The role of machine learning in digital marketing. SAGE Open 2021;11:1 12. [12] Hofmann P, Samp C, Urbach N. Robotic process automation. Electron Mark [Internet] 2020;30(1):99 106. Available from: https://doi.org/10.1007/s12525-01900365-8. [13] Bergman J, Lind J. Robot vacuum cleaner; 2019. [14] Biggi G, Stilgoe J. Artificial intelligence in self-driving cars research and innovation: a scientometric and bibliometric analysis. SSRN Electron J 2021;. [15] Goksu M, Cavus N. Fake news detection on social networks with artificial intelligence tools: systematic literature review. Adv Intell Syst Comput 2020;1095:47 53 AISC. [16] Currie G, Hawk KE, Rohren E, Vial A, Klein R. Machine learning and deep learning in medical imaging: intelligent imaging. J Med Imaging Radiat Sci 2019;50 (4):477 87. [17] Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts HJWL. Artificial intelligence in radiology. Nat Rev Cancer 2018;18(8):500 10. [18] Hogarty DT, Su JC, Phan K, Attia M, Hossny M, Nahavandi S, et al. Artificial intelligence in dermatology—where we are and the way to the future: a review. Am J Clin Dermatol 2020;21(1):41 7.

References

[19] Ting DSW, Pasquale LR, Peng L, Campbell JP, Lee AY, Raman R, et al. Artificial intelligence and deep learning in ophthalmology. Br J Ophthalmol 2019;103 (2):167 75. [20] Karnes WE, Johnson DA, Berzin TM, Gross SA, Vargo JJ, Sharma P, et al. A polyp worth removing: a paradigm for measuring colonoscopy quality and performance of novel technologies for polyp detection. J Clin Gastroenterol 2021;55(9):733 9. [21] Hassan C, Spadaccini M, Iannone A, Maselli R, Jovani M, Chandrasekar VT, et al. Performance of artificial intelligence in colonoscopy for adenoma and polyp detection: a systematic review and metaanalysis. Gastrointest Endosc 2021;93(1):77 85 e6. [22] Ebigbo A, Mendel R, Probst A, Manzeneder J, Prinz F, De Souza LA, et al. Realtime use of artificial intelligence in the evaluation of cancer in Barrett’s oesophagus. Gut. 2020;69(4):615 16. [23] Cao JS, Chen MY, Zhang B, Hu JH, Li SJ, Feng X, et al. Artificial intelligence in gastroenterology and hepatology: status and challenges. World J Gastroenterol 2021;27(16):1664 90. [24] Gubatan J, Levitte S, Patel A, Balabanis T, Wei MT, Sinha SR. Artificial intelligence applications in inflammatory bowel disease: emerging technologies and future directions. World J Gastroenterol 2021;27(17):1920 35. [25] Triester SL, Leighton JA, Leontiadis GI, Fleischer DE, Hara AK, Heigh RI, et al. A metaanalysis of the yield of capsule endoscopy compared to other diagnostic modalities in patients with obscure gastrointestinal bleeding. Am J Gastroenterol 2005;100 (11):2407 18. [26] Wang A, Banerjee S, Barth BA, Bhat YM, Chauhan S, Gottlieb KT, et al. Wireless capsule endoscopy. Gastrointest Endosc 2013;78(6):805 15. [27] Hassan AR, Haque MA. Computer-aided gastrointestinal hemorrhage detection in wireless capsule endoscopy videos. Comput Methods Prog Biomed 2015;. [28] Aoki T, Yamada A, Kato Y, Saito H, Tsuboi A, Nakada A, et al. Automatic detection of blood content in capsule endoscopy images based on a deep convolutional neural network. J Gastroenterol Hepatol 2020;. [29] Tsuboi A, Oka S, Aoyama K, Saito H, Aoki T, Yamada A, et al. Artificial intelligence using a convolutional neural network for automatic detection of small-bowel angioectasia in capsule endoscopy images. Dig Endosc J Jpn Gastroenterol Endosc Soc 2020;32(3):382 90. [30] Leenhardt R, Vasseur P, Li C, Saurin JC, Rahmi G, Cholet F, et al. A neural network algorithm for detection of GI angiectasia during small-bowel capsule endoscopy. Gastrointest Endosc 2019;89(1):189 94. [31] Zhou T, Han G, Li BN, Lin Z, Ciaccio EJ, Green PH, et al. Quantitative analysis of patients with celiac disease by video capsule endoscopy: a deep learning method. Comput Biol Med 2017;85:1 6. [32] Wang X, Qian H, Ciaccio EJ, Lewis SK, Bhagat G, Green PH, et al. Celiac disease diagnosis from videocapsule endoscopy images with residual learning and deep feature extraction. Comput Methods Prog Biomed 2020;187:105236. [33] Saito H, Aoki T, Aoyama K, Kato Y, Tsuboi A, Yamada A, et al. Automatic detection and classification of protruding lesions in wireless capsule endoscopy images based on a deep convolutional neural network. Gastrointest Endosc 2020;92 (1):144 51 e1.

9

10

CHAPTER 1 Artificial intelligence: machine learning, deep learning

[34] Wu X, Chen H, Gan T, Chen J, Ngo C-W, Peng Q. Automatic hookworm detection in wireless capsule endoscopy images. IEEE Trans Med Imaging 2016;35 (7):1741 52. [35] Ferreira JPS, de Mascarenhas Saraiva MJ, da Q e C, Afonso JPL, Ribeiro TFC, Cardoso HMC, Ribeiro Andrade AP, et al. Identification of ulcers and erosions by the novel Pillcamt Crohn’s capsule using a convolutional neural network: a multicentre pilot study. J Crohn’s Colitis [Internet] 2022;16(1):169 72. Available from: https://doi.org/10.1093/ecco-jcc/jjab117. [36] Mascarenhas Saraiva M, Ribeiro T, Afonso J, Andrade P, Cardoso P, Ferreira J, et al. Deep learning and device-assisted enteroscopy: automatic detection of gastrointestinal angioectasia. Medicina (Kaunas) 2021;(12):57.

CHAPTER

Wireless capsule endoscopy: concept and modalities

2

Pablo Cortegoso Valdivia1 and Marco Pennazio2 1

Gastroenterology and Endoscopy Unit, University Hospital of Parma, University of Parma, Parma, Italy 2 University Division of Gastroenterology, City of Health and Science University Hospital, University of Turin, Turin, Italy

Background The concept of a wireless ingestible device with recording capabilities for the study of digestive segments that were out of reach of conventional endoscopy marked a milestone over 20 years ago [1]. LED light could eventually be shed inside the black box, and its flickering allowed to drive the management of many pathological conditions via high-quality images of the entire small bowel (SB) mucosa obtained in a virtually noninvasive modality. The disruption of this dogma has led to giant steps in the development of capsule endoscopy (CE), through the evolution and adaptation of new platforms and devices that nowadays allow specific answers to specific questions. A subsequent breakthrough was determined by the introduction of CE devices that were specifically conceived for the evaluation of gastrointestinal (GI) segments other than the SB, such as the esophagus, stomach, and colon. CE exploration now potentially allows a panenteric whole gut exploration, although several issues persist [2] (Table 2.1).

Types of capsules The first generation of capsule endoscopes for the study of the SB was marketed by Given Imaging Ltd. (Yokneam, Israel) under the brand name PillCam with the acronym M2A (mouth to anus). The capsules consisted of a single-headed 11 3 26-mm device with LED lights and a camera able to record 2 frames per second (fps) with a 140-degree view angle and a battery life of 8 h. The system was updated with a second- and afterward, a third-generation capsule (SB3) in 2013, which is now able to provide images with a higher resolution and an adaptive frame rate (afr) from 2 to 6 fps according to the progression speed of the capsule throughout the bowel. Several other companies, such as Olympus (Tokyo, Japan), Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00008-3 © 2023 Elsevier Inc. All rights reserved.

11

12

CHAPTER 2 Wireless capsule endoscopy: concept and modalities

Table 2.1 Pros and cons of capsule endoscopy. Pros

Cons

Good safety profile No ionizing radiations High-quality images Limited invasivity

Inability to take biopsies or to perform therapeutic interventions Possible incomplete visualization of the digestive segment under study Difficult to distinguish between innocent bulges and SB submucosal masses Low specificity (especially for SB inflammatory findings)

SB, Small bowel.

Intromedic (Seoul, South Korea), Jinshan (Chongqing, China), and Capsovision (Saratoga, CA, USA), produce capsules for the study of the SB. In 2006, a double-headed capsule for the detection of colon pathology was introduced to the market [3]. The main difference of this capsule with the SB capsules is the presence of a second camera designed to increase the detection of pathological findings across the mucosal folds of the colon. The first generation of this system provided a 156-degree view angle, with no afr. The colon capsules are provided with a sleep mode that turns the recording off around 3 min after ingestion and reactivates when the capsule is in the SB to save battery for a complete colonic inspection. A second-generation colon CE (CCE) was subsequently released in 2010, with a wider view angle (172 degrees) and an afr from 4 to 35 fps [4]. The potential value of the CE as a panenteric tool for gut exploration was conceived for the first time with the introduction of the PillCam Crohn’s capsule (Medtronic, Dublin, Ireland) in 2016, a modified colon capsule with no sleep mode relying on a dedicated software platform for the study of both the SB and the colon in patients affected with inflammatory bowel disease (IBD) [5]. Although technically feasible and extremely appealing in terms of clinical potential [6], this approach is far from being regularly applied in everyday practice [7]. Future developments in noninvasive GI examination are looking into the use of magnetically controlled capsules for the study of the stomach or a combined gastric and SB assessment [8 10]. Nevertheless, although preliminary results are promising in terms of procedural completion rate (100%), further evidence is needed. An overview of commercially available CE models is shown in Table 2.2.

Indications Suspected small bowel bleeding According to the latest definition, a GI bleeding whose origin is not determined by initial endoscopic evaluation of the upper and lower GI tract is defined as a suspected SB bleeding (SSBB) [11]. SSBB the most frequent indication for SBCE accounts for 5% of all GI bleedings. Accordingly, an SSBB whose origin is not

Table 2.2 Overview of capsule endoscopy models. Company

Model

Dimensions (mm)

Weight (g)

Battery life (h)

Image sensor

Illumination

Field of view

Depth of field (mm)

Image sampling rate (fps)

Adaptive frame rate

Real-time monitoring

Medtronic (Dublin, Ireland)

PillCam SB3

11 3 26

3.0

8

Front

4 LEDs

156 degrees

0 30

2 6

Yes

Yes

PillCam Colon 2 PillCam Crohn’s Endo Capsule 10

11.6 3 32.3

2.9

10 12

Front and rear

4 35

Yes

Yes

2.9

10 12

Front and rear

0 30

4 35

Yes

Yes

11 3 26

3.3

12

Front

172 degrees 172 degrees 160 degrees

0 30

11.6 3 32.3

4 double LEDs 4 double LEDs 4 LEDs

0 20

2

No

Yes

Mirocam MC 1600

10.8 3 24.5

3.25

12

Front

6 LEDs

170 degrees

0 30

6

No

Yes

Mirocam MC 2000 OMOM HD

10.8 3 30.1

3.5

12

Front and rear

Yes

12

Front

0 50

6 (3 each on side) 2 10

No

3.0

170 degrees 172 degrees

0 30

11 3 25.4

6 double LEDs 4 LEDs

Yes

Yes

11 3 31

4

15

Lateral and circumferential

16 LEDs

360 degrees lateral

0 18

Up to 5

Yes

No

Olympus (Tokyo, Japan) Intromedic (Seoul, South Korea)

Jinshan (Chongqing, China) Capsovision (Saratoga, CA, USA)

CapsoCam Plus

14

CHAPTER 2 Wireless capsule endoscopy: concept and modalities

determined by SBCE is defined as obscure. Vascular lesions are encountered in 60% 70% of the cases, followed by Crohn’s disease (CD) ulcers, SB tumors, and Meckel’s diverticulum [12]. In this setting, SBCE has a diagnostic yield (rate of examinations in which a potentially bleeding lesion is detected) of 50% 60% [13,14]. Although performing SBCE in a short time after the occurrence of the bleeding may be challenging in everyday practice due to logistic issues [15], it has been demonstrated that the shorter the interval between the overt GI bleeding episode and SBCE the higher the diagnostic yield [16 18]. One of the main advantages of SBCE in bleeding patients is its high negative predictive value: patients can be safely managed with a wait-and-see strategy as long as SBCE does not detect any bleeding (or potentially bleeding) lesion [19]. Of note, a recent prospective study has shown the potential of a panenteric approach of early CE in patients presenting with melena and a negative esophagogastroscopy (EGD), with a diagnostic yield of 83.3% [20]; further investigations in this specific setting are warranted. Finally, SBCE may select patients likely to benefit from subsequent evaluation and intervention, such as a device-assisted enteroscopy (DAE). SBCE can accurately select the correct route of insertion for DAE (oral versus anal) based on the relative location of abnormal SBCE findings in terms of SB transit time, thus making subsequent DAE more efficient [21].

Small bowel tumors Although rare, SB tumors account for up to 5% of all patients undergoing SBCE [22], with a higher incidence in young patients with persistent iron-deficiency anemia or SSBB [23]. Although SBCE has shown a high diagnostic yield compared with other imaging techniques in this setting [24], the detection of SB tumors with SBCE may be challenging due to several factors [25]: (1) SB tumors are frequently located in the proximal bowel in which the capsule transit is faster and biliary secretions/debris are abundant, (2) SB tumors may show with mucosal bulging, which can be harsh to discriminate from innocent bulges—in this case, the interpretation of findings can be guided by validated visual indices [26,27].

Hereditary polyposis syndromes SBCE has a paramount role in the follow-up of patients affected by hereditary polyposis syndromes, especially familial adenomatous polyposis (FAP) and Peutz Jeghers syndrome (PJS). Although CE surveillance protocols for FAP patients are not strictly determined [28], PJS patients should undergo follow-up examinations with a 1- to 3-year interval according to the disease phenotype [29,30]. As these patients need to undergo extensive lifelong surveillance, SBCE is a reasonable alternative to magnetic resonance enterography (MRE). Although computed tomography enterography seems to have a better definition than SBCE in locating SB polyps and defining their size, it is not advisable as a routinary surveillance imaging technique due to ionizing radiations exposure.

Indications

Celiac disease The role of SBCE in patients with suspected celiac disease should be limited to those unwilling or unable to undergo EGD [23]. There are two emerging indications for SBCE examination in equivocal cases of celiac disease: (1) patients with biopsy-proven atrophy of the villi but negative serology, (2) patients with normal duodenal biopsies but positive serology (antitissue transglutaminase IgA and/or EmA). In the first scenario (in seronegative villous atrophy), SBCE is a prominent tool as it may provide useful information for alternative diagnoses, possibly having an impact on the patient’s management in up to 70% of the cases [31,32]; in the second scenario, the clinical relevance of the use of SBCE is more uncertain and mostly related to the possibility of evaluating slight mucosal alterations in the distal duodenum/jejunum [33,34]. Conversely, SBCE has a role in excluding preneoplastic or neoplastic complications in nonresponsive patients, such as refractory celiac disease, ulcerative jejuno-ileitis, enteropathy-associated T-cell lymphoma, or SB adenocarcinoma [35]. In this clinical scenario, the sequential approach of SBCE followed by DAE in case of suspect findings appears justified nowadays [23].

Crohn’s disease SB involvement is present in up to 10% of CD patients, making CE exploration highly recommended in cases where the disease is suspected but colonoscopy is negative. Although the negative predictive value in detecting SB inflammatory lesions is high (especially in the ileum, in which the capsule transit is slower), the diagnostic yield of CE in this setting is extremely dependent on the patient’s pretest probability of being affected by CD [36,37]. According to the ICCE criteria, this probability may be estimated by the combination of radiological and clinical data and by specific markers such as fecal calprotectin (its values directly correlate with a positive CE examination in suspected CD) [38,39]. A panenteric exploration of both SB and colon has been advised and explored due to the characteristics of CD [2]. In CD patients, SBCE may provide elements for a definitive diagnosis and staging, evaluating mucosal healing, and monitoring the evolution of the disease in response to medical therapy [40,41]. Considering that CD is a favoring condition for capsule retention due to the possible development of stenoses in the SB, using a patency capsule before CE is strongly advised (even in the case of a negative SB cross-sectional imaging) [42].

Colon examination The simultaneous recording of images by the double heads of the CCE allows the detection of colonic pathology with high sensitivity, overcoming the risk of missing lesions behind haustral folds. The role of CCE as a filter examination between fecal immunochemical testing and optical colonoscopy has been evaluated by a

15

16

CHAPTER 2 Wireless capsule endoscopy: concept and modalities

recent metaanalysis, in which sensitivity and specificity for polyps .6 mm were 87% and 88%, respectively [43]. Similar results (sensitivity 87% and specificity 87% for polyps .6 mm—87% and 95% for polyps .10 mm) were confirmed by another study, in which CCE was directly compared with optical colonoscopy [44]—this approach is also well tolerated by the patients, even though no real preference for one modality over the other has been demonstrated [45]. Another scenario for CCE use is represented by incomplete colonoscopy, as advised by recent guidelines [46]. Nevertheless, the high level of CCE performance encounters major drawbacks mostly related to bowel cleanliness. In a French real-life population study [47], CCE failed to show reliability for completion of the examination with less than 50% of procedures with excellent or good bowel preparation. Although a high concordance with optical colonoscopy was shown for lesions with high-grade dysplasia or cancer, the results are far from optimal. Suboptimal rates of completion and adequate cleanliness in CCE were also highlighted in a recent systematic review with metaanalysis [48], confirming the need for prospective studies with the aim of defining more suitable preparation regimens.

Future perspectives The excitement for new possibilities and perspectives in the field of CE is far from over. From new devices to new indications (a possible role in colorectal cancer detection in screening populations) [49], current fields of exploration are now mainly into artificial intelligence (AI) technologies. Outcomes such as the reduction of CE reading times (and thereby accuracy due to fatigue) [50], the growth in diagnostic accuracy, and the automated characterization of detected lesions are now starting to take shape in reality [51]. Although the journey is long (as it is through the bowel!), the future is LED-bright!

References [1] Iddan G, Meron G, Glukhovsky A, Swain P. Wireless capsule endoscopy. Nature 2000;405:417. Available from: https://doi.org/10.1038/35013140. [2] Cortegoso Valdivia P, Elosua A, Houdeville C, Pennazio M, Ferna´ndez-Urie´n I, Dray X, et al. Clinical feasibility of panintestinal (or panenteric) capsule endoscopy: a systematic review. Eur J Gastroenterol Hepatol 2021;33:949 55. Available from: https:// doi.org/10.1097/MEG.0000000000002200. [3] Eliakim R, Fireman Z, Gralnek I, Yassin K, Waterman M, Kopelman Y, et al. Evaluation of the PillCam colon capsule in the detection of colonic pathology: results of the first multicenter, prospective, comparative study. Endoscopy 2006;38:963 70. Available from: https://doi.org/10.1055/s-2006-944832.

References

[4] Eliakim R, Yassin K, Niv Y, Metzger Y, Lachter J, Gal E, et al. Prospective multicenter performance evaluation of the second-generation colon capsule compared with colonoscopy. Endoscopy 2009;41:1026 31. Available from: https://doi.org/10.1055/ s-0029-1215360. [5] Eliakim R, Spada C, Lapidus A, Eyal I, Pecere S, Ferna´ndez-Urie´n I, et al. Evaluation of a new pan-enteric video capsule endoscopy system in patients with suspected or established inflammatory bowel disease feasibility study. Endosc Int Open 2018;06:E1235 46. Available from: https://doi.org/10.1055/a-0677-170. [6] Vuik FER, Moen S, Nieuwenburg SAV, Schreuders EH, Kuipers EJ, Spaander MCW. Applicability of colon capsule endoscopy as pan-endoscopy: from bowel preparation, transit, and rating times to completion rate and patient acceptance. Endosc Int Open 2021;09:E1852 9. Available from: https://doi.org/10.1055/a-1578-1800. [7] Rondonotti E, Pennazio M. Colon capsule for panendoscopy: a narrow window of opportunity. Endosc Int Open 2021;09:E1860 2. Available from: https://doi.org/ 10.1055/a-1548-6572. [8] Geropoulos G, Aquilina J, Kakos C, Anestiadou E, Giannis D. Magnetically controlled capsule endoscopy vs. conventional gastroscopy: a systematic review and meta-analysis. J Clin Gastroenterol 2021;55:577 85. Available from: https://doi.org/ 10.1097/MCG.0000000000001540. [9] Xiao Y-F, Wu Z-X, He S, Zhou Y-Y, Zhao Y-B, He J-L, et al. Fully automated magnetically controlled capsule endoscopy for examination of the stomach and small bowel: a prospective, feasibility, two-centre study. Lancet Gastroenterol Hepatol 2021;6:914 21. Available from: https://doi.org/10.1016/S2468-1253(21)00274-0. [10] Zhu J-H, Pan J, Xu X-N, Liu Y-W, Qian Y-Y, Jiang X, et al. Noncontact magnetically controlled capsule endoscopy for infection-free gastric examination during the COVID-19 pandemic: a pilot, open-label, randomized trial. Endosc Int Open 2022;10:E163 71. Available from: https://doi.org/10.1055/a-1648-2238. [11] Gerson LB, Fidler JL, Cave DR, Leighton JA. ACG clinical guideline: diagnosis and management of small bowel bleeding. Am J Gastroenterol 2015;110:1265 87. Available from: https://doi.org/10.1038/ajg.2015.246. [12] Soncini M, Girelli CM, de Franchis R, Rondonotti ESBCE Lombardia Study GroupOn behalf AIGO, SIED and SIGE Lombardia. Small-bowel capsule endoscopy in clinical practice: has anything changed over 13 years? Dig Dis Sci 2018;63:2244 50. Available from: https://doi.org/10.1007/s10620-018-5101-9. [13] Liao Z, Gao R, Xu C, Li Z-S. Indications and detection, completion, and retention rates of small-bowel capsule endoscopy: a systematic review. Gastrointest Endosc 2010;71:280 6. Available from: https://doi.org/10.1016/j.gie.2009.09.031. ˙ [14] Cortegoso Valdivia P, Skonieczna-Zydecka K, Elosua A, Sciberras M, Piccirelli S, Rullan M, et al. Indications, detection, completion and retention rates of capsule endoscopy in two decades of use: a systematic review and meta-analysis. Diagnostics (Basel) 2022;12:1105. Available from: https://doi.org/10.3390/diagnostics12051105. [15] Rondonotti E, Spada C, Pennazio M, de Franchis R, Cadoni S, Girelli C, et al. Adherence to European Society of Gastrointestinal Endoscopy recommendations of endoscopists performing small bowel capsule endoscopy in Italy. Dig Liver Dis 2019;51:818 23. Available from: https://doi.org/10.1016/j.dld.2018.11.031. [16] Pennazio M, Santucci R, Rondonotti E, Abbiati C, Beccari G, Rossini FP, et al. Outcome of patients with obscure gastrointestinal bleeding after capsule

17

18

CHAPTER 2 Wireless capsule endoscopy: concept and modalities

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26]

[27]

[28]

endoscopy: report of 100 consecutive cases. Gastroenterology 2004;126:643 53. Available from: https://doi.org/10.1053/j.gastro.2003.11.057. O’Hara F, McNamara D. Small-bowel capsule endoscopy-optimizing capsule endoscopy in clinical practice. Diagnostics (Basel) 2021;11:2139. Available from: https:// doi.org/10.3390/diagnostics11112139. Estevinho MM, Pinho R, Fernandes C, Rodrigues A, Ponte A, Gomes AC, et al. Diagnostic and therapeutic yields of early capsule endoscopy and device-assisted enteroscopy in the setting of overt GI bleeding: a systematic review with metaanalysis. Gastrointest Endosc 2022;95:610 25. Available from: https://doi.org/ 10.1016/j.gie.2021.12.009. Yung DE, Koulaouzidis A, Avni T, Kopylov U, Giannakou A, Rondonotti E, et al. Clinical outcomes of negative small-bowel capsule endoscopy for small-bowel bleeding: a systematic review and meta-analysis. Gastrointest Endosc 2017;85:305 17. Available from: https://doi.org/10.1016/j.gie.2016.08.027. Mussetto A, Arena R, Fuccio L, Trebbi M, Tina Garribba A, Gasperoni S, et al. A new panenteric capsule endoscopy-based strategy in patients with melena and a negative upper gastrointestinal endoscopy: a prospective feasibility study. Eur J Gastroenterol Hepatol 2021;33:686 90. Available from: https://doi.org/10.1097/ MEG.0000000000002114. ˙ Cortegoso Valdivia P, Skonieczna-Zydecka K, Pennazio M, Rondonotti E, Marlicz W, Toth E, et al. Capsule endoscopy transit-related indicators in choosing the insertion route for double-balloon enteroscopy: a systematic review. Endosc Int Open 2021;09:E163 70. Available from: https://doi.org/10.1055/a-1319-1452. Johnston C, Yung D, Joshi A, Plevris J, Koulaouzidis A. Small bowel malignancy in patients undergoing capsule endoscopy at a tertiary care academic center: case series and review of the literature. Endosc Int Open 2017;05:E463 70. Available from: https://doi.org/10.1055/s-0043-106186. Pennazio M, Spada C, Eliakim R, Keuchel M, May A, Mulder C, et al. Small-bowel capsule endoscopy and device-assisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Clinical Guideline. Endoscopy 2015;47:352 86. Available from: https://doi.org/ 10.1055/s-0034-1391855. Rondonotti E, Pennazio M, Toth E, Menchen P, Riccioni M, De Palma G, et al. Small-bowel neoplasms in patients undergoing video capsule endoscopy: a multicenter European study. Endoscopy 2008;40:488 95. Available from: https://doi.org/ 10.1055/s-2007-995783. Ross A, Mehdizadeh S, Tokar J, Leighton JA, Kamal A, Chen A, et al. Double balloon enteroscopy detects small bowel mass lesions missed by capsule endoscopy. Dig Dis Sci 2008;53:2140 3. Available from: https://doi.org/10.1007/s10620-007-0110-0. Rosa B, Margalit-Yehuda R, Gatt K, Sciberras M, Girelli C, Saurin J-C, et al. Scoring systems in clinical small-bowel capsule endoscopy: all you need to know!. Endosc Int Open 2021;09:E802 23. Available from: https://doi.org/10.1055/a-1372-4051. Sciberras M, Gatt K, Elli L, Scaramella L, Riccioni ME, Marmo C, et al. OP065 Score reproducibility and reliability in differentiating small bowel subepithelial masses from innocent bulges. Endoscopy 2022;54:S1 303. Syngal S, Brand RE, Church JM, Giardiello FM, Hampel HL, Burt RW. ACG clinical guideline: genetic testing and management of hereditary gastrointestinal cancer

References

[29]

[30]

[31]

[32]

[33]

[34]

[35]

[36]

[37]

[38]

[39]

[40]

[41]

syndromes. Am J Gastroenterol 2015;110:223 62. Available from: https://doi.org/ 10.1038/ajg.2014.435. van Leerdam ME, Roos VH, van Hooft JE, Dekker E, Jover R, Kaminski MF, et al. Endoscopic management of polyposis syndromes: European Society of Gastrointestinal Endoscopy (ESGE) Guideline. Endoscopy 2019;51:877 95. Available from: https:// doi.org/10.1055/a-0965-0605. Wagner A, Aretz S, Auranen A, Bruno MJ, Cavestro GM, Crosbie EJ, et al. The management of Peutz Jeghers syndrome: European Hereditary Tumour Group (EHTG) guideline. J Clin Med 2021;10:473. Available from: https://doi.org/10.3390/ jcm10030473. Chetcuti Zammit S, Schiepatti A, Aziz I, Kurien M, Sanders DS, Sidhu R. Use of small-bowel capsule endoscopy in cases of equivocal celiac disease. Gastrointest Endosc 2020;91:1312 21. Available from: https://doi.org/10.1016/j.gie.2019.12.044. Luja´n-Sanchis M, Pe´rez-Cuadrado-Robles E, Garcı´a-Lledo´ J, Juanmartin˜ena Ferna´ndez J-F, Elli L, Jime´nez-Garcı´a V-A, et al. Role of capsule endoscopy in suspected celiac disease: a European multi-centre study. World J Gastroenterol 2017;23:703. Available from: https://doi.org/10.3748/wjg.v23.i4.703. Kurien M, Evans KE, Aziz I, Sidhu R, Drew K, Rogers TL, et al. Capsule endoscopy in adult celiac disease: a potential role in equivocal cases of celiac disease? Gastrointest Endosc 2013;77:227 32. Available from: https://doi.org/10.1016/j.gie.2012.09.031. Lidums I, Cummins AG, Teo E. The role of capsule endoscopy in suspected celiac disease patients with positive celiac serology. Dig Dis Sci 2011;56:499 505. Available from: https://doi.org/10.1007/s10620-010-1290-6. Atlas DS, Rubio-Tapia A, Van Dyke CT, Lahr BD, Murray JA. Capsule endoscopy in nonresponsive celiac disease. Gastrointest Endosc 2011;74:1315 22. Available from: https://doi.org/10.1016/j.gie.2011.05.049. Mergener K, Ponchon T, Gralnek I, Pennazio M, Gay G, Selby W, et al. Literature review and recommendations for clinical application of small-bowel capsule endoscopy, based on a panel discussion by international experts. Endoscopy 2007;39:895 909. Available from: https://doi.org/10.1055/s-2007-966930. Rosa B, Moreira MJ, Rebelo A, Cotter J. Lewis score: a useful clinical tool for patients with suspected Crohn’s disease submitted to capsule endoscopy. J Crohns Colitis 2012;6:692 7. Available from: https://doi.org/10.1016/j.crohns.2011.12.002. Yousuf H, Aleem U, Egan R, Maheshwari P, Mohamad J, McNamara D. Elevated faecal calprotectin levels are a reliable non-invasive screening tool for small bowel Crohn’s disease in patients undergoing capsule endoscopy. Dig Dis 2018;36:202 8. Available from: https://doi.org/10.1159/000485375. Kopylov U, Yung DE, Engel T, Avni T, Battat R, Ben-Horin S, et al. Fecal calprotectin for the prediction of small-bowel Crohn’s disease by capsule endoscopy: a systematic review and meta-analysis. Eur J Gastroenterol Hepatol 2016;28:1137 44. Available from: https://doi.org/10.1097/MEG.0000000000000692. Maaser C, Sturm A, Vavricka SR, Kucharzik T, Fiorino G, Annese V, et al. ECCOESGAR Guideline for Diagnostic Assessment in IBD Part 1: Initial diagnosis, monitoring of known IBD, detection of complications. J Crohns Colitis 2019;13:144 164K. Available from: https://doi.org/10.1093/ecco-jcc/jjy113. Sturm A, Maaser C, Calabrese E, Annese V, Fiorino G, Kucharzik T, et al. ECCOESGAR Guideline for Diagnostic Assessment in IBD Part 2: IBD scores and general

19

20

CHAPTER 2 Wireless capsule endoscopy: concept and modalities

[42]

[43]

[44]

[45]

[46]

[47]

[48]

[49]

[50]

[51]

principles and technical aspects. J Crohns Colitis 2019;13:273 84. Available from: https://doi.org/10.1093/ecco-jcc/jjy114. Pasha SF, Pennazio M, Rondonotti E, Wolf D, Buras MR, Albert JG, et al. Capsule retention in Crohn’s disease: a meta-analysis. Inflamm Bowel Dis 2020;26:33 42. Available from: https://doi.org/10.1093/ibd/izz083. Kjølhede T, Ølholm AM, Kaalby L, Kidholm K, Qvist N, Baatrup G. Diagnostic accuracy of capsule endoscopy compared with colonoscopy for polyp detection: systematic review and meta-analyses. Endoscopy 2021;53:713 21. Available from: https://doi.org/10.1055/a-1249-3938. Mo¨llers T, Schwab M, Gildein L, Hoffmeister M, Albert J, Brenner H, et al. Secondgeneration colon capsule endoscopy for detection of colorectal polyps: systematic review and meta-analysis of clinical trials. Endosc Int Open 2021;09:E562 71. Available from: https://doi.org/10.1055/a-1353-4849. Deding U, Cortegoso Valdivia P, Koulaouzidis A, Baatrup G, Toth E, Spada C, et al. Patient-reported outcomes and preferences for colon capsule endoscopy and colonoscopy: a systematic review with meta-analysis. Diagnostics (Basel) 2021;11:1730. Available from: https://doi.org/10.3390/diagnostics11091730. Spada C, Hassan C, Bellini D, Burling D, Cappello G, Carretero C, et al. Imaging alternatives to colonoscopy: CT colonography and colon capsule. European Society of Gastrointestinal Endoscopy (ESGE) and European Society of Gastrointestinal and Abdominal Radiology (ESGAR) Guideline Update 2020. Endoscopy 2020;52:1127 41. Available from: https://doi.org/10.1055/a-1258-4819. Benech N, Vinet O, Gaudin J-L, Benamouzig R, Dray X, Ponchon T, et al. Colon capsule endoscopy in clinical practice: lessons from a national 5-year observational prospective cohort. Endosc Int Open 2021;09:E1542 8. Available from: https://doi. org/10.1055/a-1526-0923. Bjoersum-Meyer T, Skonieczna-Zydecka K, Cortegoso Valdivia P, Stenfors I, Lyutakov I, Rondonotti E, et al. Efficacy of bowel preparation regimens for colon capsule endoscopy: a systematic review and meta-analysis. Endosc Int Open 2021;09:E1658 73. Available from: https://doi.org/10.1055/a-1529-5814. Vuik FER, Nieuwenburg SAV, Moen S, Spada C, Senore C, Hassan C, et al. Colon capsule endoscopy in colorectal cancer screening: a systematic review. Endoscopy 2021;53:815 24. Available from: https://doi.org/10.1055/a-1308-1297. Beg S, Card T, Sidhu R, Wronska E, Ragunath K, Ching H-L, et al. The impact of reader fatigue on the accuracy of capsule endoscopy interpretation. Dig Liver Dis 2021;53:1028 33. Available from: https://doi.org/10.1016/j.dld.2021.04.024. Dray X, Iakovidis D, Houdeville C, Jover R, Diamantis D, Histace A, et al. Artificial intelligence in small bowel capsule endoscopy current status, challenges and future promise. J Gastroenterol Hepatol 2021;36:12 19. Available from: https://doi.org/ 10.1111/jgh.15341.

CHAPTER

Capsule endoscopy: wide clinical scope

3

Pilar Esteban Delgado1, Renato Medas2,3, Eunice Trindade4 and Enrique Pe´rez-Cuadrado Martı´nez1 1

Small Bowel Unit, Endoscopy Department, University Hospital Morales Meseguer, Murcia, Spain 2 Gastroenterology Department, Centro Hospitalar Universita
rio de Sa˜o Joa˜o, Porto, Portugal 3 Faculty of Medicine, University of Porto, Porto, Portugal 4 Unit of Pediatric Gastroenterology and Nutrition, Centro Hospitalar Universita
rio de Sa˜o Joa˜o, Porto, Portugal

Body Capsule endoscopy (CE) was first described in 2000 [1] revolutionizing the diagnosis of small intestine (SI) pathology. Since then constant technological improvements have been implemented in successive years, leading to a revolution in the management of SI pathology. CE was the first wireless endoscopy without a fiber device attached to the outside a capsule that traveled through the digestive tract, although in an uncontrolled manner, recording several hours of the mucosa or inner layer, for which it had to emit light and save enough battery for operation (illuminate/record). CE received approval from the Food and Drug Administration (FDA) for use in the US in 2001 and in 2003 it was approved as a first-line method for the study of SI, representing a paradigm shift in the study of SI pathology. CE consists of the ingestion of a small device equipped with a camera, which advances along the digestive tract due to peristalsis. The device acquires images in real time, which it transmits to sensors placed on the abdominal wall of the patient and subsequently to an external recorder by radiofrequency. At the end of the examination, these images are transferred to a workstation where they are reviewed and evaluated for diagnosis. The interpretation of the images obtained by CE continues to be a challenge for gastroenterologists because it requires a high level of concentration and dedication and considerable reading time. Moreover, there are chances that the fatigued human eye overlooks discrete pathologies visible only in a few frames. The reading can be modulated with higher or lower speed of images per second; logically, at a higher speed, there are more possibilities of losing pathological images that sometimes occur in one or a few frames. Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00004-6 © 2023 Elsevier Inc. All rights reserved.

21

22

CHAPTER 3 Capsule endoscopy: wide clinical scope

The panoramic CE with 360-degree vision was introduced in 2013 [2]. As an advantage, it provides 4 cameras, two at each end of the CE, a 12-h battery life, and a data storage system, so it does not require a sensory belt for data recording. However, as a disadvantage this type of capsule does not have an adapted frame rate and requires the recovery of the CE to process the images. The interpretation of panoramic images requires gastroenterologists or endoscopists with extensive experience in the technique. Several studies that have compared standard CE with panoramic CE have shown greater diagnostic yields [3], while other studies did not report significant differences [4,5]. In recent years, technological improvements in software and artificial intelligence have made significant progress in the field of medicine, including the field of gastroenterology and endoscopy. Machine learning has managed to extract images and combine them with computational neural networks (CNNs), classifying and selecting pathological images to facilitate the detection and interpretation of SI pathologies. Many of the technological improvements provide advances that affect the diagnostic impact, achieving several objectives: • • • •

reduced reading time improved diagnostic sensitivity reduced interreader subjectivity improved data processing

In this chapter, we will review the indications of the endoscopic capsule and the improvement in the diagnostic impact after the implementation of artificial intelligence.

Indications in capsule endoscopy The classic indications in CE are middle digestive hemorrhage of suspected obscure gastrointestinal bleeding (OGIB); Crohn’s disease (CD) assessment; screening for polyposis syndromes—familial adenomatous polyposis (FAP) and Peutz Jeghers syndrome (PJS); evaluation of refractory celiac disease; suspected SI tumors; and graft versus host disease (GVHD).

Suspected middle digestive hemorrhage: obscure gastrointestinal bleeding OGIB was defined by the AGA (American Gastroenterology Association) as bleeding of unknown origin after performing a gastroscopy and colonoscopy regulated with quality criteria, which is accompanied by a positive fecal occult blood test (FOBT), anemia, or visible bleeding.

Indications in capsule endoscopy

It is estimated that approximately 5% of digestive hemorrhages occur in SI and are not accessible by gastroscopy and conventional colonoscopy [6]. Most of the lesions that cause OGIB are located in the SI and rarely in the stomach, colon, biliary tract, or pancreas. Before CE, fiber-optic enteroscopy improved the diagnosis of SI, but it could be only partially explored and not fully. As a result of the introduction of CE, the diagnosis of OGIB has improved. Subsequently, enteroscopes with assisted technology could be developed, starting with the doubleballoon enteroscope, then the single-balloon enteroscope, and recently the automated spiral enteroscope. Although enteroscopes with assisted technology can explore the entire SI (especially when the two routes, oral and anal, are combined), they do not allow invasive techniques of long duration, which consume a large part of a daily work schedule with human resources requirement, considering that the anesthesia team collaborates often. Symptoms associated with OGIB can be indolent, such as positive FOBT, iron deficiency anemia, manifested as melena, and hematochezia, with possible hemodynamic compromise. OGIB can be classified into two clinical forms: (1) open OGIB manifested in the form of hematochezia or melena, and (2) occult OGIB defined by the persistence of recurrent iron deficiency anemia or repeated positivity of occult blood in feces. CE is the first-line diagnostic tool to study OGIB, with the diagnostic impact of the technique being 64% [7]. The highest diagnostic profitability is obtained in patients with Hb , 10 mg/dL and/or with drops in Hb of more than 4 g/dL and performing CE as close as possible to the hemorrhagic event within the first 3 days [8]. The indicator of suspicion of blood is a common tool used to automatically label suspicious images of possible bleeding in the reading system. The CNN systems could improve the detection of bleeding compared with the indicator of suspicion of blood, for which recent studies [9] have evaluated 27,847 images of CE. As a result, greater sensitivity, specificity, and precision of 96.63%, 99.96%, and 99.89%, respectively, were obtained in the CNN system compared with the indicator of suspected blood (76.92%, 99.82%, and 99.35%). The implementation of artificial intelligence in CE could be a promising tool for improving diagnostic yields since it helps to recognize image patterns. A recent metaanalysis that included 19 retrospective studies [10] evaluated the applications of artificial intelligence for the accuracy of the diagnostic detection of CE, which was above 90% for the majority of studies and pathologies, with sensitivity and specificity for diagnosis in patients with mean digestive bleeding being 0.98 (95% CI, 0.96 0.99) and 0.99 (95% CI, 0.97 0.99), respectively. However, these retrospective studies had a high risk of bias, as such prospective studies with a greater number of patients are needed. The use of machine learning for the detection of pathologies and localized lesions in SI continues to be continuously researched and constantly implemented for technological improvements [11,12].

23

24

CHAPTER 3 Capsule endoscopy: wide clinical scope

The etiology of OGIB originating in SI can be secondary to a large number of pathologies, which are summarized in Table 3.1. The guidelines recommend performing CE as soon as possible with open OGIB, preferably in the first 14 days (Figs. 3.1 3.3). In cases of active OGIB, it has been shown that the performance of CE in the first 24 72 h is safe and has a diagnostic yield of approximately 70%, presenting an impact on the management of patients. Another clinical scenario that may occur is that of active and severe OGIB in which the patient presents signs of hemodynamic instability; in this situation, urgent and real-time CE could be evaluated. Urgent CE with real-time visualization can lead to urgent enteroscopy. Through this technique, the visualization of CE can also help decision-making. Limited data are available on the indication and urgent reading of CE, as well as the indication of enteroscopy and access route in cases of pathological findings, especially in relation to severe OGIB. A retrospective study by Pe´rezCuadrado Robles et al. [13] evaluated this combined approach in 27 patients with severe MGIB (mid gastrointestinal bleeding) who underwent CE with real-time visualization. With this, Dieulafoy lesions (n 5 11, 40.7%), angioectasias (n 5 7, 25.9%), tumors (n 5 4, 14.8%), diverticula (n 5 3, 11.1%), and ulcers (n 5 2, 7.4%) were diagnosed. In total, 23 lesions were susceptible to endoscopy and could be treated by enteroscopy in more than three-quarters of the cases.

Table 3.1 Most frequent etiologies of obscure gastrointestinal bleeding (OGIB). OGIB in small intestine Under 50 years of age • Inflammatory bowel disease • Meckel’s diverticulum • Dieulafoy lesion • Hereditary syndromes • Rendu Osler Weber • Ehlers Danlos Older than 50 years • Angiodysplasias • Injuries secondary to NSAIDs • Small intestine tumors • Small intestine diverticula • Enteropathy due to portal hypertension • Ectopic varicose veins • Kaposi sarcoma NSAIDs: non-steroidal anti-inflammatory drugs.

Indications in capsule endoscopy

FIGURE 3.1 Active bleeding from an injury of a vascular lesion.

The diagnostic yield of CE increases if it is performed near the bleeding episode [14]. The synergism between CE and enteroscopy is unquestionable, and clinical management in a specialized unit in SI can improve diagnostic performance. The indication of an urgent CE and the possibility of viewing in real time, in selected cases and with open and severe OGIB, requires highly trained health teams and agile accessibility to therapeutic techniques such as urgent enteroscopy and urgent selective arteriography.

Crohn’s disease CE plays an important role in the assessment of CD [15]. It has shown a high diagnostic yield in patients with suspected CD and patients with known CD and can be particularly useful in identifying superficial mucosal lesions not detected in radiological studies or by conventional endoscopy. In the 2004 European Society of Gastrointestinal Endoscopy (ESGE) clinical guide for the use of CE [16] the study of inflammatory bowel disease (IBD) appeared to be one of the recently accepted indications for CE. In the guidelines published in 2006 [17] by both the ESGE and the ASGE (American Society of Gastrointestinal Endoscopy), IBD is presented as the second accepted indication after OGIB.

25

26

CHAPTER 3 Capsule endoscopy: wide clinical scope

FIGURE 3.2 Active bleeding from a distal small bowel loop.

Therefore CE plays an important role in CD in several clinical scenarios— suspicion of the disease, extent of the disease, and monitoring of the treatment of CD.

Suspected Crohn’s disease CD is a chronic inflammatory disease characterized by mucosal and transmural inflammation of any segment of the gastrointestinal tract. Up to 66% of cases present involvement of the SI at the time of diagnosis, and 30% of patients present exclusive involvement of the SI. In approximately 90% of patients with CD of the SI, the involvement includes the terminal ileum; therefore ileocolonoscopy is considered the first-line examination when suspected of CD and is sufficient to establish the diagnosis in most patients. However, the absence of lesions in the terminal ileum does not rule out the diagnosis of CD since, in 10% of patients, the involvement is exclusive to the jejunum/proximal ileum, which makes the use of other methods of SI study necessary. The lesions compatible with CD visualized by CE include, according to the Capsule Endoscopy Structured Terminology (CEST), aphthae, erosions, ulcers (Fig. 3.4), pseudopolyps, fistulas, and strictures. Other lesions, such as erythema,

Indications in capsule endoscopy

FIGURE 3.3 Active bleeding from an angiodysplasia.

nodularity, denudation, or petechiae, are not considered to be related to inflammation of the mucous membranes. Even so, the imaging findings of CE are not pathognomonic of CD and should be adequately assessed in the clinical context and analytical markers.

Extent of Crohn’s disease In patients with known CD, regardless of the findings in the ileocolonoscopy, it is recommended to perform additional studies to evaluate the extent of the disease since the finding of proximal lesions has prognostic and management implications in the follow-up of these patients. In the different guidelines, the initial performance of radiological studies such as CT enterography or MR enterography is recommended, and the use of CE is reserved for patients with inconclusive radiological findings.

Monitoring Crohn’s disease activity Currently, we apply two endoscopic indices that quantify the inflammatory activity of CD by CE: the CE CD activity index (CECDAI) and the Lewis score. Both have been prospectively validated and allow the objective evaluation of the severity of the disease.

27

28

CHAPTER 3 Capsule endoscopy: wide clinical scope

FIGURE 3.4 Ulcer in the jejunum: Kaposi syndrome.

The CECDAI divides the SI into two segments: proximal and distal. In each of them, inflammation, extent of the disease, and the presence of strictures were evaluated, as shown in Table 3.2. The total score is the sum of both segments, which ranges between 0 and 35. There is no threshold. Specifically, the higher the score, the greater the severity. Score in both segments (proximal and distal): (A 3 B) 1 C Total score: Proximal segment score 1 distal segment score. The Lewis score divides the SI into three equal parts. Each of them quantifies the presence of edema, ulcers, and strictures, as shown in Table 3.3. With the score obtained, a degree of activity can be established: score ,135 indicates normal mucosa or nonsignificant inflammation, between 135 and 790 indicates moderate inflammation, and $ 790 indicates severe inflammation. This score has been used more widely in clinical practice as an automatic calculation tool available in the Rapid Reader workstation of PillCam capsules. Mucosal healing has been related to the increase in the rate of remission of the disease without steroids, a longer time to relapse after the withdrawal of the drug, and a lower rate of hospitalization and surgery. Due to the impact on the course of CD, achieving mucosal healing is one of the objectives in the treatment and management of these patients.

Indications in capsule endoscopy

Table 3.2 Capsule endoscopy Crohn’s disease activity index (CECDAI) activity index. CECDAI activity index Inflammation

Extent

Stricture

0: 1: 2: 3: 4: 5: 0: 1: 2: 3: 0: 1: 2: 3:

None Mild-to-moderate edema/hyperemia/denudation Severe edema/hyperemia/denudation Bleeding, exudate, aphthae, erosion, small ulcer (,0.5 cm) Median ulcer (0.5 2 cm), pseudopolyp Large ulcer ( . 2 cm) No involvement/Normal examination Focal disease (single segment affected) Patchy disease (2 3 affected segments) Diffuse disease ( . 3 affected segments) None Single-passed Multiple-passed Obstruction

Table 3.3 Lewis score. Lewis score Lesions assessed in the proximal, middle, and distal third of the small intestine • Appearance of the villi score A. Normal: 0; Edema 1 B. Short segment (,10%): 8; Long segment (11% 50%): 12; Entire third (. 50%): 20 C. Simple involvement: 1; Patchy: 14; Diffuse: 17 • Ulcers score A. None: 0; Single: 3; Few (2 7): 5; Multiple ($ 8): 10 B. Short segment (,10%): 5; Long segment (11% 50%): 10; Entire third (. 50%): 15 C. C. 1/4: 9; 1/4 1/2: 12; . 1/2: 18 (Proportion of image occupied by largest ulcer) Stricture (global assessment of entire small intestine) 1. None: 0 Single: 14 Multiple: 20 2. Nonulcerated: 2 Ulcerated: 24 3. Nonstricture: 7 Stricture: 10 Calculation of the score: • Score of each third: Appearance of villi (A 3 B 3 C) 1 Ulcers (A 3 B 3 C) • Stricture score (A 3 B 3 C) Final Lewis score: Score of the most affected third 1 Stricture score

Ileocolonoscopy continues to be the gold standard for the assessment of mucosal healing since it provides a direct assessment of the healing. However, the limitation of Ileocolonoscopy is that it is not able to assess the proximal sections of

29

30

CHAPTER 3 Capsule endoscopy: wide clinical scope

SI. Other noninvasive methods, such as fecal calprotectin, have shown a good correlation with mucosal activity, but this correlation is decreased in patients with SI involvement. Likewise, imaging tests for the study of SI, such as CT enterography and MR enterography, may not correctly visualize incipient mucosal lesions, mainly in proximal sections. Therefore CE can play an important role in the assessment of mucosal healing in patients with CD. Currently, the impact of assisted artificial intelligence in CE is being evaluated and has been evaluated for the detection of lesions in CD with SI involvement, showing improvements in both the diagnosis and monitoring of the disease and assessment of mucosal healing [18,19]. The CNN systems can accurately differentiate SI strictures from ulcerated lesions in the entire range of involvement and severity [20], facilitating the automated detection and classification of CE findings in CD. The main problem of CE in patients with CD is that since they can have SI strictures, which are sometimes asymptomatic, the CE can be retained in the digestive tract for more than 2 weeks after ingestion. This is a design problem due to the size of the CE. Therefore in patients with pathologies that could present SI strictures such as CD, we can previously use a similar but resorbable capsule called the Agile Patency capsule, which has a similar shape like that of the conventional capsule and is designed to discard the SI stricture. The Agile Patency capsule is composed of a central rod detectable by radiofrequency and a body formed mainly by lactose covered with a polyamide film. At each end there is a temporary counter and an uncovered area where disintegration begins at 36 40 h, leaving only the central rod (1.6 mm in diameter), which is capable of traversing very small stenoses. It does not have the capacity to capture images. For the realization of the Agile capsule, no previous preparation is necessary. After ingestion the patient is instructed to observe the stool and collect the capsule when it is expelled for delivery to the designated people for evaluation. If it is expelled intact, it is considered that there is no contraindication for the performance of the conventional endoscopic capsule.

Evaluation of refractory celiac disease In patients with celiac disease and with good control of their disease, asymptomatic, and with antitransglutaminase antibodies within normal levels, endoscopic follow-up would not be indicated. However, endoscopic follow-up could be considered in those who present clinical symptoms, unexplained iron deficiency anemia, or in the cases of primary CE diagnosis. In the context of celiac disease, it is recommended to conduct a study with CE in different scenarios [21]: •

Patients with celiac disease and unjustified alarm symptoms such as anemia, weight loss, or bleeding; and in patients with refractory celiac disease, mainly

Indications in capsule endoscopy

• •

type II, to rule out complications such as lymphoma and ulcerative jejunitis, or the possible relationship with other enteropathies such as CD Suspected CD and impossibility or refusal of the patient to perform EGD Cases with positive serology and normal histology since they allow the assessment of more distal sections in search of villous atrophy

The main value of CE in the context of celiac disease is to detect complications in patients who present alarm symptoms or persistence of clinical manifestations despite a strict gluten-free diet. If CE shows pathological findings (Fig. 3.5) enteroscopy will be indicated to rule out the appearance of ulcerative jejunitis, lymphomas, and other neoplasms. Currently, there is not enough data to show the profitability of endoscopic capsules in the study of the extent of celiac disease or in the assessment of the response to a gluten-free diet.

Screening of polyposis syndromes: familial adenomatous polyposis and Peutz Jeghers syndrome Familial adenomatous polyposis Screening and follow-up of SI lesions in patients with FAP are important in patients with duodenal lesions with Spiegelman classification III/IV, since they

FIGURE 3.5 Jejunojejunal intussusception in a patient with celiac disease and recurrent episodes of abdominal pain.

31

32

CHAPTER 3 Capsule endoscopy: wide clinical scope

have a greater tendency to have SI lesions, especially in the proximal sections [22]. The ESGE recommends a review of the upper digestive tract with frontal and lateral vision endoscopes at the time of diagnosis [23]. However, the potential for malignancy of adenomatous lesions affecting the proximal sections of the SI [24 26] has yet to be determined, which seems to be of low significance, except in patients with ileostomies and ileoanal anastomoses. Therefore there is little evidence in determining the follow-up intervals with CE and when to indicate enteroscopy. CE has shown superiority with respect to radiological techniques for the detection of polypoid lesions [27 29] (Fig. 3.6). Regarding the diagnosis of more distal lesions, MR enterography and CE have shown similar detection rates in polyps .15 mm in diameter in patients with polyposis syndromes. However, CE has a higher detection rate of 5-mm polyps, and MR enterography presented the advantage of a better localization and measurement of the lesions, as well as extramural information that may be of interest for patients with FAP due to the possibility of desmoid tumors. In conclusion, in patients with FAP with Spiegelman III/IV, when the study of SI is clinically indicated, distal lesions of SI can be evaluated with CE or by imaging with MR enterography.

FIGURE 3.6 Ileal villous adenoma with abundant mucus.

Indications in capsule endoscopy

Peutz Jeghers syndrome Patients with PJS have polypoid lesions in SI in 96% of cases [30], so early diagnosis and periodic monitoring should be recommended. In patients with PJS, screening for SI lesions is indicated with the main objective of reducing the rate of complications related to polyps at this level, especially intussusception and hemorrhage. There is no unanimous agreement regarding the most appropriate diagnostic technique since the diagnostic sensitivity of CT enterography, MR enterography, and CE depend largely on the experience of each center, availability of the technique, and patient preference [23]. For the early detection of polypoid lesions smaller than 10 mm in PJS, the most sensitive technique, superior to studies with barium and MR enterography, is CE [29,31 33]. A retrospective study conducted in tertiary centers evaluated the correlation between CE and enteroscopy in patients with SI polyposis by PJS, showing that CE is a very useful tool to define the number, location, and size of polyps, helping to predict the difficulty of polypectomy during enteroscopy [34]. However, enteroscopy seems to be more sensitive than CE in defining the size and location of lesions [34]. It is recommended to monitor patients with PJS with CE or MR enteroscopy based on the availability, experience of the center, and patient preference every 2 3 years (Fig. 3.7).

Suspected small intestine tumors SI tumors are rare and represent between 3% and 6% of all gastrointestinal tumors and only 1% 3% of malignant neoplasms. The most frequent histological subtypes are adenocarcinoma (30% 45%), neuroendocrine tumors (24% 40%), lymphomas (10% 20%), and sarcomas (10% 15%). The clinical manifestations are not very specific, which can delay the diagnosis, and are usually diagnosed in the context of the study of an open or occult OGIB or as findings in a radiological imaging study. The use of CE for the detection of SI tumors is recommended in the context of open OGIB or iron deficiency anemia (Fig. 3.8). In the case of suspicion of a tumor visualized in previous imaging tests, an enteroscopy should be considered. False negatives of CE in the diagnosis of SI tumors may be due to poor preparation, rapid transit, or lower detection capacity of submucosal tumors that can be confused with benign protrusions of the mucosa. To differentiate submucosal tumors from benign folds and increase the detection rate, SPICE score (smooth, protruding lesions index on CE) has been described in Table 3.4.

Graft versus host disease Allogeneic hematopoietic stem cell transplantation (AHSCT) has been a widely used treatment in recent years for the management of various hematological

33

34

CHAPTER 3 Capsule endoscopy: wide clinical scope

FIGURE 3.7 Jejunojejunal intussusception in a patient with Peutz Jeghers syndrome.

diseases. The disparity between donor and recipient antigens can produce a reaction mediated by donor T lymphocytes against certain recipient tissues GVHD, which is a major cause of morbidity and mortality (10% 90% depending on the source of progenitors, type of donor, and type of AHSCT). The acute form (3 months after AHSCT) mainly affects the skin, liver, and digestive tract (SI GVHD), and the chronic form is multiorgan [35]. SI GVHD occurs in 30% 80% of cases, between 15% and 20% in its severe form [35], with SI involvement being the most refractory to treatment [36]. SI is the organ most affected by gastrointestinal GVHD. Its clinical, anatomical, and endoscopic presentation is heterogeneous and often nonspecific, which makes its diagnosis difficult. The diagnosis of SI GVHD (Fig. 3.9) makes it essential to initiate adequate treatment, but it can be difficult due to its location and nonspecificity in the clinic. The differential diagnosis should be performed with the toxicity of immunosuppressive drugs, the toxicity of the conditioning itself, and infectious causes such as cytomegalovirus or Clostridium difficile [37]. CE has been described as a useful and complementary tool adding to the conventional endoscopic study in these cases, with a diagnostic yield of 54%.

Indications in capsule endoscopy

FIGURE 3.8 Previous patient (Fig 3.7), varicose veins near a large polypoid formation (right).

Table 3.4 Spice index. Spice index Poor definition of boundary with surrounding mucosa Diameter greater than height Intestinal lumen visible in images in which it appears Image of lesion lasts more than 10 min

No 1 1 0 0

Yes 0 0 1 1

Calculation of SPICE: Sum of the scores of the four sections. SPICE $ 2 higher probability of submucosal tumor. SPICE: Smooth, protruding lesions index on capsule endoscopy.

However, the data in the literature are scarce, and the diagnostic accuracy of this technique is not well defined. CE may have a relevant role in the early diagnosis of SI GVHD and in the subsequent management of these patients since it directs the insertion route of enteroscopy to obtain an anatomopathological diagnosis and informs the extent of the disease. According to these results and previous studies in the literature [38,39] CE could be considered the new gold standard in the diagnosis of mild

35

36

CHAPTER 3 Capsule endoscopy: wide clinical scope

FIGURE 3.9 Graft versus host disease.

forms of GVHD or in forms of proximal involvement, especially when the lesions are more evident (multiple ulcers or strictures), regardless of the histological study. It is important to define the ideal time of its performance to increase its sensitivity since this disease is very dynamic, which can lead to false negatives. In this sense, CE can lead to an earlier diagnosis than histopathological study, which could translate into a therapeutic impact in these patients who often present severe forms. Finally, the unification of diagnostic criteria by CE in GVHD is an increasing a priority for which artificial intelligence may have an answer in the near future by identifying endoscopic patterns of the disease.

Capsule endoscopy clinical scope in pediatrics Over the last years CE established itself as a valuable endoscopic modality in the adult population. This modality has gained a special interest in pediatrics rapidly due to its minimal invasiveness and not requiring anesthesia or ionizing radiation, representing a great advantage over other imaging modalities. However, available data is not as extensive as in the adult population [40].

Capsule endoscopy clinical scope in pediatrics

In 2001 CE was approved by the FDA for small bowel evaluation in adults. In 2004 CE use was expanded to the pediatric population, being approved for patients 10 18 years of age. Later, in 2009, its use was also approved for 2 years of age and older. In the same year, a patency capsule was also approved for the same population [41]. Since then, CE use has expanded progressively and is nowadays considered an important diagnostic method in pediatric gastroenterology. More recently, panenteric CE has also been developed, aiming to evaluate both the small and large bowel in a single procedure [42].

Indications Indications for CE are similar in children and adults. International guidelines approved CE use for the evaluation of OGIB (including iron deficiency), CD, polyposis syndromes, small bowel tumors, and malabsorptive syndromes [23 43]. Despite similar indications, the relative frequency of those indications is substantially different between children and adults [40] (Table 3.5).

Occult gastrointestinal bleeding As with the adult population, OGIB accounts only for 5% of cases of overall gastrointestinal bleeding in children (overt and occult) [23]. After negative endoscopy and colonoscopy, OGIB requires a small bowel investigation usually with CE. OGIB is also a frequent clinical indication for CE in the pediatric population, being the second most common indication for the procedure (15%) after IBD (63%). However, this varies according to age. In children under 8 years of age OGIB remains the most common indication for CE (36%) [44]. Balloon-assisted enteroscopy (BAE) is another endoscopic method for small bowel investigation in Table 3.5 Indications for capsule endoscopy. Pediatric patients

Patients under 8 years of age

Adult patients

15

36

66

63

24

10

10 8 4 61

14

11 3 10 59

Indications (%) Bleeding and/or anemia Inflammatory bowel disease Abdominal pain Polyps/neoplasms Other Positive findings (%)

25 67

Source: Adapted from Cohen SA. The potential applications of capsule endoscopy in pediatric patients compared with adult patients. Gastroenterol Hepatol (N Y) 2013;9(2):92 7.

37

38

CHAPTER 3 Capsule endoscopy: wide clinical scope

this population, usually performed as a second-line intervention because of its high invasiveness and having the potential advantage of both diagnostic and therapeutic properties [45]. Most available data is derived from retrospective studies, with a wide heterogeneity between the studies’ design. In 2011 a metaanalysis reported a diagnostic yield of 42.4%, further validated in subsequent studies [46]. Some factors may influence the diagnostic yield of CE. Active bleeding and performing the procedure within 2 weeks after bleeding (ideally in the first 3 days) were associated to a higher diagnostic yield [47]. The most frequent CE findings in children evaluated for OGIB are vascular lesions (34.0%), followed by IBD (25.5%), polyps (8.5%), and gastroduodenal lesions not previously identified in upper endoscopy (5.7%). More rare findings (,5%) include nonspecific enteritis, tumors, Meckel’s diverticulum, and villous atrophy.

Inflammatory bowel disease The peak incidence of IBD occurs between 15 and 30 years of age. However in up to 10% of patients IBD presents during childhood or adolescence. Differently from adults, children are more likely to present with more severity due to extensive intestinal involvement and rapid clinical progression [48]. However, ileocolonoscopy with terminal ileum biopsies remains the gold standard for CD diagnosis. Nevertheless, CE is a valuable complementary diagnostic method in IBD, particularly for CD and in cases of unclassified IBD. Evaluation of suspected or established CD is the most common pediatric indication for CE in children. Similar to adults, CE also has an important role in children for CD, evaluation of the extension of small bowel involvement, and treatment monitoring [49,50]. Compared with other diagnostic modalities, CE may be superior to MRE for the detection of early mucosal injuries and proximal bowel lesions [51]. In cases of pediatric IBD with pure colonic involvement, the classification as CD or as ulcerative colitis can be very challenging. CE offers the possibility to identify small bowel lesions allowing the reclassification as CD in a significant number of patients [52]. Several endoscopic scoring systems have been developed to diagnose CD and monitor treatment response. Lewis score and CECDAI are the two main scores available for CD [53,54]. Although both scores have been used in the pediatric population, CECDAI seems to be better reproducible for the assessment of intestinal inflammation [55]. More recently, Oliva et al. developed the CE CD index [56]. This novel score considers the number of ulcers, size of the largest ulcer, affected surface, and the presence or absence of stenosis in both the small and large intestines (Table 3.6). In a study involving 312 children, CE CD presented as a simple, reliable, and reproducible score for evaluation of SB inflammation [56].

Capsule endoscopy clinical scope in pediatrics

Table 3.6 Capsule endoscopy Crohn’s disease index. Parameter

Score

Number of ulcers

None—0 1 3—1 4 10—2 . 10—3 None—0 Aphthous—1 , 1/4 image size—2 . 1/4 image size—3 Unaffected segment—0 # 10%—1 11% 50%—2 . 50%—3 None—0 Single: passed—1 Multiple: passed—2 Obstruction—3

Size of the largest ulcer

Surface involved

Stenosis

The usefulness of panenteric CE has been addressed for CD. A 2019 prospective study with children concluded that panenteric CE is a reliable method to monitor mucosal healing and deep remission in a treat-to-target strategy for pediatric patients with CD [57].

Polyposis syndromes PJS is the most common polyposis syndrome during childhood. PJS is an inherited autosomal dominant syndrome characterized by multiple gastrointestinal polyps, mucocutaneous pigmentation, and cancer predisposition [58]. Similarly, in the adult population, most GI polyps are located in the small bowel (50%), followed by the stomach (36%), and the colon (21%) [59]. Thus in addition to upper endoscopy and colonoscopy, CE is a valuable diagnostic tool for the diagnosis of small bowel polyps and tumors. International guidelines recommend small bowel surveillance with CE or MRE no later than 8 years of age, earlier if symptomatic, and repeated every 3 years [60]. However, the reported polyp detection rate is higher with CE compared with MRE for smaller polyps (,15 mm) [37]. Endoscopic polypectomy should be offered for small bowel polyps larger than 1.5 2 cm or smaller polyps if symptomatic [60]. Some drawbacks of CE in PJS are difficulty or inability of younger children (,8 years) to swallow the capsule; lower detection rate for polyps in the duodenum and proximal jejunum due to fast transit; frequent bubbles in proximal SB; inability to accurately localize polyps in the SB and poorly assess the polyp size. Capsule retention risk is not higher compared to other indications [61].

39

40

CHAPTER 3 Capsule endoscopy: wide clinical scope

Contrary to adults, pediatric guidelines do not make any recommendation regarding CE utility for small bowel evaluation of patients with FAP [62].

Other indications CE may be helpful for other childhood disorders, despite the lack of robust data supporting its use. In patients with refractory celiac disease, CE may be helpful to exclude complications such as ulcerative jejunitis or small bowel lymphoma [63], but these are not usual complications in pediatric patients. Despite its proven performance as good sensitivity and specificity for celiac disease diagnosis, it is not recommended for diagnosis purposes [64]. Some authors report that it may be useful in children who refuse upper endoscopy [65]. CE may also be useful for the evaluation of protein-losing enteropathy of unknown etiology. Primary intestinal lymphangiectasia or CD could be diagnosed in this clinical context [66]. Early diagnosis of gastrointestinal GVHD after AHSCT can be achieved with CE, leading to early treatment modifications [67]. Despite a nonnegligible percentage of CE that are performed for abdominal pain evaluation, CE is not routinely recommended in the absence of other clinical and/or laboratory findings [68]. In patients with negative inflammatory markers, the diagnostic yield is very low (21.4%) [69].

Limitations of endoscopic capsule Despite the advantage of recording even longer videos than with conventional endoscopy, there are lesions or structures that are difficult to explore, due to their morphology or anatomy, without directing the optics to them in real time. This is the case, for example, of the papilla in the duodenum, which requires a lateral approach, although it can sometimes be detected with CE (Figs. 3.10 and 3.11), or the difficulty of detecting a double lumen as occurs with Meckel’s diverticulum, although present in 2% of humans, with a much lower sensitivity of CE for its detection (Fig. 3.12). The appendix in the colon is generally outside the scope of the CE optics in this larger organ, although it can sometimes be identified (Figs. 3.13 and 3.14). It is also possible to identify foreign bodies within the SI lumen, such as clips placed by enteroscope (CE indication for control) (Fig. 3.15) and others, such as catheters that have penetrated SI as a complication after placement (Fig. 3.16).

Challenges in pediatrics Swallowing the capsule A patient’s ability to swallow the capsule may be challenging at any age, but this is especially so in young children and in the elderly. A stepwise approach in which children swallow progressive larger candies may be used to teach them to

Limitations of endoscopic capsule

FIGURE 3.10 Normal Vater’s papilla.

overcome this difficulty [70]. Despite this, some children remain unable or unwilling to swallow capsules. In other cases, patients with swallow and/or motility disorders are also unable to swallow capsules. To overcome this problem, capsules can be placed directly into the duodenum through upper endoscopy. The placement of capsules into the duodenum instead of the stomach is preferable to avoid the risk of incomplete small bowel visualization due to prolonged gastric retention and consequent delay in the passage to the duodenum. A dedicated disposable CE delivery device (AdvanCE, US Endoscopy), consisting of a catheter with a sheath diameter of 2.5 mm that can be preloaded through the operative channel of a standard gastroscope, may be used for this purpose [71]. Polypectomy snares or foreign body baskets can also be used as an alternative, although tracheal capsule aspiration and mucosal trauma are known complications [72].

Bowel cleansing The diagnostic yield of CE is directly affected by small bowel cleansing due to the inability to flush or suction fluids or gas during the procedure [73]. The optimal preparation regimen (purgative agents vs clear liquid fasting diet) remains

41

42

CHAPTER 3 Capsule endoscopy: wide clinical scope

FIGURE 3.11 Vater’s papilla after ampullar resection.

unclear [74]. When purgative agents are used, the timing of bowel preparation may also be important, achieving better results if ingested in a short period before CE or even after the capsule reaches the small bowel [75]. A fasting period of 12 h, or more, is also associated with better mucosal visualization [76]. Unfortunately, only one prospective study addresses this problem in the pediatric population. A randomized clinical trial by Oliva et al. concluded that a low dose of polyethylene glycol solution (25 mL/kg) in the evening before the CE plus 20 mL (376 mg) of oral simethicone 30 min before the procedure improved the visibility of small bowel mucosa [77].

Capsule retention The most frequent complication of CE in pediatrics is capsule retention, defined as missed expulsion of the capsule within 2 weeks after the procedure or the need for intervention before this time. In the case of high suspicion or known small bowel obstruction, CE is contraindicated. The presence of previous abdominal surgery, radiotherapy, and CD are associated with the increased risk of capsule retention.

Limitations of endoscopic capsule

FIGURE 3.12 Ileal Meckel´s diverticulum.

FIGURE 3.13 Orifice of the vermiform appendix.

43

44

CHAPTER 3 Capsule endoscopy: wide clinical scope

FIGURE 3.14 Appendix orifice with inflammatory changes.

An analysis from Cohen et al., involving a metaanalysis [46] and two prospective studies [78,79], reported an overall capsule retention rate of 2.3% in children (n 5 22/1013 procedures). Small bowel retention occurred in most cases (n 5 18), followed by gastric retention (n 5 4). Endoscopic retrieval was possible in 5 cases, including all the cases of gastric retention and one case of retention in an ileal pouch. Nevertheless, thirteen cases required surgical removal. Previous studies described a similar risk of capsule retention between children and adults for different clinical indications (OGIB: 1.4% vs 1.2%, CD: 2.2% vs 2.6%, polyps/neoplastic lesions: 1.2% vs 2.1%), suggesting that clinical indication, more than age, may influence the retention rate [40,80]. More recently, a metaanalysis regarding CD reported pediatric retention rates lower than in adults (1.64% vs 3.49%, respectively) [81]. Capsule patency (Agile; Given Imaging) was an important advance in this field. A retrospective (n 5 23) and a prospective (n 5 18) study involving pediatric patients with suspected IBD allowed to perform CE successfully, without retention, in all cases except for one patient [78,82]. Thus capsule patency is a useful procedure since it allows the identification of patients at risk for capsule retention, particularly in known or suspected CD.

Limitations of endoscopic capsule

FIGURE 3.15 A clip in a lesion: capsule endoscopy control after enteroscopy.

FIGURE 3.16 Foreign body: lumboperitoneal catheter perforated to small intestine.

45

46

CHAPTER 3 Capsule endoscopy: wide clinical scope

The main drawback of CE in children could be difficulty in swallowing, as previously discussed, and the risk of capsule retention related to patient age and weight since there are no studies regarding the lowest age for CE use [65].

Conclusions •

•

•

•

•

•

•

CE allows a direct and noninvasive exploration of the SI mucosa with a low rate of complications, which has been a great advance in the study of pathologies at this level and has positioned it as the exploration of choice in certain situations. The main indications are the study of gastrointestinal bleeding of obscure origin and IBD, which seems to have an increasingly important role not only in the diagnosis but also in the follow-up. Other important indications are the monitoring of polyposis syndromes and the suspicion of SI tumors, where CE has been shown to be useful in the early diagnosis of polypoid and tumor lesions. The role of CE in GVHD is more limited, but it can be relevant in early diagnosis, informing the extent, and in directing biopsy samples for histological study in patients with distal SI involvement. Indications for CE are similar between children and adults. Nevertheless, the relative frequency of those indications is substantially different between children and adults. The limitations in the diagnosis of CE are fundamentally due to its lack of orientation in real time to identify lateralized anatomical structures or with more detailed pathologies such as vascular ones. Despite being described as a minimally invasive endoscopic procedure, CE may present some challenges in children.

References [1] Iddan G, Meron G, Glukhovsky A, Swain P. Wireless capsule endoscopy. Nature 2000;405:417. [2] Friedrich K, Gehrke S, Stremmel W, Sieg A. First clinical trial of a newly developed capsule endoscope with panoramic side view for small bowel: a pilot study. J Gastroenterol Hepatol 2013;28(9):1496 501. [3] Pioche M, Vanbiervliet G, Jacob P, Duburque C, Gincul R, Filoche B, et al. Prospective randomized comparison between axial- and lateral-viewing capsule endoscopy systems in patients with obscure digestive bleeding. Endoscopy 2014;46 (6):479 84. [4] Chetcuti Zammit S, McAlindon ME, Sidhu R. Panoramic vs axial small bowel capsule endoscopy in overt obscure gastrointestinal bleeding. Dig Endosc 2020;32(5):823.

References

[5] Branchi F, Ferretti F, Orlando S, et al. Small-bowel capsule endoscopy in patients with celiac disease, axial vs. lateral/panoramic view: results from a prospective randomized trial. Dig Endosc 2020;32(5):778 84. [6] Lewis BS. Small intestinal bleeding. Gastroenterol Clin North Am 1994;23:67 91. [7] Teshima CW, Kuipers EJ, van Zanten SV, et al. Double balloon enteroscopy and capsule endoscopy for obscure gastrointestinal bleeding: an updated meta-analysis. J Gastroenterol Hepatol 2011;26:796 801. [8] Enns RA, Hookey L, Armstrong D, et al. Clinical practice guidelines for the use of video capsule endoscopy. Gastroenterology 2017;152:497 514. [9] Aoki T, Yamada A, Kato Y, Saito H, Tsuboi A, Nakada A, et al. Automatic detection of blood content in capsule endoscopy images based on a deep convolutional neural network. J Gastroenterol Hepatol 2020;35(7):1196 200. [10] Soffer S, Klang E, Shimon O, Nachmias N, Eliakim R, Ben-Horin S, et al. Deep learning for wireless capsule endoscopy: a systematic review and meta-analysis. Gastrointest Endosc 2020;92:831 9. [11] Vieira PM, Freitas NR, Lima VB, Costa D, Rolanda C, Lima CS. Multi-pathology detection and lesion localization in WCE videos by using the instance segmentation approach. Artif Intell Med 2021;119:102141. [12] Iakovidis DK, Georgakopoulos SV, Vasilakakis M, Koulaouzidis A, Plagianakos VP. Detecting and locating gastrointestinal anomalies using deep learning and iterative cluster unification. IEEE Trans Med Imaging 2018;37:2196 210. [13] Pe´rez-Cuadrado Robles E, Bebia Conesa P, Esteban Delgado P, et al. Emergency double-balloon enteroscopy combined with real-time viewing of capsule endoscopy: a feasible combined approach in acute overt-obscure gastrointestinal bleeding? Dig Endosc 2015;27:338 44. [14] Ribeiro I, Pinho R, Rodrigues A, et al. Obscure gastrointestinal bleeding: which factors are associated with positive capsule endoscopy findings. Rev Esp Enferm Dig 2015;107:334. [15] Luja´n-Sanchis M, Sanchis-Artero L, Larrey-Ruiz L, Pen˜o-Mun˜oz L, Nu´n˜ez-Martı´nez P, Castillo-Lo´pez P, et al. Current role of capsule endoscopy in Crohn’s disease. World J Gastrointest Endosc 2016;8(17):572 83. [16] Rey JF, Gay G, Kruse A, Lambert RESGE Guidelines Committee. European Society of Gastrointestinal Endoscopy guideline for video capsule endoscopy. Endoscopy 2004;36(7):656 8. [17] Rey JF, Ladas S, Alhassani A, Kuznetsov KESGE Guidelines Committee. European Society of Gastrointestinal Endoscopy (ESGE). Video capsule endoscopy: update to guidelines (May 2006). Endoscopy 2006;38(10):1047 53. [18] Klang E, et al. Deep learning algorithms for automated detection of Crohn’s disease ulcers by video capsule endoscopy. GIE 2020;91:606 13. [19] Klang E, et al. Automated detection of Crohn’s disease intestinal strictures on capsule endoscopy images using deep neural networks. J Crohns Colitis 2021;15(5):749 56. [20] Barash Y, et al. Ulcer severity grading in video capsule images of patients with Crohn’s disease: an ordinal neural network solution. GIE 2021;93:187 92. [21] Perez-Cuadrado-Robles E, Lujan-Sanchis M, Elli L, Juan Martinena-Fernandez JF, Garcıa-Lledo J, Ruano-Dıaz L, et al. Role of capsule endoscopy in alarm features and non-responsive celiac disease: a European multicenter study. Dig Endosc 2018;30(4):461 6.

47

48

CHAPTER 3 Capsule endoscopy: wide clinical scope

[22] Sulbaran M, Campos FG, Ribeiro Jr U, et al. Risk factors for advanced duodenal and ampullary adenomatosis in familial adenomatous polyposis: a prospective, singlecenter study. Endosc Int Open 2018;6(5):E531 40. Available from: https://doi.org/ 10.1055/a-0577-2650. [23] Pennazio M, Spada C, Eliakim R, et al. Small-bowel capsule endoscopy and deviceassisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Clinical Guideline. Endoscopy 2015;47(04):352 86. Available from: https://doi.org/10.1055/s-0034-1391855. [24] Ruys AT, Alderlieste YA, Gouma DJ, et al. Jejunal cancer in patients with familial adenomatous polyposis. Clin Gastroenterol Hepatol 2010;8:731 3. [25] Tescher P, Macrae FA, Speer T, et al. Surveillance of FAP: a prospective blinded comparison of capsule endoscopy and other GI imaging to detect small bowel polyps. Hered Cancer Clin Pract 2010;8:3. [26] Katsinelos P, Kountouras J, Chatzimavoudris G, et al. Wireless capsule endoscopy in detecting small intestinal polyps in familial adenomatous polyposis. World J Gastroenterol 2009;15:6075 9. [27] Rondonotti E, Pennazio M, Toth E, et al. Small-bowel neoplasms in patients undergoing video capsule endoscopy: a multicenter European study. Endoscopy 2008;40:488 95. [28] Plum N, May A, Manner H, et al. Small-bowel diagnosis in patients with familial adenomatous polyposis: comparison of push enteroscopy, capsule endoscopy, ileoscopy, and enteroclysis. Z Gastroenterol 2009;47:339 46. [29] Caspari R, von Falkenhausen M, Krautmacher C, et al. Comparison of capsule endoscopy and magnetic resonance imaging for the detection of polyps of the small intestine in patients with familial adenomatous polyposis or with Peutz Jeghers’ syndrome. Endoscopy 2004;36:1054 9. [30] Schreibman IR, Baker M, Amos C, McGarrity TJ. The hamartomatous polyposis syndromes: a clinical and molecular review. Am J Gastroenterol 2005;100(February (2)):476 90. [31] Postgate A, Hyer W, Phillips R, et al. Feasibility of video capsule endoscopy in the management of children with Peutz Jeghers syn- drome: a blinded comparison with barium enterography for the detection of small bowel polyps. J Pediatr Gastroenterol Nutr 2009;49:417 23. Available from: https://doi.org/10.1097/MPG.0b013e31818f0a1f. [32] Brown G, Fraser C, Schofield G, et al. Video capsule endoscopy in Peutz Jeghers syndrome: a blinded comparison with barium follow-through for detection of smallbowel polyps. Endoscopy 2006;38:385 90. [33] Mata A, Llach J, Castells A, et al. A prospective trial comparing wireless capsule endoscopy and barium contrast series for small-bowel surveillance in hereditary GI polyposis syndromes. Gastrointest Endosc 2005;61:721 5. [34] Rahmi G, Samaha E, Lorenceau-Savale C, et al. Small bowel polypectomy by double balloon enteroscopy: correlation with prior capsule endoscopy. World J Gastrointest Endosc 2013;5:219 25. [35] Ferrara JL, Levine JE, Reddy P, et al. Graft-vs-host disease (Review). Lancet 2009;373:1550 61. [36] Gooley TA, Chien JW, Pergam SA, et al. Reduced mortality after allogeneic hematopoietic-cell transplantation. N Engl J Med 2010;363:2091 -10.

References

[37] Chakrabarti S, Collingham KE, Stevens R, et al. Isolation of viruses from stools in stem cell transplant recipients: a prospective surveillance study. Bone Marrow Transplant 2000;25:277 82. [38] Pe´rez-Cuadrado-Robles E, Perrod G, Rahmi G. Usefulness of capsule endoscopy in the diagnosis of gastrointestinal graft-vs-host disease. Rev Gastroenterol Mex (Engl Ed) 2021;86(3):213 14. [39] Blanco-Velasco G, Palos-Cuellar R, Dominguez-Garcia MR, et al. Utility of capsule endoscopy in the diagnosis of gastrointestinal graft-vs-host disease. Rev Gastroenterol Mex 2020;86. Available from: https://doi.org/10.1016/j.rgmx.2020. 06.005. [40] Cohen SA. The potential applications of capsule endoscopy in pediatric patients compared with adult patients. Gastroenterol Hepatol (N Y) 2013;9(2):92 7. [41] Cohen SA, Oliva S. Capsule endoscopy in children. Front Pediat 2021;9. [42] Hong SN, et al. Recent advance in colon capsule endoscopy: what’s new? Clin Endosc 2018;51(4):334 43. [43] Wang A, et al. Wireless capsule endoscopy. Gastrointest Endosc 2013;78(6):805 15. [44] Walsh CM. Assessment of competence in pediatric gastrointestinal endoscopy. Curr Gastroenterol Rep 2014;16(8):401. [45] Romano C, et al. Pediatric gastrointestinal bleeding: perspectives from the Italian Society of Pediatric Gastroenterology. World J Gastroenterol 2017;23(8):1328 37. [46] Cohen SA, Klevens AI. Use of capsule endoscopy in diagnosis and management of pediatric patients, based on meta-analysis. Clin Gastroenterol Hepatol 2011;9 (6):490 6. [47] Ge ZZ, et al. Best candidates for capsule endoscopy for obscure gastrointestinal bleeding. J Gastroenterol Hepatol 2007;22(12):2076 80. [48] Vernier-Massouille G, et al. Natural history of pediatric Crohn’s disease: a population-based cohort study. Gastroenterology 2008;135(4):1106 13. [49] IBD Working Group of the ESPGHN. Inflammatory bowel disease in children and adolescents: recommendations for diagnosis the Porto criteria. J Pediatr Gastroenterol Nutr 2005;41(1):1 7. [50] Naspghn, et al. Differentiating ulcerative colitis from Crohn disease in children and young adults: report of a working group of the North American Society for Pediatric Gastroenterology, Hepatology, and Nutrition and the Crohn’s and Colitis Foundation of America. J Pediatr Gastroenterol Nutr 2007;44(5):653 74. [51] Casciani E, et al. MR enterography vs capsule endoscopy in paediatric patients with suspected Crohn’s disease. Eur Radiol 2011;21(4):823 31. [52] Ouahed J, et al. Role of wireless capsule endoscopy in reclassifying inflammatory bowel disease in children. J Pediatr 2013;89(2):204 9. [53] Gralnek IM, et al. Development of a capsule endoscopy scoring index for small bowel mucosal inflammatory change. Aliment Pharmacol Ther 2008;27(2):146 54. [54] Niv Y, et al. Validation of the Capsule Endoscopy Crohn’s Disease Activity Index (CECDAI or Niv score): a multicenter prospective study. Endoscopy 2012;44 (1):21 6. [55] Omori T, et al. Comparison of Lewis score and capsule endoscopy Crohn’s disease activity index in patients with Crohn’s disease. Dig Dis Sci 2020;65(4):1180 8. [56] Oliva S, et al. Assessment of a new score for capsule endoscopy in pediatric Crohn’s disease (CE-CD). Endosc Int Open 2021;9(10):E1480 -e1490.

49

50

CHAPTER 3 Capsule endoscopy: wide clinical scope

[57] Oliva S, et al. A treat to target strategy using panenteric capsule endoscopy in pediatric patients with Crohn’s disease. Clin Gastroenterol Hepatol 2019;17(10):2060 7 e1. [58] van Lier MGF, et al. High cancer risk in Peutz Jeghers syndrome: a systematic review and surveillance recommendations. Am J Gastroenterol 2010;105(6):1258 64 author reply 1265. [59] Goldstein SA, Hoffenberg EF. Peutz Jegher syndrome in childhood: need for updated recommendations? J Pediatr Gastroenterol Nutr 2013;56(2):191 5. [60] Latchford A, et al. Management of Peutz Jeghers syndrome in children and adolescents: a position paper from the ESPGHAN polyposis working group. J Pediatr Gastroenterol Nutr 2019;68(3):442 52. [61] Caspari R, et al. Comparison of capsule endoscopy and magnetic resonance imaging for the detection of polyps of the small intestine in patients with familial adenomatous polyposis or with Peutz Jeghers’ syndrome. Endoscopy 2004;36(12):1054 9. [62] Hyer W, et al. Management of familial adenomatous polyposis in children and adolescents: position paper from the ESPGHAN polyposis working group. J Pediatr Gastroenterol Nutr 2019;68(3):428 41. [63] Culliford A, et al. The value of wireless capsule endoscopy in patients with complicated celiac disease. Gastrointest Endosc 2005;62(1):55 61. [64] Rokkas T, Niv Y. The role of video capsule endoscopy in the diagnosis of celiac disease: a meta-analysis. Eur J Gastroenterol Hepatol 2012;24(3):303 8. [65] Argu¨elles-Arias F, et al. Guideline for wireless capsule endoscopy in children and adolescents: a consensus document by the SEGHNP (Spanish Society for Pediatric Gastroenterology, Hepatology, and Nutrition) and the SEPD (Spanish Society for Digestive Diseases). Rev Esp Enferm Dig 2015;107(12):714 31. [66] Gortani G, Maschio M, Ventura A. A child with edema, lower limb deformity, and recurrent diarrhea. J Pediatr 2012;161(6):1177. [67] Neumann S, et al. Wireless capsule endoscopy for diagnosis of acute intestinal graftvs-host disease. Gastrointest Endosc 2007;65(3):403 9. [68] Bardan E, et al. Capsule endoscopy for the evaluation of patients with chronic abdominal pain. Endoscopy 2003;35(8):688 9. [69] Katsinelos P, et al. Diagnostic yield and clinical impact of wireless capsule endoscopy in patients with chronic abdominal pain with or without diarrhea: a Greek multicenter study. Eur J Inter Med 2011;22:e63 6. Available from: https://doi.org/ 10.1016/j.ejim.2011.06.012). [70] Yoo JH, Tarbox J, Granpeesheh D. Using stimulus fading to teach a young child with autism to ingest wireless capsule endoscopy. Gastrointest Endosc 2008;67 (7):1203 4. [71] Keuchel M, et al. Endoscopic placement of the video capsule with the AdvanCE delivery device. Gastrointest Endosc 2006;63(5):AB185. [72] Fritscher-Ravens A, et al. The feasibility of wireless capsule endoscopy in detecting small intestinal pathology in children under the age of 8 years: a multicentre European study. Gut 2009;58(11):1467 72. [73] Rokkas T, et al. Does purgative preparation influence the diagnostic yield of small bowel video capsule endoscopy?: A meta-analysis. Am J Gastroenterol 2009;104 (1):219 27.

References

[74] Yung DE, et al. Systematic review and meta-analysis: is bowel preparation still necessary in small bowel capsule endoscopy? Expert Rev Gastroenterol Hepatol 2017;11 (10):979 93. [75] Xavier S, et al. Bowel preparation for small bowel capsule endoscopy the later, the better!. Dig Liver Dis 2019;51(10):1388 91. [76] Catalano C, et al. Video capsule endoscopy: is bowel preparation necessary? J Investig Med 2016;64(6):1114 17. [77] Oliva S, et al. Small bowel cleansing for capsule endoscopy in paediatric patients: a prospective randomized single-blind study. Digestive and Liver Disease 2014;46 (1):51 5. [78] Cohen SA, et al. Pediatric capsule endoscopy: review of the small bowel and patency capsules. J Pediatr Gastroenterol Nutr 2012;54(3):409 13. [79] Gralnek IM, et al. Small bowel capsule endoscopy impacts diagnosis and management of pediatric inflammatory bowel disease: a prospective study. Dig Dis Sci 2012;57(2):465 71. [80] Liao Z, et al. Indications and detection, completion, and retention rates of smallbowel capsule endoscopy: a systematic review. Gastrointest Endosc 2010;71 (2):280 6. [81] Pasha SF, et al. Capsule retention in Crohn’s disease: a meta-analysis. Inflamm Bowel Dis 2020;26(1):33 42. [82] Cohen SA, et al. The use of a patency capsule in pediatric Crohn’s disease: a prospective evaluation. Dig Dis Sci 2011;56(3):860 5.

51

This page intentionally left blank

CHAPTER

The role of capsule endoscopy in diagnosis and clinical management of obscure gastrointestinal bleeding

4

Nayantara Coelho-Prabhu1, Shabana F. Pasha2 and Jonathan Leighton2 1

Mayo Clinic, Rochester, MN, United States Division of Gastroenterology and Hepatology, Mayo Clinic, Scottsdale, AZ, United States

2

Introduction Gastrointestinal (GI) bleeding is exceedingly common, with an incidence of about 1 in 1000 population [1 4]. GI bleeding can be classified as upper GI bleeding when the bleeding source is above the level of the duodenal ampulla or lower GI bleeding when it is below the ileocecal valve. The sites of upper GI bleeding and lower GI bleeding in the GI anatomy can be examined using upper endoscopy and colonoscopy, respectively. However, about 5% of all GI bleeding arises from the small bowel between the ligament of Treitz and the ileocecal valve [5 8], traditionally termed obscure GI bleeding, and more recently called suspected small bowel bleeding [9]. With advanced technology, including capsule endoscopy (CE), triple-phase computed technography (CT) enterography (CTE), and deviceassisted deep enteroscopy, 40% 75% of these cases have an identifiable etiology and are no longer obscure [10 14]. Before we proceed, we will define a few more terms related to GI bleeding. When there are visible signs of bleeding, such as symptoms of hematemesis, coffee-ground emesis, hematochezia, or melena, it refers to overt GI bleeding. Occult GI bleeding refers to presentation with iron deficiency anemia and/or positive fecal occult blood on testing. Both overt and occult GI bleeding should first be evaluated by upper endoscopy in the form of either esophagogastroduodenoscopy (EGD) or push enteroscopy and/or colonoscopy. If these tests are conclusively negative for the etiology of bleeding, then a small bowel source should be suspected and examined. The timing of CE versus other small bowel testing will be discussed further in the chapter. CE allows noninvasive visualization of the mucosa of the GI tract using wireless or wire-free technology. It was first FDA approved in the USA in 2001. Since Given’s first commercialization of SBCE in 2001, capsule endoscope Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00006-X © 2023 Elsevier Inc. All rights reserved.

53

54

CHAPTER 4 The role of capsule endoscopy in obscure GI bleeding

systems currently in use worldwide include PillCam SB3 (Medtronic, Dublin, Ireland), MiroCam (Intromedic, Seoul, Korea), CapsoCam (Capso-Vision, Saratoga, NY, USA), EndoCapsule (Olympus, Tokyo, Japan), OMOM Capsule (Jinshan science and technology, Chongqing, China). Originally developed for imaging the small bowel, there are now specialized versions of CE designed for imaging of the upper GI tract and colon, specifically for Crohn’s disease. In this chapter, we will focus on the applications of small bowel CE. The outline of this chapter will first describe the utility of CE in suspected small bowel bleeding, then its practical aspects of use, and finally, the current data on applications of artificial intelligence to CE.

Suspected small bowel bleeding The diagnostic yield of CE in obscure GI bleeding ranges from 27% to 92.3% [13,15 21]. The factor shown to affect yield most in GI bleeding is clinical presentation. A study of 100 patients with obscure GI bleeding found the diagnostic yield of CE was 92.3% with ongoing overt bleeding, 44.2% with previous overt bleeding, and 12.9% in patients with iron deficiency anemia alone [22]. Multiple studies have also shown that the highest yield is within the first 48 72 h after an overt bleeding episode but remains high up to 2 weeks later [13,23 26]. Hence, in patients with overt GI bleeding, early CE significantly increases the diagnostic yield and, as shown later in the document, improves outcomes as well. Other factors that have shown to increase yield of capsule include need for blood transfusion [27 29], inpatient status [18,30], increasing age [18,31], male sex [18,30], use of anticoagulation or liver disease [31] (Fig. 4.1). Patients with a negative CE have a low risk of rebleeding. In a metaanalysis of 26 studies and 3657 patients the pooled rate of rebleeding after a negative capsule study was only 0.19 [32]. However, the effect was most pronounced in studies with a short follow-up, considered as 2 years in this analysis, and longer-term data may be helpful. There was no difference in rebleeding rates after overt and occult GI bleeding. Repeat CE performed for rebleeding has been shown to have a diagnostic yield of 35% 75% [33 35]. A change in clinical presentation from overt to occult bleeding and a decrease in hemoglobin by more than 4 g/dL were predictive of increased diagnostic yields [35]. The impact of antithrombotic and anticoagulant medications on capsule findings in obscure GI bleeding is significant. A systematic review and metaanalysis of 14 studies and 1023 antithrombotic users and 2359 nonusers found that antithrombotic users had a higher odds ratio of 1.98 of having a positive capsule study [36]. In secondary analyses, both antiplatelet agents and anticoagulants had a statistically significant increase in capsule positivity. Significantly, more patients on antithrombotic medications had rebleeding. The increase in risk of recurrent bleeding in patients taking warfarin has been confirmed in other studies [37,38]. Direct oral anticoagulants

Suspected small bowel bleeding

FIGURE 4.1 CE images of an arterio-venous malformation in proximal jejunum with active bleeding in a 72-year-old male, chronically anticoagulated with coumadin (warfarin), presenting with dark tarry stool and 4 g drop in hemoglobin. An example of overt obscure bleeding. CE, Capsule endoscopy.

(DOACs) have also been shown to be associated with a higher risk of bleeding lesions noted on CE in suspected small bowel bleeding [39,40]. The yield of CE has been compared with other diagnostic modalities. Pooled analysis of 14 prospective studies comparing CE to push enteroscopy showed a diagnostic yield of 56% for capsule versus 26% for push enteroscopy [41]. A randomized controlled trial of 79 patients with suspected small bowel bleeding showed a yield of 72.5% for capsule versus 48.7% for push enteroscopy (P , .05), even though there were no differences in ongoing bleeding [20]. Device-assisted deep enteroscopy has been compared with CE. Multiple metaanalyses have shown similar diagnostic yields between the two modalities [21,42 45]. The advantage of capsule over deep enteroscopy is the ability to visualize the entire length of the small bowel in a single study, while the disadvantage is that no therapy can be applied. CE has been shown to be superior to small bowel barium radiography for obscure GI bleeding, with diagnostic yields of 30% 42% for capsule and 6% 7% for barium radiography [41,46]. In brisk overt bleeding, CT angiography can be performed to help localize the source of the bleeding [12]. CTE is considered to be complementary to CE in the assessment of GI bleeding. In a metaanalysis of 7 studies and 279 patients, CTE

55

56

CHAPTER 4 The role of capsule endoscopy in obscure GI bleeding

had a pooled yield of 40% compared with 53% for CE [44]. In multiple studies, CE was found to be superior for the diagnosis of mucosal vascular abnormalities [44,47 52], while CTE or MR enterography (MRE) had higher diagnostic yields for small bowel tumors and masses and inflammatory wall changes [53] [49,51,52]. Cross-sectional imaging such as CTE and MRE can also be used to screen patients suspected of small bowel strictures prior to capsule placement to minimize the risk of capsule retention. The findings on CE can be used to guide the direction and modality of deep enteroscopy. Antegrade enteroscopy is performed when the abnormality is noted at less than 60% of transit time between pylorus and cecum and retrograde if the lesion is more distal, especially more than 75% [54,55] (Figs. 4.2 4.4). Also, in a metaanalysis of 10 studies where both CE and double-balloon enteroscopy (DBE) were performed the yield of DBE was 75% after a previously positive CE [21]. It was only 27% when CE was negative. Guidelines from the ACG, AGA, ASGE, and ESGE recommend CE as the next step after bidirectional endoscopy in patients with overt obscure GI bleeding as soon as possible after the bleeding event. In patients with occult obscure GI bleeding, it is recommended in select cases, such as those with recurrent anemia despite adequate iron replacement [13,56].

FIGURE 4.2 CE images from the ileum of a 24-year-old female presenting with fatigue and found to have iron deficiency anemia. Denied non steroidal anti-inflammatory drug (NSAID) use. Retrograde double-balloon enteroscopy confirms Crohn’s disease. CE, Capsule endoscopy.

FIGURE 4.3 CE images from the proximal duodenum of a 54-year-old female presenting with diarrhea. Esophagogastroduodenoscopy (EGD) with duodenal biopsies from the second part of the duodenum were negative. Images show villous blunting. Follow-up extended enteroscopy confirms diagnosis of celiac disease. CE, Capsule endoscopy.

FIGURE 4.4 CE images from a 42-year-old female with a history of migraines and non steroidal antiinflammatory drug (NSAID) use presenting with obscure occult GI bleeding. Images show erosions throughout the small bowel. As seen in this figure, they can be very subtle. CE, Capsule endoscopy.

58

CHAPTER 4 The role of capsule endoscopy in obscure GI bleeding

CE does have a miss rate of about 10%, especially for single mass lesions [57]. In a retrospective analysis of 183 patients with obscure GI bleeding, CE missed 67% of small bowel tumors eventually found by balloon enteroscopy [58]. The highest miss rate is for lesions in the proximal small bowel [59]. However, cross-sectional imaging such as CTE has been shown to have significantly greater sensitivity for these lesions. Hence, it should be considered a complementary examination to CE especially in younger patients or those with persistent obscure bleeding [53].

Timing of capsule endoscopy All patients with GI bleeding must first be hemodynamically stabilized. Any blood thinning medications must be appropriately managed [10,60]. In certain cases a second look endoscopy, including push enteroscopy and/or colonoscopy, can increase the yield of the diagnosis. This happens especially if the initial examinations are suboptimal because of the quality of preparation. Common diagnoses within reach of standard endoscopy include Cameron’s erosions associated with large hiatal hernias, fundal varices, gastric antral vascular ectasias, ulcers especially at the pylorus or duodenal angle, subtle celiac disease, and malignancy [61 66]. Colonoscopy performed for GI bleeding should always include an examination of the terminal ileum, even if blood is noted throughout the colon unless a definitive source is identified in the colon. Ileal Crohn’s disease can often present as iron deficiency anemia but very rarely with overt bleeding. Fresh or old blood in the terminal ileum suggests a small bowel source of GI bleeding and should lead to further evaluation of the small bowel, especially if EGD is negative. In patients presenting with suspected overt small bowel bleeding, who are hemodynamically unstable, urgent radiographic assessment is performed. This can be ideally in the form of CT angiography (CTA) or a tagged red blood cell scan, followed by a mesenteric angiogram depending on the availability at the treating institution. Of these, multiple guidelines recommend multiphasic CT scanning as the preferred test [10,12]. This can help localize the site of bleeding, and the patient can be referred for embolization [67 69]. Patients who have suspected overt small bowel bleeding, but are hemodynamically stable, can be assessed using CE after confirming the absence of contraindications to placement. In patients presenting with occult suspected small bowel bleeding and negative upper and lower endoscopy, CE is the first-line test if no contraindications exist [10]. Multiple guidelines recommend CE within 14 days of overt obscure GI bleeding [13,26]. Singh et al. performed a retrospective study on 144 inpatients with overt small bowel bleeding and found diagnostic yields on day 1 to be 55%, on day 2 to be 48%, on day 3 to be 29%, on day 4 to be 27%, and on day 5 to be 18% [24]. Another study demonstrated a capsule yield of 91% when performed within 2 weeks of obscure bleeding and only 34% after 2 weeks [23]. Early CE (on days 0, 1, or 2) compared with delayed (on days 3 7) has been associated

Contraindications and complications of capsule endoscopy

with a higher likelihood of endoscopic treatment and shorter hospital length of stay and charges, confirming a reduction in hospitalization resource utilization [70]. This test also has high positive (94% 97%) and negative (83% 100%) predictive values in GI bleeding [11,22]. CE findings led to a change in management either medically or endoscopically or surgically in 37% 87% of patients in multiple retrospective cohorts [22,71 73]. Emergent CE has been reported as a tool to triage patients rapidly. In 20 patients who presented with presumed upper GI bleeding after a negative EGD, CE correctly guided further therapeutic procedures in 85% [74]. In multiple studies where CE has been performed in the emergency department in patients with acute upper GI bleeding, there has been good patient tolerance, high sensitivity, decreased length of stay, and decreased hospital admissions [75 78]. The presence of bowel contents can decrease the yield of CE, especially in the presence of active bleeding, blood clots, and/or altered motility. The use of bowel preparation agents prior to CE has been suggested to ameliorate this problem. Current data, including metaanalyses and randomized controlled studies, have not demonstrated increased diagnostic yields but have confirmed improved visualization [79 82]. Current guidelines recommend the use of a bowel purgative prior to placement of a capsule to improve visualization [83]. In our practice, we recommend 12 h of fasting and a 2 L of polyethylene glycol preparation the day before the CE. In patients with delayed small bowel transit or narcotic usage, for example, prokinetic agents may be beneficial [84]. Two metaanalyses have found simethicone, at a dose between 80 and 200 mg, significantly decreases small bowel bubbles and foam and improves visualization [85,86]. A recent randomized controlled trial comparing high (1125 mg) to standard (300 mg) volume simethicone found no difference in visualization during CE [87].

Contraindications and complications of capsule endoscopy The potential complication of CE is capsule retention [88,89]. This is defined as retention of the capsule within the GI tract for 2 weeks or longer or requiring intervention for retrieval. Most patients are asymptomatic and do not present with obstructive symptoms. A systematic review and metaanalysis of 402 studies and 108,079 procedures published in 2020, demonstrated a retention rate of 0.93% [90]. Crohn’s disease was the most common reason for retention, with a retention rate of 4.63% in this population [91] (Fig. 4.5). The retention rate has decreased significantly over the years. This is likely due to the usage of a patency capsule when a small bowel obstruction is suspected [92,93]. It is recommended to utilize a patency capsule before placement of a diagnostic capsule in any patient in whom there is potential for retention, such as patients with suspected or known inflammatory bowel disease, chronic nonsteroidal antiinflammatory drug (NSAID) use, or radiation disease, or prior small bowel surgery [83].

59

60

CHAPTER 4 The role of capsule endoscopy in obscure GI bleeding

FIGURE 4.5 CE images of a stricture in a patient with known small bowel Crohn’s disease. Capsule was retained and required to be removed via antegrade double-balloon enteroscopy. CE, Capsule endoscopy.

Other complications include the inability to swallow the capsule due to swallowing disorders and resultant aspiration. In the systematic review, 0.75% of patients were unable to swallow the capsule, which can be related to old age, neurological disorders, and a weak cough reflex [90]. This can be circumvented by endoscopic placement of the capsule into the duodenum during EGD. Pregnancy is another relative contraindication to CE as safety data are missing [88]. Incomplete small bowel examination is another complication that occurs at a rate of 20% 30% [94,95]. Risk factors for incomplete examination include delayed gastric transit, poor bowel preparation, small bowel surgery, narcotic use, and decreased physical activity [95,96]. Direct duodenal capsule placement via endoscope can overcome gastroparesis. The presence of implantable cardiac devices is listed as a contraindication to CE by manufacturers except for CapsoCam. However, multiple published studies and a recent metaanalysis have shown no adverse events in this patient population [97 100]. Hence, CE is considered safe and efficacious in patients with implantable cardiac devices, including pacemakers, defibrillators, and left ventricular assist devices [100,101].

Artificial intelligence in capsule endoscopy

Advanced technologies in capsules CapsoCam SV-1 is the latest capsule system to be approved in 2013. It is built with 4 cameras around the middle of the capsule, which affords a lateral 360degree panoramic view of the small bowel mucosa. Another unique feature is that it stores the images within the capsule, thus eliminating the need for an external device. In a randomized study, it has been shown to have a higher diagnostic yield than PillCam SB3 [102]. The latest generation capsule is CapsoCam Plus. Some capsule systems have the ability to add light filters to the images. Blue light imaging and flexible spectral imaging color enhancement allow enhanced visualization of surface mucosal and vascular patterns and thus aim to enhance lesion detection and characterization. A metaanalysis of this technology showed no improvement in the detection of angioectasias or ulcers and erosions over white light, but FICE Setting 1 did improve the delineation of these lesions [103]. Newer technologies aim to control the movement of the capsule through the small bowel. These include internally propelled capsules and externally magnetically driven capsules [104]. In patients with refractory or recurrent iron deficiency anemia, a combined upper and small bowel capsule examination (MiroCam Navi—magnetically assisted CE) had a higher diagnostic yield than upper endoscopy alone. Robot-guided magnetic capsules are also being developed and studied [105]. None of these capsules are currently approved for use in the USA [106].

Artificial intelligence in capsule endoscopy CE interpretation requires significant time commitment and expertise in interpretation [88,107]. Hence, it is an excellent area of focus for the implementation of artificial intelligence methods to enhance output. Multiple recent metaanalyses have confirmed very high pooled sensitivity (95.5% 98%) and specificity (95.8% 99%) for CE in the detection of the bleeding source, including erosions, ulcers, polyps, and active bleeding [108 110]. A deep learning CNN to identify small bowel angioectasias was developed using 2237 CE images of angioectasia [111]. It has an AUC of 0.998 to detect small bowel angioectasias and sensitivity and specificity of 98.8% and 98.4%, respectively. Another large study developed and validated on 6970 patients’ CE images was able to identify abnormalities with 99.90% sensitivity in a per-lesion analysis, better than conventional gastroenterologists reading, which had a per-lesion sensitivity of 76.89% [112]. It also reduced the mean reading time per patient from 96.6 5.9 min. As this technology continues to improve, it has the potential to significantly improve the yield of CE for GI bleeding.

61

62

CHAPTER 4 The role of capsule endoscopy in obscure GI bleeding

References [1] Ghassemi KA, Jensen DM. Lower GI bleeding: epidemiology and management. Curr Gastroenterol Rep 2013;15:333. [2] Longstreth GF. Epidemiology of hospitalization for acute upper gastrointestinal hemorrhage: a population-based study. Am J Gastroenterol 1995;90:206 10. [3] Wuerth BA, Rockey DC. Changing epidemiology of upper gastrointestinal hemorrhage in the last decade: a nationwide analysis. Dig Dis Sci 2018;63:1286 93. [4] Friedman LS, Martin P. The problem of gastrointestinal bleeding. Gastroenterol Clin North Am 1993;22:717 21. [5] Katz LB. The role of surgery in occult gastrointestinal bleeding. Semin Gastrointest Dis 1999;10:78 81. [6] Lau WY, Fan ST, Wong SH, et al. Preoperative and intraoperative localisation of gastrointestinal bleeding of obscure origin. Gut 1987;28:869 77. [7] Longstreth GF. Epidemiology and outcome of patients hospitalized with acute lower gastrointestinal hemorrhage: a population-based study. Am J Gastroenterol 1997;92:419 24. [8] Szold A, Katz LB, Lewis BS. Surgical approach to occult gastrointestinal bleeding. Am J Surg 1992;163(90-2):92 3 discussion. [9] Kuo JR, Pasha SF, Leighton JA. The Clinician’s guide to suspected small bowel bleeding. Am J Gastroenterol 2019;114:591 8. [10] Committee ASoP, Gurudu SR, Bruining DH, et al. The role of endoscopy in the management of suspected small-bowel bleeding. Gastrointest Endosc 2017;85:22 31. [11] Delvaux M, Fassler I, Gay G. Clinical usefulness of the endoscopic video capsule as the initial intestinal investigation in patients with obscure digestive bleeding: validation of a diagnostic strategy based on the patient outcome after 12 months. Endoscopy 2004;36:1067 73. [12] Gerson LB, Fidler JL, Cave DR, et al. ACG clinical guideline: diagnosis and management of small bowel bleeding. Am J Gastroenterol 2015;110:1265 87. [13] Pennazio M, Spada C, Eliakim R, et al. Small-bowel capsule endoscopy and device-assisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Clinical Guideline. Endoscopy 2015;47:352 76. [14] Sun B, Rajan E, Cheng S, et al. Diagnostic yield and therapeutic impact of doubleballoon enteroscopy in a large cohort of patients with obscure gastrointestinal bleeding. Am J Gastroenterol 2006;101:2011 15. [15] Ell C, Remke S, May A, et al. The first prospective controlled trial comparing wireless capsule endoscopy with push enteroscopy in chronic gastrointestinal bleeding. Endoscopy 2002;34:685 9. [16] Jang HJ, Choi MH, Park CH, et al. Comparison of double balloon enteroscopy and capsule endoscopy in patients with suspected small bowel diseases. Gastrointest Endoscopy 2006;63. [17] Lee BJ, Chun HJ, Koo JS, et al. [Analysis of the factors that affect the diagnostic yield of capsule endoscopy in patients with obscure gastrointestinal bleeding]. Korean J Gastroenterol 2007;49:79 84. [18] Lepileur L, Dray X, Antonietti M, et al. Factors associated with diagnosis of obscure gastrointestinal bleeding by video capsule enteroscopy. Clin Gastroenterology Hepatology 2012;10:1376 80.

References

[19] Leung WK, Ho SSM, Suen B-Y, et al. Capsule endoscopy or angiography in patients with acute overt obscure gastrointestinal bleeding: a prospective randomized study with long-term follow-up. Am J Gastroenterology 2012;107:1370 6. [20] Segarajasingam DS, Hanley SC, Barkun AN, et al. Randomized controlled trial comparing outcomes of video capsule endoscopy with push enteroscopy in obscure gastrointestinal bleeding. Can J Gastroenterology Hepatology 2015;29:85 90. [21] Teshima CW, Kuipers EJ, van Zanten SV, et al. Double balloon enteroscopy and capsule endoscopy for obscure gastrointestinal bleeding: an updated meta-analysis. J Gastroenterol Hepatol 2011;26:796 801. [22] Pennazio M, Santucci R, Rondonotti E, et al. Outcome of patients with obscure gastrointestinal bleeding after capsule endoscopy: report of 100 consecutive cases. Gastroenterology 2004;126:643 53. [23] Bresci G, Parisi G, Bertoni M, et al. The role of video capsule endoscopy for evaluating obscure gastrointestinal bleeding: usefulness of early use. J Gastroenterol 2005;40:256 9. [24] Singh A, Marshall C, Chaudhuri B, et al. Timing of video capsule endoscopy relative to overt obscure GI bleeding: implications from a retrospective study. Gastrointest Endosc 2013;77:761 6. [25] Yamada A, Watabe H, Kobayashi Y, et al. Timing of capsule endoscopy influences the diagnosis and outcome in obscure-overt gastrointestinal bleeding. Hepatogastroenterology 2012;59:676 9. [26] Shim KN, Moon JS, Chang DK, et al. Guideline for capsule endoscopy: obscure gastrointestinal bleeding. Clin Endosc 2013;46:45 53. [27] Estevez E, Gonzalez-Conde B, Vazquez-Iglesias JL, et al. Diagnostic yield and clinical outcomes after capsule endoscopy in 100 consecutive patients with obscure gastrointestinal bleeding. Eur J Gastroenterol Hepatol 2006;18:881 8. [28] Esaki M, Matsumoto T, Yada S, et al. Factors associated with the clinical impact of capsule endoscopy in patients with overt obscure gastrointestinal bleeding. Dig Dis Sci 2010;55:2294 301. [29] Carey EJ, Leighton JA, Heigh RI, et al. A single-center experience of 260 consecutive patients undergoing capsule endoscopy for obscure gastrointestinal bleeding. Am J Gastroenterol 2007;102:89 95. [30] Robinson CA, Jackson C, Condon D, et al. Impact of inpatient status and gender on small-bowel capsule endoscopy findings. Gastrointest Endosc 2011;74:1061 6. [31] Sidhu R, Sanders DS, Kapur K, et al. Factors predicting the diagnostic yield and intervention in obscure gastrointestinal bleeding investigated using capsule endoscopy. J Gastrointestin Liver Dis 2009;18:273 8. [32] Yung DE, Koulaouzidis A, Avni T, et al. Clinical outcomes of negative small-bowel capsule endoscopy for small-bowel bleeding: a systematic review and meta-analysis. Gastrointest Endosc 2017;85(305-317):e2. [33] Bar-Meir S, Eliakim R, Nadler M, et al. Second capsule endoscopy for patients with severe iron deficiency anemia. Gastrointest Endosc 2004;60:711 13. [34] Jones BH, Fleischer DE, Sharma VK, et al. Yield of repeat wireless video capsule endoscopy in patients with obscure gastrointestinal bleeding. Am J Gastroenterol 2005;100:1058 64. [35] Viazis N, Papaxoinis K, Vlachogiannakos J, et al. Is there a role for second-look capsule endoscopy in patients with obscure GI bleeding after a nondiagnostic first test? Gastrointest Endosc 2009;69:850 6.

63

64

CHAPTER 4 The role of capsule endoscopy in obscure GI bleeding

[36] Tziatzios G, Gkolfakis P, Papanikolaou IS, et al. Antithrombotic treatment is associated with small-bowel video capsule endoscopy positive findings in obscure gastrointestinal bleeding: a systematic review and meta-analysis. Dig Dis Sci 2019;64:15 24. [37] Iwamoto J, Mizokami Y, Shimokobe K, et al. The clinical outcome of capsule endoscopy in patients with obscure gastrointestinal bleeding. Hepatogastroenterology 2011;58:301 5. [38] Niikura R, Yamada A, Hirata Y, et al. Role of warfarin as a predictor of recurrent bleeding after negative small-bowel capsule endoscopy. Gastrointest Endosc 2018;88:574 574 e2. [39] Macedo Silva V, Freitas M, Arieira C, et al. Direct oral anticoagulants are associated with potentially bleeding lesions in suspected mid-gastrointestinal bleeding. Scand J Gastroenterol 2021;1 7. [40] Yamaoka M, Imaeda H, Hosoe N, et al. Small-bowel lesions in patients taking direct oral anticoagulants detected using capsule endoscopy. Gastroenterol Res Pract 2020;2020:7125642. [41] Triester SL, Leighton JA, Leontiadis GI, et al. A meta-analysis of the yield of capsule endoscopy compared to other diagnostic modalities in patients with obscure gastrointestinal bleeding. Am J Gastroenterol 2005;100:2407 18. [42] Chen X, Ran ZH, Tong JL. A meta-analysis of the yield of capsule endoscopy compared to double-balloon enteroscopy in patients with small bowel diseases. World J Gastroenterol 2007;13:4372 8. [43] Pasha SF, Leighton JA, Das A, et al. Double-balloon enteroscopy and capsule endoscopy have comparable diagnostic yield in small-bowel disease: a meta-analysis. Clin Gastroenterol Hepatol 2008;6:671 6. [44] Wang Z, Chen JQ, Liu JL, et al. CT enterography in obscure gastrointestinal bleeding: a systematic review and meta-analysis. J Med Imaging Radiat Oncol 2013;57:263 73. [45] Zhang Q, He Q, Liu J, et al. Combined use of capsule endoscopy and double-balloon enteroscopy in the diagnosis of obscure gastrointestinal bleeding: meta-analysis and pooled analysis. Hepatogastroenterology 2013;60:1885 91. [46] Laine L, Sahota A, Shah A. Does capsule endoscopy improve outcomes in obscure gastrointestinal bleeding? Randomized trial vs. dedicated small bowel radiography. Gastroenterology 2010;138:1673 80 e1; quiz e11-2. [47] Bocker U, Dinter D, Litterer C, et al. Comparison of magnetic resonance imaging and video capsule enteroscopy in diagnosing small-bowel pathology: localizationdependent diagnostic yield. Scand J Gastroenterol 2010;45:490 500. [48] Golder SK, Schreyer AG, Endlicher E, et al. Comparison of capsule endoscopy and magnetic resonance (MR) enteroclysis in suspected small bowel disease. Int J Colorectal Dis 2006;21:97 104. [49] Khalife S, Soyer P, Alatawi A, et al. Obscure gastrointestinal bleeding: preliminary comparison of 64-section CT enteroclysis with video capsule endoscopy. Eur Radiol 2011;21:79 86. [50] Leighton JA, Triester SL, Sharma VK. Capsule endoscopy: a meta-analysis for use with obscure gastrointestinal bleeding and Crohn’s disease. Gastrointest Endosc Clin N Am 2006;16:229 50. [51] Rajesh A, Sandrasegaran K, Jennings SG, et al. Comparison of capsule endoscopy with enteroclysis in the investigation of small bowel disease. Abdom Imaging 2009;34:459 66.

References

[52] Wiarda BM, Heine DG, Mensink P, et al. Comparison of magnetic resonance enteroclysis and capsule endoscopy with balloon-assisted enteroscopy in patients with obscure gastrointestinal bleeding. Endoscopy 2012;44:668 73. [53] Huprich JE, Fletcher JG, Fidler JL, et al. Prospective blinded comparison of wireless capsule endoscopy and multiphase CT enterography in obscure gastrointestinal bleeding. Radiology 2011;260:744 51. [54] Gay G, Delvaux M, Fassler I. Outcome of capsule endoscopy in determining indication and route for push-and-pull enteroscopy. Endoscopy 2006;38:49 58. [55] Li X, Chen H, Dai J, et al. Predictive role of capsule endoscopy on the insertion route of double-balloon enteroscopy. Endoscopy 2009;41:762 6. [56] Enns RA, Hookey L, Armstrong D, et al. Clinical practice guidelines for the use of video capsule endoscopy. Gastroenterology 2017;152:497 514. [57] Lewis BS, Eisen GM, Friedman S. A pooled analysis to evaluate results of capsule endoscopy trials. Endoscopy 2005;37:960 5. [58] Ross A, Mehdizadeh S, Tokar J, et al. Double balloon enteroscopy detects small bowel mass lesions missed by capsule endoscopy. Dig Dis Sci 2008;53:2140 3. [59] Zagorowicz ES, Pietrzak AM, Wronska E, et al. Small bowel tumors detected and missed during capsule endoscopy: single center experience. World J Gastroenterol 2013;19:9043 8. [60] Veitch AM, Radaelli F, Alikhan R, et al. Endoscopy in patients on antiplatelet or anticoagulant therapy: British Society of Gastroenterology (BSG) and European Society of Gastrointestinal Endoscopy (ESGE) guideline update. Endoscopy 2021;53:947 69. [61] Chak A, Koehler MK, Sundaram SN, et al. Diagnostic and therapeutic impact of push enteroscopy: analysis of factors associated with positive findings. Gastrointest Endosc 1998;47:18 22. [62] Fry LC, Bellutti M, Neumann H, et al. Incidence of bleeding lesions within reach of conventional upper and lower endoscopes in patients undergoing doubleballoon enteroscopy for obscure gastrointestinal bleeding. Aliment Pharmacol Ther 2009;29:342 9. [63] Kitiyakara T, Selby W. Non-small-bowel lesions detected by capsule endoscopy in patients with obscure GI bleeding. Gastrointest Endosc 2005;62:234 8. [64] Lara LF, Bloomfeld RS, Pineau BC. The rate of lesions found within reach of esophagogastroduodenoscopy during push enteroscopy depends on the type of obscure gastrointestinal bleeding. Endoscopy 2005;37:745 50. [65] Sidhu R, Sanders DS, McAlindon ME. Does capsule endoscopy recognise gastric antral vascular ectasia more frequently than conventional endoscopy? J Gastrointestin Liver Dis 2006;15:375 7. [66] Singh H, Nugent Z, Demers AA, et al. Rate and predictors of early/missed colorectal cancers after colonoscopy in Manitoba: a population-based study. Am J Gastroenterol 2010;105:2588 96. [67] He B, Yang J, Xiao J, et al. Diagnosis of lower gastrointestinal bleeding by multislice CT angiography: a meta-analysis. Eur J Radiol 2017;93:40 5. [68] Mirsadraee S, Tirukonda P, Nicholson A, et al. Embolization for non-variceal upper gastrointestinal tract haemorrhage: a systematic review. Clin Radiol 2011;66:500 9. [69] Wu LM, Xu JR, Yin Y, et al. Usefulness of CT angiography in diagnosing acute gastrointestinal bleeding: a meta-analysis. World J Gastroenterol 2010;16:3957 63.

65

66

CHAPTER 4 The role of capsule endoscopy in obscure GI bleeding

[70] Wood A.R., Ham S.A., Sengupta N., et al. Impact of early video capsule endoscopy on hospitalization and post-hospitalization outcomes: a propensity score-matching analysis. Dig Dis Sci 2021. [71] Barnett CB, Dipalma JA, Olden KW. Capsule endoscopy: impact on patient management. Gastroenterol Hepatol (N Y) 2007;3:124 6. [72] Redondo-Cerezo E, Perez-Vigara G, Perez-Sola A, et al. Diagnostic yield and impact of capsule endoscopy on management of patients with gastrointestinal bleeding of obscure origin. Dig Dis Sci 2007;52:1376 81. [73] Ben Soussan E, Antonietti M, Herve S, et al. Diagnostic yield and therapeutic implications of capsule endoscopy in obscure gastrointestinal bleeding. Gastroenterol Clin Biol 2004;28:1068 73. [74] Schlag C, Menzel C, Nennstiel S, et al. Emergency video capsule endoscopy in patients with acute severe GI bleeding and negative upper endoscopy results. Gastrointest Endosc 2015;81:889 95. [75] Gralnek IM, Ching JY, Maza I, et al. Capsule endoscopy in acute upper gastrointestinal hemorrhage: a prospective cohort study. Endoscopy 2013;45:12 19. [76] Marya NB, Jawaid S, Foley A, et al. A randomized controlled trial comparing efficacy of early video capsule endoscopy with standard of care in the approach to nonhematemesis GI bleeding (with videos). Gastrointest Endosc 2019;89:33 43 e4. [77] Meltzer AC, Ali MA, Kresiberg RB, et al. Video capsule endoscopy in the emergency department: a prospective study of acute upper gastrointestinal hemorrhage. Ann Emerg Med 2013;61:438 43 e1. [78] Sung JJ, Tang RS, Ching JY, et al. Use of capsule endoscopy in the emergency department as a triage of patients with GI bleeding. Gastrointest Endosc 2016;84:907 13. [79] Gkolfakis P, Tziatzios G, Dimitriadis GD, et al. Meta-analysis of randomized controlled trials challenging the usefulness of purgative preparation before small-bowel video capsule endoscopy. Endoscopy 2018;50:671 83. [80] Rahmi G, Cholet F, Gaudric M, et al. Effect of different modalities of purgative preparation on the diagnostic yield of small bowel capsule for the exploration of suspected small bowel bleeding: a multicenter randomized controlled trial. Am J Gastroenterol 2022;117:327 35. [81] Wu S, Gao YJ, Ge ZZ. Optimal use of polyethylene glycol for preparation of small bowel video capsule endoscopy: a network meta-analysis. Curr Med Res Opin 2017;33:1149 54. [82] Yung DE, Rondonotti E, Sykes C, et al. Systematic review and meta-analysis: is bowel preparation still necessary in small bowel capsule endoscopy? Expert Rev Gastroenterol Hepatol 2017;11:979 93. [83] Rondonotti E, Spada C, Adler S, et al. Small-bowel capsule endoscopy and deviceassisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Technical Review. Endoscopy 2018;50:423 46. [84] Koulaouzidis A, Giannakou A, Yung DE, et al. Do prokinetics influence the completion rate in small-bowel capsule endoscopy? A systematic review and meta-analysis. Curr Med Res Opin 2013;29:1171 85. [85] Kotwal VS, Attar BM, Gupta S, et al. Should bowel preparation, antifoaming agents, or prokinetics be used before video capsule endoscopy? A systematic review and meta-analysis. Eur J Gastroenterol Hepatol 2014;26:137 45.

References

[86] Wu L, Cao Y, Liao C, et al. Systematic review and meta-analysis of randomized controlled trials of simethicone for gastrointestinal endoscopic visibility. Scand J Gastroenterol 2011;46:227 35. [87] Sey M, Yan B, McDonald C, et al. A randomized controlled trial of high volume simethicone to improve visualization during capsule endoscopy. PLoS One 2021;16:e0249490. [88] Committee AT, Wang A, Banerjee S, et al. Wireless capsule endoscopy. Gastrointest Endosc 2013;78:805 15. [89] Liao Z, Gao R, Xu C, et al. Indications and detection, completion, and retention rates of small-bowel capsule endoscopy: a systematic review. Gastrointest Endosc 2010;71:280 6. [90] Wang YC, Pan J, Liu YW, et al. Adverse events of video capsule endoscopy over the past two decades: a systematic review and proportion meta-analysis. BMC Gastroenterol 2020;20:364. [91] Pasha SF, Pennazio M, Rondonotti E, et al. Capsule retention in Crohn’s disease: a meta-analysis. Inflamm Bowel Dis 2020;26:33 42. [92] Herrerias JM, Leighton JA, Costamagna G, et al. Agile patency system eliminates risk of capsule retention in patients with known intestinal strictures who undergo capsule endoscopy. Gastrointest Endosc 2008;67:902 9. [93] Spada C, Spera G, Riccioni M, et al. A novel diagnostic tool for detecting functional patency of the small bowel: the given patency capsule. Endoscopy 2005;37:793 800. [94] Rondonotti E, Herrerias JM, Pennazio M, et al. Complications, limitations, and failures of capsule endoscopy: a review of 733 cases. Gastrointest Endosc 2005;62:712 16 quiz 752, 754. [95] Westerhof J, Weersma RK, Koornstra JJ. Risk factors for incomplete small-bowel capsule endoscopy. Gastrointest Endosc 2009;69:74 80. [96] Shibuya T, Mori H, Takeda T, et al. The relationship between physical activity level and completion rate of small bowel examination in patients undergoing capsule endoscopy. Intern Med 2012;51:997 1001. [97] Bandorski D, Holtgen R, Stunder D, et al. Capsule endoscopy in patients with cardiac pacemakers, implantable cardioverter defibrillators and left heart assist devices. Ann Gastroenterol 2014;27:3 8. [98] Leighton JA, Sharma VK, Srivathsan K, et al. Safety of capsule endoscopy in patients with pacemakers. Gastrointest Endosc 2004;59:567 9. [99] Leighton JA, Srivathsan K, Carey EJ, et al. Safety of wireless capsule endoscopy in patients with implantable cardiac defibrillators. Am J Gastroenterol 2005;100:1728 31. [100] Tabet R, Nassani N, Karam B, et al. Pooled analysis of the efficacy and safety of video capsule endoscopy in patients with implantable cardiac devices. Can J Gastroenterol Hepatol 2019;2019:3953807. [101] Harris LA, Hansel SL, Rajan E, et al. Capsule endoscopy in patients with implantable electromedical devices is safe. Gastroenterol Res Pract 2013;2013:959234. [102] Zwinger LL, Siegmund B, Stroux A, et al. CapsoCam SV-1 vs PillCam SB 3 in the detection of obscure gastrointestinal bleeding: results of a prospective randomized comparative multicenter study. J Clin Gastroenterol 2019;53:e101 6. [103] Yung DE, Boal Carvalho P, Giannakou A, et al. Clinical validity of flexible spectral imaging color enhancement (FICE) in small-bowel capsule endoscopy: a systematic review and meta-analysis. Endoscopy 2017;49:258 69.

67

68

CHAPTER 4 The role of capsule endoscopy in obscure GI bleeding

[104] Nam SJ, Lee HS, Lim YJ. Evaluation of gastric disease with capsule endoscopy. Clin Endosc 2018;51:323 8. [105] Jiang B, Qian YY, Pan J, et al. Second-generation magnetically controlled capsule gastroscopy with improved image resolution and frame rate: a randomized controlled clinical trial (with video). Gastrointest Endosc 2020;91:1379 87. [106] Ching HL, Hale MF, Kurien M, et al. Diagnostic yield of magnetically assisted capsule endoscopy vs. gastroscopy in recurrent and refractory iron deficiency anemia. Endoscopy 2019;51:409 18. [107] Koulaouzidis A, Iakovidis DK, Karargyris A, et al. Optimizing lesion detection in small-bowel capsule endoscopy: from present problems to future solutions. Expert Rev Gastroenterol Hepatol 2015;9:217 35. [108] Mohan BP, Khan SR, Kassab LL, et al. High pooled performance of convolutional neural networks in computer-aided diagnosis of GI ulcers and/or hemorrhage on wireless capsule endoscopy images: a systematic review and meta-analysis. Gastrointest Endosc 2021;93:356 64 e4. [109] Qin K., Li J., Fang Y., et al. Convolution neural network for the diagnosis of wireless capsule endoscopy: a systematic review and meta-analysis. Surg Endosc 2021. [110] Soffer S, Klang E, Shimon O, et al. Deep learning for wireless capsule endoscopy: a systematic review and meta-analysis. Gastrointest Endosc 2020;92:831 839 e8. [111] Tsuboi A, Oka S, Aoyama K, et al. Artificial intelligence using a convolutional neural network for automatic detection of small-bowel angioectasia in capsule endoscopy images. Dig Endosc 2020;32:382 90. [112] Ding Z, Shi H, Zhang H, et al. Gastroenterologist-level identification of smallbowel diseases and normal variants by capsule endoscopy using a deep-learning model. Gastroenterology 2019;157:1044 54 e5.

CHAPTER

The role of capsule endoscopy in diagnosis and clinical management of inflammatory bowel disease

5

Isabel Garrido1,2, Patrı´cia Andrade1,2,3, Susana Lopes1,2,3 and Guilherme Macedo1,2,3 1

Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal 2 World Gastroenterology Organization Porto Training Center, Porto, Portugal 3 Faculty of Medicine, University of Porto, Porto, Portugal

Introduction Crohn’s disease and ulcerative colitis are the principal forms of inflammatory bowel disease (IBD). Both represent chronic inflammation of the gastrointestinal tract, with distinct clinical and pathological features. The diagnosis of IBD is based on a combination of clinical, biochemical, radiological, endoscopic, and histological findings [1]. IBD is marked by frequent relapses, which usually require repeated investigations. In the past decade, there have been major advances in investigations, pharmacological, nonpharmacological, and surgical interventions for both ulcerative colitis and Crohn’s disease. Capsule endoscopy (CE) has revolutionized our ability to visualize the small bowel mucosa [2]. This modality is currently a valuable tool for the diagnosis of small bowel Crohn’s disease as well as for monitoring disease activity. CE has also been assessed as a tool for detecting postoperative recurrence for those who have undergone an intestinal resection. The evolution of colon capsule endoscopy (CCE) has expanded the application of this technology further [3]. Indeed, this advancement in capsule technology as a minimally invasive tool to assess the whole gastrointestinal tract has sparked interest opening the possibility of its use for the panenteric assessment of Crohn’s disease. In addition, the use of CCE to assess the activity of ulcerative colitis has also been described. This chapter contains an overview of the current and future clinical applications of capsule endoscopy in IBD.

Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00001-0 © 2023 Elsevier Inc. All rights reserved.

69

70

CHAPTER 5 The role of capsule endoscopy in IBD

Crohn’s disease Crohn’s disease is characterized by a transmural inflammation that may involve any portion of the luminal gastrointestinal tract, from the oral cavity to the perianal area [1]. The most common symptoms are diarrhea and abdominal pain. However, clinical manifestations can be very heterogeneous, depending on the disease location and phenotype. Fecal calprotectin and serum C-reactive protein are useful markers to detect and monitor inflammation. Cross-sectional imaging techniques provide information about the bowel wall and extra-enteric soft tissues and can classify disease phenotype. The endoscopic hallmark of Crohn’s disease is the patchy distribution of inflammation (Fig. 5.1), and mucosal biopsies usually show focal inflammation, crypt distortion, and/or granulomas. Ileocolonoscopy and upper gastrointestinal endoscopy have well-established roles in assessing disease activity and therapeutic intervention. However, the small bowel is one of the most common areas affected in patients with Crohn’s disease, which is often inaccessible to conventional endoscopy. At the time of diagnosis, up to 30% of patients have only small bowel involvement [4]. The advent of CE and balloon-assisted and spiral enteroscopy is revolutionizing the management of small bowel Crohn’s disease [5]. Capsule endoscopy provides high-quality endoluminal images of the small bowel, is less invasive than conventional endoscopic techniques, and is usually well tolerated by patients.

Ulcerative colitis Ulcerative colitis is a chronic inflammatory disease characterized by mucosal inflammation starting distally in the rectum with continuous extension proximally through the colon. The diagnosis requires a lower digestive tract endoscopic

FIGURE 5.1 Endoscopic appearance of Crohn’s disease—colon with discontinuous segments of edema, friability, ulcerations (A) and stenosis (B).

Ulcerative colitis

examination with histologic confirmation [6]. Endoscopic features of inflammation include loss of vascular markings, granularity, friability of the mucosa, erosions, deep ulcerations and spontaneous bleeding in the setting of severe inflammation (eliminate the “and” before deep ulcerations) (Fig. 5.2). Monitoring mucosal status may require colonoscopy to be repeated at different stages of the disease. Indeed, treating to the target of mucosal healing has been proposed for medical treatment of ulcerative colitis, because it may alter the course of the disease and reduce the need for hospitalization or surgery [7]. In addition, colorectal cancer is a well-described complication of ulcerative colitis, and the predominant approach has been secondary prevention via colonoscopy screening and surveillance. Although at this stage CCE cannot be recommended to replace endoscopy for the detection and follow-up of ulcerative colitis, some pilot studies indicate that CCE has a potential role in monitoring mucosal healing [8]. Indeed, if CCE can be shown to be useful in ulcerative colitis, this will open up the potential for monitoring mucosal healing with minimal discomfort and procedural risk.

FIGURE 5.2 Endoscopic appearance of ulcerative colitis—colon with erythema and absence of vascular patterns (A), erosions (B), spontaneous bleeding and ulceration (C).

71

72

CHAPTER 5 The role of capsule endoscopy in IBD

Capsule endoscopy in suspected Crohn’s disease Suspected Crohn’s disease should be investigated with ileocolonoscopy, including segmental colonic and ileal biopsies, and imaging to assess the location and extent of small bowel disease [1]. CE should be performed when IBD is still suspected despite normal cross-sectional imaging, such as in the case of unexplained anemia, severe malnutrition, and inconsistency between symptoms and other imaging findings. In fact, capsule endoscopy is a useful adjunct in the diagnosis of patients with Crohn’s disease restricted to the small bowel since it allows for direct visualization of the mucosa of the entire small intestine. It is able to identify mucosal lesions compatible with Crohn’s disease in patients in whom conventional endoscopic and small bowel radiographic imaging modalities have been nondiagnostic, especially in the proximal small bowel (Fig. 5.3) [9]. Several metaanalyses have examined the diagnostic yield of CE in the evaluation of patients with suspected Crohn’s disease and showed that it is superior to small bowel barium studies, computed tomography enterography, and ileocolonoscopy, with an incremental diagnostic yield of 32%, 47%, and 22%, respectively [10]. This is of prognostic significance, as detection of proximal small bowel disease in patients with Crohn’s disease has been associated with poorer clinical outcomes. Many symptoms of Crohn’s disease such as diarrhea, abdominal pain, and bloating can be attributed to a multitude of etiologies other than active inflammation. Therefore negative CE results are also of clinical importance as they prevent unnecessary and expensive initiation of an antiinflammatory regimen for another condition such as irritable bowel syndrome. Indeed, CE has a negative predictive value of 96%, essentially ruling out small bowel Crohn’s disease [11]. On the other hand, Solem et al. assessed the accuracy of CE, computed tomography enterography, and ileocolonoscopy in detecting active small bowel Crohn’s disease and showed that the specificity of CE was significantly lower than the other tests [12]. The mucosal features of small bowel Crohn’s disease that may be seen at capsule endoscopy include erythema, aphthous ulceration, loss of villi, villous edema, mucosal fissures, and strictures [5]. However, these findings are not specific to Crohn’s disease and may be seen in patients with other types of small bowel enteropathy, such as intestinal Behc¸et’s disease, nonsteroidal antiinflammatory drugs (NSAIDs) induced enteropathy, and lymphoma. Therefore, CE should be reserved for cases in which ileocolonoscopy plus small bowel radiography is not diagnostic, but there is a high rate of Crohn’s disease suspicion (suggestive clinical presentation and raised fecal calprotectin) [13]. Although there are no validated diagnostic criteria for the diagnosis of Crohn’s disease, the most commonly used diagnostic criterion in practice constitutes the presence of more than 3 small bowel ulcerations in the absence of NSAIDs ingestion for at least 1 month before the examination [14]. There are 2 validated indexes available, the Capsule Endoscopy Crohn’s Disease Activity Index (CECDAI) [15] and the Lewis score [16], which assess the disease location and activity of small bowel involvement. The CECDAI was

Capsule endoscopy in suspected Crohn’s disease

FIGURE 5.3 Capsule endoscopy images showing multiple erosions (A) aphthous ulcers (B) and large ulcers (C) in the small bowel compatible with the diagnosis of Crohn’s disease.

validated in a multicenter prospective study of patients with isolated small bowel Crohn’s disease, and the following 3 endoscopic parameters based on the transit time of the capsule were evaluated: inflammation, the extent of the disease, and strictures for both the proximal and the distal segments of the small bowel (Table 5.1). The Lewis score is another scoring system based on the evaluation of 3 endoscopic parameters: villous appearance, ulcers, and strictures (Table 5.2). The small bowel is divided into 3 equal parts, and a subscore is determined for each tertile. The Lewis score is the sum of the worst-affected tertile plus the stenosis score. Both the scoring systems are incorporated into the software platform of the capsules and assist in the quantification of small bowel inflammatory burden and diagnosis of Crohn’s disease.

73

74

CHAPTER 5 The role of capsule endoscopy in IBD

Table 5.1 Capsule endoscopy Crohn’s disease activity index (CECDAI). Parameter

Score and descriptor

A. Inflammation

0—None 1—Mild to moderate (edema, hyperemia, or denudation) 2—Severe (edema, hyperemia, or denudation) 3—Bleeding, exudate, erosion aphthae, ulcers ,0.5 cm 4—Pseudopolyp, ulcers 0.5 2 cm 5—Ulcers .2 cm None 1—Single segment (focal disease) 2—2 3 segments (patchy disease) 3— . 3 segments (diffuse disease) 0—None 1—Single-passed 2—Multiple-passed 3—Obstruction (nonpassage)

B. Extent of disease

C. Stricture

CECDAI 5 proximal segment (A 3 B 1 C) 1 distal segment (A 3 B 1 C). Clinical or endoscopic remission: CEDAI ,4.

Table 5.2 Lewis score. Parameter Villous appearance

Ulcers

Stenosis

Descriptor or number 0—Normal 1—Edematous

0—Normal 3—Single 5—Few (2 7) 10—Multiple ($8) 0—None 14—Single 20—Multiple

Longitudinal extent

Descriptor

8—Short-segment (,10%) 12—Long-segment (11% 50%) 20—Whole tertile ( . 50%) 5—Short-segment (,10%) 10—Long-segment (11% 50%) 15—Whole tertile ( . 50%)

1—Single 14—Patchy 17—Diffuse

2—Nonulcerated 24—Ulcerated

9— , 1/4 12—1/4 1/2 18— . 1/2 7—Transversed 10—Not transversed

Score total 5 worst-affected tertile villous appearance and ulcers plus stenosis score. Clinically insignificant inflammation: Lewis score ,135, mild inflammation: Lewis score 5 135 790, moderateto-severe inflammation: Lewis score .790.

A recent study comparing the two scores found that Lewis score 135 and 790 were equivalent to CECDAI values of 4.9 and 6.9, respectively [17]. There was a strong correlation between the two scores, but the CECDAI was more reflective of extensive inflammation and high clinical activity. Nevertheless, there is no gold-standard or best scoring system widely accepted in practice to date.

Capsule endoscopy in patients with established Crohn’s disease

Capsule endoscopy in patients with established Crohn’s disease CE has some advantages over other modalities for assessing inflammatory activity in patients with an established diagnosis of Crohn’s disease. It has the potential to identify the presence of active diseases that may not be evident from conventional biomarkers or to identify mucosal lesions that are not visible on radiological imaging. Long et al. reported on the outcomes of 86 patients with Crohn’s disease undergoing CE. Severe findings, defined as multiple aphthous ulcers or stenosis, as compared to minimal or no inflammatory change, were associated with the addition of new medication and also with the likelihood of surgery in the 3 months following the examination [18]. Similarly, in a study of 53 patients with Crohn’s disease restricted to the small bowel, moderate-to-severe inflammation (Lewis score of $ 790) was associated with an increased risk of corticosteroid therapy and hospitalization during a mean follow-up period of 42 months [19]. It appears therefore that the severity of inflammation as quantified by the Lewis score may predict a more aggressive course of the disease in patients with Crohn’s disease [20]. Clinical symptoms can correlate poorly with the activity of Crohn’s disease. Kopylov assessed the inflammatory burden in the small bowel in patients with Crohn’s disease in clinical remission, defined as those with a Crohn’s Disease Activity Index score of ,150 [21]. A total of 84.6% of patients in clinical remission had significant mucosal inflammation of the small bowel (Lewis score of .135). C-reactive protein and fecal calprotectin are inflammatory biomarkers frequently used to assess and monitor the activity of IBD. Fecal calprotectin has a stronger correlation with mucosal inflammation, with a reported sensitivity and specificity for the detection of mucosal disease of 70% 100% and 44% 100%, respectively [22]. Several studies have investigated the degree to which findings at CE correlate with inflammatory biomarkers. Niv et al. assessed the correlation between laboratory and clinical markers of disease activity and findings at CE in patients with active Crohn’s disease [23]. No correlation was demonstrated between the Lewis score and C-reactive protein. In addition, a poor correlation between the Lewis score and clinical symptoms was assessed by the Crohn’s Disease Activity Index and Inflammatory Bowel Disease Questionnaire. Koulaouzidis evaluated 70 patients in whom isolated small bowel Crohn’s disease was suspected [24]. A moderate correlation between fecal calprotectin and the Lewis score (r 5 0.448) was reported. When the analysis was restricted to patients with a fecal calprotectin of ,100 a strong correlation was reported (r 5 0.68). On the other hand, there was no significant correlation between CECDAI and fecal calprotectin. In a multicenter cross-sectional study assessing 187 patients undergoing CE, significant small bowel inflammation (defined as a Lewis score .790) correlated poorly with the elevation of fecal calprotectin, C-reactive protein, or a

75

76

CHAPTER 5 The role of capsule endoscopy in IBD

combination of both markers (r 5 0.2) [25]. On the basis of these data, the use of biomarkers as a triage tool would have missed some patients with moderate-toseverely inflamed small bowel. Mucosal healing, defined as the absence of visible endoscopic inflammation, has become established as an important endpoint for the treatment of Crohn’s disease. It has been associated with improvements in the quality of life and in clinically relevant outcomes, including frequency of hospitalization, rates of surgery, and sustained steroid-free remission [26]. Conventional ileocolonoscopy is the current gold-standard modality for the assessment of mucosal healing. Although a gold standard for small bowel mucosal healing in Crohn’s disease has yet to be established, a Lewis score of ,135 is accepted as representing clinically insignificant inflammation [20,27], as a Lewis score of 135 usually represents the presence of at least 1 small bowel ulcer. Several studies have shown that CE significantly changed the therapeutic management of Crohn’s disease patients, even in those with long-term disease. Dussault et al. performed a retrospective study on 71 patients to assess Crohn’s disease and showed that the findings at CE led to a change in medical therapy in 54% of patients within three months of the investigation [28]. Similarly, in a study that included 86 patients with Crohn’s disease, alteration in therapy occurred in 62% of patients as a consequence of findings from CE [18]. In 40%, this took the form of a new antiinflammatory medication, the most common of which was a corticosteroid. Finally, Santos-Antunes et al. showed, in a retrospective study of 106 patients, that CE determined changes in the treatment of 40% of patients [29]. Indeed, only 21% remained free of immunosuppressors after CE compared with 44% before the procedure. The authors also revealed that a higher Lewis score was associated with therapeutic modifications.

Assessment of postoperative recurrence Recurrence of small bowel Crohn’s disease in the neoterminal ileum following surgical resection can be demonstrated in 73% of the patients within 1 year of ileocolonic resection (Fig. 5.4) [30]. Of these patients, 80% are symptom-free. Early detection of endoscopic recurrence is important for starting biological therapy and the prevention of clinical recurrence and the need to repeat surgery. Therefore, IBD experts advocate routine endoscopic assessment postoperatively and offer a step-up in treatment to those with significant recurrence [31]. Although the standard method for assessing this is ileocolonoscopy, CE is an attractive monitoring modality for postoperative patients, providing a noninvasive and accurate visualization of the entire small bowel, including the neoterminal ileum. In fact, some authors have already reported that CE detected lesions in the small bowel beyond the reach of ileocolonoscopy in up to two-thirds of patients [32].

Role of capsule endoscopy in reclassification

FIGURE 5.4 Postoperative recurrence of Crohn’s disease in the neoterminal ileum (A) and ileocolic anastomosis (B).

Role of capsule endoscopy in reclassification of inflammatory bowel disease The term IBD unclassified (IBDU) is conventionally used to classify patients in whom ulcerative colitis and Crohn’s disease cannot be distinguished based on endoscopic and histological assessments [1]. At least 30% of these patients will be reclassified as Crohn’s disease during the course of their illness, usually after the identification of small bowel lesions [33]. The reclassification of the diagnosis is of particular relevance to patients in whom the formation of an ileoanal pouch is being considered, as rates of chronic pouchitis, fistula formation, and pouch failure are higher in patients with Crohn’s disease compared with patients with ulcerative colitis [34]. Several small studies have evaluated the utility of CE for the reclassification of IBDU. Mow et al. described the use of CE in patients with an established diagnosis of IBD who had previously undergone a radiological assessment of the small bowel [14]. Of 21 patients, 12 patients (57.1%) with ulcerative colitis or IBDU were reclassified as having Crohn’s disease after CE. Similarly, Mehdizadeh et al. reported that 19 of 120 (15.8%) patients with IBDU or ulcerative colitis were found to have CE findings consistent with Crohn’s disease. In both these studies, the reclassification of patients as having Crohn’s disease was based on the identification of inflammatory lesions within the small bowel. CE also has a role in classifying disease type in the pediatric population, affecting patient management and altering health outcomes. Cohen performed CE in 28 patients with an average of 4.2 6 3 years after the original IBD

77

78

CHAPTER 5 The role of capsule endoscopy in IBD

diagnosis [35]. The authors showed that, following CE examination, 4 of 5 patients with ulcerative colitis and 1 of 2 patients with IBDU (total 71%) had their disease reclassified to Crohn’s disease based upon newly diagnosed small bowel mucosal lesions. Moreover, 13 of 21 (62%) patients with Crohn’s disease were found to have more extensive small bowel disease at the time of CE.

Colon capsule endoscopy The CCE was developed in 2006 to allow a noninvasive visualization of the colon [36]. A prospective study with first-generation CCE (CCE-1) demonstrated that the sensitivity of CCE-1 for detecting colonic lesions was low compared with colonoscopy, and the results were unsatisfactory [37]. Thus, second-generation CCE (CCE-2) was subsequently developed and reported in 2009. CCE-2 is equipped with 2 high-resolution cameras providing a viewing angle of 172 degrees in front and back, which sense the moving speed of the capsule endoscope and capture 4 to 35 images per second (Fig. 5.5B). A prospective European multicenter study showed that the detection rate of colon polyps of .5 mm using CCE-2 was almost equivalent to colonoscopy [38]. More recently, the PillCam Crohn’s Capsule (PCC) was developed, which is a panenteric video capsule system that allows visualization of the small and the large bowel (Fig. 5.5C). Bowel preparation in CCE is designed to optimize mucosal examination and excretion rates. The usual bowel cleansing regimen includes 4 L of polyethylene glycol, split into 2 doses [39]. During the procedure, further boosters based on sodium phosphate are used to enhance the propulsion of the capsule through the small bowel and colon. Currently, CCE is primarily utilized in screening for colonic neoplasia, particularly in situations such as incomplete colonoscopy. Although CCE has several limitations, including high preparation volume and cost, the demand in clinical settings is gradually increasing. Indeed, CCE is minimally invasive and especially useful for patients with IBD and ulcerative colitis in particular.

FIGURE 5.5 Commercial wireless capsule endoscopes: PillCam SB3 by Medtronic (A), PillCam Colon 2 by Medtronic (B) and PillCam Crohn’s by Medtronic (C).

Colon capsule endoscopy in Crohn’s disease

Colon capsule endoscopy in Crohn’s disease Since PCC captures images of the entire gastrointestinal tract, its potential role in assessing inflammation of Crohn’s disease has prompted interest (Fig. 5.6). It is also considered to be useful for evaluating postoperative cases (Fig. 5.7). Some studies have already confirmed the safety and feasibility of PCC as a panenteric tool for patients with Crohn’s disease. In 2015 D’Haens et al. compared PCC and colonoscopy to assess disease severity in patients with active colonic Crohn’s disease [40]. The authors showed that there was a substantial agreement between the Crohn’s Disease Endoscopic Index of Severity score calculated using both modalities (intraclass correlation coefficient 5 0.65). The greatest agreement between

FIGURE 5.6 PillCam Crohn’s Capsule images—small bowel Crohn’s disease (A) and colon Crohn’s disease (B).

79

80

CHAPTER 5 The role of capsule endoscopy in IBD

FIGURE 5.7 PillCam Crohn’s Capsule showing ileocolic perianastomotic ulcers.

colonoscopy and PCC was observed in the ileum with a trend toward poorer agreement near the distal colon. Another study compared PCC with magnetic resonance enterography and small intestine contrast ultrasonography in pediatric Crohn’s disease patients [35]. It was demonstrated that PCC was superior to the other techniques for the detection of colonic Crohn’s disease lesions with sensitivity and specificity of 89% and 100% and positive and negative predictive values of 100% and 91%, respectively. In 2021 Yamada et al. compared PCC with double-balloon endoscopy as the gold standard [41]. The study evaluated the presence of ulcer scars, erosions, and ulcers in both the small and the large bowel. For the 60 small bowel segments, sensitivities were 84.2%, 95.5%, and 90.0%, respectively, and the specificities were 63.4%, 86.8%, and 87.5%, respectively. For the 64 large bowel segments, the sensitivities were 80.0%, 90.0%, and 83.3%, and specificities were 84.7%, 72.2%, and 77.6%, respectively. The relatively low specificities of PCC for the detection of erosions and ulcers in the colon were thought to reflect the quality of bowel preparation, with adherent feces often mistaken for erosions. With the current availability of PCC, which allows for the visualization of colonic inflammatory disease activity, Niv et al. extended the validated CECDAI score to include the colon, introducing the novel CECDAIic score, allowing for

Colon capsule endoscopy in ulcerative colitis

an objective panenteric assessment of Crohn’s disease inflammatory activity [42]. It evaluates the inflammation, the extent of the disease, and the strictures in the proximal small bowel, distal small bowel, right colon, and left colon (Table 5.3). The currently available data support the use of the CECDAIic score to objectively assess the inflammatory disease activity in the small bowel and colonic Crohn’s disease [43]. The CECDAIic is reproducible, with a high degree of interobserver agreement for the overall score and also for its individual subscores. Although the PCC is conceivably an appropriate endoscopic method for mapping and grading established Crohn’s disease, further studies are needed to support its role in a treat-to-target strategy for disease management and monitoring [3]. Preliminary research on the role of artificial intelligence in the detection of erosions and ulcers using PillCam Crohn’s capsule has shown promising results and may not only improve accuracy but also reduce reading times meriting further investigation [44].

Colon capsule endoscopy in ulcerative colitis CCE was also evaluated for diagnosis and monitoring of ulcerative colitis (Fig. 5.8). In one of the largest studies, 100 patients with suspected or confirmed ulcerative colitis were assessed with CCE and colonoscopy [8]. CCE had a sensitivity and specificity for the detection of colonic inflammation of 89% and 75%, respectively. In the study by Ye et al., 25 patients were evaluated for presence and Table 5.3 Total score 5 (A1 3 B1 1 C1) 1 (A2 3 B2 1 C2) 1 (A3 3 B3 1 C3) 1 (A4 3 B4 1 C4). Parameter

Score and descriptor

A—Inflammation

0—None 1—Mild to moderate (edema, hyperemia, or denudation) 2—Severe (edema, hyperemia, or denudation) 3—Bleeding, exudate, erosion aphthae, ulcers ,0.5 cm 4—Pseudopolyp, ulcers 0.5 2 cm 5—Ulcers .2 cm 0—None 1—Single segment (focal disease) 2—2 3 segments (patchy disease) 3— . 3 segments (diffuse disease) 0—None 1—Single-passed 2—Multiple-passed 3—Obstruction (nonpassage)

B—Extent of disease

C—Stricture

(1) Proximal small bowel; (2) distal small bowel; (3) right colon; (4) left colon.

81

82

CHAPTER 5 The role of capsule endoscopy in IBD

FIGURE 5.8 Colon capsule endoscopy images of patients with ulcerative colitis—mucosa with edema, erythema, and multiple ulcers.

severity (Mayo Score) of ulcerative colitis by CCE and conventional colonoscopy [45]. A significant correlation in the severity (k 5 0.751) and extent (k 5 0.522) of ulcerative colitis between the CCE and conventional colonoscopy was demonstrated. In addition, there were no remarkable adverse events during the study. Similarly, Hosoe et al. reported that endoscopic scores determined by CCE had a strong correlation with scores obtained by conventional colonoscopy [46]. However, there are several limitations in the use of CCE to assess ulcerative colitis. First, the disease may only involve the distal colon, and an incomplete CCE examination would fail to identify inflammatory pathology in these patients. The inability to obtain biopsy specimens is a further limitation. Therefore, its role in ulcerative colitis would not encompass surveillance for monitoring of dysplasia and cancer surveillance or scenarios in which biopsies to exclude superadded cytomegalovirus infection are required [47]. In 2018, Hosoe et al. developed an endoscopic severity score for ulcerative colitis using CCE [48]. The descriptors used in the Capsule Scoring of Ulcerative Colitis (CSUC) score including vascular pattern, bleeding, erosions, and ulcers are shown in Table 5.4. The CSUC score is obtained by adding the proximal and distal colons scores (0 14 points). CSUC may be used as a predictor for the risk of relapse during clinical remission. In a retrospective observational study, patients were more likely to maintain clinical remission for a year if they had a CSUC score of # 1. In addition, a CSUC $ 1 was shown to be a predictor of relapse (area under the curve of 0.82, sensitivity of 83.3%, and specificity of 58.6%) [49]. It is unlikely that CCE will replace fecal biomarkers and ileocolonoscopy as the first-line investigation for UC. However, it remains a viable alternative, for

Cost-effectiveness of colon capsule endoscopy

Table 5.4 Capsule scoring of ulcerative colitis descriptors and definitions. Descriptor Vascular pattern Bleeding

Likert scale anchor points Normal (0) Patchy obliteration (1) Obliterated (2) None (0) Mild (1) Severe (2)

Erosions and ulcers

None (0) Erosions (1) Superficial ulcer (2) Deep ulcer (3)

Definition Normal vascular pattern Obliterated area # 30% Obliterated area . 30% No visible blood detected by image reading software No bleeding picture detected by image reading software # 10 No bleeding picture detected by image reading software . 10 Normal mucosa, no visible erosions or ulcers Tiny (# 5 mm) defects in the mucosa Larger ( . 5 mm) defects in the mucosa Larger ( . 5 mm) and deeper excavated defects in the mucosa, with a slightly raised edge

example, for those with indeterminate results of fecal biomarkers but are reluctant or at high risk for ileocolonoscopy. In addition, CCE is an attractive minimally invasive method for disease monitoring, which includes assessments of mucosal healing in patients in clinical remission.

Cost-effectiveness of colon capsule endoscopy in inflammatory bowel disease Health economics is an important consideration in the management of chronic illnesses. Although several studies have investigated the cost-effectiveness of CCE in colorectal cancer screening, there is a paucity of publications on IBD. In one of the largest studies, 4000 simulated Crohn’s disease patients were investigated for the cost-effectiveness of panintestinal capsule endoscopy within the British National Health Service [50]. The authors estimated the annual mean cost per patient to be 2191 GBP for those receiving standard care (typically utilizing ileocolonoscopy and/or magnetic resonance enterography), with a 20-year estimate of 42,266 GBP. The cost of care using CCE for disease assessment was lower, with a per annum cost of 1960 GBP and a mean 20-year cost of 38,433 GBP. Similar results were demonstrated in one US study, where 4000 simulated Crohn’s disease patients were analyzed. Common monitoring practice (ileocolonoscopy and imaging) was compared with Crohn’s capsule monitoring strategies [51]. Over 20 years, the use of the CCE reduced costs (313,367 USD vs. 320,015 USD), increased life expectancy (18.15 vs. 17.9 years), and increased the quality of life

83

84

CHAPTER 5 The role of capsule endoscopy in IBD

(8.7 vs. 8.0 quality-adjusted life years). Despite the limitations of such simulated analyses, the studies suggested that the use of the CCE would be cost-effective for monitoring Crohn’s disease patients. Unfortunately, no data is available on the cost-effectiveness of CCE in ulcerative colitis.

Complications of capsule endoscopy The main complication of capsule endoscopy is capsule retention, defined as the failure of the video capsule to pass through the gastrointestinal tract after 2 weeks without medical intervention. It is more common in patients who performed the procedure for suspected or definite Crohn’s disease. A systematic review, which included 2538 CE procedures, revealed a capsule retention rate of 2.6% in patients with definite or suspected Crohn’s disease compared with an overall retention rate of 1.4% [52]. In patients with a retained capsule due to a Crohn’s inflammatory stricture (Fig. 5.9), a short course of steroids may enable the capsule to pass spontaneously. However, some patients may require endoscopy or surgery to retrieve the capsule [53]. Therefore, strategies to reduce the risk of capsule retention should be carried out in those with small bowel strictures, previous abdominal surgeries, and a prior history of small bowel obstruction. Conventional small bowel imaging, such as computed tomographic enterography and magnetic resonance enterography, are useful adjuncts to identify small bowel features that may contraindicate the use of capsule endoscopy. Nevertheless, capsule retention may still occur if small bowel imaging misses clinically significant structuring disease. In a retrospective study of 50 patients with a confirmed diagnosis of

FIGURE 5.9 Capsule endoscopy (A) and PillCam Crohn’s Capsule (B) showing Crohn’s inflammatory strictures.

New research areas for future

Crohn’s disease, 6% had capsule retention despite normal cross-sectional small bowel imaging studies and no history of obstructive symptoms [54]. A patency capsule was developed for use as a prescreening tool to reduce the risk of capsule retention in patients undergoing capsule endoscopy. It is the same size and shape as the video capsule. However, it contains a radiofrequency emitter that can be detected by a hand-held scanner. If the capsule is retained in the gastrointestinal tract, it disintegrates into small, mostly soft, fragments that can easily pass through strictures (Fig. 5.10). On the other hand, if the patency capsule is successfully excreted or not detectable on the scanner at 30 h postingestion, it is usually safe to perform the diagnostic capsule endoscopy. Indeed, video capsule retention is a rare occurrence after a negative patency capsule test (0.6%) [25,55]. A prospective study assessed with computed tomography, which included 54 patients with established Crohn’s disease, reported a patency capsule small bowel retention rate of 9% [56]. Recently, a metaanalysis estimated a CE retention rate of 4.63% in established Crohn’s disease and 2.35% in suspected Crohn’s disease. This metaanalysis included some studies that enrolled patients previously tested with a patency capsule [57]. Thus current guidelines advise the use of a patency capsule prior to capsule endoscopy in those at a higher risk of capsule retention [13,57]. The handful of cases of perforation reported in patients undergoing investigation with CE has largely occurred in patients with capsule retention and an established diagnosis of Crohn’s disease [54]. Aspiration of the video capsule occurs rarely and has been reported in 1 in 800 examinations.

New research areas for future Technological enhancements in the future may potentially lead to a further expansion of the indications for capsule endoscopy in IBD. These improvements may include the development of maneuverable systems for targeted inspection of the gastrointestinal tract, using magnetic fields for steering of a video capsule with

FIGURE 5.10 Patency capsule intact (A), partially dissolved (B), and completely dissolved (C).

85

86

CHAPTER 5 The role of capsule endoscopy in IBD

magnetic inclusions [58]. An additional significant limitation of the capsule endoscopy is the lack of sampling ability, diminishing its usefulness for monitoring neoplasms and colonic or small bowel dysplasia. Technological advances that are under development include tissue diagnosis capabilities, fluid aspiration, and therapeutic capabilities [59]. Automated software for the detection of inflammation and further development of designated IBD systems may further increase the clinical utility of this modality. Capsule reading times limit the number of procedures that can be assessed by a single reader. This might soon be addressed by the use of artificial intelligence. Machine learning has been used to train computers to analyze video capsule recordings to diagnose and assess disease severity [60]. In the near future, it is likely that automation will play an integral part in the clinical practice of gastroenterology.

Conclusion CE has evolved into an important complementary tool for investigating the small bowel in patients with suspected or established Crohn’s disease. It is a minimally invasive and well-tolerated test with a high diagnostic yield. Its place in the monitoring of Crohn’s disease and the implications of CE findings for the treatment of Crohn’s disease are becoming better understood. The more recent development of CCE has expanded the potential applications of capsule endoscopy to include the assessment of ulcerative colitis and to provide a panenteric assessment of patients with Crohn’s disease.

References [1] Lamb CA, Kennedy NA, Raine T, et al. British Society of Gastroenterology consensus guidelines on the management of inflammatory bowel disease in adults. Gut 2019;68 (Suppl 3):s1 s106. [2] Collins PD. Video capsule endoscopy in inflammatory bowel disease. World J Gastrointest Endosc. 2016;8(14):477 88. [3] Halder W, Laskaratos FM, El-Mileik H, et al. Review: Colon capsule endoscopy in inflammatory bowel disease. Diagnostics (Basel). 2022;12(1). [4] Hall B, Holleran G, McNamara D. Small bowel Crohn’s disease: an emerging disease phenotype? Dig Dis 2015;33(1):42 51. [5] Bourreille A, Ignjatovic A, Aabakken L, et al. Role of small-bowel endoscopy in the management of patients with inflammatory bowel disease: an international OMEDECCO consensus. Endoscopy 2009;41(7):618 37. [6] Rubin DT, Ananthakrishnan AN, Siegel CA, et al. ACG clinical guideline: ulcerative colitis in adults. Am J Gastroenterol 2019;114(3):384 413.

References

[7] Pineton de Chambrun G, Peyrin-Biroulet L, Le´mann M, et al. Clinical implications of mucosal healing for the management of IBD. Nat Rev Gastroenterol Hepatol. 2010;7(1):15 29. [8] Sung J, Ho KY, Chiu HM, et al. The use of Pillcam Colon in assessing mucosal inflammation in ulcerative colitis: a multicenter study. Endoscopy 2012;44 (8):754 8. [9] Rodrigues-Pinto E, Cardoso H, Rosa B, et al. Development of a predictive model of Crohn’s disease proximal small bowel involvement in capsule endoscopy evaluation. Endosc Int Open. 2016;4(6):E631 6. [10] Dionisio PM, Gurudu SR, Leighton JA, et al. Capsule endoscopy has a significantly higher diagnostic yield in patients with suspected and established small-bowel Crohn’s disease: a meta-analysis. Am J Gastroenterol 2010;105(6):1240 8. [11] Hall B, Holleran G, Costigan D, et al. Capsule endoscopy: high negative predictive value in the long term despite a low diagnostic yield in patients with suspected Crohn’s disease. United European Gastroenterol J. 2013;1(6):461 6. [12] Solem CA, Loftus EV, Fletcher JG, et al. Small-bowel imaging in Crohn’s disease: a prospective, blinded, 4-way comparison trial. Gastrointest Endosc 2008;68 (2):255 66. [13] Pennazio M, Spada C, Eliakim R, et al. Small-bowel capsule endoscopy and deviceassisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Clinical Guideline. Endoscopy 2015;47(4):352 76. [14] Mow WS, Lo SK, Targan SR, et al. Initial experience with wireless capsule enteroscopy in the diagnosis and management of inflammatory bowel disease. Clin Gastroenterol Hepatol 2004;2(1):31 40. [15] Niv Y, Ilani S, Levi Z, et al. Validation of the capsule endoscopy Crohn’s disease activity index (CECDAI or Niv score): a multicenter prospective study. Endoscopy 2012;44(1):21 6. [16] Monteiro S, Boal Carvalho P, Dias de Castro F, et al. Capsule endoscopy: diagnostic accuracy of Lewis score in patients with suspected Crohn’s disease. Inflamm Bowel Dis 2015;21(10):2241 6. [17] Omori T, Kambayashi H, Murasugi S, et al. Comparison of Lewis score and capsule endoscopy Crohn’s disease activity index in patients with Crohn’s disease. Dig Dis Sci 2020;65(4):1180 8. [18] Long MD, Barnes E, Isaacs K, et al. Impact of capsule endoscopy on management of inflammatory bowel disease: a single tertiary care center experience. Inflamm Bowel Dis 2011;17(9):1855 62. [19] Dias de Castro F, Boal Carvalho P, Monteiro S, et al. Lewis score prognostic value in patients with isolated small bowel Crohn’s disease. J Crohns Colitis. 2015;9 (12):1146 51. [20] Santos A, Silva MA, Cardoso H, et al. Lewis score: a useful tool for diagnosis and prognosis in Crohn’s disease. Rev Esp Enferm Dig 2020;112(2):121 6. [21] Kopylov U, Yablecovitch D, Lahat A, et al. Detection of small bowel mucosal healing and deep remission in patients with known small bowel Crohn’s disease using biomarkers, capsule endoscopy, and imaging. Am J Gastroenterol 2015;110(9):1316 23. [22] Lewis JD. The utility of biomarkers in the diagnosis and therapy of inflammatory bowel disease. Gastroenterology 2011;140(6):1817 26 e2.

87

88

CHAPTER 5 The role of capsule endoscopy in IBD

[23] Niv E, Fishman S, Kachman H, et al. Sequential capsule endoscopy of the small bowel for follow-up of patients with known Crohn’s disease. J Crohns Colitis 2014;8 (12):1616 23. [24] Koulaouzidis A, Douglas S, Rogers MA, et al. Fecal calprotectin: a selection tool for small bowel capsule endoscopy in suspected IBD with prior negative bi-directional endoscopy. Scand J Gastroenterol 2011;46(5):561 6. [25] Kopylov U, Nemeth A, Koulaouzidis A, et al. Small bowel capsule endoscopy in the management of established Crohn’s disease: clinical impact, safety, and correlation with inflammatory biomarkers. Inflamm Bowel Dis 2015;21(1):93 100. [26] Dulai PS, Levesque BG, Feagan BG, et al. Assessment of mucosal healing in inflammatory bowel disease: review. Gastrointest Endosc 2015;82(2):246 55. [27] Gralnek IM, Defranchis R, Seidman E, et al. Development of a capsule endoscopy scoring index for small bowel mucosal inflammatory change. Aliment Pharmacol Ther 2008;27(2):146 54. [28] Dussault C, Gower-Rousseau C, Salleron J, et al. Small bowel capsule endoscopy for management of Crohn’s disease: a retrospective tertiary care centre experience. Dig Liver Dis 2013;45(7):558 61. [29] Santos-Antunes J, Cardoso H, Lopes S, et al. Capsule enteroscopy is useful for the therapeutic management of Crohn’s disease. World J Gastroenterol 2015;21 (44):12660 6. [30] Rutgeerts P, Geboes K, Vantrappen G, et al. Predictability of the postoperative course of Crohn’s disease. Gastroenterology 1990;99(4):956 63. [31] Jones GR, Kennedy NA, Lees CW, et al. Systematic review: the use of thiopurines or anti-TNF in post-operative Crohn’s disease maintenance progress and prospects. Aliment Pharmacol Ther 2014;39(11):1253 65. [32] Bourreille A, Jarry M, D’Halluin PN, et al. Wireless capsule endoscopy vs ileocolonoscopy for the diagnosis of postoperative recurrence of Crohn’s disease: a prospective study. Gut 2006;55(7):978 83. [33] Eliakim R. The impact of wireless capsule endoscopy on gastrointestinal diseases. South Med J 2007;100(3):235 6. [34] Fazio VW, Ziv Y, Church JM, et al. Ileal pouch-anal anastomoses complications and function in 1005 patients. Ann Surg 1995;222(2):120 7. [35] Oliva S, Cucchiara S, Civitelli F, et al. Colon capsule endoscopy compared with other modalities in the evaluation of pediatric Crohn’s disease of the small bowel and colon. Gastrointest Endosc 2016;83(5):975 83. [36] Eliakim R, Fireman Z, Gralnek IM, et al. Evaluation of the PillCam Colon capsule in the detection of colonic pathology: results of the first multicenter, prospective, comparative study. Endoscopy 2006;38(10):963 70. [37] Van Gossum A, Munoz-Navas M, Fernandez-Urien I, et al. Capsule endoscopy vs colonoscopy for the detection of polyps and cancer. N Engl J Med 2009;361(3):264 70. [38] Spada C, Hassan C, Munoz-Navas M, et al. Second-generation colon capsule endoscopy compared with colonoscopy. Gastrointest Endosc 2011;74(3):581 9 e1. [39] Spada C, Hassan C, Galmiche JP, et al. Colon capsule endoscopy: European Society of Gastrointestinal Endoscopy (ESGE) Guideline. Endoscopy 2012;44(5):527 36. [40] D’Haens G, Lo¨wenberg M, Samaan MA, et al. Safety and feasibility of using the second-generation Pillcam Colon capsule to assess active colonic Crohn’s disease. Clin Gastroenterol Hepatol 2015;13(8):1480 6 e3.

References

[41] Yamada K, Nakamura M, Yamamura T, et al. Diagnostic yield of colon capsule endoscopy for Crohn’s disease lesions in the whole gastrointestinal tract. BMC Gastroenterol 2021;21(1):75. [42] Spada C, Hassan C, Bellini D, et al. Imaging alternatives to colonoscopy: CT colonography and colon capsule. European Society of Gastrointestinal Endoscopy (ESGE) and European Society of Gastrointestinal and Abdominal Radiology (ESGAR) Guideline - Update 2020. Endoscopy 2020;52(12):1127 41. [43] Tabone T, Koulaouzidis A, Ellul P. Scoring systems for clinical colon capsule endoscopy-all you need to know. J Clin Med 2021;10(11). [44] Ferreira JPS, Mascarenhas Saraiva MJ, Afonso JPL, et al. Identification of ulcers and erosions by the novel Pillcam Crohn’s capsule using a convolutional neural network: a multicentre pilot study. J Crohn’s Colitis 2022;16(1):169 72. [45] Ye CA, Gao YJ, Ge ZZ, et al. PillCam colon capsule endoscopy vs conventional colonoscopy for the detection of severity and extent of ulcerative colitis. J Dig Dis 2013;14(3):117 24. [46] Hosoe N, Matsuoka K, Naganuma M, et al. Applicability of second-generation colon capsule endoscope to ulcerative colitis: a clinical feasibility study. J Gastroenterol Hepatol 2013;28(7):1174 9. [47] Kopylov U, Seidman EG. Role of capsule endoscopy in inflammatory bowel disease. World J Gastroenterol 2014;20(5):1155 64. [48] Hosoe N, Nakano M, Takeuchi K, et al. Establishment of a novel scoring system for colon capsule endoscopy to assess the severity of ulcerative colitis-capsule scoring of ulcerative colitis. Inflamm Bowel Dis 2018;24(12):2641 7. [49] Matsubayashi M, Kobayashi T, Okabayashi S, et al. Determining the usefulness of capsule scoring of ulcerative colitis in predicting relapse of inactive ulcerative colitis. J Gastroenterol Hepatol 2021;36(4):943 50. [50] Lobo A, Torrejon Torres R, McAlindon M, et al. Economic analysis of the adoption of capsule endoscopy within the British NHS. Int J Qual Health Care 2020;32 (5):332 41. [51] Saunders R, Torrejon Torres R, Konsinski L. Evaluating the clinical and economic consequences of using video capsule endoscopy to monitor Crohn’s disease. Clin Exp Gastroenterol 2019;12:375 84. [52] Liao Z, Gao R, Xu C, et al. Indications and detection, completion, and retention rates of small-bowel capsule endoscopy: a systematic review. Gastrointest Endosc 2010;71 (2):280 6. [53] Cave D, Legnani P, de Franchis R, et al. ICCE consensus for capsule retention. Endoscopy 2005;37(10):1065 7. [54] Cotter J, Dias de Castro F, Moreira MJ, et al. Tailoring Crohn’s disease treatment: the impact of small bowel capsule endoscopy. J Crohns Colitis 2014;8 (12):1610 15. [55] Silva M, Cardoso H, Macedo G. Patency capsule safety in Crohn’s disease. J Crohns Colitis 2017;11(10):1288. [56] Silva M, Cardoso H, Cunha R, et al. Evaluation of small-bowel patency in Crohn’s disease: prospective study with a patency capsule and computed tomography. GE Port J Gastroenterol 2019;26(6):396 403. [57] Pasha SF, Pennazio M, Rondonotti E, et al. Capsule retention in Crohn’s disease: a meta-analysis. Inflamm Bowel Dis 2020;26(1):33 42.

89

90

CHAPTER 5 The role of capsule endoscopy in IBD

[58] Keller J, Fibbe C, Rosien U, et al. Recent advances in capsule endoscopy: development of maneuverable capsules. Expert Rev Gastroenterol Hepatol 2012;6 (5):561 6. [59] Koulaouzidis A, Rondonotti E, Karargyris A. Small-bowel capsule endoscopy: a tenpoint contemporary review. World J Gastroenterol 2013;19(24):3726 46. [60] Mascarenhas M, Afonso J, Andrade P, et al. Artificial intelligence and capsule endoscopy: unravelling the future. Ann Gastroenterol 2021;34(3):300 9.

CHAPTER

Artificial intelligence for automatic detection of blood and hematic residues

6

Gerardo Blanco Sr, Oscar Mondragon and Omar Solo´rzano Department of Endoscopy, Hospital de Especialidades, Centro Me´dico Nacional Siglo XXI, Instituto Mexicano del Seguro Social, Mexico City, Mexico

Small bowel bleeding (SBB) accounts for 5% 10% of all gastrointestinal bleeding, making it an uncommon event [1]. Endoscopic tests available for the diagnosis and management of the SBB are capsule endoscopy (CE) and device-assisted enteroscopy. The CE is considered the first-line investigation method for small bowel evaluation when a patient is stable and with no evidence of obstruction [2] since it is a non-invasive and patient-friendly method that can visualize the entire small bowel mucosa [3]. The diagnostic yield of the CE depends on the time when the test is performed. Pennazio et al. reported that the diagnostic yield of the CE can be as high as 92.3% in patients with ongoing obscure overt bleeding and as low as 12.9% in patients with previous overt bleeding [4]. However, the possibility of observing the presence of active bleeding during a CE test is even lower. A study noted that its investigators detected the presence of active bleeding or hematic residues in 42 of 199 CE (21%) performed in patients with suspected SBB [5]. In another article, reported by one of the authors of this chapter, active bleeding or blood debris was identified in only 22 of 223 CE (9.8%) placed for suspected SBB [6]. Since its development, the software for the interpretation and reading of the CE sought an artificial intelligence (AI) model for the detection of active bleeding or blood remains in the small bowel. Given, as part of its RAPID CE reading software, develops technical features to make CE video analysis easier and shorter with the intention of increasing diagnostic yields. The first software feature designed was the suspected blood indicator (SBI), which is an image selection feature that detects video frames with red pixels that possibly represent areas of hemorrhage in the gastrointestinal tract. Other software tools introduced were QuickView and the electronic chromoendoscopy [7,8]. The SBI feature is activated the moment the first duodenal image is selected. Since 2003, more than 20 studies have been carried out that assess the usefulness of the SBI in SBB [9]. A metaanalysis realized in 2017 included 16 studies with 2040 patients that underwent 2049 CE examinations. The study found that for any bleeding or potentially bleeding lesions the SBI has a sensitivity of 0.553 Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00010-1 © 2023 Elsevier Inc. All rights reserved.

91

92

CHAPTER 6 Artificial intelligence for detection of blood

(95% confidence interval (CI), 0.510 0.596), specificity of 0.578 (95% CI, 0.547 0.608), and diagnostic odds ratio (DOR) of 12.354 (95% CI, 3.297 46.297), with an area under the curve (AUC) of 0.878. But when the SBI was used only for active bleeding, the sensitivity was 0.988 (95% CI, 0.956 0.999), specificity was 0.646 (95% CI, 0.610 0.680), and a DOR of 229.89 (95% CI, 20.748 2547.3), with an AUC of 0.993. The study concluded that although SBI has a limited validation in CE reading, it has a good sensitivity supporting its use in SBB [10]. Another Portuguese study from 2017, which is not included in the recently mentioned metaanalysis, evaluated CE ingested by 281 patients and analyzed the sensitivity, specificity, positive predictive value, and negative predictive value (NPV) of the SBI in hemorrhage and P2 lesions. The results for active hemorrhage were 96.6%, 17.1%, 11.8%, and 97.7%, respectively, while for P2 lesions the results were 39.5%, 51.2%, 19%, and 74.5%, respectively. They concluded that although SBI sensitivity for potential bleeding lesions was low, it effectively detected active SBB with very high sensitivity and NPV [11]. The study performed in 2018, carried out on the usefulness of the SBI in the identification of active gastrointestinal bleeding, found that two or more contiguous SBI markers had a sensitivity of 96.5%, while the presence of eight contiguous SBI markers had a sensitivity and specificity of 100% (P , .001) for active gastrointestinal bleeding [12] (Fig. 6.1).

Artificial intelligence The first developments of AI in fixed images were based on the scrutiny of colors and textures. Blood does not have a typical shape or texture; the color of blood may vary depending on the position of the camera, active hemorrhage, and intentional content [13]. Some methods of hemorrhage detection are: • • •

Detection of areas of hemorrhage in videos from endoscopic capsules using textures based on chrominance [14]. Algorithm of analysis of the main components of joint diagonalization combined with the vector of color coherence [15]. Algorithm of automatic segmentation for the detection of hemorrhage in images in the hyperspectral imaging (HSI) color space using the channel of intensity to extract texture characteristics [16].

These methods have evolved for the recognition of images through algorithms that identify a problem or clinical situation. AI tries to imitate biological functions. Data is processed to imitate the visual process of the cerebral cortex— image, retina, cerebral cortex—and, finally, recognition of the image. The process is cyclical, with its own or supervised management. Supervised learning uses

FIGURE 6.1 Rapid v8 software with SBI. (A) One SBI marker alone with no evidence of small bowel bleeding in the image. (B) and (C) More than eight SBI markers together with active small bowel bleeding in the images. SBI, Suspected blood indicator.

94

CHAPTER 6 Artificial intelligence for detection of blood

tagged data or a known response; on the other hand, unsupervised learning is for automated groupings of similar data with common characteristics. Cycling of this method refers to recognizing new information, processing it with entry algorithms, and returning to the initial part of the cycle to recognize new characteristics of the problem, which help to process previously unrecognized information to resolve new cases. These techniques of machine learning are known as artificial neural networks (ANNs) and support vector machines (SVMs) [17].

Support vector machines SVM is a machine learning method that classifies data sets into different categories through the creation of a linear structure. The separation of data into different categories allows the machine to classify newly entered data based on previously entered data [18].

Artificial neural network ANN is a machine learning method in which multiple interconnected layers are programmed to process data with a specific pattern, which are fed among them so that the software can be trained to carry out a specific task. The concept is based on the synaptic function of the human brain, with interaction between many neurons on various levels. McCulloch and Pitts first proposed this concept in 1943. ANN can be used in the analysis of images with a specific pattern: (1) An image is composed of pixels, (2) pixels are analyzed based on specific algorithms, (3) data is combined and processed, and (4) the result is finally offered. The result can be the categorization or classification of an object in the imagen, or it may be the detection of a specific characteristic, such as the detection of active bleeding in an endoscopic capsule imagen (Fig. 6.2). In 1980 a method was created for the recognition of patterns, which combined the characteristics of the entry pattern, leading to the creation of the convolutional neural network (CNN) [19].

Convolutional neural network CNN is a machine learning technique that organizes its connections into 3 or more layers, unlike traditional ANN, which has a neural network of two dimensions (layers). The occult layers in CNN process the data with a nonlinear focus. This technique of machine learning is known as deep learning. In this way, CNN can extract characteristics from an image and process the data more efficiently

Convolutional neural network

FIGURE 6.2 Sample bleeding images with their corresponding ground truth (pixel-level detection).

95

96

CHAPTER 6 Artificial intelligence for detection of blood

and autonomously. The steps of a CNN model can be summarized into the following: (1) extraction of the characteristics of the image, (2) reduction of dimensions to create multiple layers of the image, (3) relating the layers of the image, and finally (4) classifying them. The interpretation of endoscopic capsules aided by systems based on selflearning is a concept that has developed recently [17,20,21]. Some examples of the CNN systems are:

ESNavi ESNavi is a platform in the cloud developed by Ankon Technologies Co. Ltd. It automatically detects various anomalies in endoscopic capsule images. It identified 4206 lesions in 3280 patients in validation images. The auxiliary method based on CNN identified lesions with a sensitivity of 99.88% in the analysis per patient (95% CI, 99.67 99.96) and a sensitivity of 99.90% in the analysis by lesions (95% CI, 99.74 99.97). Conventional reading identified anomalies with a sensitivity of 74.57% (95% CI, 73.05 76.03) in the analysis by patients and 76.89% in the analysis by lesions (95% CI, 75.58 78.15). The mean time for reading the CE images of one patient was 96.6 6 22.53 min per conventional reading and 5.9 6 2.23 min for auxiliary reading based on CNN (P , .001) [22].

SSD 1 ResNet50 The authors compared CNN (SSD 1 ResNet50) and QuickView mode. Per-patient abnormality detection for CNN was significantly higher than for QuickView (99% vs. 89%, P , .001). CNN detection rates for mucosal erosions, angioectasias, bulging lesions, and blood content were 100% (94 of 94), 97% (28 of 29), 99% (80 of 81), and 100% (23 of 23), respectively; for QuickView modality it was 91%, 97%, 80%, and 96%, respectively [23].

Inception-Resnet-V2 Inception-Resnet-V2 is an AI method that uses images categorized in a binary fashion. The AI algorithm was compared with the reading of physicians in training. A group of students showed reduced reading time compared with a group of experts (rate of detection of lesions, P 5 .057; time of reading, P 5 .343). AI based on the Inception-Resnet-V2 model was trained with binary classified images according to clinical relevance. The performance of AI was compared with the two groups of reviewers with different degrees of experience. AI selected 67,008 (31.89%) images with a probability of more than 0.8 to detect lesions in 210,100 images from 20 selected endoscopic capsule videos. Using the reading aided by AI, reviewers in both groups showed higher rates of detection of lesions compared with conventional reading (experts: 34.3% 73.0%; P 5 .029. students: 24.7% 53.1%; P 5 .029). The result for the students was similar to the

References

improvement observed in the group of experts (P 5 .057). In addition, the reading model aided by AI significantly cut down reading time for the students (1621.0 746.8 min; P 5 .029). The authors concluded that the reading model aided by AI can detect various relevant lesions and reduce reading time [24].

Recent outcomes of artificial intelligence in detecting active bleeding and hematic residues A metaanalysis that included 19 studies using computer-aided diagnosis for the detection of hemorrhage found pooled AUC, sensitivity, and specificity of 0.99 (95% CI, 0.98 0.99), 0.96 (95% CI, 0.94 0.97), and 0.97 (95% CI, 0.95 0.99), respectively [25]. Speaking specifically on the detection of blood and hematic residues, Ghosh et al. [26] used a CNN-based deep learning framework to identify bleeding and nonbleeding CE images with an identification of bleeding zones of 94.42%. Hajabdollahi et al. [27] also proposed a simple CNN structure for the detection of bleeding zones in CE images. The results obtained were an accuracy of 98.9% with a sensitivity of 94.8%, and a specificity of 99.1%. Rathnamala et al. [28] proposed a different system based on Gaussian mixture model superpixels for bleeding detection that achieved 99.88% accuracy, 99.83% sensitivity, and 100% specificity, showing that the system has very few classification errors. CNN system has also been used for the automatic detection of blood in colon CE images. Saraiva et al. [29] showed that AI can detect blood with a 99.8% sensitivity, a 93.2% specificity, and a 93.8% and 99.8% positive and NPV, respectively, in colon CE images. With what is seen in this chapter, we can conclude that AI, in its various forms, for the detection of blood and hematic residues is here to stay. AI will allow us to optimize the functioning of the CE, significantly increasing its diagnostic yields.

Acknowledgments The authors would like to thank Ph.D. Arturo Minor Martı´nez and Ramo´n Eduardo Cortina for their contributions to the ground truth of the capsule endoscopy small bowel bleeding images.

References [1] Gerson LB, Fidler JL, Cave DR, et al. ACG clinical guideline: diagnosis and management of small bowel bleeding. Am J Gastroenterol 2015;110(9):1265 87.

97

98

CHAPTER 6 Artificial intelligence for detection of blood

[2] Zammit SC, Sidhu R. Small bowel bleeding: cause and the role of endoscopy and medical therapy. Curr Opin Gastroenterol 2018;34(3):165 74. [3] Hosoe N, Takabayashi K, Ogata H, et al. Capsule endoscopy for small-intestinal disorders: current status. Dig Endosc 2019;31(5):498 507. [4] Pennazio M, Santucci R, Rondonotti E, et al. Outcome of patients with obscure gastrointestinal bleeding after capsule endoscopy: report of 100 consecutive cases. Gastroenterology 2004;126(3):643 53. [5] Tal AO, Filmann N, Makhlin K, et al. The capsule endoscopy “suspected blood indicator” (SBI) for detection of active small bowel bleeding: no active bleeding in case of negative SBI. Scand J Gastroenterol 2014;49(9):1131 5. ´ lvarez-Licona NE. Small bowel transit [6] Blanco-Velasco G, Pe´rez-Rodrı´guez M, A time of capsule endoscopy as a factor for the detection of lesions in potential small bowel bleeding. Rev Esp Enferm Dig 2019;111(9):696 8. [7] Koulaouzidis A, Rondonotti E, Karargyris A. Small-bowel capsule endoscopy: a tenpoint contemporary review. World J Gastroenterol 2013;19(24):3726 46. [8] D’Halluin PN, Delvaux M, Lapalus MG, et al. Does the “Suspected Blood Indicator” improve the detection of bleeding lesions by capsule endoscopy? Gastrointest Endosc 2005;61(2):243 9. [9] Liangpunsakul S, Mays L, Rex DK. Performance of given suspected blood indicator. Am J Gastroenterol 2003;98(12):2676 8. [10] Yung DE, Sykes C, Koulaouzidis A. The validity of suspected blood indicator software in capsule endoscopy: a systematic review and meta-analysis. Expert Rev Gastroenterol Hepatol 2017;11(1):43 51. [11] Boal Carvalho P, Magalha˜es J, Dias DE, Castro F, et al. Suspected blood indicator in capsule endoscopy: a valuable tool for gastrointestinal bleeding diagnosis. Arq Gastroenterol 2017;54(1):16 20. [12] Han S, Fahed J, Cave DR. Suspected blood indicator to identify active gastrointestinal bleeding: a prospective validation. Gastroenterol Res 2018;11(2):106 11. [13] Pogorelov K, Suman S, Azmadi Hussin F, et al. Bleeding detection in wireless capsule endoscopy videos—color vs texture features. J Appl Clin Med Phys 2019;20 (8):141 54. [14] Li B, Meng MQH. Computer-aided detection of bleeding regions for capsule endoscopy images. IEEE Trans Biomed Eng 2009;56(4):1032 9. [15] Liu DY, Gan T, Rao NN, et al. Identification of lesion images from gastrointestinal endoscope based on feature extraction of combinational methods with and without learning process. Med Image Anal 2016;32:281 94. [16] Tuba E, Tuba M, Jovanovic R. An algorithm for automated segmentation for bleeding detection in endoscopic images. in Neural Networks (IJCNN). In: International joint conference; 2017. p. 4579 86. [17] Tziortziotis I, Laskaratos FM, Coda S. Role of artificial intelligence in video capsule endoscopy. Diagnostics 2021;11(7):1192. [18] Noble WS. What is a support vector machine? Nat Biotechnol 2006;24(12):1565 7. [19] Kim SH, Lim YJ. Artificial intelligence in capsule endoscopy: a practical guide to its past and future challenges. Diagnostics 2021;11(9). [20] Soffer S, Klang E, Shimon O, et al. Deep learning for wireless capsule endoscopy: a systematic review and meta-analysis. Gastrointest Endoscopy 2020;92(4):831 9. [21] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521(7553):436 44.

References

[22] Ding Z, Shi H, Zhang H, et al. Gastroenterologist-level identification of small-bowel diseases and normal variants by capsule endoscopy using a deep-learning model. Gastroenterology 2019;157(4):1044 54. [23] Aoki T, Yamada A, Kato Y, et al. Automatic detection of various abnormalities in capsule endoscopy videos by a deep learning-based system: a multicenter study. Gastrointest Endoscopy 2021;93(1):165 73. [24] Park J, Hwang Y, Nam JH, et al. Artificial intelligence that determines the clinical significance of capsule endoscopy images can increase the efficiency of reading. PLoS One 2020;15(10):e0241474. [25] Bang CS, Lee JJ, Baik GH. Computer-aided diagnosis of gastrointestinal ulcer and hemorrhage using wireless capsule endoscopy: systematic review and diagnostic test accuracy meta-analysis. J Med Internet Res 2021;23:e33267. [26] Ghosh T, Chakareski J. Deep transfer learning for automated intestinal bleeding detection in capsule endoscopy imaging. J Digit Imaging 2021;34:404 17. [27] Hajabdollahi M, Esfandiarpoor R, Najarian K, et al. Low complexity CNN structure for automatic bleeding zone detection in wireless capsule endoscopy imaging. Annu Int Conf IEEE Eng Med Biol Soc 2019;2019:7227 30. [28] Rathnamala S, Jenicka S. Automated bleeding detection in wireless capsule endoscopy images based on color feature extraction from Gaussian mixture model superpixels. Med Biol Eng Comput 2021;59:969 87. [29] Mascarenhas Saraiva M, Ferreira JPS, Cardoso H, et al. Artificial intelligence and colon capsule endoscopy: automatic detection of blood in colon capsule endoscopy using a convolutional neural network. Endosc Int Open 2021;9:E1264 8.

99

This page intentionally left blank

CHAPTER

Artificial intelligence in capsule endoscopy for detection of ulcers and erosions

7

Shabana F. Pasha1 and Jean-Christophe Saurin2 1

Division of Gastroenterology and Hepatology, Mayo Clinic, Scottsdale, AZ, United States Department of Gastroenterology, E. Herriot Hospital, Claude Bernard University, Lyon, France

2

Introduction Capsule endoscopy (CE) is a disruptive endoscopic modality that uses miniature disposable capsules for noninvasive and physiologic imaging of the small bowel (SB) mucosa. With CE, we now have the capacity to directly visualize SB vascular, inflammatory, neoplastic, and other lesions, which was not possible earlier. However, the technology remains far from perfect with long manual reading times, human error of missed lesions, interobserver variation in interpretation of lesions, no control on locomotion, and lack of therapeutic capability. Since its initial conception, there have been ongoing advances to help reduce and ideally eliminate some of these inherent limitations. With the active exploration of artificial intelligence (AI) and other innovative ways to further refine and expand its role in the diagnosis and management of SB disorders, the future of CE remains exciting. Erosions and ulcers can be seen in the SB and/or colon in various inflammatory disorders, including Crohn’s disease and ulcerative colitis, NSAID and other medication related injuries, celiac disease-related ulcerative jejunoileitis, vasculitis, ischemia, infections, idiopathic disorders, and even certain neoplasms. Inflammatory lesions present a unique diagnostic dilemma as they range from solitary to multiple lesions, vary in size and appearance, and may be subtle and difficult to differentiate from surrounding artifacts and debris. Moreover, fibrostenotic complications from SB inflammation increase the risk of capsule retention [1,2]. Although studies are currently in progress, a lack of data is still there on the utility and actual benefit of AI in the evaluation of erosions and ulcerations and whether its use in clinical practice leads to improved diagnosis, reproducibility of endoscopic classification, or reduction in CE reading times. There are many potential AI applications in the future for CE in inflammatory disorders, including automated and improved detection of inflammatory lesions, interpretation and Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00017-4 © 2023 Elsevier Inc. All rights reserved.

101

102

CHAPTER 7 Artificial intelligence for detection of ulcers and erosions

diagnosis of the underlying disorder, evaluation for dysplasia and malignancy, prediction of fibrostenotic complications, and auto-documentation of reports. However, these algorithms will require robust testing and validation before they can be reliably applied in clinical practice.

Capsule endoscopes and current challenges Capsule endoscopes (currently available and in process of development) There are five SBCE systems available worldwide, four of which are FDA approved. The PillCam SB3 (Medtronic Inc., Minneapolis, USA), Endocapsule (Olympus, Centervalley Pennsylvania, USA), MiroCam (Intromedic, Seoul, Korea), and OMOM (Jinshan Science and Technology, Chongqing China) all have a single camera that provides a luminal view of the SB, and they utilize wireless technology, either radiofrequency or human body communication, to transmit images from the capsule to a data recorder. CapsoCam Plus (CapsoVision Inc., Saratoga, USA) has four side-facing cameras for a circumferential mucosal view, with all the images being stored by a wire-free technology on board the capsule. This is the only capsule that requires retrieval by patients [3,4]. To improve SB examination completion rates, the newer generation capsule endoscopes have an extended battery life ranging from 12 to 15 h. Some of the capsules have an adaptive frame rate technology that modifies image capture to conserve battery in areas of slow transit and increase image capture rate in areas of rapid transit to reduce the possibility of missed lesions. CapsoCam Plus has an inbuilt tool that adjusts brightness based on the proximity of the capsule to the SB mucosa to optimize illumination and mucosal visualization. Table 7.1 shows a comparison of the different SB capsule technologies. Improved algorithms in CE software allow the identification and deselection of duplicate images to improve the efficiency of the capsule read. The newer generation capsules have various Table 7.1 Flexible spectral imaging color enhancement (FICE) setting 1 2 used in small bowel capsule endoscopy, wavelengths in nanometers for red, green, and blue channels [5]. Mode

Red (nm)

Green (nm)

Blue (nm)

FICE 1 FICE 2 FICE 3

595 420 595

540 520 570

535 530 415

Source: Yung DE, Carvalho PB, Giannakou A, et al. Clinical validity of flexible spectral imaging color enhancement (FICE) in small bowel capsule endoscopy: a systematic review and metaanalysis. Endoscopy 2017;49:258 69.

Capsule endoscopes and current challenges

inbuilt technologies for imaging enhancement, which are described later in the chapter. SBCE

PillCam SB3

Endocapsule

MiroCam

OMOM

CapsoCam Plus

Dimensions Field of view ( )

26 3 11 156

26 3 11 160

24 3 11 170

28 3 13 140

31 3 11 360

Technology Frames/s

Radiofrequency 2 6

Radiofrequency 2

Radiofrequency 3

Electromagnetic 2

Wirefree 20

Battery life (h) Smart illumination View

12

12

12

10

15 Yes

Luminal

Luminal

Luminal

Luminal

340 3 340

512 3 512

320 3 320

640 3 480

Circumferential mucosal 221 3 184

Resolution (pixels)

The second-generation PillCam Colon Capsule (PCCE2) (Medtronic, Minneapolis, USA) is an 11x13 mm capsule with two luminal facing cameras and 172-degree visualization with each camera. The system has an inbuilt cursor and software for polyp size estimation [6]. A panenteric capsule (PillCam Crohn), similar to PillCam Colon 2, is being evaluated for SB and colon examination in patients with inflammatory bowel disease (IBD). It has an IBD-dedicated software in which the SB is divided into three and the colon into two arbitrary segments. The software allows the recognition of the most common lesion (MCL), the most severe lesion (MSL), and the extent of involvement [7]. The NaviCam SB system (Ancon Medical Technologies Co., Ltd.) has a 27 3 11.8-mm capsule with an implanted magnet within its dome, battery life of up to 12 h, and an adaptive frame rate of 0.5 of up to 12 frames/s. In addition to the capsule, locator, and data recorder, the system comprises a guidance magnet robot and proprietary ESNavi software. The guidance magnet robot is a C-arm type robot with two rotational (horizontal and vertical) and three translational degrees of freedom (forward/backward, up/down, and right/left). The capsule movement can be controlled automatically in default mode or manually by using two joysticks installed on the computer workstation [8]. Convoluted neural networks (CNNs) incorporated within the software help with the automatic detection of different SB lesion subtypes and improved efficiency of capsule readings. The data will be reviewed later in the chapter. This capsule is not yet FDA approved [9].

Current challenges in capsule endoscopy There are many gaps in current CE technology that can be successfully bridged with AI applications. A capsule endoscope takes approximately 50,000 to 70,000 images throughout the entire SB. Even after the elimination of duplicate images,

103

104

CHAPTER 7 Artificial intelligence for detection of ulcers and erosions

it remains a cumbersome process to manually review the entire study [2,10]. With the prolonged read, which typically takes 30 60 min, and monotonous view of similar frames, readers are prone to eye fatigue and could miss even significant lesions if these happen to be present only on a few, or sometimes a single frame. While it has been suggested that readers take short breaks every 15 20 min to maintain focused attention, this may not be feasible in a busy practice [11]. Individual variations in the speed and frame view (single, dual, or quadruple) adopted by different readers may have an impact on the quality and efficiency of the overall read. Moreover, most readers are not cross-trained to read capsules, and those accustomed to reading luminal or forward-facing capsules may find it challenging to adjust to a circumferential view and vice versa. Another limitation is the low image resolution of CE compared with standard endoscopy and colonoscopy [12]. With low pixelation, the utility of image enhanced endoscopy (IEE) to improve the detection and characterization of lesions with CE remains far from optimal [5]. For inflammatory disorders, specifically, there is wide interobserver variability in interpretation of CE findings. This is largely due to the lack of a common language to describe the appearance and document the severity of inflammatory lesions. A nomenclature description (ND) has previously been proposed based on a Delphi consensus of 80% to include aphthoid erosion, superficial ulceration, deep ulceration, stenosis, edema, hyperemia, and denudation. However, there are several limitations, including the lack of applicability of ND for nonspecific inflammatory findings seen with non-IBD disorders, lack of an accurate way to measure the extent and severity of ulcers and erosions, challenges with describing multiple findings on a single frame, and inability to accurately differentiate superficial from deep ulcers [13]. Based on earlier data as well as more recent studies, there is only moderate agreement for larger ulcers and cobblestoning and poor agreement for diminutive ulcers, erosions, and other SB inflammatory parameters [14 16].

Capsule endoscopy scoring systems for small bowel inflammation CE scoring systems allow for more standardized reporting of SB findings with structured terminology, improved interobserver agreement in the interpretation of lesions, and a more uniform assessment of the extent and severity of SB involvement. These scoring systems were developed based on a priori definition and agreement on SB inflammatory findings among readers. It is well described that high-quality training data, labeled and unlabeled, is essential for accuracy in the performance of machine learning (ML) and deep learning (DL) models and should be robust, relevant, and representative of all important data points and features [17]. Therefore, standardization of CE reporting is a key factor in the

Capsule endoscopy scoring systems for small bowel inflammation

development of effective AI models to avoid ambiguity in computer-aided detection and diagnosis, clinical interpretation, and management of SB findings. There are three main CE scoring systems, two of which were specifically developed for Crohn’s disease (CD) (Lewis score and Capsule Endoscopy Crohn’s Disease Activity Index, CECDAI), while the Saurin classification was originally proposed to stratify SB lesions based on their bleeding risk [18 21]. There are several other scores, including Ohmiya, Yano-Yamamoto, Rhemitt, ORBIT, and Nikimura, which have been proposed for risk stratification of patients based on different clinical parameters and underlying SB lesions [22]. CD scoring systems are useful for the diagnosis of suspected CD, assessment of known CD activity in patients with persistent symptoms, objective measurement of disease severity, and monitoring response to medical therapy [23].

Lewis score The Lewis score is the most widely used validated scoring system to assess SB CD activity on CE and is incorporated into RAPID (Medtronic) software. The severity of inflammation is categorized using endoscopic parameters of villous edema, ulcerations, and stenosis in three SB tertiles based on capsule transit time. The score is calculated using the sum of the worst affected tertile plus the score of stenosis in the entire SB. A Lewis score of ,135 has been shown to correlate with an overall normal SB examination without clinically significant inflammation, a score of $ 135 and ,790 with mild inflammation, and a score of $ 790, with moderate-to-severe inflammation [19]. In a study of 70 patients, there was a high correlation between Lewis scores for each tertile and global score and an excellent interobserver agreement between investigators and a central reader [24]. Another retrospective study found the Lewis score to be useful for the diagnosis of suspected CD, with 82.6% of patients with a Lewis score . 135 being diagnosed with CD compared with 12.1% of those with a Lewis score ,135 (P , .05), with a sensitivity and specificity of 82.6% and 87.9%, respectively [25]. A prospective study also showed that a Lewis score is useful in objectively monitoring patients with known CD, with a baseline Lewis score of .350 predicting relapse in 6 24 months [area under the curve (AUC) 0.79, 95% CI 0.66 0.88; P , .0001; hazard ratio 10.7, 3.8 30.3]. In addition, increase in Lewis score of 383 points or more from baseline was 100% predictive of a Crohn exacerbation within 6 months (AUC 0.79, 0.65 0.89; P 5 .011) [26]. The RAPID (Medtronic) software allows automatic calculation of the Lewis score and therefore eliminates subjective variation in the interpretation of disease activity. New algorithms are being developed that incorporate capsule speed in addition to overall transit time for a more accurate estimation of tertiles and to reduce the overrepresentation of findings from SB segments where capsule transit is delayed. This represents one of the first global assessments of a CE video using an informatics algorithm. This software tool requires objective scientific evaluation before being utilized in clinical practice.

105

106

CHAPTER 7 Artificial intelligence for detection of ulcers and erosions

Capsule endoscopy Crohn’s disease activity index The CECDAI is based on CE parameters of inflammation (A, 0 5 points), disease extent (B, 0 3 points), and strictures (C, 0 3 points) for proximal and distal segments of SB. The score is calculated ass a sum of proximal and distal scores: proximal ([A1 3 B1] 1 C1) 1 distal ([A2 3 B2] 1 C2) [27]. The CECDAI was validated in a prospective multicenter double-blind study that showed good correlation between endoscopists from different study centers (r 5 0.767, range 0.717 0.985, kappa 0.66; P , .001). Involvement of the proximal SB was found in 62% of patients supporting the role of the capsule in determining the extent of SB involvement and monitoring response to treatment. Although there are no validated levels of CECDAI to determine the severity of inflammation, a retrospective study reported that values of 3.8 and 5.8 correlated with Lewis score thresholds of 135 and 790, respectively. And although the Lewis score and CECDAI did not correlate reliably with fecal calprotectin levels, the Lewis score performed better than CECDAI in ruling out active inflammation in patients with fecal calprotectin ,100 μg/g [28]. With the advent of colon and panenteric capsules, a modified CECDAI for the SB and colon (CECDAIic/NIV) and a panenteric Crohn’s capsule score (PCCS)/ Eliakim score (ES) have also been proposed [29 31]. Future goals probably include AI tools that evaluate a complete SB or colonic capsule study using these criteria to reproduce the score independently of human estimation to avoid the usually low interobserver reproducibility of these evaluations, especially when considering nonexpert readers.

Capsule endoscopy software enhancements to improve detection of inflammatory lesions Image enhanced endoscopy IEE uses the principle of narrowing the bandwidth of white light endoscopic (WLE) images with an arithmetic process and spectral estimation technology [32]. It is routinely utilized as a tool to improve the detection as well as diagnosis of upper GI tract and colorectal neoplasia. IEE might be a useful adjunct to WLE examination on CE to improve detection of subtle SB lesions, accurately characterize lesions based on their appearance, differentiate lesions from surrounding debris and bubbles, and improve mucosal visualization in the presence of inadequate preparation, bilious, and bloody fluid. Interestingly, however, studies have failed to show the benefit of IEE in the SB, especially for inflammatory lesions. This appears to be mainly related to the pixelation of CE images and low resolution in contrast to the high-resolution imaging currently provided with upper endoscopes and colonoscopes [33]. Flexible spectral imaging color enhancement is a digital algorithm that uses arithmetic processing of white light endoscopy (WLE) images to emphasize

Capsule endoscopy software enhancements

certain light wavelengths selectively. Single wavelengths of red, blue, and green (RBG) are selected to display a composite color-enhanced image with improved visualization of mucosal pit patterns, surface architecture, and microvasculature [5,34,35]. CE-FICE (flexible spectral imaging color enhancement) is integrated into RAPID (Medtronic) software. Several studies have found no significant benefit of FICE over WLE for the detection of inflammatory lesions. In a metaanalysis of 10 studies that compared FICE to WLE, FICE1 showed some improvements in the detection of angioectasias, while no improvement in the detection of mucosal ulcers/erosions with any of the three modes [5]. Blue filter (BF) or blue mode is another enhanced imaging technique incorporated in RAPID (Medtronic) software. This modality allows color coefficient shift of light in the short wavelength range (490 430 nm) superimposed onto a white (RBG) light image [36]. In a retrospective study, 167 selected SB images from video sequences of 52 consecutive patients were viewed by two certified gastroenterologists using WLE, FICE 1, 2, and 3 settings and BF and categorized based on the visibility of blood vessels, mucosal surface contrast, and demarcation of lesion borders. Compared with WLE there was an improvement with BF in 93% of images for ulcers/ aphthae, no change in 5.8%, and worse image quality in 3.3%. With FICE 1, an improvement was observed in 36.6%, no change in 9%, and worse in 54%, while the use of FICE 2 and 3 resulted in worse image quality in 83% and 90%, respectively. Similarly, for villous edema, there was improvement with BF in 73.5%, but worse image quality with FICE 1, 2, and 3 of 70.6%, 79.5%, and 85.3%, respectively. Although BF performed better than FICE, there was only moderate interobserver agreement for these findings [36]. Other studies have also failed to show the benefit of FICE and BF over WLE for the evaluation of inflammatory lesions. The Augmented Live-Body Image Color-Spectrum Enhancement (ALICE) chromoendoscopy system is incorporated in the MiroCam (Intromedic, Seoul, Korea) software. There is only limited data to suggest that this system improves the visibility of flat and depressed lesions like angioectasias, erosions, and ulcers [37]. Another image-enhancing modality in MiroCam that uses three color modes (CM1, CM2, and CM3) has not been found to be beneficial over WLE based on a retrospective review of 100 SB lesions, which included 62 erosions and ulcerations [38]. Improvements in CE image resolution coupled with robust ML tools are therefore essential not only to improve automatic detection and classification of SB lesions but, more importantly, to reproducibly quantify the burden of inflammatory lesions as a baseline at the time of diagnosis and for comparison and monitoring during symptom exacerbations and treatment. One interesting question for future development of automatic diagnostic and severity algorithms would be to evaluate the AI process and learning in white light versus color enhancement, as only human identification capacity has been evaluated in prior studies.

107

108

CHAPTER 7 Artificial intelligence for detection of ulcers and erosions

Artificial intelligence and its application in capsule endoscopy AI leverages computer vision and robust data sets to mimic problem-solving and decision-making capabilities of the human mind. Computer vision utilizes three main tasks of image classification based on the presence or absence of the object of interest, detection of the object using a region of interest (box or circle), and segmentation to delineate the pixel-wise borders of the object. AI algorithms that incorporate ML, DL, and data analytics enable intelligent decision-making from either historical and/or real-time data [39,40]. ML, both supervised and unsupervised, is an application of AI that uses algorithms to iteratively learn, describe, and improve data to predict better outcomes. Supervised ML uses tagged or annotated training data sets to learn a function by mapping selected variables from the data onto a qualitative or quantitative output. It is then tested on a different dataset for validation. Unsupervised ML uses algorithms to analyze and cluster unlabeled data sets by discovering hidden patterns and data groupings without human intervention [41]. DL is a subset of ML with a neural network that has three or more layers. Artificial neural networks (ANNs) comprise multiple layers of interconnected nodes or neurons between the input and output layers. ANNs use data inputs, weights, and biases to accurately recognize, classify, and describe objects within the data. Most ANNs have forward propagation from input to output. CNNs are a type of ANN that mimic the human visual cortex and have the capacity to detect and analyze image patterns and features to allow the detection and recognition of objects. Convolutional and pooling layers and backward propagation allows calculation and correction of errors at each neuron [42,43]. Some of the limitations in developing an AI model include the small size of available training and testing data sets and the overfitting and underfitting of the data. Data augmentation is a way to artificially expand the number of training images with transformations such as random image rotation, scaling, and skewing. Overfitting occurs when the AI model fits too closely with the training data and is therefore unable to generalize to new data. It should be suspected if there is a low error rate with the training data and a high variance in the testing data. With underfitting, on the other hand, there is inadequate training of the model to determine a meaningful relation between the input and output variables. This is indicated by a high error in both training and testing data sets. Optimum training is therefore a balance wherein the amount of training is just sufficient to maintain a low error and variance rate in both the training and testing data sets. Several CNN architectures are now available, including the original AlexNet, which comprise five convolutional layers, and newer models, including VGGNet, Inception V1 to V4 (27 layers), ResNet (18, 50, 152, or up to 1202 layers), and DenseNet (40, 100, 121, and 169 layers) [44 48].

Artificial intelligence for detection of small bowel ulcerations

The most common functions of AI in endoscopic evaluation are computeraided detection (CADe) and computer-aided diagnosis (CADx) [49,50]. Both of these functions have wide applicability in upper GI endoscopy and colonoscopy, including early diagnosis of esophageal and gastric cancer, colon polyp and colorectal cancer detection and characterization, prediction of pathology, and automated documentation of endoscopic diagnostics and therapeutics [51,52]. While earlier research studies have focused on support vector machines (SVMs) and multilayer perceptron networks, CNN is now the main DL algorithm adopted in endoscopic imaging, including CE [53]. A third function that would be of major interest in IBD would be the global severity assessment of disease, reproducing CE indices, similar to preparation scores that have been validated for SB global preparation assessment [54].

Artificial intelligence for detection of small bowel ulcerations and erosions There are many potential AI applications in the endoscopic evaluation of IBD and other inflammatory disorders, which include automatic detection of ulcers and erosions to reduce miss rate and minimize reading time; differentiation between CD, ulcerative colitis, and other disorders, including NSAID or medication-related inflammation, autoimmune disorders such as celiac disease-related jejunoileitis, autoimmune enteropathy, Behcet’s disease or vasculitis, ischemia, and malignancy; and objective evaluation of symptoms and monitoring response to therapy (Fig. 7.1). Some of the challenges in accurately determining disease activity with CE lies in the lack of a gold standard for comparison as it is neither practical nor feasible to obtain biopsies throughout the SB and cross-sectional imaging is a much less sensitive modality than CE for assessment of inflammation limited to the

FIGURE 7.1 Artificial intelligence in automatic detection of small bowel inflammation. Courtesy: Dr. Miguel Mascarenhas Saraiva.

109

110

CHAPTER 7 Artificial intelligence for detection of ulcers and erosions

mucosa. In addition, there is only a moderate correlation of noninvasive inflammatory markers, including fecal calprotectin, with CE CD activity indices [55 57]. These different applications will require different performance characteristics of the AI system. For instance, quantification of the size and number of ulcers/ erosions and SB extent of involvement is much more relevant when determining disease severity and prognosis and for monitoring response to treatment, while estimating the size and number of ulcerations and erosions may not be as critical as their appearance and characteristics to establish or rule out a diagnosis of IBD [57]. DL for automatic detection of ulcers and erosions with CE is still at an early proof of concept stage and uses data sets with still frames manually selected by expert readers for analysis. Robust testing and validation of CNNs for automatic detection of inflammatory lesions using unlabeled full-length capsule videos is essential before DL tools can be reliably utilized in clinical practice.

Automatic detection of ulcers and erosions Automatic ulcer detection with CE has been previously studied using extraction of color and texture feature patterns through a variety of techniques, including SVMs, Gabor filters and textural descriptors, and others, which were tedious, required manual feature extraction, and utilized traditional ML methods with significant limitations [58 61]. More recent studies have shown that CNN is superior to these older ML techniques for the detection of both ulcers and erosions [62] (Table 7.2). Iakovidis et al. were the first investigators to propose a supervised methodology to detect different SB lesion subtypes using a color feature-based pattern Table 7.2 Summary of studies evaluating convoluted neural network (CNN) for ulcers and erosions. Study

Design

Fan 2018

Retrospective

Capsule

Images

Algorithm

Results

21,160

AlexNet

Ulcers: Sensitivity 96.8% Specificity 94.79% Accuracy 95.16 Erosions: Sensitivity 93.67% Specificity 95.98% Accuracy 95.34% (Continued)

Automatic detection of ulcers and erosions

Table 7.2 Summary of studies evaluating convoluted neural network (CNN) for ulcers and erosions. Continued Study

Design

Capsule

Images

Algorithm

Results

Aoki 2019

Ulcer detection

15,800

Ulcer detection

CNN based on SSD Xception CNN

Sensitivity 88.2% Specificity 90.9%

Klang 2019

PillCam SB2 and Grade 3 PillCam SB3

Wang 2019

Ulcer detection

Ankon

Xception CNN

Ferreira 2022

Ulcer detection

PillCam Crohn’s capsule

Training: 15,781 ulcer frames and 17,138 normal frames Validation: 2040 ulcer frames and 2319 normal frames Test: 4917 ulcer frames and 5007 normal frames 24,675 SB and colon

Barash 2021

Ulcer severity grading

PillCam SB3

17,640

ResNet

Ding 2019

Detection of abnormal SB findings

Ankon

58,235

ResNet

17,640

Xception model trained on ImageNet

AUC, Area under the curve; SB, small bowel; SSD, Single Shot MultiBox Detector.

Sensitivity 96.8% Specificity 96.6% Sensitivity 89.71% Specificity 90.48% Accuracy 90.10% Threshold value 0.6706

Sensitivity 98% Specificity 99% PPV 96.6% NPV 99.5% Accuracy 98.8% AUC 1.00 Classification accuracy of the algorithm: 0.91 (95% CI 0.867 0.954) for Grade 1 vs. Grade 3 ulcers 0.78 (95% CI, 0.716 0.844) for Grade 2 vs. Grade 3 0.624 (95% CI 0.547 0.701) for Grade 1 vs. Grade 2 99.73 (98.28 99.99) for all SB 100 (99.90 100) (for ulcers)

111

112

CHAPTER 7 Artificial intelligence for detection of ulcers and erosions

recognition method. The average performance in terms of the area under the receiver-operating characteristic curve reached 89.2% 6 0.9%. The best average performance was obtained for angiectasias (0.975) and nodular lymphangiectasias (0.963). However, this technique was suboptimal for the detection of ulcerations (76.2 6 10%) [62]. Fan et al. used AlexNET CNN to evaluate two independent models for the detection of ulcers and erosions. Their data set included 3250 ulcer images and 5000 normal images, and 4910 erosion images and 8000 normal images for these models. These were divided into training and validation parts at an image level. The training and testing set were used for model training, while a verification set was used for evaluating the performance of the trained model. There was no overlap in these three data sets. The CNN model showed high performance with a sensitivity of 91.35% and 93.33%, specificity of 95.83% and 96.61%, accuracy of 95.43% and 96.32%, and AUC value of 0.9805 and 0.9904, respectively, for ulcers and erosions. There was a misclassification of 5% of images, with false negative results due to subtle lesions that were difficult to differentiate from normal tissue and false positives due to the presence of yellowish debris, bubbles, and blood vessels. Poor image quality was also a limiting factor in the accurate differentiation of lesions from normal tissues. When compared with the traditional method of using grayscale histogram and SVM classifier, CNN clearly outperformed the former in the detection of both ulcers and erosions [61]. Aoki et al. developed an automated detection system for inflammatory lesions using CNN. They used the Single Shot MultiBox Detector (SSD), which is a CNN with 16 or more layers. Ulcers and erosions in the training set were independently and manually annotated by two expert endoscopists, and consensus was determined later. The images were fed into the SSD architecture through the Caffe DL framework (Berkeley Vision and Learning Center). The CNN was trained to recognize annotated areas within the bounding boxes as erosions or ulcerations and the other areas as background. The training set included 5360 images, with 5 of these requiring modification at consensus. The trained CNN shaped the erosions and ulcerations in the training set with bounding boxes, and output the probability score of the inflammatory lesion from 0 to 1. The higher the probability score, the higher assurance of CNN that the region contained an ulcer or erosion. Validation of CNN was performed using single images. When the overlapped area between the CNN and true box covered more than 70%, or when at least one of multiple CNN boxes within the true box detected an erosion or ulcer, it was considered that the finding was correctly detected. The test set included 10,440 SB images with 440 images of erosions and ulcers. The trained set required 233 s to evaluate the test images, with an AUC of 0.958 (CI 0.947 0.968). The CNN model had a sensitivity, specificity, and accuracy of 88.2% (CI 84.8% 91%), 90.9% (CI 90.3% 91.4%), and 90.8% (CI 90.2% 91.3%), respectively, at a cut off value of 0.481 for the probability score. More than half of the false-positive lesions were due to foam, debris, and vascular dilatation and appeared to be easily distinguishable from

Automatic detection of ulcers and erosions

inflammatory lesions. The size of the lesion did not affect false negative images in the presence of bubbles, debris, or bile, with even larger lesions being missed with images that had inadequate preparation. More robust straining using larger data sets with artifacts and suboptimal preparations might improve CNN performance in these scenarios [63]. The application of AI to detect SB ulcers at an individual patient level was evaluated in a retrospective study of 17,640 CE images from 49 patients with CD. Images were collected from CD patients and controls and were labeled by an expert gastroenterologist as ulceration or normal. A CNN was trained on fivefold randomly split images, with each fold comprising 80% training images and 20% testing images. Ten experiments were conducted with images from n-1 patients used to train a network, and images from a different individual patient were used to test the network. There were 7391 images with mucosal ulcers and 10,249 images of normal mucosa. With this CNN model, there was accurate and fast automated detection of mucosal ulcers, and individual patient-level analysis provided high and consistent diagnostic accuracy with shortened reading time. The AUC was 0.99, with accuracies of 95.4% 96.7% for randomly split images. The AUC for individual patient experiments was 0.94 0.99, with a median time of 3.5 min to analyze the entire SB video [57]. A systematic review and metaanalysis by Soffer et al., which included five studies, including the ones described above, showed that the CNN models had an overall high sensitivity of 95% (89% 98%) and specificity of 94% (90% 96%) for automatic ulcer detection with CE. The results confirm that the CNN is superior to the SVM and other ML technologies that have been evaluated earlier for the detection of inflammatory lesions [48]. Ferreira et al. performed a retrospective multicenter proof of concept study to develop and validate a CNN model for the automatic detection of SB and colon ulcers with the panenteric PillCam Crohn Capsule (PCC). A total of 24,675 frames were extracted from 59 PCC studies, with 5300 containing ulcers or erosions and the remainder showing normal enteric or colonic mucosa. The training dataset comprised 80% of the total image pool. The remaining 20% (4935) of images used to test the model were composed of 1060 (21.5%) images with ulcers and erosions and 3875 (78.5%) images with normal enteric or colonic mucosa. The model had an overall accuracy of 98.8%, with a sensitivity of 98.0%, specificity of 99.0%, positive and negative predictive values of 96.6% and 99.5%, respectively. CNN read the entire validation dataset in 72 s with an average rate of 68 frames/s [64]. A separate study by Wang et al. evaluated a second glance (secG) system to evaluate a large dataset of images from the Ankon capsule for the automatic detection of ulcerations. They used 15,781 ulcer frames from 753 ulcer cases and 17,138 normal frames for training. The validation dataset consisted of 2040 ulcer frames and 2319 normal frames, while the test set included 4917 ulcer frames and 5007 normal frames. They found that the detection of ulcers was highly related to the size, with a sensitivity of greater than 92% for ulcers larger than 1% of the

113

114

CHAPTER 7 Artificial intelligence for detection of ulcers and erosions

full image size and 85% for ulcers smaller than 1% of the full image size. The overall sensitivity, specificity, and accuracy were 89.71%, 90.48%, and 90.10% at a threshold value of 0.6706 [65].

Grading of ulcers and erosions severity There is only one study by Barash et al. that has studied the utility of CNN to automatically grade ulcer severity on PCC and determine CD activity [66]. They included 17,640 CE images of patients with CD, with 7391 mucosal ulcers and 10,249 normal images. Ulcers were classified independently by expert readers as Grade 1 (small and superficial ulcers), Grade 2 (intermediate size and depth), and Grade 3 (prominent ulcers with fissuring, circumferential, cobblestone, or kissing morphology). There was an overall agreement between the experts for only 31% of images. The interreader agreement when differentiating between Grade 1 and Grade 3 ulcers was 76% (752 of 989 images), 40% (185 of 457 images) for differentiating between Grade 1 and Grade 2, and 36% (220 of 608 images) for differentiating between Grade 2 and Grade 3 ulcers. A consensus reading by three capsule readers was performed on another set of 1490 images that were used to train and test the CNN network (1242 for the training process and 248 for testing the network performance). The overall agreement between the consensus reading and the automatic algorithm was 67% (166 of 248). There was excellent accuracy when comparing Grade 1 ulcerations with Grade 3 ulcerations (91% agreement with AUC of 0.958, and specificity and sensitivity of 0.91% and 0.91%, respectively). When comparing Grade 1 and Grade 3 (mild and severe) with Grade 2 (intermediate) ulcerations, the performance was substantially lower. Between Grade 1 and Grade 2, the overall agreement between the automatic grading and the consensus reading was 65% (84 of 128) with AUC of 0.565 and optimal specificity and sensitivity of 0.34% and 0.71%, respectively; while the overall agreement between Grade 2 and Grade 3 ulcers was 79% (90 of 113) with AUC of 0.939 and specificity and sensitivity of 0.73% and 0.91%, respectively.

Artificial intelligence in next-generation capsule endoscopes The NaviCam SBCE system (ANX Robotica) incorporates ProScan intelligent reading support to eliminate reading time and automatically identify different subtypes of SB lesions using CNNs. In a large multicenter study, data from 6970 patients (113,426,569 SB images) were used for the training phase (1970) and the validation phase (5000) of the CNN-based auxiliary reading model. There was no overlap between the data sets. Abnormal images were defined in two categories clinically significant abnormal lesions and normal variants. Twenty gastroenterologists evaluated the images with conventional analysis and CNN-based auxiliary analysis.

References

If there was disagreement between the conventional analysis and the CNN model, the image was re-examined by the gastroenterologists to confirm or reject the CNN categorization. The CNN-based model identified abnormalities with 99.88% sensitivity in the per-patient analysis (95% CI 99.67 99.96) and 99.90% in the per-lesion analysis (95% CI 99.74 99.97). In comparison, conventional reading identified abnormalities with a sensitivity of 74.57% (95% CI 73.05 76.03) in the per-patient analysis and 76.89% (95% CI 75.58 78.15) in the per-lesion analysis. Sensitivity for ulcer detection was 99.73% with CNN-based versus 98.12% (P 5 .0339) with conventional reading on a per-patient analysis, 99.73 versus 98.12 on a per-lesion analysis, with a specificity of 100% with both models. Mean reading time was significantly shorter with CNN reading (5.9 6 2.23 min compared with 96.6 6 22.53 min, P , .001). However, the methodology of the study (the time used to note and report each lesion was included in the observer reading time) precludes any firm conclusion about this considerable difference. The first main limitation to the CNN model is that it currently does not have the ability to classify SB abnormalities [9]. The second limitation was the lack of relevance classification of the different lesions described so that one possibility would be that AI is excellent at identifying clinically insignificant images, but we do not really know its ability, for example, in tumor or ulcer detection, would necessitate a dedicated study with a selection of patients with usually difficult lesions (tumors, ulcerations, stenoses).

Conclusions Time and accuracy are of the essence as we move toward a new future of enhanced diagnosis and management in endoscopy, IBD, and other gastrointestinal disorders. As the role of AI continues to grow in these realms, it is important to ensure that the process is as easy and automated as possible, with fool-proof measures to check the accuracy of the information provided by these augmentation tools. Most of the studies on AI-guided CE, especially for inflammatory lesions, are still in the proof-of-concept stage, and there is much work to be done before these tools are ready for implementation in clinical practice. Now is the time for gastroenterologists and data science engineers to widely share ideas, discuss missing gaps in technology, and use our existing compiled knowledge and databases to come up with AI tools that will enhance patient care and allow us to meaningfully utilize our time and intelligence for interpretation and individualized management decisions for our patients.

References [1] McCain JD, Pasha SF, Leighton JA. Role of capsule endoscopy in inflammatory bowel disease. Gastrointest Endosc Clin North Am 2021;31:345 61. [2] Cave DR, Hakimian S, Patel K. Current controversies concerning capsule endoscopy. Dig Dis Sci 2019;64:3040 7.

115

116

CHAPTER 7 Artificial intelligence for detection of ulcers and erosions

[3] Zammit SC, Sidhu R. Capsule endoscopy-recent developments and future directions. Expert Rev Gastroenterol Hepatol 2021;15:127 37. [4] Rondonotti E, Spada C, Adler S, et al. Small bowel capsule endoscopy and device assisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Technical Review. Endoscopy 2018;50:423 46. [5] Yung DE, Carvalho PB, Giannakou A, et al. Clinical validity of flexible spectral imaging color enhancement (FICE) in small-bowel capsule endoscopy: a systematic review and meta-analysis. Endoscopy 2017;49:258 69. [6] Eliakim R, Yassin K, Niv Y, et al. Prospective multicenter performance evaluation of the second generation colon capsule compared with colonoscopy. Endoscopy 2009;41:1026 31. [7] Goran L, Negreanu AM, Stemate A, et al. Capsule endoscopy: current state and role in Crohn’s disease. World J Gastroenterol 2018;16:184 92. [8] Jiang X, Oan J, Li Z-S, Liao Z. Standard examination procedure of magnetically controlled capsule endoscopy. Video GIE 2019;4:239 43. [9] Ding Z, Shi H, Zhang H, et al. Gastroenterologist-level identification of small bowel diseases and normal variants by capsule endoscopy using a deep-learning model. Gastroenterology 2019;157:1044 54 e5. [10] Wang A, et al. Wireless capsule endoscopy ASGE Technology Committee Gastrointest Endosc 2013;78:805 15. [11] Rondonotti E, Pennazio M, Toth E, Kouaouzidis A. How to read small bowel capsule endoscopy: a practical guide for everyday use. Endosc Int Open 2020;8:E1220 4. [12] Ciuti G, Menciassi A, Dario P. Capsule endoscopy: from current achievements to open challenges. IEEE Rev Biomed Eng 2011;4:59 72. [13] Leenhardt R, et al. Nomenclature and semantic descriptions of ulcerative and inflammatory lesions seen in Crohn’s disease in small bowel capsule endoscopy: an international Delphi consensus statement. U Eur Gastroenterol J 2020;8(1):99 107. [14] Mergener K, Ponchon T, Gralnek I, et al. Literature review and recommendations for clinical application of small-bowel capsule endoscopy, based on a panel discussion by international experts. Endoscopy 2007;39:895 909. [15] De Leusse A, Landi B, Edery J, et al. Video capsule endoscopy for investigation of obscure gastrointestinal bleeding: feasibility, results and interobserver agreement. Endoscopy 2005;37:617 21. [16] Esaki M, Matsumoto T, Ohmiya N, et al. Capsule endoscopy findings for the diagnosis of Crohn’s disease: a nationwide case-control study. J Gastroenterol 2019;54:249 60. [17] Nichols JA, Chan HWH, Baker MA. Machine learning: applications of artificial intelligence to imaging and diagnosis. Biophys Rev 2019;11:111 18. [18] Korman LY, Delvaux M, Gay G, et al. Capsule endoscopy structured terminology (CEST): proposal of a standardized and structured terminology for reporting capsule endoscopy procedures. Endoscopy 2005;37:951 9. [19] Gralnek IM, Defranchis S, Seidman E, et al. Development of a capsule endoscopy scoring index for small bowel mucosal inflammatory change. Aliment Pharm Ther 2008;27:146 54. [20] Gal E, Geller A, Fraser S, et al. Assessment and validation of the new capsule endoscopy Crohn’s disease activity index (CECDAI). Dig Dis Sci 2008;53:1933 7.

References

[21] Suarin JC, Delvaux M, Gaudin JL, et al. Diagnostic value of endoscopic capsule in patients with obscure digestive bleeding: blinded comparison with video pushenteroscopy. Endoscopy 2003;35:576 84. [22] Rosa B, Margalit-Yahuda R, Gatt K, et al. Scoring systems in clinical small-bowel capsule endoscopy: all you need to know. Endosc Int Open 2021;9:E802 23. [23] Rosa B, Pinho R, de Ferro SM, et al. Endoscopic scores for evaluation of Crohn’s disease activity at small bowel capsule endoscopy: general principles and current applications. Portuguese J Gastroenterol 2015;23:36 41. [24] Cotter J, Dias de Castro F, Magalhaes J, et al. Validation of the Lewis score for the evaluation of small bowel Crohn’s disease activity. Endoscopy 2015;47:330 5. [25] Rosa B, Moreira MJ, Rebelo A, et al. Lewis Score: a useful clinical tool for patients with suspected Crohnʼs disease submitted to capsule endoscopy. J Crohnʼs Colitis 2012;6:692 7. [26] Ben-Horin S, Lahat A, Amitai MM, et al. Assessment of small bowel mucosal healing by video capsule endoscopy for the prediction of short-term and long-term risk of Crohnʼs disease flare: a prospective cohort study. Lancet Gastroenterol Hepatol 2019;4:519 28. [27] Niv Y, Ilani S, Hershkowitz M, et al. Validation of the capsule endoscopy Crohn’s disease activity index (CECDAI or Niv score): a multicenter prospective study. Endoscopy 2012;44:21 6. [28] Koulaouzidis A, Douglas S, Plevris JN. Lewis score correlates more closely with fecal calprotectin than Capsule Endoscopy Crohn’s Disease Activity Index. Dig Dis Sci 2012;57:987 93. [29] Arieira C, Magalha˜es R, Dias de Castro F, et al. CECDAIic a new useful tool in pan-intestinal evaluation of Crohnʼs disease patients in the era of mucosal healing. Scand J Gastroenterol 2019;54:1326 30. [30] Eliakim R, Spada C, Lapidus A, et al. Evaluation of a new pan-enteric video capsule endoscopy system in patients with suspected or established inflammatory bowel disease feasibility study. Endosc Int Open 2018;6:E1235 46. [31] Eliakim R, Yablecovitch D, Lahat A, et al. A novel PillCam Crohnʼs capsule score (Eliakim score) for quantification of mucosal inflammation in Crohn’s disease. U Eur J Gastroenterol 2020;8:544. [32] Pohl J, May A, Rabenstein T, et al. Computed virtual chromoendoscopy: a new tool for enhancing tissue surface structures. Endoscopy 2007;39:80 3. [33] Ogata N, Ohtsuka K, Ogawa M, et al. Image enhanced endoscopy improves the identification of small intestinal lesions. Diagnostics 2021;11:2122. [34] Mishkin D.S., Chuttani R., Croffie J., ASGE technology status evaluation report: wireless capsule endoscopy 2006; 63:539 45. [35] Van Gossum A. Image-enhanced capsule endoscopy for characterization of small bowel lesions. Best Pract Res Clin Gastroenterol 2015;29:525 31. [36] Krystallis C, Koulaouzidis A, Douglas S, Plevris JN. Chromoendoscopy in small bowel capsule endoscopy: blue mode or Fuji Intelligent Color Enhancement? Dig Liver Dis 2011;43:953 7. [37] Ryu C, Song J, Lee M, et al. Does capsule endoscopy with ALICE improve visibility of small bowel lesions? Gastrointest Endosc 2013;77:AB46. [38] Ribeiro da Silva J, Pinho R, Rodrigues A, et al. Evaluation of the usefulness of virtual chromoendoscopy with different color modes in the MiroCam® system for characterization of small bowel lesions. Portuguese J Gastroenterol 2018;25:222 9.

117

118

CHAPTER 7 Artificial intelligence for detection of ulcers and erosions

[39] Klang E. Deep learning and medical imaging. J Thorac Dis 2018;10:1325 8. [40] Darrell W. A blueprint for the future of AI. Brookings What is artificial intelligence? 2018. (brookings.edu). [41] Janiesch C, Zschech P, Kai H. MAchine learning and deep learning. Electron Mark 2021;31:685 95. [42] What is Deep Learning? | IBM [IBM Cloud Education 2020]. 2010. [43] Alzubaidi L, Zhang J, Humaidi AJ, et al. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. J Big Data 2021;8:53. [44] Simonyan K., Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:14091556. 2014. [45] Szegedy C., Ioffe S., Vanhoucke V., et al. Inception-v4, inception-resnet and the impact of residual connections on learning. Thirty-First AAAI Conference on Artificial Intelligence; 2017. [46] He K., Zhang X., Ren S., et al. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition; 2016. p. 770-8. [47] Iandola F., Moskewicz M., Karayev S., et al. Densenet: implementing efficient convnet descriptor pyramids. 2014. arXiv:1404.1869v1. [48] Soffer S, Klang E, Shimon O, et al. Deep learning for wireless capsule endoscopy: a systematic review and meta-analysis. Gastrointest Endosc 2020;92:831 9. [49] Yang YJ, Bang CS. Application of artificial intelligence in gastroenterology. World J Gastroenterol 2019;25:1666 83. [50] Alagappan M, Glissen Brown JR, Mori Y, et al. Artificial intelligence in gastrointestinal endoscopy: The future is almost here. World J Gastrointest Endosc 2018;10:239 49. [51] Abadir A.P., Mohammed F.I., William K., et al. Artificial intelligence in gastrointestinal endoscopy 2020;53: 132 41. [52] Pannala R, Krishnan K, Melson J, et al. Artificial intelligence in capsule endoscopy. Video GIE 2020;5:599 613. [53] Li B, Meng Q, et al. Computer based detection of bleeding and ulcer in wireless capsule endoscopic images by chromaticity moments. Comput Biol Med 2009;39:141 7. [54] Leenhardt R, Souchaud M, et al. A neural network-based algorithm for assessing the cleanliness of small bowel during capsule endoscopy. Endoscopy. 2021;932 6. [55] Kopylov U, Klang E, Yablecovitch D, et al. Magnetic resonance enterography versus capsule endoscopy activity indices for quantification of small bowel inflammation in Crohn’s disease. Ther Adv Gastroenterol 2016;9:655 63. [56] Tham YS, Yung DE, Fay S, et al. Fecal calprotectin for detection of postoperative endoscopic recurrence in Crohn’s disease: a systematic review and meta-analysis. Ther Adv Gastroenterol 2018;11 1756284818785571. [57] Koulaouzidis A, Sipponen T, Nemeth A, et al. Association between fecal calprotectin levels and small bowel inflammation in capsule endoscopy: a multicenter retrospective study. Dig Dis Sci 2016;61:2033 40. [58] Klang E, Barash Y, Margalit RY, et al. Deep learning algorithms for automated detection of Crohn’s disease ulcers by capsule endoscopy. Gastrointest Endosc 2020;91:606 13. [59] Kundu A., Bhattacharjee A., Fattah S., Shahnaz C. Automatic ulcer detection scheme using grayscale histogram from wireless capsule endoscopy. In: Proc. 2016 IEEE Int. WIE Conf. on electrical and computer engineering; 2016. p. 242 5.

Further reading

[60] Li B, Meng MQ-H. Texture analysis for ulcer detection in capsule endoscopy images. Image Vis Comput 2009;27:1336 42. [61] Li B., Qi L., Meng M.Q.-H., Fan Y. Using ensemble classifier for small bowel ulcer detection in wireless capsule endoscopy images. In: Proc. Int. Conf. on IEEE robotics and biomimetics; 2009. p. 2326 31. [62] Yi S, Jiao H, Leighton JA, Pasha SF, et al. A novel software platform for the automatic detection of small bowel ulcers. Gastrointest Endosc 2014;77 AB 172-3. [63] Iakovidis DK, Koulaouzidis A. Automatic lesion detection in capsule endoscopy based on color saliency: closer to an essential adjunct for reviewing software. Gastrointest Endosc 2014;80:877 83. [64] Ferreira JPS, Saraiva MJM, Afonso JPL, et al. Identification of ulcers and erosions by the novel Pillcamt Crohn’s capsule using a convolutional neural network: a multicentre pilot study. J Crohn Colitis 2022;169 72. [65] Wang S, Xing Y, Zhang L, et al. A systematic evaluation and optimization of automatic detection of ulcers in wireless capsule endoscopy on a large dataset using deep convolutional neural networks. Phys Med Biol 2019;64:235014. [66] Barash Y, Azaria L, Soffer S, et al. Ulcer severity grading in video capsule images of patients with Crohn’s disease: an ordinal neural network solution. Gastrointest Endosc 2021;93:187 92.

Further reading Aoki T, Yamada A, Aoyama K, et al. Automatic detection of erosions and ulcerations in wireless capsule endoscopy images based on a deep convolutional neural network. Gastrointest Endosc 2019;89:357 63. Shanhui F, et al. Computer aided detection of small intestinal ulcer and erosion in wireless capsule endoscopy images. Phys Med Biol 2018;63:165001.

119

This page intentionally left blank

CHAPTER

Artificial intelligence for protruding lesions

8

Xavier Dray1,2, Aymeric Histace2, Alexander Robertson3 and Santi Segui4 1

Center for Digestive Endoscopy, Saint Antoine Hospital, APHP, Sorbonne University, Paris, France 2 ENSEA, CNRS, ETIS UMR 8051, CY Cergy Paris University, Cergy, France 3 Leicester General Hospital, Leicester, United Kingdom 4 Department of Mathematics and Computer Science, University of Barcelona, Barcelona, Spain

Introduction Although in essence it is one long tube, very little of the gastrointestinal (GI) tract is entirely smooth, and projections into the lumen are seen throughout. There are normal protruding landmarks in the healthy GI tract, such as gastric rugae, pylorus, the papilla, a carpet of small bowel (SB) villi, the ileocaecal valve, haustra, and the anal cushion. In conjunction with peristalsis and normal variations, these can look atypical and misleading but can generally be recognized by the experienced capsule endoscopy (CE) reader. Unexpected projections, however, can be associated with symptoms such as bleeding, obstruction, or the development of malignancy, making accurate and reliable recognition of protruding lesions vital. When searching for the source of symptoms, including suspected bleeding, for which a CE has been arranged, an abnormal projecting growth of tissue, or polyp, raises concerns as a potential cause. Ulceration and inflammatory pathology are often associated with swelling, and therefore a degree of luminal protrusion is one of the ways in which these are recognized by either the human reader or artificial intelligence (AI) program. Likewise, vascular lesions will often protrude, which is important, for example, in the recognition and prognostication in varices. Mucosal, or submucosal abnormalities and masses, can be entirely benign, have malignant potential, or already be malignant. Those arising from the mucosa will generally show a disrupted and irregular surface pattern, with a change in color compared with the surrounding mucosa and a more irregular projection. Submucosal lesions will have preserved surface mucosa, veiled and often appearing only as a subtle bulge, making these more easily missed. In clinical practice, these are important for several reasons, including the local mass effect, which can lead to obstruction of the GI tract or disruption of the mucosal or vascular surface, resulting in bleeding. Recognition of neoplastic

Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00005-8 © 2023 Elsevier Inc. All rights reserved.

121

122

CHAPTER 8 Artificial intelligence for protruding lesions

lesions may allow recognition of a cancer or alert the patient to a premalignant condition that can be treated to avoid cancer. When investigating the GI tract, capsule technology has many advantages, including ease and comfort. Colonic screening is offered in many countries with the aim of early detection and removal of premalignant polyps to reduce cancerrelated deaths [1]. Colon cancer is becoming more common in younger patients [2,3], and internationally there is a drive to push the screening age for colorectal cancer (CRC) down [4]. However, with this increased service requirement, it is clear that many health systems are unable to reach the targets expected by the public. Unfortunately, current approaches are not working, rates of missed adenoma on colonoscopy are high and interval cancers present frequently. When compared with colonoscopy, colon CE (CCE) is more acceptable to patients [5] and a good option for many. The outward appearance of a CE device is that of a large pill or capsule. The main components inside the capsule shell are a camera, including an image acquisition system (transparent glass dome, lens, LEDs, and camera unit), a computer chip, and batteries. Most devices also have a bidirectional communication unit, allowing images to be transferred to a sensor, carried on a belt and connected to a data recorder. Some devices save the images onto the computing chip to be uploaded after the capsule is retrieved by the patient from their stool (without the need for a belt or a recorder). The size (usually 24.5 32.3 mm long with a diameter of 10.8 13.0 mm), number and position of cameras (1 4, axial or lateral), optical field, focal length, resolution, definition, contrast, brightness, capture rate, and tool for bidirectional communication may vary according to the targeted organ in the GI tract and according to manufacturers. Although CE is considered a safe investigation, some basic contraindications should be ruled out before prescription. CE should not be performed in patients with an inability to swallow, as tracheal obstruction is an exceptional but immediately life-threatening complication. Patients should be checked for conditions or symptoms putting them at risk of GI obstruction, as CE retention is the most frequent adverse event, occurring in approximately 1% of cases. Clinical consequences are unusual; the need for endoscopic or surgical retrieval of the capsule is even rarer. Discrete protruding lesions (unlike diffuse colonic changes, such as colitis) are often seen in only one or two frames, which can be blurry in rapid transit. This may result in a high miss rate, regardless of experience in reading CE, with one study suggesting masses and polyps having a detection rate of only 46% on CE [6]. AI, which is a rapidly expanding field because of huge investment from researchers and industry, is being used in the detection, characterization, and risk stratification of protruding lesions seen on CE. Initially, AI is likely to be complementary to clinical inputs from the human reader but has the potential to reduce some of the burdens of time when analyzing reports of around 60,000 recorded frames. It is speculated that ultimately AI will accurately analyze the images, diagnose the disease, and triage patients to further investigation or intervention, if required. In this chapter, we will discuss protruding lesions of the GI tract and

State-of-the-art clinical aspects

explore the AI systems and technologies under development for CE to aid physicians in diagnosing them.

State-of-the-art technological aspects Since the introduction of CE, several computer-aided diagnosis (CAD) systems have been proposed to reduce its inherent drawbacks in clinical settings. In the literature, several AI-based CAD systems are specially designed to detect suspicious or abnormal CE images. Most of these methods are aimed at reducing visualization and diagnosis time by detecting specific GI events with high-performance machine learning systems [7]. Several solutions have been presented to classify, segment, and characterize the abnormalities of protruding lesions. Polyp detection has been an active research topic (Table 8.1). Early methods from the community are based on a classic computer vision approach, where a set of hand-crafted features are extracted from the images, and supervised learning is used to classify or segment the images. Recently, several deep learning solutions have been presented. Although these models have shown promise, they are generally trained and validated using internal, small, or biased datasets, which does not guarantee generalizable results in clinical practice. Moreover, there is no agreement on a common evaluation methodology to allow the community to compare techniques. Most of these methods have been developed and validated as fully automatic systems, suitable for image detection systems but not fully informative for CAD systems in medical applications. Several techniques have been proposed for training the models, such as data augmentation [13], deep metric-learning [14,16], or self-supervised learning [18] to overcome the issue of limited data.

State-of-the-art clinical aspects Esophagus As elsewhere in the GI tract, intraluminal protrusions are common in the esophagus and are not always pathological. Most diagnostic procedures of the esophagus are performed with the aim of excluding serious or dangerous pathology, with esophageal cancers being a primary concern. Although several malignant tumors affect the esophagus, including adenocarcinoma, squamous cell carcinoma (SCC), sarcomatoid carcinoma, melanoma, sarcoma, lymphoma, and metastatic tumors, the most common are adenocarcinoma and SCC, with the bulk of AI-based diagnostic research in the esophagus aimed at these. Esophageal cancer is common (the eighth most common internal cancer [20]) and has a high mortality rate, which is often due to late presentation. However, what is of concern is that the miss rate for esophageal SCC during endoscopy is high [21]. The list of benign

123

124

CHAPTER 8 Artificial intelligence for protruding lesions

Table 8.1 Comparison of existing artificial intelligence (AI) methods for protruding lesions in capsule endoscopy. First author, year (reference)

AI model

Target

Li, 2009 [8] Zhao, 2011 [9] Li, 2012 [10] Yuan, 2014 [11] Yuan, 2017 [12] Yuan, 2018 [13] Guo, 2019 [14] Yuan, 2020 [15] Laiz, 2020 [16] Saito, 2020 [17] Pascual, 2022 [18] Gilabert, 2022 [19]

Unknown Unknown

Dataset

Evaluation metrics

Videos

Polyp

Control images

Polyp Polyp

2 2

150

150

A B C A D E

Unknown Unknown

Polyp Polyp

10 10

600 430

600 430

A A B C

Unknown

Polyp

35

CNN

Polyp

62

CNN

Polyp

CNN

Polyp

CNN CNN CNN UI with CNN model

A G 1.5k

1.5k

A B F H

585

2.2k

A I

80

1.2k

6k

A B C F H

Polyp

120

2.1k

1.3M

A E

Protruding lesions Polyp

379

30k

20k

A G

120

2.1k

200k

A E

Polyp

18

52 unique

500k

B Time

Metrics, the legend used is: Accuracy (A), Sensitivity—Recall—TPR (B), Specificity—TNR (C), ROC (D), AUC (E), Precision (F), Confusion Matrix (G), F1-Score (H), Cohen’s Kappa score (I). CNN, Convolutional neural network; UI, User Interface.

protruding lesions is far longer, including papillomas, cysts, fibrovascular or inflammatory polyps, strictures, or rings. Submucosal lesions can include extrinsic compression from normal intrathoracic structures, tumors, and vascular structures (such as esophageal varices), and are again easily confused or overlooked entirely on endoscopy. Given the wide range of pathology projecting into the esophageal lumen, the nonspecific nature of symptoms experienced, and the complete lack of symptoms in many, medical investigation is often required. The gold standard in investigation is esophagogastroduodenoscopy (OGD), which allows controlled visualization and biopsy. CE has been explored extensively as an alternative, and is superior in several ways, most notably viewed as less invasive and more comfortable for patients. In the diagnosis of protruding lesions, however, an

State-of-the-art clinical aspects

esophageal investigation by a capsule is less desirable due to the risk of capsule retention and obstruction if the protruding lesion is causing a significant narrowing of the lumen. Head-to-head studies of CE with OGD have been used to evaluate varices. Sensitivity for diagnosing esophageal varices in metaanalysis and Cochrane review is 83% and 73.7%, and therefore CE is not viewed as adequate as a screening strategy [22,23]. It may be useful for those who decline OGD, as the mortality from variceal bleeding remains high at around 20% [24]. AI research in esophageal optical and endoscopic diagnostics is primarily centered around OGD images of malignancy, which are surpassing nonexpert endoscopists with a constructed convolutional neural network (CNN) system for image recognition showing a 98% sensitivity for esophageal cancer cases. The OGD images can also distinguish between superficial and invasive cancers with an accuracy of 98%, based on recognition with white light imaging and narrow band imaging (NBI) [25]. Metaanalysis of AI for the identification of esophageal neoplasia would suggest a sensitivity and specificity of 94% and 88% for cancer and neoplasia in image-based analysis and 93% and 85% in patient-based studies. Although 21 studies were included, the authors noted a significant publication bias [26]. The false negatives, that is, the cancers being missed by AI software during endoscopy, were often due to poor image quality or overlying inflammation. Image quality is an issue in esophageal capsules due to the rapid transit and undirected pictures compared with the targeted high-definition pictures taken during OGD for AI learning. To combat this issue, modifications to capsules, including increasing the frame rate to counter rapid esophageal transits, have been made. The PillCam ESO3 (Medtronic, Dublin, Ireland) for example takes 35 frames per second but remains inconsistent and is not suitable as a first-line investigation. A recent study using PillCam ESO3 recognizes Barrett’s esophagus in 82% of those with a .2 cm segment (when compared with OGD) and a lower percentage with shorter segment disease. All patients found the capsule more convenient [27], which would make it unreliable for more subtle protruding lesions. The use of string attachments that hold the capsule in the esophagus resulted in high sensitivity and specificity for esophageal varices (96% and 100%), as reported in a relatively small study [28]. String attachments have also been used in dysphagia or Barrett’s surveillance [29 31], and would also allow retrieval if investigating for protruding esophageal lesions and the capsule was to be retained. Magnetically assisted CE (MACE) also allows the capsule to be held in the esophagus, but this still gave a sensitivity of 73.3% for esophageal varices [32]. Due to these drawbacks, there is little evidence of AI in CE diagnosing protruding esophageal lesions, which lags behind its use in other regions of the GI tract.

Stomach The stomach, located between the esophagus and the SB, is an asymmetric, J-shaped organ with a greater curvature on the left and a lesser curvature on the right-hand side. From proximal to distal, the human stomach is divided into four

125

126

CHAPTER 8 Artificial intelligence for protruding lesions

parts: (1) the cardia (where inputs enter from the esophagus), (2) the fundus (the upper part of the stomach with folds secreting enzymes and hydrochloric acid), (3) the antrum (acting like a vacuum pump to empty the fundus downward), and (4) the pylorus (an anatomical sphincter). The stomach is a muscular, distensible organ that can usually expand to hold about one liter of food or more during meals and then churn and break this down before emptying through the pylorus into the SB (where digestion and absorption occur). OGD is currently the gold standard procedure to access, visualize, biopsy, and treat (when needed) the stomach. During this procedure, with or without anesthesia, trained physicians inflate the stomach with air or carbon dioxide to maximize the surface of the gastric wall that can be examined. AI solutions are under development to assist physicians in detecting and characterizing the lesions that can be seen in the stomach. Protruding lesions of the stomach are numerous and varied (Table 8.2). Most developments in AI for upper GI endoscopy aim to improve the accuracy of early gastric cancer (EGC) detection [33]. Gastric cancer is a major public health issue; in Asia, for example, there is a 90% 5-year survival rate when discovered early, but 30% when found in the advanced stages [34]. Despite ongoing technical progress in virtual coloring and magnification, the miss rate of EGC during upper GI endoscopy is suggested to be between 10% and 25% [34,35]. One of the most advanced AI products to date, the EndoAngel system (Wuhan, China), offers AI integrating solutions for upper GI endoscopy by combining both a navigation tool (monitoring blind spots) and a detection tool (with 92% sensitivity for EGC) [36]. There is also flourishing research in AI tools to delineate and predict the invasive depth of gastric cancers and to detect infection by Helicobacter pylori, chronic gastric atrophy, and gastric intestinal metaplasia (all three of late being major determinants of gastric cancers) [37]. Predicting the depth of invasion of gastric cancer (staging) is central, as it establishes the suitability for removal by endoscopic submucosal dissection, thus avoiding surgically invasive procedures for superficial lesions [37]. Gastric CE has been recently introduced into clinical practice. The main technological challenge for long has been navigation within the stomach. Commercially available MACE systems allow a comprehensive examination of the gastric mucosal surface [38]. The main benefit of this technique is that it is more comfortable and convenient for patients compared with the gastroscope (with or without general anesthesia). In a multicenter study enrolling 350 patients, MACE showed a sensitivity of 90.4% and a specificity of 94.7% in detecting focal gastric lesions (polyps, ulcers, submucosal tumors, xanthoma, diverticulum) [38]. This approach gained interest during the COVID-19 pandemic, as gastric cancer screening with a noncontact technique (CE) was considered low risk compared with a contact (OGD) procedure [39]. In addition, gastric capsules can also provide images of the SB distal to the view of conventional endoscopes. Still, several technical limitations of MACE compared with OGD should be noted: (1) capsules do not have any lavage or suction capabilities, (2) image resolution is

State-of-the-art clinical aspects

Table 8.2 Most frequent protruding lesions of the stomach. Protruding lesions

Prevalence

Site

Most frequent features, although not specific

Benign lesions Epithelial types Fundic gland polyp

42% (of benign lesions) Sporadic (induced by proton-pump inhibitors) or syndromic (familial adenomatous polyposis)

Fundus

Hyperplastic polyp

37% (of benign lesions) associated with Helicobacter pylori gastritis, cirrhosis

Adenoma

5% (of benign lesions) Sporadic (H. pylori gastritis) or syndromic (familial adenomatous polyposis and Lynch syndrome)

More frequent in the antrum than in the fundus Minimal neoplastic Potential but associated with synchronous cancers More frequent in the antrum than in the fundus Possible transformation into adenocarcinoma

Hamartoma

3% (of benign lesions) Sporadic or syndromic (Peutz Jeghers syndrome, juvenile polyposis, Cowden disease)

More frequent in the fundus than in the antrum

Round Well limited Translucent Beaded Small Multiple Red Elongated glands Sometimes lobulated. Sometimes ulcerated Can be large Adjacent H. pylori gastritis White Flat or slightly elevated or sessile Glandular pattern Single or few Adjacent H. pylori gastritis Very similar to hyperplastic polyps But with normal adjacent gastric mucosa

Nonepithelial types Inflammatory polyp

2% (of benign lesions)

Antrum Very low neoplastic potential

Isocolour (submucosal tumor) Round Sometimes erosive Single Can be large Adjacent H. pylori gastritis (Continued)

127

128

CHAPTER 8 Artificial intelligence for protruding lesions

Table 8.2 Most frequent protruding lesions of the stomach. Continued Protruding lesions

Prevalence

Site

Most frequent features, although not specific

Xanthelasma

0.5% (of benign lesions)

Fundus or antrum

White or yellow Round Small Slightly elevated Single Adjacent H. pylori gastritis

Submucosal tumors Heterotopic pancreas

1% (of benign lesions)

Antrum, greater curvature

Others Leiomyoma Neurinoma Lipoma

Isocolour Single Umbilicate Can be large Normal adjacent gastric mucosa Mass with normal gastric mucosa overlay Isocolour Single Normal adjacent gastric mucosa

Malignant lesions Primary malignancies Adenocarcinoma

90% of malignancies

Fundus or antrum

Neuroendocrine tumor

1% 2% of malignancies

Fundus/body

Irregular tubular structures Ulcerated, exophytic Sometimes infiltrating Large Nodule Umbilication Discoloration Infiltrating, exophytic. Medium size (1 2 cm) Multiple (type I and II) or single (type III) lesions (Continued)

State-of-the-art clinical aspects

Table 8.2 Most frequent protruding lesions of the stomach. Continued Protruding lesions

Lymphomas

Prevalence

Site

5% 7%, related to H. pylori infection

B lymphocytes

Most frequent features, although not specific Possible mucosal atrophy of the fundus when associated with auto-immune chronic atrophic gastritis (70% 80% of cases) Nodular or infiltrative Focal atrophy or ulcerated

Sarcomas GIST

1% 3% of malignancies

Fundus or antrum

Kaposi

Rare (mostly related to AIDS and HHV8 infection)

Fundus or antrum

Isocolour Mass Single Normal mucosal overlay, or ulcerated Can be large Blue, purple Nodules Multiple

GIST, Gastrointestinal stromal tumor.

low, without (or very limited) the possibility to use virtual chromoendoscopy or zoom, (3) capsules cannot sample or treat lesions yet, although some prototypes are under development. In the study by Xia et al., a CNN-based algorithm was trained on 820,000 1 images from 697 patients who had a MACE with the NaviCam system (Ankon Technologies, China) [40]. The images were labeled in this training dataset as polyps, submucosal tumors, xanthoma, but not gastric cancer (as only four cases were identified in this series). An independent dataset including 200,000 1 images from 100 different patients was used for validation. The validation dataset encompassed 17 polyps, five submucosal tumors, and 11 xanthomas. The perpatient sensitivity, specificity, and diagnostic accuracy of the AI system for recognizing the image in each category were: 96.5%, 94.8%, and 94.9% for polyps; 87.2%, 95.3%, and 95.2% for submucosal tumor; 90.6%, 96.9%, and 96.9% for xanthoma [40]. The same group has recently evaluated their system for real-time detection of gastric anatomical landmarks (avoiding blind spots during navigation) and lesions [41]. Performances were excellent, demonstrating a 100% sensitivity for gastric polyp/submucosal tumor detection on 50 patients. An example of detection of a gastric polyp by an AI algorithm is given in Fig. 8.1.

129

130

CHAPTER 8 Artificial intelligence for protruding lesions

FIGURE 8.1 Gastric adenoma detected and characterized as polypoid lesion by Axaro prototype (Augmented Endoscopy, Paris, France). A level of confidence in diagnosis (73%) is provided.

Small bowel The SB is a 3- to 6-m long, 3-cm wide organ positioned between the stomach and the colon. At a further distance from the mouth or anus, it has long been poorly accessible for diagnostic tools. Previously, radiological imaging was limited to SB follow-through, during which oral contrast, either barium or water-soluble contrast, was given orally before consecutive X-ray studies were performed. Since then, cross-sectional imaging (CT-scan, MRI), possibly combined with water or contrast for SB distension through a nasojejunal tube (named enteroclysis), has arisen, allowing more precise radiological investigations. In the late 1990s the advent of CE allowed direct visualization of the SB. Device-assisted enteroscopy (DAE) is far more labor-intensive and invasive than CE. To gain access to lesions in the SB, overtubes, balloons (single or double), or spirals (sometimes motorized) are manipulated by highly skilled operators in patients under general anesthesia. Overall, given that CE and cross-sectional imaging are noninvasive, they are prefered for diagnosis (and both can be combined), whereas indications of DAE are restricted to challenging diagnosis requiring biopsies or direct treatment.

State-of-the-art clinical aspects

SBCE may be offered in either inpatient or outpatient settings. Preparation include an overnight fast and may be associated with a 500 2000 mL purgative drink to cleanse the SB [42]. The capsule advances through the digestive tract due to autonomous gut motility (peristalsis). In the absence of GI obstruction, the capsule passes down the esophagus in a few seconds, and through the stomach in a mean time of 20 min. The patient can drink water 2 h after capsule ingestion and have a meal 2 h after that. The mean SB transit is 170 230 min [42], during which a mean number of 10,000 intestinal images are captured. The overall procedure lasts 8 h or more. In 90% of cases, the SB examination is complete, which means the capsule reached the colon before the battery life ended. In two-thirds of the cases, SBCE is performed for obscure GI bleeding (OGIB) (either overt or occult) after normal gastroscopy and colonoscopy. Suspected SB Crohn’s disease is another indication of SBCE, accounting for approximately 15% of procedures. Other indications (although not fully validated) are the assessment of established Crohn’s disease, the suspicion of refractory celiac disease, or the staging of polyposis (such as Peutz Jeghers syndrome). In some rarer cases, a CE investigation is performed for the assessment of a suspected SB polyp identified during cross-sectional imaging. Although the SB comprises 90% of the mucosal surface area of the GI tract, it accounts for only 3% 5% of the tumors and for 2% of the cancers that arise from it, with an incidence lower than 1 per 100,000. These lesions account for 4% of CE findings [43] and are usually diagnosed in the sixth decade. Although rare, these polyps (whether epithelial or subepithelial) vary considerably (Table 8.3) with around 40 different histological types described, one-third are benign [43]. In addition, some inflammatory lesions (fibroid or ulcerated with edematous edges), vascular structures (varices, blebs), as well as ectopic (pancreatic, gastric, or endometriosis) or hyperplastic tissues can have a tumor-like, protruding appearance, although they are neither true epithelial or mesenchymal polyps. Some are well-characterized, harmless, often incidental findings, such as benign lipomas, lymphangiectasias, or chylous cysts for example, and do not require further investigation. Others can be symptomatic (pain, anemia, bleeding, obstruction, intussusception, volvulus, perforation) or have obvious malignant features (such as an ulcerated GI stromal tumor or carcinoids) and call for prompt management (possibly endoscopic or surgical resection). Many protruding SB lesions are in a gray zone where their relation to the patient’s symptoms (often scarce) is unclear, and where they may be a precursor of a malignant lesion (adenoma, for instance) and need medical treatment, or may be malignant (lymphoma, for instance) and need surgical intervention, or are doubtful thus calling for a DAE for biopsy sampling. The SB protruding lesions represent a real challenge in CE: (1) in terms of detection, a protruding lesion (malignant or precursor of a malignant lesion) can be present in only a few abnormal frames in a sequence of 10,000 or more, with a critical risk of oversight, (2) in terms of characterization, where malignant lesions and their precursors should not be overlooked, and where benign lesions should not be overinvestigated.

131

132

CHAPTER 8 Artificial intelligence for protruding lesions

Table 8.3 Most frequent and/or relevant protruding lesions in the small bowel. Protruding lesionsa

Prevalence

Origin

Most frequent features, although not specific

Benign lesions Inflammatory lesions Inflammatory polyps

Rare

Inflammatory

Fibroid polyps

Rare

Inflammatory

Red Erosive Next to sutures and strictures Red Erosive

Ectopic tissues Gastric heterotopia

Rare

Congenital

Pancreatic heterotopia

Rare

Congenital

White center, red halo Sessile Isocolour Sessile, smooth Sometimes depressed or ulcerated

Endometriosis Hyperplasia Hyperplastic polyps

Rare

Brunneromas

7%

Nodular lymphoid hyperplasia

Frequent in children and young adults

Hyperplasia

Lamina propria and submucosa

Isocolour Round, smooth Sometimes lobulated Small Duodenum Isocolour Multiple, small Rarely large, pedunculated, ulcerated Duodenal bulb Whitish Small Multiple (aggregated) Distal jejunum and ileum (Continued)

State-of-the-art clinical aspects

Table 8.3 Most frequent and/or relevant protruding lesions in the small bowel. Continued Protruding lesionsa

Prevalence

Origin

Most frequent features, although not specific

Hamartomas Hamartomas

Rare Sporadic or syndromic (Peutz Jeghers, juvenile polyposis)

Adenoma

1% Sporadic or syndromic (familial adenomatous polyposis and Lynch syndrome)

Multiple tissue components abnormally combined

Isocolour Smooth Pedunculated Lobulated

Epithelial neoplasms Epithelial (crypt) Possible transformation into adenocarcinoma

Regular tubular structures White Flat Duodenum

Mesenchymal tumors Hemangioma

0.05% Sporadic or syndromic (blue rubber bleb nevus syndrome)

Lymphangioma

3% 20%

Leiomyoma

Rare

Lipoma

3% 15%

Mesenchymal (from the mucosa to the serosa) Cavernous, capillary, or mixed Mesenchymal (submucosa)

Mesenchymal (smooth muscle cells) Distinguished from GIST, when needed, by immunochemistry Mesenchymal (adipocytes)

Red, blue, or purple Soft Bleeding Yellow and white engorged villi or Plaque with superficial capillaries (known as chylous cyst) Flat or raised Multiple Gray-white Round, smooth Normal mucosal overlay

Yellow Round, smooth Often sessile, mobile Normal mucosal overlay (Continued)

133

134

CHAPTER 8 Artificial intelligence for protruding lesions

Table 8.3 Most frequent and/or relevant protruding lesions in the small bowel. Continued Protruding lesionsa

Prevalence

Origin

Most frequent features, although not specific

Neurofibroma

Rare (mostly related to neurofibromatosis type 1)

Mesenchymal (neural and connective tissue)

Isocolour Round, smooth Sometimes ulcerated

Malignant lesions Primary malignancies Adenocarcinoma Rare (47% of malignancies)

Epithelium

Neuroendocrine tumors

Rare (28% of malignancies)

Neuroendocrine cells

Rare (12% of malignancies)

Cajal cells

Rare (mostly related to AIDS and HHV8 infection) Very rare Rare (12% of malignancies) T-cell lymphomas often arising from celiac disease

Endothelial cells

Sarcomas GIST

Kaposi

Hemangiosarcoma Lymphomas

Endothelial cells B (90%) or T (10%) lymphocytes

Irregular tubular structures Ulcerated, infiltrating, exophytic Large Duodenum, jejunum Nodule Umbilication Discoloration Infiltrating, exophytic Sometimes multiple Ileum Red Smooth or lobulated Duodenum, jejunum Blue, purple Multiple Dark red or blue Varied 11 1 Nodular or infiltrative Focal atrophy or ulcerated Wall thickening, ulcerated strictures (Continued)

State-of-the-art clinical aspects

Table 8.3 Most frequent and/or relevant protruding lesions in the small bowel. Continued Protruding lesionsa

Prevalence

Origin

Most frequent features, although not specific Localized when primary Diffused when immunoproliferative

Metastases Melanoma

Rare (and sometimes primary)

Skin

Others

Very rare

Colon, stomach, uterus, ovaries, bladder, breast, bronchi. . .

Typically black Sometimes amelanotic Sometimes multiple Varied 11 1 Exophytic or nodular

Infiltration by extraintestinal malignancies Very rare

Neighboring organs

Red Infiltrating Ulcerated

GIST, Gastrointestinal stromal tumor. a More than 40 histological types have been described, only the most frequent/relevant types are described here.

It is estimated that the diagnostic yield of SBCE is 75% for epithelial lesions smaller or equal to 10 mm and 91% for larger lesions. For subepithelial lesions, the diagnostic yields are 83% and 78%, respectively [44]. Similarly, the miss rate of ulcerated SB neoplasm by CE is 18.9% [45]. Computer-aided diagnostics are under development to assist physicians in the detection of protruding lesions. Some examples are provided in Figs. 8.2 8.4. It is likely that these diagnostic tools will be able to characterize such lesions (tumors vs. tumor-like lesions and epithelial versus subepithelial) in the future. According to a recent systematic review by Kim et al., based on PRISMA methodology, 11 studies published in extenso in English have evaluated the performances of CAD systems for the detection of protruding lesions in SBCE [46]. All solutions were trained on still images from internal datasets. Four targeted SB polyps, five targeted SB tumors, and one used the generic term SB protruding lesions. These are presented in Table 8.4. The study by Saito et al., published in 2020, clearly outperforms the previous ones in terms of datasets: 30,584 images

135

136

CHAPTER 8 Artificial intelligence for protruding lesions

FIGURE 8.2 Small bowel adenoma detected and characterized as a polypoid lesion by Axaro prototype (Augmented Endoscopy, Paris, France). A level of confidence in diagnosis (99%) is provided.

of protruding lesions from 292 patients for training and 7507 images of protruding lesions from 73 patients for evaluation, in addition to 10,000 control, normal images from 20 patients. A formal diagnosis was based on the pathological examination of surgical or endoscopic specimens in more than 80% of cases. A CNN showed an overall 90.72% sensitivity for protruding lesions detection on still frames. Of note, the sensitivity varied according to the type of lesion: 86.5% for polyps, 92.0% for nodules, 95.8% for epithelial tumors, 77.0% for submucosal tumors, and 94.4% for venous structures. At the patient level (n 5 73), the detection rate was 98.6%. The CNN also managed to correctly classify more than 80% of nodules and epithelial tumors. In addition, two large, controlled trials detailed the diagnostic performance of commercially available CAD systems for the detection of such lesions (Table 8.5). Both studies demonstrated that AI-assisted reading outperforms conventional reading (CR) for any type of lesions in terms of both sensitivity and time taken. In the first study by Ding et al., a CNN-based solution (Proscan, Ankon Technologies Co, China) was trained on 158,235 images from 1970 patients, and it was then assessed on the CE examinations of 5000 different patients [55]. Overall, both per-lesion and per-patient analysis reached a 99.9% sensitivity

State-of-the-art clinical aspects

FIGURE 8.3 Small bowel Peutz Jeghers polyp detected and characterized as polypoid lesion by Axaro prototype (Augmented Endoscopy, Paris, France). A level of confidence in diagnosis (64%) is provided.

FIGURE 8.4 Small ileal polyp in young adult patient with Lynch syndrome by Proscan (Ankon Technologies Co, Ltd, Shanghai, China).

137

Table 8.4 Studies evaluating the performances of computer-aided diagnosis systems for the detection of protruding lesions in small bowel capsule endoscopy on still images databases (adapted from [46]). First author, year, (reference)

AI model

Images of protruding/ control lesions in training dataset

Target

Images of protruding/ control lesions in test dataset

Sensitivity (%)

Specificity (%)

PPV (%)

NPV (%)

Accuracy (%)

Li 2009, [8]

Feature analysis with MLP BoW modelSVM Texture analysis with SVM Texture analysis with SVM Texture analysis with k-NN, SVM or MLP Texture analysis with NN Texture analysis with SVM

150/150

Polyps

150/150

89.33

95.38

95.71

88.57

92.14

25/50

Polyps

50/100

66.00

95.00

86.84

84.82

85.33

Unclear

Polyps

10/40

100.00

67.50

43.48

100.00

74.00

550/550

Tumors

50/50

90.00

98.00

97.83

90.74

94.00

450/450

Tumors

150/150

92.00

88.67

89.03

91.72

90.33

700/2300

Tumors

700/2300

93.86

93.09

80.51

98.03

93.27

540/540

Tumors

60/60

85.00

81.67

82.26

84.48

83.33

Hwang, 2011, [47] Karagyris, 2011, [48] Li, 2011, [49]

Li, 2011, [50]

Barbosa, 2012, [51] Li, 2012, [10]

Li, 2012,[52]

Constantinescu, 2015, [53] Kundu, 2020, [54] Saito, 2020, [17]

Texture analysis with SVM Texture analysis with NN Discriminant analysis with SVM CNN

540/540

Tumors

60/60

88.33

96.67

96.36

89.23

92.50

Unclear

Polyps

32/58

93.75

91.38

85.71

96.36

92.22

30/1617

Tumors

30/1617

86.67

91.96

16.67

99.73

91.86

30584/0

Protruding lesions

7507/20000

90.72

79.81

77.13

91.97

84.49

AI, Artificial intelligence; BoW, bag-of-words; CNN, convolutional neural network; MLP, multilayer perceptron; SVM, support vector machine.

140

CHAPTER 8 Artificial intelligence for protruding lesions

Table 8.5 Studies evaluating the performances of computer-aided diagnosis systems for the detection of protruding lesions in small bowel capsule endoscopy on video recordings. First author, year, (reference)

Company, capsule, AI solution

Lesions

CR detection, n (%)

CAD detection (%)

P

Ding et al., 2019, [55]

Ankon Technologies Co., Proscan

Polyps

204 (3.44%) 128 (2.16%) 120 (2.02%)

261 (4.40%) 227 (3.83%) 257 (4.34%)

.007

311 (10.73%) 545 (18.81%) 104 (3.59%) 254 (8.76%)

414 (14.29%) 676 (23.33%) 131 (4.52%) 327 (11.28%)

,.01

Xie et al., 2022, [56]

Jinshan, Omom, SmartScan

Protruding lesions Lymphatic nodular hyperplasia Venous structure Nodule Mass/Tumor Polyp(s)

,.0001 ,.0001

,.01 ns ,.01

AI, Artificial intelligence; CAD, computer-aided diagnosis; CR, conventional reading.

(compared with 76.9% and 74.6%, respectively, with CR). The video sequence mean reading time was dramatically reduced from 96.6 to 5.9 min. The number and types of protruding lesions were not detailed in the study protocol. However, sensitivity for protruding lesions detection was significantly raised from 56.1% with CR to 99.6% with the CNN-based auxiliary reading, on both per-patient and per-lesion analyses, still with a 100% specificity. Similar performances were observed for polyps and lymphangiectasia. Of note, this system currently offers no option for the characterization of lesions. Most recently, a similar study by Xie et al. was performed on a CNN algorithm named SmartScan (Chongqing Jinshan Science & Technology Group Co., Ltd., China) [56]. The system was trained on 2927 CE procedures (with approximately 15 million images) from 29 medical centers using SmartScan to detect and characterize 17 different types of findings (according to structured terminology). The performance was then assessed on an independent dataset of 2898 CE examinations from 22 medical centers. Overall, the AI-assisted reading was significantly more sensitive than CR (i.e., for any type of lesion, 79.3% vs. 70.7%), and more specifically for the four different types of protruding lesions (venous structures, nodules, mass/tumor, and polyps) (Table 8.5). In other words, the CNN-based algorithm missed 4.1% of findings versus 23.9% with CR (central, experts reading) (Table 8.5). The mean reading time with SmartScan was 5.4 min, compared with 51.4 min with CR.

State-of-the-art clinical aspects

Colon The colonic lumen is comparatively anatomically irregular, with normal physiological spasms, folds, and compression from intra-abdominal structures mimicking pathology on still capsule images. The ileocaecal valve can be mistaken for a polyp or tumor, and unlike CT colonography or invasive colonoscopy, which allows exact positioning and further characterization of findings, this can lead to a false positive being reported on CE. The rapid rise in CCE is driven primarily by cancer diagnostics. The survival rate for colon cancer is vastly improved by early diagnosis and better yet identification of precancerous changes and prevention. This forms the basis of national screening in many countries, resulting in a potentially vast and time-consuming requirement for CCE reporting. AI is becoming an appealing solution in this field, and this is reflected in the comparatively large body of research. In clinical practice, modifications to the capsule design, which include longer battery life, pausing with a sleep mode in the stomach, variable frame rates with movement, and software to reduce the number of duplicate frames given, to optimize exploration of the colon are already present. Due to relative stasis during colonic transit these changes have increased completion rates and cut the time for manual review of CCE recordings. Automated analysis and removal of poor-quality frames without disregarding significant findings speed up analysis [57]. Flexible spectral imaging color enhancement (FICE) (Fujifilm, Tokyo, Japan) on CCE increases polyp detection [58]. Using FICE, it is possible to differentiate between adenomatous and hyperplastic polyps (91.2% sensitivity and 88.2% accuracy) [59]. Deep learning systems for real-time polyp detection, such as the commercially available GI-Genius (Medtronic), which are already available in conventional colonoscopy, have been shown to increase polyp detection (from 40.3% to 54.8%) when randomized against conventional colonoscopy without AI, which did not increase withdrawal time [60]. Although this increased yield was predominantly driven by smaller polyps, there was improved detection in all neoplasia types. CAD systems have also been shown to directly reduce the number of missed adenomas in a prospective tandem colonoscopy study, with the implication that it will reduce interval cancers [61]. Although not commercially available, similar developments in CCE for polyp detection also show high levels of accuracy [13,16]. Deep learning for polyp recognition in CCE was found to be 98% accurate [12], suggesting this is rapidly approaching commercial viability, just as it has for conventional colonoscopy. There are several remaining clinical concerns regarding polyp detection by CCE. Sessile serrated lesions (SSLs) minimally protrude and are suggested to account for 15% 30% of CRCs and a much higher proportion of interval cancers [62], suggesting these are missed on colonoscopy. CCE is reliable at detecting adenomas but risks missing SSLs (sensitivity of only 29% and 33% for SSLs of .6 and .10 mm seen on colonoscopy [63]). Pathologically, SSLs and hyperplastic polyps are difficult to tell apart, hampering the training of AI and CAD [63,64].

141

142

CHAPTER 8 Artificial intelligence for protruding lesions

When cancers are found, CAD systems have as high as 89.4% sensitivity in identifying whether there is invasion [65,66], although these are not used in clinical practice. Real-time CAD analysis of polyps, allowing an optical biopsy prior to resection with an accuracy of 94% from unaltered videos, outperforms nonexpert endoscopists [67 69]. Diminutive rectal polyps are a common finding, and there is an ongoing debate over the safety of an assess and discard policy. CNNs can distinguish polyps of ,5 mm into adenomatous or hyperplastic with greater accuracy than non-NBI-trained endoscopists and reduce the time required for assessment by all endoscopists [70]. As the demand for CRC screening already exceeds capacity, AI-interpreted CCE would provide a potential solution [71]. Given the increasing access to CCE, the high yield from the procedure, and the high population prevalence of colonic polyps, there is a risk that this will increase the burden on interventional endoscopy services. According to studies such as Repici et al., the incidence of polyps in the screening population may be over 50%, making the requirement for therapeutics overwhelming if all of these require removal [60]. As such accurate assessment of polyps, or cancers, using AI on CCE would reduce the number of follow-up procedures required and allow more appropriate triage (be it surgery, endoscopy, or surveillance), thus improving the patient pathway [72].

Perspectives on challenges and developments The integration of AI into the interpretation of CE procedures is highly expected by the community of readers [73]. Still, there are several barriers to this development. First and foremost, large databases are required to develop AI systems for CE. These databases are needed not only to identify but also to characterize the many types of protruding lesions found in the GI tract. Outside the colon, these lesions are relatively infrequent, making data collection (and the subsequent steps) challenging. Whatever is developed for CE, the established polyp analysis AI programs are based on high-definition images, allowing detailed assessment of the mucosal surface and pit patterns, which will be rarely possible with the oftenpartial views of lesions on the untargeted pictures taken by capsule devices. These are often blurry due to rapid movement, limiting interpretation. CE databases should be collected and labeled from a variety of patients and centers to ensure they are heterogeneous and varied enough. It is also desirable to obtain agnostic-device models, which would require images captured from different capsule devices and brands. In addition, these databases should be split into multiple sets of sufficient size for different tasks (training, tuning, testing). Merging databases poses medicolegal issues regarding ownership and data sharing between regions and countries [74]. Consecutive images of the same lesion should avoid overfitting, thus limiting the growth of these databases. Labeling is demanding and calls for expertise when applied to the specific appearances of GI protruding lesions. Indeed, inter- and intra-observer reproducibility is

Conflict of interest

moderate regarding these types of lesions, even among experts [75]. Several initiatives, even by expert CE readers, are attempting to limit the variability in language and interpretation [76]. Training AI systems for CE diagnostics based on gastroscopy or colonoscopy images (with pathological analysis as ground truth) is the most reasonable option to overcome these limitations, but it is hard to apply to SBCE (where pathological data are not always available). Data augmentation (transforming images by means of cropping, rotating, mirroring, etc.) and generative adversarial networks (creating untrue but realistic images) are other alternative options to alleviate the need for additional original images, as currently proposed for applications in colonoscopy [77]. After databases and AI solutions are developed, evaluation metrics must be defined to evaluate the clinical impact of the system. Here, again, expert consensus will certainly be required when pathology is not available. The localization of detected findings is another challenging area of research. It is likely that AI will be able to determine the organ (esophagus/stomach/SB/colon) from which the lesion arises, based on the background and surrounding images. Further research is required to determine the segment of the organ (ascending/transverse/descending colon, for example) where the defect is located [78].

Conclusion AI-based solutions for the detection of GI protruding lesions in SBCE are already available [55,56]. Similar initiatives are underway for the stomach and colon capsule as well. With tremendous improvements in robotics and miniaturization, it is likely to be possible in the coming years to detect protruding lesions in outpatients by making them swallow a panenteric capsule [79]. It will, however, take time before this becomes the standard of care for GI cancer screening or diagnostics, as this depends not only on research efforts but also on the acceptability of this approach by patients and physicians. AI-based characterization and localization of the varied protruding lesions in the GI tract are challenges that will run in the longer term. Success will strongly depend on our ability to gather large sets of high-quality images and clinical data.

Conflict of interest Xavier Dray and Aymeric Histace are cofounders and shareholders of Augmented Endoscopy. Xavier Dray has received lecture fees from Bouchara Recordati, Fujifilm, Medtronic, MSD, and Pfizer, and has acted as a consultant for Alfasigma, Boston Scientific, Norgine, and Pentax. All other authors have declared no conflict of interest.

143

144

CHAPTER 8 Artificial intelligence for protruding lesions

References [1] Zauber AG, Winawer SJ, O’Brien MJ, et al. Colonoscopic polypectomy and longterm prevention of colorectal-cancer deaths. N Engl J Med 2012;366:687 96. [2] Yong KK, Kyaw M, Chadwick G, et al. Increasing incidence of young-onset colorectal cancers in the UK and rising mortality in rectal cancers. Gut 2020;69:2267 8. [3] Vuik FE, Nieuwenburg SA, Bardou M, et al. Increasing incidence of colorectal cancer in young adults in Europe over the last 25 years. Gut 2019;68:1820 6. [4] Wolf AMD, Fontham ETH, Church TR, et al. Colorectal cancer screening for average-risk adults: 2018 guideline update from the American Cancer Society: ACS Colorectal Cancer Screening Guideline. CA Cancer J Clin 2018;68:250 81. [5] Ismail MS, Murphy G, Semenov S, et al. Comparing Colon Capsule Endoscopy to colonoscopy; a symptomatic patient’s perspective. BMC Gastroenterol 2022;22:31. [6] Zheng Y, Hawkins L, Wolff J, et al. Detection of lesions during capsule endoscopy: physician performance is disappointing. Am J Gastroenterol 2012;107:554 60. [7] Takada K, Yabuuchi Y, Kakushima N. Evaluation of current status and near future perspectives of capsule endoscopy: Summary of Japan Digestive Disease Week 2019. Dig Endosc 2020;32:529 31. [8] Li B., Meng M.Q.-H., Xu L. A comparative study of shape features for polyp detection in wireless capsule endoscopy images. In: 2009 annual international conference of the IEEE engineering in medicine and biology society; 2009. p. 3731 34. [9] ZhaoQ., Meng M.Q.-H. Polyp detection in wireless capsule endoscopy images using novel color texture features. In: 2011 9th world congress on intelligent control and automation. Taipei: IEEE; 2011. p. 948 52. [10] Li B-P, Meng MQ-H. Comparison of several texture features for tumor detection in CE images. J Med Syst 2012;36:2463 9. [11] Yuan Y., Meng M.Q.-H. A novel feature for polyp detection in wireless capsule endoscopy images. In: 2014 IEEE/RSJ international conference on intelligent robots and systems. Chicago, IL, USA: IEEE; 2014. p. 5010 15. [12] Yuan Y, Meng MQ-H. Deep learning for polyp recognition in wireless capsule endoscopy images. Med Phys 2017;44:1379 89. [13] Yuan Y., Qin W., Ibragimov B., et al. RIIS-DenseNet: rotation-invariant and image similarity constrained densely connected convolutional network for polyp detection. In: Medical image computing and computer assisted intervention MICCAI 2018 21st international conference, 2018, Proceedings. Springer Verlag. p. 620 28. [14] Guo X, Yuan Y. Triple ANet: adaptive abnormal-aware attention network for WCE image classification. In: Shen D, Liu T, Peters TM, et al., editors. Medical Image Computing and Computer Assisted Intervention MICCAI. Cham: Springer International Publishing; 2019. p. 293 301. [15] Yuan Y, Qin W, Ibragimov B, et al. Densely connected neural network with unbalanced discriminant and category sensitive constraints for polyp recognition. IEEE Trans Autom Sci Eng 2020;17:574 83. [16] Laiz P, Vitria` J, Wenzek H, et al. WCE polyp detection with triplet based embeddings. Comput Med Imaging Graph 2020;86:101794. [17] Saito H, Aoki T, Aoyama K, et al. Automatic detection and classification of protruding lesions in wireless capsule endoscopy images based on a deep convolutional neural network. Gastrointest Endosc 2020;92:144 51. e1.

References

[18] Pascual G, Laiz P, Garcı´a A, et al. Time-based self-supervised learning for wireless capsule endoscopy. Comput Biol Med 2022;146:105631. [19] Gilabert P., Vitria` J., Laiz P., et al. Artificial intelligence to improve polyp detection and screening time in colon capsule endoscopy. Preprint, In Review. Epub ahead of print 31 January; 2022. doi: 10.21203/rs.3.rs-1278962/v1. [20] Oesophageal Cancer Statistics | World Cancer Research Fund International. WCRF international, ,https://www.wcrf.org/cancer-trends/oesophageal-cancer-statistics/. [accessed 14.06.22]. [21] Ishihara R, Takeuchi Y, Chatani R, et al. Original article: Prospective evaluation of narrow-band imaging endoscopy for screening of esophageal squamous mucosal high-grade neoplasia in experienced and less experienced endoscopists: NBI for esophageal neoplasia. Dis Esophagus 2010;23:480 6. [22] McCarty TR, Afinogenova Y, Njei B. Use of wireless capsule endoscopy for the diagnosis and grading of esophageal varices in patients with portal hypertension: a systematic review and meta-analysis. J Clin Gastroenterol 2017;51:174 82. [23] Colli A, Gana JC, Turner D, et al. Capsule endoscopy for the diagnosis of oesophageal varices in people with chronic liver disease or portal vein thrombosis. Cochrane Database Syst Rev 2014;. Available from: https://doi.org/10.1002/14651858. CD008760.pub2 Epub ahead of print 1. [24] Stokkeland K, Brandt L, Ekbom A, et al. Improved prognosis for patients hospitalized with esophageal varices in Sweden 1969 2002. Hepatology 2006;43:500 5. [25] Horie Y, Yoshio T, Aoyama K, et al. Diagnostic outcomes of esophageal cancer by artificial intelligence using convolutional neural networks. Gastrointest Endosc 2019;89:25 32. [26] Bang CS, Lee JJ, Baik GH. Computer-aided diagnosis of esophageal cancer and neoplasms in endoscopic images: a systematic review and meta-analysis of diagnostic test accuracy. Gastrointest Endosc 2021;93:1006 15. e13. [27] Duvvuri A, Desai M, Vennelaganti S, et al. Diagnostic accuracy of a novel third generation esophageal capsule as a non-invasive detection method for Barrett’s esophagus: a pilot study. J Gastroenterol Hepatol 2021;36:1222 5. [28] Ramirez FC, Hakim S, Tharalson EM, et al. Feasibility and safety of string wireless capsule endoscopy in the diagnosis of esophageal varices. Am J Gastroenterol 2005;100:1065 71. [29] Ramirez FC, Akins R, Shaukat M. Screening of Barrett’s esophagus with stringcapsule endoscopy: a prospective blinded study of 100 consecutive patients using histology as the criterion standard. Gastrointest Endosc 2008;68:25 31. [30] Liao Z, Gao R, Xu C, et al. Sleeve string capsule endoscopy for real-time viewing of the esophagus: a pilot study (with video). Gastrointest Endosc 2009;70:201 9. [31] Chen Y-Z, Pan J, Luo Y-Y, et al. Detachable string magnetically controlled capsule endoscopy for complete viewing of the esophagus and stomach. Endoscopy 2019;51:360 4. [32] Beg S, Card T, Warburton S, et al. Diagnosis of Barrett’s esophagus and esophageal varices using a magnetically assisted capsule endoscopy system. Gastrointest Endosc 2020;91:773 81 e1. [33] Fu X-Y, Mao X-L, Chen Y-H, et al. The feasibility of applying artificial intelligence to gastrointestinal endoscopy to improve the detection rate of early gastric cancer screening. Front Med 2022;9:886853.

145

146

CHAPTER 8 Artificial intelligence for protruding lesions

[34] Katai H, Ishikawa T, Akazawa K, et al. Five-year survival analysis of surgically resected gastric cancer cases in Japan: a retrospective analysis of more than 100,000 patients from the nationwide registry of the Japanese Gastric Cancer Association (2001 2007). Gastric Cancer 2018;21:144 54. [35] Yoshida N, Doyama H, Yano T, et al. Early gastric cancer detection in high-risk patients: a multicentre randomised controlled trial on the effect of second-generation narrow band imaging. Gut 2021;70:67 75. [36] He X, Wu L, Dong Z, et al. Real-time use of artificial intelligence for diagnosing early gastric cancer by magnifying image-enhanced endoscopy: a multicenter diagnostic study (with videos). Gastrointest Endosc 2022;95:671 8. e4. [37] Tokat M, van Tilburg L, Koch AD, et al. Artificial intelligence in upper gastrointestinal endoscopy. Dig Dis Basel Switz 2021;. Available from: https://doi.org/10.1159/ 000518232 Epub ahead of print 21 July. [38] Liao Z, Hou X, Lin-Hu E-Q, et al. Accuracy of magnetically controlled capsule endoscopy, compared with conventional gastroscopy, in detection of gastric diseases. Clin Gastroenterol Hepatol 2016;14:1266 73. e1. [39] Zhu J-H, Pan J, Xu X-N, et al. Noncontact magnetically controlled capsule endoscopy for infection-free gastric examination during the COVID-19 pandemic: a pilot, open-label, randomized trial. Endosc Int Open 2022;10:E163 71. [40] Xia J, Xia T, Pan J, et al. Use of artificial intelligence for detection of gastric lesions by magnetically controlled capsule endoscopy. Gastrointest Endosc 2021;93:133 9 e4. [41] Pan J, Xia J, Jiang B, et al. Real-time identification of gastric lesions and anatomical landmarks by artificial intelligence during magnetically controlled capsule endoscopy. Endoscopy 2022;. Available from: https://doi.org/10.1055/a-1724-6958 Epub ahead of print 26 January. [42] Rahmi G, Cholet F, Gaudric M, et al. Effect of different modalities of purgative preparation on the diagnostic yield of small bowel capsule for the exploration of suspected small bowel bleeding: a multicenter randomized controlled trial. Am J Gastroenterol 2022;117:327 35. [43] de Latour RA, Kilaru SM, Gross SA. Management of small bowel polyps: a literature review. Best Pract Res Clin Gastroenterol 2017;31:401 8. [44] Honda W, Ohmiya N, Hirooka Y, et al. Enteroscopic and radiologic diagnoses, treatment, and prognoses of small-bowel tumors. Gastrointest Endosc 2012;76:344 54. [45] Lewis BS, Eisen GM, Friedman S. A pooled analysis to evaluate results of capsule endoscopy trials. Endoscopy 2005;37:960 5. [46] Kim HJ, Gong EJ, Bang CS, et al. Computer-aided diagnosis of gastrointestinal protruded lesions using wireless capsule endoscopy: a systematic review and diagnostic test accuracy meta-analysis. J Pers Med 2022;12:644. [47] Hwang S. Bag-of-visual-words approach to abnormal image detection in wireless capsule endoscopy videos. In: Bebis G, Boyle R, Parvin B, et al., editors. Advances in visual computing. Berlin, Heidelberg: Springer; 2011. p. 320 7. [48] Karargyris A, Bourbakis N. Detection of small bowel polyps and ulcers in wireless capsule endoscopy videos. IEEE Trans Biomed Eng 2011;58:2777 86. [49] Li B, Meng MQ-H. Contourlet-based features for computerized tumor detection in capsule endoscopy images. Ann Biomed Eng 2011;39:2891. [50] Li B, Meng MQ-H, Lau JYW. Computer-aided small bowel tumor detection for capsule endoscopy. Artif Intell Med 2011;52:11 16.

References

[51] Barbosa DC, Roupar DB, Ramos JC, et al. Automatic small bowel tumor diagnosis by using multi-scale wavelet-based analysis in wireless capsule endoscopy images. Biomed Eng Online 2012;11:3. [52] Li B, Meng MQ-H. Tumor recognition in wireless capsule endoscopy images using textural features and SVM-based feature selection. IEEE Trans Inf Technol Biomed 2012;16:323 9. [53] Constantinescu A, Ionescu M, Iov˘anescu V, et al. A computer-aided diagnostic system for intestinal polyps identified by wireless capsule endoscopy. Rom J Morphol Embryol 2016;57(3):979 84. [54] Kundu AK, Fattah SA, Wahid KA. Multiple linear discriminant models for extracting salient characteristic patterns in capsule endoscopy images for multi-disease detection. IEEE J Transl Eng Health Med 2020;8:1 11. [55] Ding Z, Shi H, Zhang H, et al. Gastroenterologist-level identification of small-bowel diseases and normal variants by capsule endoscopy using a deep-learning model. Gastroenterology 2019;157:1044 54. e5. [56] Xie X, Xiao Y, Zhao X, et al. Development and validation of an artificial intelligence model for small bowel capsule endoscopy video review. JAMA Netw Open 2022;5(7):e2221992. [57] Biniaz A, Zoroofi RA, Sohrabi MR. Automatic reduction of wireless capsule endoscopy reviewing time based on factorization analysis. Biomed Signal Process Control 2020;59:101897. [58] Xavier S, Monteiro S, Boal Carvalho P, et al. Sa1944 chromoendoscopy lightening the way for colorectal polyps’ detection in colon capsule. Gastrointest Endosc 2018;87:AB259. [59] Nakazawa K, Nouda S, Kakimoto K, et al. The differential diagnosis of colorectal polyps using colon capsule endoscopy. Intern Med 2021;60:1805 12. [60] Repici A, Badalamenti M, Maselli R, et al. Efficacy of real-time computer-aided detection of colorectal neoplasia in a randomized trial. Gastroenterology 2020;159:512 20. e7. [61] Wang P, Liu P, Glissen Brown JR, et al. Lower adenoma miss rate of computeraided detection-assisted colonoscopy vs routine white-light colonoscopy in a prospective tandem study. Gastroenterology 2020;159:1252 61. e5. [62] East JE, Atkin WS, Bateman AC, et al. British Society of Gastroenterology position statement on serrated polyps in the colon and rectum. Gut 2017;66:1181 96. [63] Rex DK, Adler SN, Aisenberg J, et al. Accuracy of capsule colonoscopy in detecting colorectal polyps in a screening population. Gastroenterology 2015;148:948 57. e2. [64] Hann A, Meining A. Artificial intelligence in endoscopy. Visc Med 2021;37:471 5. [65] Takeda K, Kudo S, Mori Y, et al. Accuracy of diagnosing invasive colorectal cancer using computer-aided endocytoscopy. Endoscopy 2017;49:798 802. [66] Ito N, Kawahira H, Nakashima H, et al. Endoscopic diagnostic support system for cT1b colorectal cancer using deep learning. Oncology 2019;96:44 50. [67] Byrne MF, Chapados N, Soudan F, et al. Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy using a deep learning model. Gut 2019;68:94 100. [68] Kominami Y, Yoshida S, Tanaka S, et al. Computer-aided diagnosis of colorectal polyp histology by using a real-time image recognition system and narrow-band imaging magnifying colonoscopy. Gastrointest Endosc 2016;83:643 9. [69] Min M, Su S, He W, et al. Computer-aided diagnosis of colorectal polyps using linked color imaging colonoscopy to predict histology. Sci Rep 2019;9:2881.

147

148

CHAPTER 8 Artificial intelligence for protruding lesions

[70] Jin EH, Lee D, Bae JH, et al. Improved accuracy in optical diagnosis of colorectal polyps using convolutional neural networks with visual explanations. Gastroenterology 2020;158:2169 79. e8. [71] Robertson AR, Segui S, Wenzek H, et al. Artificial intelligence for the detection of polyps or cancer with colon capsule endoscopy. Ther Adv Gastrointest Endosc 2021;14 263177452110202. [72] Medtronic. Cutting edge technology requires cutting edge solutions. That’s why we are proud to announce our partnership with @amazon to bring PillCamTM Genius to market. VP & GM of GIH, Gio Di Napoli, explains how it will work. ,https:// TwitterCom/Medtronic.; 2021 [accessed 6.03.21]. [73] Leenhardt R, Fernandez-Urien Sainz I, Rondonotti E, et al. PEACE: perception and expectations toward artificial intelligence in capsule endoscopy. J Clin Med 2021;10:5708. [74] Dray X, Toth E, de Lange T, et al. Artificial intelligence, capsule endoscopy, databases, and the Sword of Damocles. Endosc Int Open 2021;09:E1754 5. [75] Afecto E, Pinho R, Gomes C, et al. Evaluation of a new composite score combining SPICE and protrusion angle scores to distinguish submucosal lesions from innocent bulges. Rev Espanola Enfermedades Dig Organo Of Soc Espanola Patol Dig 2022;114:151 5. [76] Leenhardt R, Koulaouzidis A, McNamara D, et al. A guide for assessing the clinical relevance of findings in small bowel capsule endoscopy: analysis of 8064 answers of international experts to an illustrated script questionnaire. Clin Res Hepatol Gastroenterol 2021;45:101637. [77] Yoon D, Kong H-J, Kim BS, et al. Colonoscopic image synthesis with generative adversarial network for enhanced detection of sessile serrated lesions using convolutional neural network. Sci Rep 2022;12:261. [78] Herp J, Deding U, Buijs MM, et al. Feature point tracking-based localization of colon capsule endoscope. Diagn Basel Switz 2021;11:193. [79] Dray X, Koulaouzidis A. Panenteric capsule endoscopy: a new soldier at the forefront of lower gastrointestinal bleeding workup and. . .beyond!. Eur J Gastroenterol Hepatol 2021;33:947 8.

CHAPTER

Artificial intelligence for vascular lesions

9

Pere Gilabert, Pablo Laiz and Santi Seguı´ Department of Mathematics and Computer Science, University of Barcelona, Barcelona, Spain

Introduction This chapter describes the impact of artificial intelligence (AI) for the detection and characterization of vascular lesions of the gastrointestinal (GI) tract using wireless capsule endoscopy (WCE).

Wireless capsule endoscopy and artificial intelligence WCE is an effective, reliable, safe, and noninvasive technology used for GI tract diagnosis of various lesions and abnormalities [1]. The clinical value of WCE has been demonstrated in a wide array of illnesses, including the evaluation of patients with suspected small bowel hemorrhage, diagnosis and monitoring of Crohn’s disease, diagnostics of ulcers, and detection of polyps and tumors. However, despite its advantages for both patients and clinicians, its use in the clinical practice has been below initial expectations, mainly because of the complexity of WCE examinations. Due to the long time required to pass through the GI tract, the resulting WCE videos can contain more than 50,000 frames, which leads to a complex, tedious, and timeconsuming job for clinical experts to perform a visual review of each study. It has been reported that even experienced readers take more than 50 min to analyze the data of each patient [2]. To alleviate these problems, and consider WCE a reliable solution, the use of AI has been claimed in several papers [3 6] as the solution to reduce the required time and cost to review WCE-generated videos as well as to increase the sensitivity and specificity of the diagnostic tests. Several solutions have been presented to detect and characterize pathological frames such as bleeding [7], ulcers [8], or polyps [9], but just a few of them have been presented and validated to work within a real-world clinical environment [10]. The use of AI-based models has been gaining importance for some years now. These models are based on machine learning algorithms and now they are almost all based on deep learning techniques. However, the development of these models is challenging. They need a large amount of labeled data to be trained and proper Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00012-5 © 2023 Elsevier Inc. All rights reserved.

149

150

CHAPTER 9 Artificial intelligence for vascular lesions

data to be validated, which, in the field of medical imaging, is sometimes difficult to accomplish. The existing public datasets are usually very limited and most researchers use private data to fulfill this shortage.

Vascular lesions in gastrointestinal tract Many types of abnormalities can occur along the GI tract that can affect the vascular architecture—arteries, veins, capillaries, and lymphatic vessels can be affected by vascular lesions. Vascular lesions can appear spontaneously at any stage of life or can be congenital, i.e., can appear during the gestation period of the parent. Early detection of these lesions is essential because although in some cases they may remain asymptomatic, in others they may cause bleeding, anemia, or, in a minority of cases, intussusception [11]. Although vascular lesions of the GI tract are an important medical problem, the definition and description of these lesions have not been widely accepted, and it is, in fact, scarce and confusing [12]. To alleviate this problem and to standardize WCE reports, a nomenclature and semantic description of small bowel capsule endoscopy vascular lesions was established within the consensus of 18 European WCE experts [12] as can be seen in Table 9.1. Angiectasia or angiodysplasia (AD) is an abnormal, tortuous, and dilated small blood vessel in the mucosal and submucosal layers of the GI tract and it appears as a clearly demarcated, bright-red, flat lesion (see Fig. 9.1). It is the most common vascular lesion of the GI tract [13] and is also reported as the most common cause of both overt and occult GI bleeding (OGIB) [14 17]. Table 9.1 Nomenclature and description of different vascular lesions proposed in Leenhardt et al. [12]. Nomenclature

Semantic description

Angiectasia/angiodysplasia

A clearly demarcated, bright-red, flat lesion, consisting of tortuous and clustered capillary dilatations, within the mucosal layer (surrounded by intestinal villi)

Can be small (a few mm) to large (a few cm) Erythematous patch

Red spot/dot

Phlebectasia Diminutive angiectasia

A small (a few mm) and flat reddish area, without any vessel appearance, within the mucosal layer (surrounded by intestinal villi) A minuscule (less than 1 mm), punctate, and flat lesion with a bright-red area, without linear or vessel appearance, within the mucosal layer (surrounded by intestinal villi) A small (few mm), flat to slightly elevated, bluish venous dilatation running below the mucosa (covered by intestinal villi) A clearly demarcated, linear, bright-red lesion, consisting of tiny nonclustered capillary dilatations, within the mucosal layer (surrounded by intestinal villi)

Datasets

FIGURE 9.1 Four typical images with angiectasia/angiodysplasia. Image obtained from Leenhardt R, et al. Nomenclature and semantic description of vascular lesions in small bowel capsule endoscopy: an international Delphi consensus statement. Endosc Int Open 2019;7.03:E372 9.

In 2010 Liao et al. [18] performed a large systematic review of all original articles relevant to WCE published between 2000 and 2008. From a total of 227 studies with 22,800 procedures, OGIB was the most common indication (66%) and AD the most common reason (50%) for OGIB on those patients. In another study with 5744 procedures, AD lesions from the small bowel were the most common indication (35%) of severe OGIB [19].

Datasets The recent publications of different open datasets on the detection of vascular lesions have prompted interest among different AI researchers. In the field of AI,

151

152

CHAPTER 9 Artificial intelligence for vascular lesions

and especially when using deep learning models, public datasets are key. Not only do they help to train better and more robust models, but they also serve to compare different methods from different research groups. Thus they are essential to evaluate the impact of published solutions. For a dataset to be good for use by any researcher, it must meet certain minimum requirements. It has to be complex, diverse and nonbiased, i.e. it has to cover all types of patients in terms of age, gender, race. In this way, the results derived from the study will be applicable to real clinical practice. In this section we aim to introduce the datasets that have been made publicly available.

KIDs dataset This is the first database of WCE images that was made publicly available [20]. Unfortunately, it is no longer accessible. It consisted of two sets of data. The first
one, known as KID (καψoυλα Interactive Database; based on Greek for “capsule”) Dataset 1, contained a total of 77 WCE images obtained from MiroCam (IntroMedic Co, Seoul, Korea). These images presented different types of abnormalities, including angioectasias, apthae, chylous cysts, polypoid lesions, villous edema, bleeding, lymphangiectasias, ulcers and stenoses. Each released image was accompanied with a manually segmented mask created by an experienced reader. A larger version of the same dataset, KID Dataset 2, was released in 2017 with 2371 WCE images.

Red lesion endoscopy dataset This dataset was published by Coelho et al. [21] in 2018. It consists of two different sets of data. On the one hand, Set 1 was created to have a high diversity of images. It contains 3295 frames of which 1131 are associated with red lesions, such as angioectasias, ADs and bleeding. Images were taken using different cameras (MiroCam, PillCam SB1, SB2 and SB3). On the other hand, Set 2 is a sequence of 600 images from a unique PillCam SB3 video. In both sets, manually annotated masks are provided showing the exact location of the lesion. One of the main drawbacks of this dataset is that it contains a large number of images with lesions from the same patient (see Figs. 9.2 and 9.3) which can be difficult to use to assess the generalization of the models.

CAD CAP 2020 Computer-Assisted Diagnosis for CAPsule Endoscopy (CAD-CAP) is a multicenter dataset approved by the French Data Protection Authority and recorded in 13 French centers [22]. Images were collected from 4174 deidentified, thirdgeneration, small bowel WCE videos (PillCam SB3). This dataset contains 5124 frames with abnormal findings, and 20,000 normal frames. A subset of this

Datasets

FIGURE 9.2 Random set of images from Red Lesion Dataset.

FIGURE 9.3 ATen images of the same patient from the Red Lesion Dataset.

dataset has been used in several competitions in the Medical Image Computing and Computer Assisted Intervention (MICCAI) conference:

GIANA—MICCAI 2017 This dataset was presented and used in the MICCAI 2017 Endoscopic Vision Challenge: Sub-Challenge Gastrointestinal Image Analysis (GIANA). The goal of the competition was to detect and localize AD lesions in WCE frames. The dataset contains a collection of 1200 images with resolution 576 3 576 pixels. These images are split into 600 images with pathology and 600 images without it. Half of them are defined for training (labels are provided) and the other half are used for testing. .

GIANA—MICCAI 2018 This new version of the dataset was presented and used in the MICCAI 2018 Endoscopic Vision Challenge: Sub-Challenge Gastrointestinal Image Analysis (GIANA). The dataset keeps the same lesions of the version from 2017 but adds a new class, inflammatory lesions. It contains 1800 images to train the model, 600 per each class, and 1800 for testing (no labels are provided).

153

154

CHAPTER 9 Artificial intelligence for vascular lesions

Table 9.2 Overview of existing capsule endoscopy datasets for vascular lesions. Dataset year

Findings

Size

Availability

Angiectasia Bleeding Inflammations Angiectasia Bleeding Inflammations 2018

77 images

Open academic

2371 images 47 videos

Open academic

Angiectasia bleeding

2164 normal 1131 abnormal

Open academic

2017

Angiectasia

By request

GIANA 2018

2018

Angiectasia inflammatory

CAD CAP [22]

2020

Fresh blood vascular lesions inflammatory ulcerative lesions

600 Normal images 600 angiectasia images 600 normal images 600 angiectasia images 600 inflammatory images 20,000 normal 718 fresh blood 3097 vascular lesions 25,124 images

KID dataset 1 [24] 2014 KID dataset 2 [24] 2017 Red Lesions [21] GIANA 2017

Angiectasia Blood Erosion Kvasir Capsule [23] 2021 Erythema Ulcers Polyp

4,741,504 images

By request

By request

Open academic

Kvasir Capsule The Kvasir Capsule [23] is the most recent publicly available and the largest and most challenging dataset. It was collected from January 2016 to January 2018 at the Vestre Viken—Bærum Hospital in Norway using the Olympus EC-S10 endocapsule. It consists of 117 videos which can be used to extract a total of 4,741,504 frames. From these frames, only 44,228 of them are labeled in one of the 14 different classes, some of them related with vascular lesions such as angiectasia (866 frames), fresh blood (446 frames) and small black stripes, that is, blood hematin (12 frames). The authors also present two splits of the annotated frames and report some metrics in each of them for researchers to compare their solutions. This is one of the first attempts to standardize the results of WCE classification models (Table 9.2).

Artificial intelligence methods for vascular lesions In this section, we present some of the most important methods that have been developed to detect and characterize vascular lesions using AI in the last few

Artificial intelligence methods for vascular lesions

Table 9.3 Some of the most important methods presented to classify, detect and segment vascular lesion using Artificial Intelligence. Authors

Year

C/S/D

Type

Datasets

Iakovidis et al. [20] Vieira et al. [25] Noya et al. [26] Coelho et al. [21] Ghosh et al. [30] Shvets et al. [27] Vieira et al. [32] Leenhardt et al. [33] Vallée et al. [34] Guo et al. [35] Tsuboi et al. [36] Gobpradit et al. [38] Smedsrud et al. [23] Guo et al. [41] Costa et al. [10] Jain et al. [44] Ribeiro et al. [45]

2014 2016 2017 2018 2018 2018 2019 2019 2019 2019 2020 2020 2020 2020 2021 2021 2021

C C1S C1S C1S S S1S C1S C C D D S D D D C1S C

ML ML ML DL DL DL ML DL DL DL DL DL DL DL ML DL DL

KIDs 1 Private KIDs 1 Private Private Red Lesions KIDs Private (1200 image) KIDs 1 Private Private (200 CE studies) GIANA 2017 and GIANA 2018 GIANA 2018, KID, Private (polyps) Private GIANA Kvasir Capsule GIANA 2018 54 Full videos Private Private

C, Classification; D, detection; S, segmentation; ML, machine learning; DL, deep learning.

years (Table 9.3). Since the presence of bleeding is mainly caused by vascular lesions, the detection of this event was one of the first tasks in many works from the literature and, later on, some works began to also detect specific lesions. In this chapter, we will cover the research that is directly focused on the detection or characterization of vascular lesions and not on solely bleeding. To our knowledge, the first attempt to automatically detect vascular lesions from WCE through AI was presented by Iakovidis et al. [20] in November 2014. The contribution of this paper was twofold, on the one hand, a simple method to automatically classify images into several lesions, such as AD or lymphangiectasia, and on the other hand, the release of the first public dataset for WCE endoscopy, known in the community as KIDs database, with the aim to aid and educate clinicians, and also for the development of computer-aided support systems. The proposed method was based on a classical computer vision approach where a set of salient points were detected and characterized using speeded-up robust features descriptor and then classified with a support vector machine classifier. In August 2016, Vieira et al. [25] presented a new method aiming to segment AD lesions using the expectation maximization (EM) algorithm with Markov random fields. Although the goal of this work was quite similar to the previous one [20], no comparison was made between them. In this case, the approach was

155

156

CHAPTER 9 Artificial intelligence for vascular lesions

presented as a segmentation method and evaluated using Dice’s coefficient over the subset of 27 AD images from the KID database. Some months later, in July 2017, a new approach was proposed by Noya et al. [26]. This new attempt to detect AD lesions, followed a similar procedure as the one in [20] but with a private dataset of data with 799 pathological frames and 849 nonpathological frames from 36 different patients. From these images, 46,693 regions were extracted. Potential regions of interest were firstly detected and characterized using hand-crafted features such as color and texture but also some statistical and morphological features. After that, they were classified using a boosted decision tree, a machine learning algorithm. The method obtained very good results but its impact was limited, probably due to the use of a private database that could not be analyzed neither compared by other researchers. There was a clear need for high-quality common datasets that allowed the scientific community to perform comparative benchmarking and validation of the proposed algorithms. To this end, in May 2017, the Gastrointestinal Image ANAalysis (GIANA) Challenge 2017 was presented as part of the MICCAI 2017 conference. The challenge was divided in two tasks: first, to detect and second, to localize, AD lesions in WCE. For this purpose, 600 normal and 600 AD images were released with their corresponding masks. Seven teams participated in the competition and, for first time, several deep learning methods were proposed to detect and localize vascular lesions in WCE. All participants obtained very high results in both tasks in terms of sensitivity and specificity and the winning solution was achieved by Shvets et al. [27] team. Their solution was based on a deep learning algorithm but, even so, there was no clear difference, in terms of performance between deep learning solutions and the classical computer vision ones. The winning solution from Shvets et al. was based on the AlbuNet-34 architecture [28] (see Fig. 9.4). In their paper, they also compare the U-Net network architecture with its improved modifications (TernausNet-11, TernausNet-16, and AlbuNet-34) and showed that AlbuNet-34 was the best one for this task. AlbuNet [28], is a modification of U-Net, which involved pretrained ResNet-34 as encoders. The conclusions drawn from this first challenge were twofold. Firstly, as we have already mentioned, the proposals based on deep learning did not represent a substantial improvement over solutions based on traditional computer vision algorithms. Secondly, and perhaps more relevant, the dataset that was released was not sufficiently challenging and thus, larger more complex datasets were needed if further improvements were to be made in this field. In June 2018, Coelho et al. [21] presented a new database commonly known as the Red Lesion Endoscopy Dataset and proposed a deep learning model based on the U-Net architecture [29] to detect and segment red lesions. Although in the paper they state that the dataset contains a diversity of lesions, most of them are related to active bleeding. The method they presented, achieved 96.83% of accuracy when trained on Set 1 and evaluated on Set 2, showing a high level of generalization.

Artificial intelligence methods for vascular lesions

FIGURE 9.4 AlbuNet architecture. Image from Shvets AA, et al. Automatic instrument segmentation in robot-assisted surgery using deep learning. In: 2018 17th IEEE international conference on machine learning and applications (ICMLA). IEEE; 2018.

Few months later, in October 2018, Ghosh et al. [30] presented a convolutional neural network (CNN) architecture which uses SegNet [31] layers. SegNet is a pixel-wise segmentation method, which works on a deep convolutional encoder decoder architecture. It shares several characteristics with U-Net but it is more time and memory efficient. Authors achieved good results when evaluating on a few KID dataset images. Further improvements were made over the following years. In 2019, three relevant works were presented. Vieira et al. [32] presented a classical computer vision model for detecting AD automatically. The authors justified the use of classical methods instead of a deep learning approach because of the limited amount of data, as well as the computational resources available. The proposed approach was based on the maximum a posteriori, using the EM algorithm. The authors stated that the proposed method outperformed other current state-of-theart algorithms, achieving sensitivity and specificity values of more than 96% on a private database with 800 WCE frames. Leenhardt et al. [33] and Valle´e et al. [34] presented both deep learning based solutions. In Leenhardt et al., they used a classical CNN method to detect small bowel AD which was trained and validated with a subset of images from the CAD CAP dataset. The study yielded excellent results, with a sensitivity of 100%, specificity of 96% and a positive predictive value of 96%, and although several details about the architecture are missing in the paper, it deserves to be considered as highly relevant taking into account the large number of CE studies used for the experiment. Valle´e et al. present a recurrent attention neural

157

158

CHAPTER 9 Artificial intelligence for vascular lesions

network architecture trained on a private dataset with 3128 WCE images collected in Hospital of Nantes. They also present results on the GIANA 2017, GIANA 2018 and CROHN-IPI datasets for comparison. Its model use a combination of three different losses to be trained, achieving a nearly perfect score in the GIANA 2018 dataset. In MICCAI 2019, Guo et al. [35] proposed the method Triple ANet. It reduces the high intraclass variability while increasing the small interclass variance. To achieve these distributions, the system uses deformable convolutions to capture the informative areas of the images. A key point of their system is the use of angular contractive loss to optimize the network. The model was tested over the GIANA 2018, KID and a private collection of 585 polyp images. The model obtains an overall accuracy of 89.41 while reaching a score of 95.26 and 91.73 in vascular lesion and inflammatory detection. Two outstanding papers were published in March 2020. On one hand, Tsuboi et al. [36] presented a CNN-based method, ShotMultiBox Detector [37], capable of detecting angiectasis in the small bowel with a sensitivity of 98.8% and specificity of 98.4%. The model was trained on a private dataset with 2237 angiectatic images from 141 different patients and it was validated using images from 48 different patients. On the other hand, Gobpradit et al. [38] presented a new model based on the AlbuNet architecture. This model obtained outstanding results on GIANA 2017 and 2018. In August 2020 Smedsrud et al. [23] presented the Kvasir Capsule dataset and a couple of benchmarks. The algorithms were based on two standard CNN architectures: DenseNet-161 [39] and ResNet-152 [40]. The models included two techniques to overcome the imbalance of the dataset. Simultaneously, Guo et al. [41] proposed a semisupervised learning method with adaptive attention on the GIANA 2018 dataset. Their approach utilizes the 1800 unlabeled test images to capture global dependencies and obtain a more robust abnormality classification. This attempt is the first work that proposes the use of this data as a solution to mitigate the lack of them. Already in 2021, Costa et al. [10] from the University of Minho presented a classical computer vision approach. They segmented the images into two different regions using the EM algorithm and, for each of the two regions, created handcrafted features. These features were classified using a multilayer perceptron. The algorithm is simple and, in this case, very effective. Moreover, to our knowledge, it is one of the first papers to evaluate its usefulness in the medical field, comparing it with traditional screening methods. Three medical experts reviewed 54 videos from start to finish, identifying 77.70% (115/148) of the angioectasias that appeared. This outperformed the review with Rapid Reader commercial software that provides tools for that task”: Suspected Blood Indicator” [42] and” Top100” [43]. In August 2021, Jain et al. [44] presented a solution to deal with multiple types of lesions from WCE videos. The solution firstly used an attention-based CNN architecture with 11 layers to classify images into four categories: polyp, vascular, inflammatory and normal. Then, if the image was classified as abnormal (polyp, vascular or inflammatory) it was next processed in a second phase using SegNet network together with GradCam11. They named this second phase as

Conclusions

anomaly detection. The authors showed that by combining the output of SegNet with the output of Grad-Cam11 the localization accuracy was improved. In fact, the proposed system achieved outstanding accuracy on the KIDs dataset. In November 2021, Ribeiro et al. [45] proposed a new solution based on the Xception network [46] to detect and discriminate vascular lesions that had the potential to lead to hemorrhage. The model was trained and validated with a large private dataset with 1129 CE studies from which 11,588 frames were extracted. 20% of the data was used as validation. It contained a total of 206 images with red spots, 207 images with angiectasia or varicose veins and 1905 images without any pathology. The authors reported sensitivities and specificities between 91% and 97% in the detection of these three classes which allowed 94.4% of the lesions to be correctly identified.

Conclusions The potential benefits of using AI for detecting small bowel vascular lesions are clear. Early detection can lead to earlier diagnosis and treatment for patients with GI bleeding, which can improve their prognosis significantly. Additionally, as this technology continues to be developed further, it may become more accurate at identifying different types of lesion morphology—potentially leading to even better patient outcomes down the road. In recent years, multiple solutions have been presented within the research community with excellent results. This interest, in part, it is thanks to the databases that have been released for scientific use. However, they have some inherent problems that make it difficult to use them and to validate the generalizability of the models. That is why we see that the authors of the models use different validation methodologies, and in many cases private databases. Despite the interesting findings from all these works, there exist some important drawbacks that need to be solved before we can see AI algorithms in the clinical use. The algorithms are being evaluated in still images rather than videos. In most of the public datasets the “clean” or “normal” images are just a subset, usually biased, from the reality of a WCE video images. In real videos, normal and pathological images are dirty, blurred, and partially occluded by content. New and more challenging datasets are needed to evaluate the proposed methods, but also a standard evaluation metrics need to be defined to allow easy and fair comparison between methods. This is a really unbalanced problem, and the used metrics need to reflect the need expected performance in the clinical practice. All this suggests that it is necessary and urgent, a unified validation strategy using all these public data. Unfortunately, we have to say that there is no method that we can consider the state of the art for the detection of vascular lesions, neither for common lesions such as AD. Several methods have been presented, showing very good results, but the lack of a high-quality public dataset difficult its fair comparison.

159

160

CHAPTER 9 Artificial intelligence for vascular lesions

References [1] Iddan G, et al. Wireless capsule endoscopy. Nature 2000;405 6785. [2] Rondonotti E, et al. How to read small bowel capsule endoscopy: a practical guide for everyday use. Endosc Inte Open 2020;8(10) E1220E1224. [3] Hwang Y, et al. Application of artificial intelligence in capsule endoscopy: where are we now? Clin Endosc 2018;51(6):547. [4] Tziortziotis I, Laskaratos F-M, Coda S. Role of artificial intelligence in video capsule endoscopy. Diagnostics 2021;11(7):1192. [5] Dray X, et al. Artificial intelligence in small bowel capsule endoscopy-current status, challenges and future promise. J Gastroenterol Hepatol 2021;36(1):12 19. [6] Robertson AR, et al. Artificial intelligence for the detection of polypsor cancer with colon capsule endoscopy. Ther Adv Gastrointestinal Endosc 2021;14 26317745211020277. [7] Fu Y, Zhang W, Mandal M, Meng MQ. Computer-aided bleeding detection in WCE video in IEEE J Biomed Health Inf 2014;18(2):636 42. [8] Yuan Y, Wang J, Li B, Meng MQH. Saliency based ulcer detection for wireless capsule endoscopy diagnosis in IEEE Trans Med Imaging 2015;34(10):2046 57. [9] Yuan Y, Li B, Meng MQH. Improved bag of feature for automatic polyp detection in wireless capsule endoscopy images in IEEE Trans Automat Sci Eng 2016;13 (2):529 35. [10] Costa D, et al. Clinical performance of new software to automatically detect angioectasias in small bowel capsule endoscopy. GE-Portuguese J Gastroenterol 2021;28 (2):87 96. [11] Gordon FH, Watkinson A, Hodgson H. Vascular malformations of the gastrointestinal tract. Best Pract Res Clin Gastroenterol 2001;15(1):41 58. [12] Leenhardt R, et al. Nomenclature and semantic description of vascularlesions in small bowel capsule endoscopy: an international Delphi consensus statement. Endosc Int Open 2019;7(03):E372 9. [13] Foutch P, Gregory DK, Rex, Lieberman DA. Prevalence and natural history of colonic angiodysplasia among healthy asymptomatic people. Am J Gastroenterol 1995;90:4 Springer Nature. [14] Foutch PG. Angiodysplasia of the gastrointestinal tract. Am J Gastroenterol 1993;88:6 Springer Nature. [15] Li F, Jonathan AL, Virender KS. Capsule endoscopyin the evaluation of obscure gastrointestinal bleeding: a comprehensive review. Gastroenterol Hepatol 2007;3(10):777. [16] Sakai E, et al. Factors predicting the presence of small bowel lesions in patients with obscure gastrointestinal bleeding. Dig Endosc 2013;25(4):412 20. [17] Regula J, Wronska E, Pachlewski J. Vascular lesions of the gastrointestinal tract. Best Pract Res Clin Gastroenterol 2008;22(2):313 28. [18] Liao Z, et al. Indications and detection, completion, and retention rates ofsmall-bowel capsule endoscopy: a systematic review. Gastrointest Endosc 2010;71(2):280 6. [19] Lecleire S, et al. Yield and impact of emergency capsule enteroscopy in severe obscure-overt gastrointestinal bleeding. Endoscopy 2012;44(04):337 42. [20] Iakovidis DK, Koulaouzidis A. Automatic lesion detection in capsule endoscopy based on color saliency: closer to an essential adjunct for reviewing software. Gastrointest Endosc 2014;80(5):877 83.

References

[21] Coelho P., et al. A deep learning approach for red lesions detection in video capsule endoscopies. In: International conference image analysis and recognition. Cham: Springer; 2018. [22] Leenhardt R, et al. CAD-CAP: a 25,000-image database serving the development of artificial intelligence for capsule endoscopy. Endosc Int Open 2020;8(03):E415 20. [23] Smedsrud PH, et al. Kvasir-Capsule, a video capsule endoscopy dataset. Sci Data 2021;8(1):1 10. [24] Koulaouzidis A, et al. KID Project: an internet-based digital video atlas of capsule endoscopy for research purposes. Endosc Int Open 2017;5(06):E477 83. [25] Vieira P.M., et al. Segmentation of angiodysplasia lesions in WCE images using a MAP approach with Markov Random Fields. In: 2016 38th annual international conference of the IEEE engineering in medicine and biology society (EMBC). IEEE; 2016. [26] Noya F., Alvarez-Gonza´lezM.A., BenitezR. Automated angiodysplasia detection from wireless capsule endoscopy. In: 2017 39th annual international conference of the IEEE engineering in medicine and biology society (EMBC). IEEE; 2017. [27] Shvets A.A., et al. Angiodysplasia detection and localization using deep convolutional neural networks. In: 2018 17th IEEE international conference on machine learning and applications (ICMLA). IEEE; 2018. [28] Shvets A.A., et al. Automatic instrument segmentation in robot-assisted surgery using deep learning. In: 2018 17th IEEE international conference on machine learning and applications (ICMLA). IEEE; 2018. [29] Ronneberger O., FischerP., BroxT. U-net: convolutional networks for biomedical image segmentation. In: International conference on medical image computing and computer-assisted intervention. Cham: Springer; 2015. [30] Ghosh T., LiL., Chakareski J. Effective deep learning for semantic segmentation based bleeding zone detection in capsule endoscopy images. In: 2018 25th IEEE international conference on image processing (ICIP). IEEE; 2018. [31] Badrinarayanan V, Kendall A, Cipolla R. SegNet: a deep convolutional encoderdecoder architecture for image segmentation. IEEE Trans Pattern Anal Mach Intell 2017;39(12):2481 95. [32] Vieira PM, et al. Automatic segmentation and detection of small bowel angioectasias in WCE images. Ann Biomed Eng 2019;47(6) 14461462. [33] Leenhardt R, et al. A neural network algorithm for detection of GI angiectasia during small-bowel capsule endoscopy. Gastrointest Endosc 2019;89(1):189 94. [34] Valle´e R., et al. Accurate small bowel lesions detection in wireless capsule endoscopy images using deep recurrent attention neural network. In: 2019 IEEE 21st international workshop on multimedia signal processing (MMSP). IEEE; 2019. [35] Guo, X. et al. Triple ANet: adaptive abnormal-aware attention network for WCE image classification. In: Medical image computing and computer assisted intervention—MICCAI. 2019. p. 293 301 [36] Tsuboi A, et al. Artificial intelligence using a convolutional neural network for automatic detection of small-bowel angioectasia in capsule endoscopy images. Dig Endosc 2020;32(3):382 90. [37] Liu W., et al. SSD: single shot multibox detector. In: European conference on computer vision. Cham: Springer; 2016. [38] Gobpradit S., Vateekul, P. Angiodysplasia segmentation on capsule endoscopy images using AlbuNet with squeeze-and-excitation blocks. In: Asian conference on intelligent information and database systems. Cham: Springer; 2020.

161

162

CHAPTER 9 Artificial intelligence for vascular lesions

[39] Huang G., et al. Densely connected convolutional networks. In: Proc. of IEEE CVPR; 2017. p. 2261 9. [40] He K., et al. Deep residual learning for image recognition. In: Proc. of IEEE CVPR; 2016. p. 770 8. [41] Guo X, et al. Semi-supervised WCE image classification with adaptive aggregated attention. Med Image Anal 2020;64. [42] Han S, et al. Suspected blood indicator to identify active gastrointestinal bleeding: a prospective validation. Gastroenterol Res 2018;11(2):106 11. [43] Freitas M, et al. Simplify to improve in capsule endoscopy-TOP 100 is a swift and reliable evaluation tool for the small bowel inflammatory activity in Crohn’s disease. Scand J Gastroenterol 2020;55(4):408 13. [44] Jain S, et al. A deep CNN model for anomaly detection and localization in wireless capsule endoscopy images. Comput Biol Med 2021;137:104789. [45] Ribeiro T, et al. Artificial intelligence and capsule endoscopy: automatic detection of vascular lesions using a convolutional neural network. Ann Gastroenterol 2021;34 (6):820. [46] Chollet F. Xception: deep learning with depth wise separable convolutions. In: Proc. of the IEEE conference on computer vision and pattern recognition; 2017.

CHAPTER

10

Artificial intelligence for luminal content analysis and miscellaneous findings

Nuno Almeida1,2 and Pedro Figueiredo1,2 1

Gastroenterology Department, Coimbra Hospital and Universitary Center, Coimbra, Portugal 2 Faculty of Medicine, University of Coimbra, Coimbra, Portugal

Introduction Since its first presentation to the medical community in 2000, capsule endoscopy (CE) has become a major tool in gastroenterology [1]. This device revolutionized the approach to small bowel (SB) pathologies since it provides a reliable and noninvasive method for complete visualization and assessment of the mucosal surface [2]. Currently CE can also be used for other gastrointestinal (GI) segments, such as the upper GI tract and the colon, but has yet to prove its added value compared with the traditional endoscopic techniques [3]. CE is now a first-line investigation tool for mid-GI bleeding and also an important method for evaluation of Crohn’s disease, diagnosis of SB tumors, and surveillance of polyposis syndromes [2,4,5]. This device can also be useful in certain cases of celiac disease [6]. CE allows capture of images without the need of a cable connection in the endoscope platform. The images collected by the device are interpreted afterwards in a dedicated workstation. So, the diagnostic yield of CE relies on lesion detection and interpretation. Lesion detection is affected by the quality and the number of images taken per second. Lesion interpretation takes considerable time and requires focused attention from the gastroenterologist [4]. However, it is recognized that the human eye is imprecise and the human attention range has limitations [7]. So, CE accuracy is limited by multiple factors such as transit time, peristaltic activity, bowel distension, bowel preparation, shadows, and the angle or direction of the camera [8]. We must also emphasize that CE has some major drawbacks: first there is no control in the image acquisition process; second, it has no additional diagnostic or therapeutic capabilities; third, there is a lack of specificity of some findings; fourth, the length of time required for the procedure is long, and generally the results are available in the followings days and that is a limitation in patients with persistent bleeding; fifth, there are some contraindications, with bowel obstruction being one of them [9].

Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00013-7 © 2023 Elsevier Inc. All rights reserved.

163

164

CHAPTER 10 Artificial intelligence for luminal content analysis

With thousands of images to interpret and long reading times, the evaluation of CE examinations can be tedious and time-consuming [10]. Furthermore, SB lesions can be present in only one or two frames and there is a possibility to be overlooked during the manual reading by physicians [5,11]. So, computer-aided diagnosis (CAD) was frequently referred in older publications as one of the potential major achievements in the future for endoscopy [12,13]. The term artificial intelligence (AI) as “the science and engineering of making intelligent machines” was first used by John McCarthy in 1956 [14]. Applying AI in medicine was attractive, but only after 2000, with the continuous growth of computational power, availability of larger datasets, and the introduction of deep learning, particularly through convolutional neural networks (NNs) (CNNs) it was possible to experience major breakthroughs in this field [14]. CNN can recognize patterns from images and is also responsible for its image feature extraction and performs the entire process up to image classification [15]. Considering gastroenterology, the application of AI has greatly expanded over the last decade. Initial applications in the field of endoscopy included CAD for the detection, differentiation, and characterization of neoplastic and nonneoplastic colon polyps [16 19]. AI for CE reading, and interpretation would be, obviously, a major achievement. There are already publications in this field of endoscopy, including metaanalysis and systematic reviews [20,21]. Diagnostic accuracy of CNN for the detection of a range of diseases in CE is impressive, and there is a potential for a major reduction in reading times and consequently in cost performance [20]. Some authors defend that AI integration in CE reading may take a stepwise introduction, from quality improvement (Stage 1) to diagnostic capabilities with no or minimal human intervention (Stage 5) [22]. In this chapter, we analyze the publications about AI in CE concerning the luminal contents, including bowel preparation evaluation, and other miscellaneous findings.

Small bowel preparation and luminal content In CE, images are retrospectively reviewed. Residues, debris, bile, chyme, and bubbles can compromise mucosal visualization and there is no possibility to remove it, as happens with traditional endoscopy. Although a specific chapter in this book is dedicated to the question of preparation, it is important to mention the potential role of AI concerning luminal content. Although optimal patient preparation has been controversial, it is now formally recommended [23]. Cleansing levels must be stated in the CE report and this is considered a minor performance measure [24]. There are validated instruments for this purpose, such as the Park, Brotz, and Koda scales, that can be used in clinical practice [24,25]. However, the application of such scores is not simple, and its automatic calculation would represent a considerable help. Evaluating the

Small bowel preparation and luminal content

appropriateness of SB preparation in each segment is not an easy task, and human dependent scores are complex and clearly prone to subjectivity [26]. It is virtually impossible for a human to objectively determine the quality of bowel cleansing for thousands of images [27]. As Ponte et al. stated in 2016, computer grading scales are based on objective measurements and may potentially overcome the disadvantages of human dependent scoring systems [26]. So, one of major achievements with AI would be objective and precise grading of bowel cleansing not only for SB but also for colon capsule. Air bubbles and residual materials can hamper the reading of CE videos (Figs. 10.1 10.6). Detection and elimination of such artifacts would, by itself, reduce the examination reading time as well as the bias and interpretation errors [28]. However, if a significant number of images is automatically eliminated due to artifacts this will have a major impact on the quality and accuracy of examination. So, an automatic, validated, and easily available cleansing evaluation tool would be more important than the automatic elimination of artifacts. Nam et al. described a deep learning based automation software for calculating SB cleansing score in 2021 [29]. These authors extracted video segments without any significant SB lesion from 72 cases obtained with the PillCam SB3. The correspondent frames were separated from extracted video segments and classified into four categories: normal-clean mucosa, bubble-dominant mucosa, bile-dominant mucosa, and debris-dominant mucosa. Then, 2 experienced CE readers reviewed these frames and scored cleansing based on the proportion of visualized mucosa from 5 (more than 90%) to 1 (less than 25%). For each cleansing score (3500 in total), 700 images were used to train a deep learning network. The trained scoring software was validated using 96 CE cases different from

FIGURE 10.1 Excellent small bowel preparation.

165

166

CHAPTER 10 Artificial intelligence for luminal content analysis

FIGURE 10.2 Presence of some bubbles in the small bowel lumen.

FIGURE 10.3 Mucosa visualization hampered by a great amount of bubbles.

those used in the training set. Extracted frames were divided into 3 equal number of segments according to the time sequence of the video: proximal, middle, and distal third. A cleansing score was assigned to every frame of the validation set and the average cleansing score (from 1.0 to 5.0) per segment and per case was calculated as the sum of cleansing scores divided by the number of frames. These automatically calculated score was then compared with the clinical ones.

Small bowel preparation and luminal content

FIGURE 10.4 Chyme in the small bowel.

FIGURE 10.5 Inadequate preparation due to debris.

Clinically adequate preparation was achieved for 91.7% (88/96) of cases. The average cleansing score for the adequate group was significantly higher than that for the inadequate group (4.0 vs 2.9, P , .001). In receiver operating characteristic (ROC) curve, a cut-off value of cleansing score at 3.25 for clinically adequate preparation had a sensitivity of 93%, a specificity of 100%, and an AUC (area under the curve) of 0.977 (95% confidence interval (CI): 0.926 0.999, P , .001). Interestingly, the average cleansing score did not differ according to the overall

167

168

CHAPTER 10 Artificial intelligence for luminal content analysis

FIGURE 10.6 Blood in the small bowel.

diagnostic yield (4.0 vs 3.8, P 5 .197), although it was significantly higher when small lesions were detected (4.3 vs 3.8, P , .001). The same authors published a similar work using still-cut frames from video segments without major abnormalities, obtained with the Mirocam platform [27]. They used 400,000 frames from 100 CE examinations to train the algorithm that was then validated in an additional 50 CE cases, comparing the performance of the algorithm with the clinical scores, as happened in the study previously mentioned [29]. Adequate preparation was achieved in 62% (31/50) of cases in the validation set. The average cleansing score for the adequate preparation group was significantly higher than that for the inadequate preparation group (3.4 vs 2.5, P , .001). ROC curve analysis indicated a cleansing score cut-off value of 2.95 for clinically adequate preparation, with a sensitivity of 81%, specificity of 84%, and AUC of 0.913 (95% confidence interval, 0.835 0.990, P , .001) [27]. The sensitivity and specificity were lower than the obtained in the preliminary study but the last one, involving a considerably higher number of images, reflects better what would be real clinical cases. Both studies pave the way for introduction of such algorithms in CE reading platforms. If the system automatically determines that preparation was inadequate it would be probably necessary to repeat CE examination. This can also serve as a quality indicator. However, it is not known if there is a real impact on diagnostic yield in examinations classified as insufficient preparation by the algorithm [27]. Noorda et al. also developed a CNN model using 563 individual frames of 576 3 576 pixels, that was then validated in a clinical setting with a total of 854 additional frames [30]. Their model obtained a high accuracy of 95.23%. In the validation procedure, they found that the agreement between their method and each of the

Small bowel preparation and luminal content

specialists individually over all images approached the interhuman agreement, yet the values fell just outside of the interhuman 95% confidence interval [30]. Pietri et al. developed an automated algorithm to evaluate the abundance of bubbles in SB CE [31]. They selected 5 groups of 40 still frames each, divided by the content in bubbles (less than 2%; 2% 10%; 10% 25%; 25% 50%; more than 50%). Two independent sets of 200 still frames (for development and validation, respectively) were independently analyzed by two experienced readers that classified it as scarce in bubbles (,10%) or abundant in bubbles (more than 10%). The authors tested four different computed algorithms and the Gray Level Co-occurrence matrix (GLCM) was considered the best one in the validation step, with sensitivity of 95.79% (95% CI: 93.01% 98.57%), specificity of 95.19% (95% CI: 92.22% 98.16%), negative predictive value of 96.12% (95% CI: 93.44% 98.8%), positive predictive value of 94.79% (95% CI: 91.71% 97.87%) and a calculation time by frame of 0.037 6 0.005 s [31]. In the development step GLCM model had an AUC of 0.9852. Such an algorithm could be useful in future prospective studies of anti-bubble agents in SB CE preparation regimens. The interest in noninformative frames in CE is not new. Noninformative frames may be defined as having invisible tissues, folds and/or lumen in CE frames and represent approximately 10,000 still images in a video [32]. In such uninformative frames major parts of the mucosa are covered by turbid fluids, bubble patterns, and/or other substances (e.g., fecal materials and unabsorbed foods). These useless frames can be further classified as highly contaminated nonbubbled (HCN) frames or significantly bubbled frames [33]. The first work concerning intestinal content was published by Vilarin˜o et al. in 2006 [34]. The authors applied Gabor filters for the characterization of the bubble-like shape of intestinal juices in fasting patients. However, the effects of residual foods, and fecal materials remained unexplored [34]. This unsupervised learning method achieved a significant reduction in visualization time, with no relevant loss of valid frames [34]. Bashar et al. went further using a support vector machine (SVM) classifier to distinguish HCN from non-HCN frames [32,33]. The authors used two steps in their method: the first step was to isolate HCN frames, and the second step was to signal bubbled frames. They used local color moments and the hue saturation value (HSV) color histogram, which characterized HCN frames. Then, a nonlinear SVM was applied to classify the frames. In second step, a Gauss Laguerre transform (GLT) (based on texture feature) was used to isolate the bubble structures [32,33]. The SVM classifier used in stage-1 to isolate HCN frames had a high preserve ratio of informative frames ( . 98%). Even so, about one fifth of the entire video was classified as invalid frame, and the video reading time would be decreased by about 20% on average. The results from the article published in 2010 confirmed that the proposed GLT-based feature in combination with local color moment or the HSV color histogram had the best recall (86.42% or 84.45%), which was better than that of the Gabor-based (78.18% or 76.29%) and wavelet based (65.43% or 63.83%) features with the same combinations of color features [32].

169

170

CHAPTER 10 Artificial intelligence for luminal content analysis

Later, Sun et al. modified this method, using local quantized histograms of classic color local binary patterns (CLBP) for the detection of bubbles, while it also replaced the SVM classifiers by a linear k-nearest neighbor (KNN). In the first stage the gastric juice information frames were removed. In the second stage, features of impurity frames with bubbles are extracted by CLBP and discrete cosine transform algorithms. In both stages, the KNN classifier was used to distinguish them and the accuracies of each stage reached as high as 99.31% and 97.54%, respectively [35]. Other methods proposed the detection of intestinal content in a single stage. Khun et al. analyzed images in terms of color and texture, using SVM an NN [36]. For the detection of informative frames, color feature provided 94.10% accuracy with SVM classifier and 93.44% accuracy with NN classifier, but with texture feature these rates declined to 73.85% and 70.41%, respectively. The time required to analyze the frames was 0.2339 s for color and 1.79676 s for texture (9-time increase in computational time) [36]. Seguı´ et al. developed and validated an automatic system for intestinal content detection as well as a segmentation method for detection of bubbles and turbid content in CE images [37]. Their system allowed the correct detection of intestinal content (turbid 1 bubbles) obtaining an overlap area with three experienced experts higher than 80% (in segmentation). For the bubble segmentation method, the overlap exceeded 90% [37]. Only a few methods have attempted to directly classify annotated regions of intestinal content instead of whole images [30]. One of these methods is the work done by Maghsoudi et al. that tried to distinguish informative from uninformative regions in a frame [38]. They proposed a multi-stage approach, where the first stage was aimed at classifying entire images, while the second classifies extracted nonoverlapping regions of 32 3 32 pixels. In the first stage they used morphological feature extraction in combination with fuzzy k-means, which was fed to an NN classifier. The second stage segmented the image with parameters based on its classification results and then extracted the regions. After extracting statistical features from those regions, they classified them through a second NN, thus obtaining classified regions. Their accuracy results were higher than 91% [38]. The main application of all these works is the classification of bowel preparation, as already stated. Recently, a multi-criterion computer-assisted algorithm to determine whether mucosal visualization was adequate on SB CE still frames revealed a sensitivity of 90.0% and specificity of 87.7%, with optimal reproducibility compared to three human expert analysis [39]. Even so, the authors extrapolated that the scrutiny time of a full 50,000 images would take 28 min [39].

Lymphangiectasia and other miscellaneous findings Detection and interpretation of protruding lesions is an important problem in CE. Although lymphangiectasia and xanthomas are easily identifiable (Figs. 10.7 and 10.8)

Lymphangiectasia and other miscellaneous findings

FIGURE 10.7 Small bowel lymphangiectasia.

FIGURE 10.8 Small bowel xanthoma.

other subepithelial lesions are difficult to detect and to distinguish from innocent bulges. Girelli et al. developed a scoring system called SPICE (smooth, protruding lesions index in CE) do differentiate real subepithelial masses from innocent bulges, and this score was already validated by other groups [40,41]. AI systems for detection of protruding lesions are extensively reviewed in another chapter. However, it is interesting to notice that some miscellaneous

171

172

CHAPTER 10 Artificial intelligence for luminal content analysis

findings such as lymphangiectasia are relatively common and must be distinguished from more relevant findings. The possibility of automatic detection of lymphangiectasia, xanthoma and lymphoid hyperplasia was already studied by Haji-Maghsoudi et al. in 2012, using still images from PillCam videos [42]. The authors used a model consisting of HSV to omit the normal tissue on the red channel. Next, they applied a sigmoid transform to select light parts in a frame and, finally, they segmented diseases by using a Canny filter [42]. The accuracy for detection of lymphangiectasia, xanthoma and lymphoid hyperplasia was 94.41%, 97.11%, and 81.73%, respectively [42]. More recently, Ding et al. trained a CNN-base algorithm using at least 1000 representative images of 10 different categories, eight of it considered significant abnormal lesions (inflammation, ulcer, polyps, protruding lesions, vascular disease, bleeding, parasite, and diverticulum) and other two as normal variants (lymphangiectasia, lymphoid hyperplasia) [43]. In the validation phase 5000 recordings (113,268,334 images) were read by both conventional reading and CNN-based auxiliary reading by 20 experienced gastroenterologists. The sensitivity for detection of lymphangiectasia by conventional and CNN-based auxiliary reading was 51.35% and 100% (P , .0001) respectively. The specificity was 100% for both [43]. Concerning lymphatic follicular hyperplasia the sensitivity was 46.69% for conventional reading and 100% for the CNN-based model (P , .0001) [43]. In the per-lesion analysis the authors also considered the diverticulum but, contrarily to other lesions, there was no increase in sensitivity for CNN-based reading (100% vs 100%) [43].

Hookworms and foreign bodies Parasitic infections can be detected by CE (Figs. 10.9 and 10.10). Hookworms are intestinal parasites that are part of the soil-transmitted helminths, a group of nematode pathogens infecting millions of people worldwide, particularly in lowincome and middle-income countries [44]. Necator americanus and Ancylostoma spp. exist in all regions of the world, including in tropical and nontropical areas [44]. Unfortunately, automatic hookworm detection is a challenging task due to poor quality of images, presence of other luminal contents such as food, stool, bile, and bubbles (Fig. 10.9), complex structure of GI, and diverse appearances in shapes, color and texture of the parasites (Fig. 10.10) [45]. However, the parallel regional information of hookworms is very beneficial for the detection of these parasites [45]. The first manuscript about the automatic detection of hookworms in CE images was presented in 2013, using images obtained with the OMOM device. The authors used a new gradient space, named hybrid color gradient. They applied their model to

Hookworms and foreign bodies

FIGURE 10.9 Parasite in the lumen (differential diagnosis with solid debris).

FIGURE 10.10 Large number of parasites in the small bowel.

700 hookworm images and 1000 normal images obtained from 10 patients. In each time 560 positive samples and 800 negative samples were randomly selected as a training set and the remainder became part of the test set. The model proposed by the authors had a sensitivity of 84.5%, a specificity of 93% and a global accuracy of 88.7% [46]. However, the dataset used by the authors was relatively small and balanced.

173

174

CHAPTER 10 Artificial intelligence for luminal content analysis

In 2016 Wu, et al. published other manuscript about automatic detection of hookworms using, sequentially, a multi scale dual matched filter, a piecewise parallel detection method, a histogram of average intensity (HAI), and finally, Rusboost to classify CE images [45]. The authors used 440,000 images from 11 infected patients. In total 4828 of the 440,000 images had hookworms and two patients were heavily infected. The proposed method had a sensitivity of 77.2%, a specificity of 77.9% and a diagnostic accuracy of 78.2% [45]. Although a good performance has been achieved, 23% of hookworm’s images were not detected. The authors concluded that temporal and spatial relationship between consecutive images should be taken into consideration and that deep learning approaches should be applied to this field of investigation. Continuing the previous work, in 2018 He et al. developed a deep hookworm detection framework (DHDF) based on its’ visual appearances and tubular patterns [47]. This model integrated two CNN networks, that is, edge extraction network and hookworm classification network, to extract the edge features and classify the hookworms, respectively. They used the same dataset of 440,000 images and the same protocol of a training set with 10 patients and a test set with one. To evaluate the performance two experienced endoscopists examined and labeled all images as the ground truth. The images containing hookworms were labeled as positive. The DHDF achieved a sensitivity, specificity, and accuracy of 84.6%, 88.6%, and 88.5%, respectively. It outperformed the HAI 1 Rusboost by an average of 10% to 20% in terms of accuracy and specificity [47]. In 2021 Gan et al. used a deep detection NN called YOLO-V4 (a deep CNN with 53 or more layers) as the main part and a small classification NN (with 3 layers) as the supplementary part [48]. The system was trained with 11,236 images of hookworm annotated for this study by two expert endoscopists. Then a validation was performed using 10,000 normal images and 529 with hookworms. The trained CNN was able to read 26 images per second. The AUC of the CNN used for detecting hookworms was 0.972 (95% CI: 0.967 0.978). At a cut-off value of 0.485 the sensitivity, specificity and accuracy of the CNN were 92.2%, 91.1%, and 91.2%, respectively [48]. Curiously, the trained CNN detected 2 true hookworms that the expert missed. This model is very promising in terms of accuracy and time-efficacy. Notably, a CNN-based auxiliary reading, proposed by Ding et al., revealed no benefit of their model in the detection of parasites when compared to conventional reading [43]. In fact, as stated previously in this chapter, the model of CNN assisted reading had higher sensitivities for detection of all abnormalities, except parasites and diverticula with a considerable reduction in the mean reading time (96.6 6 22.53 vs 5.9 6 2.23 min) [43]. Foreign body detection by CE has been previously reported in medical literature [49]. The first series to document the role of CE in detecting foreign bodies in SB was presented in the 69th annual scientific meeting of the American College of Gastroenterology [50]. To the best of our knowledge there was no publication about AI use for detection of foreign bodies in this GI segment. Since it

Acknowledgments

is an infrequent finding, and foreign bodies can be so diverse in format, color, and texture, it would probably be necessary to create a large database of images with contributes from different endoscopists.

Discussion and conclusions CE is now the gold-standard for SB endoscopic evaluation. Technologic advances such as the suspected blood indicator, adaptative frame rate and the Quick-view algorithm improved the reading experience, accuracy and diagnostic capabilities of this device [7]. However, the real value of such tools is limited. AI is now gaining a considerable attention in CE, but most studies were performed with still images or small video segments. Real video scenarios present more challenges for AI reading assistance, and the perfect algorithm/machine must have the capabilities to determine which frames are inadequate for observation, what is the preparation quality in a global evaluation, and to detect awkward/rare findings such as parasites and foreign bodies. AI is gaining an extreme importance in many aspects of our life, including multiple healthcare domains. It has therefore become obvious to caregivers, academic researchers, medical, and information technology companies that CAD solutions are needed for medical imaging, for example, GI endoscopy and more specifically CE [20,51]. NN-based solutions would alleviate physicians from the tedious task of CE reading and reviewing. The main aim of such platforms is a marked reduction of reading times, while maintaining (or even increasing) diagnostic accuracy. In a near future only 1% of frames will be reviewed by the human operator [51]. Multiple algorithms for different analysis were already presented in medical literature. Unfortunately, most of it were developed using still frames and not video sequences. There is clearly a lack of prospective trials, using unaltered images, and with randomized comparison of AI versus standard expert reading. So, a dedicated device (or software incorporated in the capsule reading platform) that can analyze a real video in all its’ components, classifying the bowel preparation, identifying landmarks and main lesions, is anxiously awaited. Until the capsule become fully autonomous in the reading and interpretation process, we think that a deep learning based AI model for differentiating abnormal from normal images with no need for specific high-end graphics processing engine, must be incorporated in all CE workstations.

Acknowledgments The authors wish to thank Dr. Elisa Gravito-Soares, Dr. Mariana Sant’Anna, and Dr. Marta Gravito-Soares for their help with the figures provided in this chapter.

175

176

CHAPTER 10 Artificial intelligence for luminal content analysis

Disclosures/transparency declaration There are no conflicts of interest to disclose.

References [1] Iddan G., Meron G., Glukhovsky A., Swain P. Wireless capsule endoscopy. Nature 2000. https://doi.org/10.1038/35013140. [2] Pennazio M, Spada C, Eliakim R, Keuchel M, May A, Mulder CJ, et al. Small-bowel capsule endoscopy and device-assisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Clinical Guideline. Endoscopy 2015;. Available from: https://doi.org/10.1055/s-00341391855. [3] Chetcuti Zammit S, Sidhu R. Capsule endoscopy—recent developments and future directions. Expert Rev Gastroenterol Hepatol 2021;15:127 37. Available from: https://doi.org/10.1080/17474124.2021.1840351. [4] Kim SH, Yang DH, Kim JS. Current status of interpretation of small bowel capsule endoscopy. Clin Endosc 2018;51:329 33. Available from: https://doi.org/10.5946/ ce.2018.095. [5] Yang YJ. The future of capsule endoscopy: the role of artificial intelligence and other technical advancements. Clin Endosc 2020;53:387 94. Available from: https:// doi.org/10.5946/ce.2020.133. [6] Hosoe N, Takabayashi K, Ogata H, Kanai T. Capsule endoscopy for small-intestinal disorders: current status. Dig Endosc 2019;31:498 507. Available from: https://doi. org/10.1111/den.13346. [7] Byrne MF, Donnellan F. Artificial intelligence and capsule endoscopy: is the truly “smart” capsule nearly here? Gastrointest Endosc 2019;89:195 7. Available from: https://doi.org/10.1016/j.gie.2018.08.017. [8] Sey MSL, Yan BM. Optimal management of the patient presenting with small bowel bleeding. Best Pract Res Clin Gastroenterol 2019;. Available from: https://doi.org/ 10.1016/j.bpg.2019.04.004. [9] Gerson LB, Fidler JL, Cave DR, Leighton JA. ACG clinical guideline: diagnosis and management of small bowel bleeding. Am J Gastroenterol 2015;. Available from: https://doi.org/10.1038/ajg.2015.246. [10] Wang A, Banerjee S, Barth BA, Bhat YM, Chauhan S, Gottlieb KT, et al. Wireless capsule endoscopy. Gastrointest Endosc 2013;78:805 15. Available from: https:// doi.org/10.1016/j.gie.2013.06.026. [11] Zheng YP, Hawkins L, Wolff J, Goloubeva O, Goldberg E. Detection of lesions during capsule endoscopy: physician performance is disappointing. Am J Gastroenterol 2012;107:554 60. Available from: https://doi.org/10.1038/ajg.2011.461. [12] Sivak MV. Gastrointestinal endoscopy: past and future. Gut 2006;55:1061 4. Available from: https://doi.org/10.1136/gut.2005.086371. [13] Eliakim R. Video capsule colonoscopy: where will we be in 2015? Gastroenterology 2010;139:1468 71. Available from: https://doi.org/10.1053/j.gastro.2010.09.026 e1.

References

[14] Kaul V, Enslin S, Gross SA. History of artificial intelligence in medicine. Gastrointest Endosc 2020;92:807 12. Available from: https://doi.org/10.1016/j. gie.2020.06.040. [15] Kim SH, Lim YJ. Artificial intelligence in capsule endoscopy: a practical guide to its past and future challenges. Diagnostics 2021;11. Available from: https://doi.org/ 10.3390/diagnostics11091722. [16] Figueiredo P, Figueiredo I, Pinto L, Kumar S, Tsai Y-H, Mamonov A. Polyp detection with computer-aided diagnosis in white light colonoscopy: comparison of three different methods. Endosc Int Open 2019;07:E209 15. Available from: https://doi. org/10.1055/a-0808-4456. [17] Kudo Se, Mori Y, Misawa M, Takeda K, Kudo T, Itoh H, et al. Artificial intelligence and colonoscopy: current status and future perspectives. Dig Endosc 2019;31:363 71. Available from: https://doi.org/10.1111/den.13340. [18] Hoogenboom SA, Bagci U, Wallace MB. Artificial intelligence in gastroenterology. The current state of play and the potential. How will it affect our practice and when? Tech Innov Gastrointest Endosc 2020;22:42 7. Available from: https://doi.org/ 10.1016/j.tgie.2019.150634. [19] Wang P, Berzin TM, Glissen Brown JR, Bharadwaj S, Becq A, Xiao X, et al. Realtime automatic detection system increases colonoscopic polyp and adenoma detection rates: a prospective randomised controlled study. Gut 2019;68:1813 19. Available from: https://doi.org/10.1136/gutjnl-2018-317500. [20] Soffer S, Klang E, Shimon O, Nachmias N, Eliakim R, Ben-Horin S, et al. Deep learning for wireless capsule endoscopy: a systematic review and meta-analysis. Gastrointest Endosc 2020;92:831 9. Available from: https://doi.org/10.1016/j. gie.2020.04.039 e8. [21] Qin K, Li J, Fang Y, Xu Y, Wu J, Zhang H, et al. Convolution neural network for the diagnosis of wireless capsule endoscopy: a systematic review and meta-analysis. Surg Endosc 2022;36:16 31. Available from: https://doi.org/10.1007/s00464-02108689-3. [22] Robertson AR, Segui S, Wenzek H, Koulaouzidis A. Artificial intelligence for the detection of polyps or cancer with colon capsule endoscopy. Ther Adv Gastrointest Endosc 2021;14:1 8. Available from: https://doi.org/10.1177/26317745211020277. [23] Emanuele Rondonotti A, Spada C, Adler S, May A, Despott EJ, Koulaouzidis A, et al. Small-bowel capsule endoscopy and device-assisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Technical Review. Endoscopy 2018;50:423 46. Available from: https://doi.org/10.1055/a-0576-0566. [24] Spada C, McNamara D, Despott EJ, Adler S, Cash BD, Ferna´ndez-Urie´n I, et al. Performance measures for small-bowel endoscopy: a European Society of Gastrointestinal Endoscopy (ESGE) Quality Improvement Initiative. U Eur Gastroenterol J 2019;7:614 41. Available from: https://doi.org/10.1177/ 2050640619850365. [25] Alageeli M, Yan B, Alshankiti S, Al-Zahrani M, Bahreini Z, Dang TT, et al. KODA score: an updated and validated bowel preparation scale for patients undergoing small bowel capsule endoscopy. Endosc Int Open 2020;08:E1011 17. Available from: https://doi.org/10.1055/a-1176-9889.

177

178

CHAPTER 10 Artificial intelligence for luminal content analysis

[26] Ponte A, Pinho R, Rodrigues A, Carvalho J. Review of small-bowel cleansing scales in capsule endoscopy: a panoply of choices. World J Gastrointest Endosc 2016;8:600 9. Available from: https://doi.org/10.4253/wjge.v8.i17.600. [27] Nam JH, Oh DJ, Lee S, Song HJ, Lim YJ. Development and verification of a deep learning algorithm to evaluate small-bowel preparation quality. Diagnostics 2021;11. Available from: https://doi.org/10.3390/diagnostics11061127. [28] Mascarenhas M, Afonso J, Andrade P, Cardoso H, Macedo G. Artificial intelligence and capsule endoscopy: unravelling the future. Ann Gastroenterol 2021;34:300 9. Available from: https://doi.org/10.20524/aog.2021.0606. [29] Nam JH, Hwang Y, Oh DJ, Park J, Kim KB, Jung MK, et al. Development of a deep learning-based software for calculating cleansing score in small bowel capsule endoscopy. Sci Rep 2021;11:1 8. Available from: https://doi.org/10.1038/s41598-02181686-7. [30] Noorda R., Neva´rez A., Colomer A., Beltra´n V.P., Naranjo V. Automatic evaluation of degree of cleanliness in capsule endoscopy based on a novel CNN architecture. Sci Rep 2020. https://doi.org/10.1038/s41598-020-74668-8. [31] Pietri O, Rezgui G, Histace A, Camus M, Nion-Larmurier I, Li C, et al. Development and validation of an automated algorithm to evaluate the abundance of bubbles in small bowel capsule endoscopy. Endosc Int Open 2018;06:E462 9. Available from: https://doi.org/10.1055/a-0573-1044. [32] Bashar MK, Kitasaka T, Suenaga Y, Mekada Y, Mori K. Automatic detection of informative frames from wireless capsule endoscopy images. Med Image Anal 2010;14:449 70. Available from: https://doi.org/10.1016/j.media.2009.12.001. [33] Bashar MK, Mori K, Suenaga Y, Kitasaka T, Mekada Y. Detecting informative frames from wireless capsule endoscopic video using color and texture features. Lect Notes Comput Sci (Including Subser Lect Notes Artif Intell Lect Notes Bioinforma) 2008;5242:603 10. Available from: https://doi.org/10.1007/978-3-540-85990-1_72. LNCS. [34] Vilarin˜o F, Spyridonos P, Pujol O, Vitria` J, Radeva P, De Iorio F. Automatic detection of intestinal juices in wireless capsule video endoscopy. Proc Int Conf Pattern Recognit 2006;4:719 22. Available from: https://doi.org/10.1109/ICPR.2006.296. [35] Sun Z, Li B, Zhou R, Zheng H, Meng MQH. Removal of non-informative frames for wireless capsule endoscopy video segmentation. IEEE Int Conf Autom Logist ICAL 2012;294 9. Available from: https://doi.org/10.1109/ICAL.2012.6308214. [36] Khub P.C., Zhuo Z., Yang L.Z., Liyuan L., Jiang L. Feature selection and classification for wireless capsule endoscopic frames. In: Second international conference on biomedical and pharmaceutical engineering, ICBPE 2009 Conference Proc.; 2009. p. 0 5. https://doi.org/10.1109/ICBPE.2009.5384106. [37] Seguı´ S, Drozdzal M, Vilarin˜o F, Malagelada C, Azpiroz F, Radeva P, et al. Categorization and segmentation of intestinal content frames from wireless capsule endoscopy. IEEE Trans Inf Technol Biomed 2012;16:1341 52. Available from: https://doi.org/10.1109/TITB.2012.2221472. [38] Maghsoudi O, Talebpour A, Sotanian-Zadeh H, Alizadeh M, Soleimani H. Informative and uninformative regions detection in WCE frames. J Adv Comput 2014;. Available from: https://doi.org/10.7726/jac.2014.1002a. [39] Oumrani S, Histace A, Abou Ali E, Pietri O, Becq A, Houist G, et al. Multicriterion, automated, high-performance, rapid tool for assessing mucosal visualization

References

[40]

[41]

[42]

[43]

[44]

[45]

[46]

[47]

[48]

[49]

[50] [51]

quality of still images in small bowel capsule endoscopy. Endosc Int Open 2019;07: E944 8. Available from: https://doi.org/10.1055/a-0918-5883. Girelli CM, Porta P, Colombo E, Lesinigo E, Bernasconi G. Development of a novel index to discriminate bulge from mass on small-bowel capsule endoscopy. Gastrointest Endosc 2011;74:1067 74. Available from: https://doi.org/10.1016/j. gie.2011.07.022. Rodrigues JP, Pinho R, Rodrigues A, Silva J, Ponte A, Sousa M, et al. Validation of SPICE, a method to differenciate small bowel submucosal lesions from innocent bulges on capsule endoscopy. Rev Esp Enfermedades Dig 2017;109:106 13. Available from: https://doi.org/10.17235/REED.2017.4629/2016. Haji-Maghsoudi O., Talebpour A., Soltanian-Zadeh H., Haji-Maghsoodi N. Segmentation of Crohn, lymphangiectasia, xanthoma, lymphoid hyperplasia and stenosis diseases in WCE. Stud Health Technol Inform 2012;180:143 7. https://doi. org/10.3233/978-1-61499-101-4-143. Ding Z, Shi H, Zhang H, Meng L, Fan M, Han C, et al. Gastroenterologist-level identification of small-bowel diseases and normal variants by capsule endoscopy using a deep-learning model. Gastroenterology 2019;157:1044 54. Available from: https://doi.org/10.1053/j.gastro.2019.06.025 e5. Clements ACA, Addis Alene K. Global distribution of human hookworm species and differences in their morbidity effects: a systematic review. Lancet Microbe 2022;3: e72 9. Available from: https://doi.org/10.1016/S2666-5247(21)00181-6. Wu X, Chen H, Gan T, Chen J, Ngo CW, Peng Q. Automatic hookworm detection in wireless capsule endoscopy images. IEEE Trans Med Imaging 2016;35:1741 52. Available from: https://doi.org/10.1109/TMI.2016.2527736. Chen H., Chen J., Peng Q., Sun G., Gan T. Automatic hookworm image detection for wireless capsule endoscopy using hybrid color gradient and contourlet transform. In: Proc 2013 sixth international conference on biomedical engineering and informatics, BMEI; 2013. p. 116 20. https://doi.org/10.1109/BMEI.2013.6746918. He JY, Wu X, Jiang YG, Peng Q, Jain R. Hookworm detection in wireless capsule endoscopy images with deep learning. IEEE Trans Image Process 2018;27:2379 92. Available from: https://doi.org/10.1109/TIP.2018.2801119. Gan T, Yang Y, Liu S, Zeng B, Yang J, Deng K, et al. Automatic detection of small intestinal hookworms in capsule endoscopy images based on a convolutional neural network. Gastroenterol Res Pract 2021;2021. Available from: https://doi.org/10.1155/ 2021/5682288. Halloran BP, Stam FJ, Van Weyenberg SJB. A pill for cholesterol and a capsule for bleeding. Dig Liver Dis 2012;44:8658. Available from: https://doi.org/10.1016/j. dld.2012.03.004. Ali T, Sachdev R, Cave D. Foreign bodies in the small bowel detected by capsule endoscopy. Am J Gastroenterol 2004;99:S61 2. Dray X, Iakovidis D, Houdeville C, Jover R, Diamantis D, Histace A, et al. Artificial intelligence in small bowel capsule endoscopy current status, challenges and future promise. J Gastroenterol Hepatol 2021;36:12 19. Available from: https://doi.org/ 10.1111/jgh.15341.

179

This page intentionally left blank

CHAPTER

Small bowel and colon cleansing in capsule endoscopy

11

Vı´tor Macedo Silva1,2,3, Bruno Rosa1,2,3, Francisco Mendes4, Miguel Mascarenhas4,5, Miguel Mascarenhas Saraiva6 and Jose´ Cotter1,2,3 1

Gastroenterology Department, Hospital da Senhora da Oliveira, Guimara˜es, Portugal Life and Health Sciences Research Institute (ICVS), School of Medicine, University of Minho, Braga, Portugal 3 ICVS/3B’s, PT Government Associate Laboratory, Braga/Guimara˜es, Portugal 4 Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal 5 Faculty of Medicine, University of Porto, Porto, Portugal 6 Manoph Gastroenterology Clinic, Porto, Portugal

2

Introduction With the ever-growing application of capsule endoscopy (CE) in everyday clinical practice, significant progress has been made in several fields, including software improvement, hardware features, and expansion of clinical indications, with innovative applications of CE still emerging [1]. Nevertheless, from the technical aspect of CE, optimal preprocedural preparation and evaluation of final cleansing quality are still some of the most controversial areas in the field. In traditional operator-dependent endoscopic techniques, the quality of visualization is highly dependent on the absence of air bubbles, bile, and intestinal debris [2]. In this chapter, we aim to construct a comprehensive review of the currently available scientific evidence concerning small bowel and colon CE (CCE) cleansing, spanning from preparation regimens to cleansing quality evaluation, to guide procedure protocols and eventually provide the tool for the optimization of CE quality parameters, consequently improving the diagnostic yield of the examination.

Small bowel capsule endoscopy preparation Optimal preparation for small bowel CE (SBCE) is a highly debated topic among CE users worldwide. Several regimens have been proposed, with a wide range of approaches being applied, such as diet modifications, laxative regimens, and purgative protocols [3]. However, there is no standardized protocol for optimal SBCE preparation. Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00015-0 © 2023 Elsevier Inc. All rights reserved.

181

182

CHAPTER 11 Small bowel and colon cleansing in capsule endoscopy

Diet and fasting Historically, Given Imaging (Yokneam, Israel), the first manufacturer of capsule endoscopes, did not endorse preprocedure purgative regimens for SBCE, with the only recommendation being a liquid diet on the day before the procedure with clear liquids only in the evening and 12-h fasting until CE ingestion [4]. This classical protocol was the main applied preparation regimen in the primordial controlled studies of the application of CE in obscure gastrointestinal bleeding and suspected small bowel Crohn’s disease. From that time onward, in the absence of studies evaluating the effect of the timing and type of food ingestion on capsule view, experts support adherence to this regimen to this day. More recently, Chen et al. [5] achieved a higher rate of small bowel cleansing with 3-day fasting combined with oral senna compared with the only-oral purgative approach, suggesting a possible benefit of longer fasting. Nevertheless, regardless of the use of additional modifiers of small bowel cleansing, the indication for a clear liquid diet on the day before SBCE, followed by 12-h fasting, has been consistently applied in most CE performing units, as recommended in international clinical guidelines [4].

Oral purgatives For a long time, several authors have assessed the effect of oral purgative solutions in small bowel cleansing and subsequent CE diagnostic yield, not only regarding the use of these solutions themselves but also regarding optimal compositions, quantity ingested, and timing of ingestion [6]. In 2007 after reviewing the value of bowel lavage prior to SBCE, the first ever expert consensus paper on SBCE concluded that “the current evidence mainly from fully published papers suggests that polyethylene-glycol (PEG) solution lavage and simethicone both positively affect mucosal visibility and perhaps also diagnostic yield” [7]. Ever since, multiple randomized controlled trials (RCT) and metaanalyses have addressed whether oral purgatives improve mucosal visibility, diagnostic yield, and completion rates, with most of them concluding that the ingestion of 2 L of PEG solution previously to CE ingestion leads to improved visibility of the small bowel mucosa [8, 9]. Recently, a case-control study by Mascarenhas-Saraiva et al. [10] compared a control protocol consisting only of suspension of iron supplementation for three days before the procedure and a diet consisting of light meals and clear liquids with 10-h fasting prior to the CE examination with an intervention group with the same dietary alterations plus a 2 L PEG-ascorbic acid consumption. The addition of the PEG-ascorbic acid preparation increased both mucosal visualization and lumen content scores, enhancing the performance of the small bowel study. On the other side, the use of oral purgatives can be associated with adverse effects (nausea, abdominal distension, or discomfort), which may impact the protocol adhesion, with a not-so-obvious benefit. A prospective RCT by Pons Beltran et al. [11] compared three different protocols before CE—one consisting of a 4 L

Small bowel capsule endoscopy preparation

clear liquid diet, another consisting of 90 mL of aqueous sodium phosphate, and a third based on 4 L of PEG solution. In fact, despite no differences in the degree of cleanliness between groups, the PEG group was associated with lower compliance by patients. Another RCT by Hansel et al. [12] compared a combined bowel preparation (consisting of ingestion of 2 L of PEG solution in the night prior to the examination, 5 mL simethicone, and 5 mL metoclopramide ingested 20 min prior to CE, and lying 30 min on the right side after swallowing the CE) with a control group with only dietary restrictions. There was no significant difference between the groups in terms of small-bowel visualization, diagnostic yield, or small bowel transit time. However, the reported discomfort was significantly different between the groups (62% vs 17%; P 5 .01) [12]. Thus it is important to measure the benefits and adverse outcomes of more rigid or voluminous preparation regimens to have good protocol compliance. Recently, other agents have been studied for SBCE preparation. An investigation was carried out about bowel secretor use in patients. A small study compared the use of single-dose linaclotide 1 h before CE in a group with the standard PEG preparation in another group. Both groups had similar rates of good/excellent or ideal preparation, small bowel time, and diagnostic yield, with no serious side effects, suggesting linaclotide as a substitute for PEG regimens [13]. Lubiprostone is another bowel secretor that acts through the activation of specific intestinal chloride channels, producing a chloride-rich secretion. Matsuura et al. [14] investigated the benefit of a lubiprostone tablet consumption 2 h before CE examination. Lubiprostone administration improved small bowel visualization while augmenting fluid secretion, with a nonsignificant reduction in small bowel transit time. Despite the clear evidence of the beneficial effects of oral purgatives in SBCE visualization, a consensus is yet to be reached concerning the optimal timing for their ingestion to achieve the best cleansing. In 2019 Xavier et al. [15] conducted a prospective randomized study comparing three distinct protocols differing in the time of PEG solution for SBCE cleansing: (1) Protocol A: Clear liquid diet the day before the examination with fasting from 8 p.m.; (2) Protocol B: Protocol A 1 PEG solution from 8 p.m. of the day before the examination; (3) Protocol C: Protocol A 1 PEG solution consumed after real-time confirmation of capsule arrival at small bowel using the real-time viewer function of the data recorder. The protocol with PEG ingestion after confirmation of CE in the small bowel outperformed the others, with an excellent/good small bowel preparation in a higher percentage of examinations (Protocol C: . 75% vs Protocol A: 38% vs Protocol B: 45%; P 5 .003). Additionally, this protocol had a higher detection of angioectasias (A: 5.4% vs B: 9.7% vs C: 27.3%, P 5 .022). In 2021, Mascarenhas-Saraiva et al. [16] addressed the same problem by evaluating three types of protocols. (1) Group 1 followed a clear liquid diet and fasting-based bowel preparation only; (2) Group 2 followed the same procedure with additional ingestion of 1 L of PEG/ascorbic acid booster when there was evidence of the capsule reaching the small intestine; and (3) Group 3 ingested only 0.5 L PEG/ascorbic acid booster after the capsule reached the small intestine.

183

184

CHAPTER 11 Small bowel and colon cleansing in capsule endoscopy

There were significant differences between Group 3 and Group 1 when addressing the cleansing score (with advantages in both visualization and obscuration subscores of the Park cleansing score) but nonsignificant differences between Group 2 and Group 3. Therefore the administration of a low dose of PEG-based booster could be an alternative for a larger volume of PEG solution in CE preparation, with higher tolerability and compliance. This could also be an advantage in patients with cardiac or renal diseases who cannot tolerate large quantities of ingestion of oral purgatives. In conclusion, to this day, the ingestion of oral purgatives for SBCE has an important value for allowing optimal mucosal visualization, with the timing of ingestion leading to the best outcomes when CE has been confirmed to have reached the small bowel.

Prokinetic drugs The usefulness of prokinetic drugs in SBCE is still a matter of debate. A metaanalysis of four randomized trials evaluating the impact of prokinetics in SBCE concluded that the nonselective usage alone was ineffective in increasing completion rates [17]. A systematic review and metaanalysis by Kotwal et al. [18] consisting of four RCTs using different prokinetics before CE (metoclopramide, mosapride, erythromycin) showed no significative difference in terms of completion rate. Therefore the routinary use of prokinetics in SBCE is not associated with better visualization outcomes. On the other hand, patients with increased risk of an incomplete SBCE study (for example, old age, delayed gastric emptying, diabetic neuropathy, severe hypothyroidism, and use of psychotropic drugs, among others) and overall those with CE remaining in the stomach for more than 1 h (prolonged gastric transit time) as confirmed by real-time viewing, may benefit from the administration of prokinetic agents such as domperidone, to reduce incomplete examinations, without influencing small bowel transit time or diagnostic yield [19].

Antifoaming agents The clear visualization of the mucosa depends on the absence of air bubbles and foam. Multiple RCTs have demonstrated that the use of antifoaming agents, such as simethicone, significantly improves the quality of small bowel mucosa visualization in SBCE by decreasing the presence of bubbles and foam [20]. Simethicone adds an extra value to oral purgative solutions. In a study by Rosa et al. [21], administration of 100 mg of simethicone 30 min prior to capsule ingestion, even after consumption of 2 L of PEG in the evening before the procedure, was associated with a higher proportion of patients achieving an excellent level of cleansing. Another study of 115 patients submitted to SBCE evaluated simethicone ingestion associated with a PEG preparation compared with a PEG

Colon capsule endoscopy preparation

protocol alone. Simethicone ingestion resulted in a significative higher quality of visualization [22]. Optimal doses of simethicone are yet to be determined. However, the available scientific evidence reports a range between 80 and 200 mg to be effective [23]. In an RCT by Sey et al. [24], a high dose of simethicone (1125 mg simethicone in 750 mL water) versus a low dose of simethicone (300 mg in 200 mL) showed no additional benefits in terms of adequate cleansing, diagnostic yield, or completion rate.

Water ingestion Investigations are being carried out about the role of water ingestion during SBCE examinations. Zeng et al. [25] studied the benefit of hourly 500 mL water ingestion starting when the capsule reaches the small intestine calling it the diving method. Water ingestion may facilitate image cleansing and wireless CE travel through the small intestine. Water ingestion was associated with better visualization scores (calculated based on the percentage of unobscured area), but without significative differences in the detection rate between the diving and the control group. Despite the potential benefits of this method, a careful approach is needed regarding the large volume of water ingestion in this long-drawn-out examination, which might be a problem in elderly patients or those with cardiac or renal diseases.

Colon capsule endoscopy preparation CCE emerged in 2006 as a novel diagnostic procedure for colorectal examination [26]. In contrast to conventional colonoscopy, CCE does not have the capacity to water lavage or lens washing, which, when associated with low image quality due to hardware and software limitations, renders the technique highly dependent on the quality of bowel preparation [27]. Studies have consistently shown that improvements in bowel cleansing increase the sensitivity and specificity of CCE for diagnostic purposes, such as the detection of colorectal polyps [28]. As in SBCE, the clinical practice in terms of preprocedural preparation regimens is diverse, with the ideal protocol still being a matter of debate.

Diet and fasting As for conventional colonoscopy, the adoption of a low-fiber diet has an important role in achieving adequate cleansing in CCE. A low-fiber diet aims to reduce the amount of debris in the final CCE video. Despite many distinct protocols and recommendations worldwide, most endoscopic units endorse a low-fiber diet with ingestion of 10 glasses of water for at least 2 days before the procedure, followed by a clear liquid exclusive diet on the day before CCE [29].

185

186

CHAPTER 11 Small bowel and colon cleansing in capsule endoscopy

Oral purgatives PEG solutions have been the most frequently used laxative agents for CCE preparation due to their lower impact on electrolyte imbalance without altering the mucosa [30]. Classically, large-volume (3 4 L of PEG solution) regimens have been proposed in most of the available studies in CCE, a recommendation that has also been supported by the European Society of Gastrointestinal Endoscopy (ESGE) guidelines (evidence level 4, recommendation grade D) [26, 28]. In most centers, dose splitting is applied to increase tolerability while achieving higher patient compliance and effective cleansing. Many combinations of dose splitting have been used in previous investigations; however, the most commonly reported ones are the 3 L 1 1 L (3 L of PEG on day -1 and 1 L on day 0) [31], and 2 L 1 2 L combinations of dose splitting [32]. The quality of cleanliness achieved with both of these combinations is similar, but direct comparative studies have yet to become available. The large amount of liquid consumption required in oral purgative regimens can be inconvenient and may eventually decrease patients’ tolerability to the protocol as well as increase the amount of the luminal liquid during the procedure. Therefore efforts have been made to reduce the total volume of the bowel preparation for CCE, with the most notable being the association of ascorbate to PEG solutions, resulting in a similar overall adequate bowel cleanliness when compared with the PEG-exclusive preparation, with significantly lower ingested volumes [33].

Boosters Contrary to conventional colonoscopy, complete visualization of the colon in CCE depends on capsule propulsion throughout the gastrointestinal tract before the end of the battery life. To overcome this barrier, boosters have been added to established regimen protocols. Sodium phosphate (NaP) has become the most commonly used booster in CCE preparation. NaP agents promote colon cleansing by osmotically drawing plasma water into the bowel lumen, the volume effect of which allows the capsule to move in a water-filled environment, aiding in capsule propulsion [29]. The booster is administered in two timed intervals: First, when the capsule reaches the small bowel, and then 3 4 h following the initial dose [34]. Different doses have been applied depending on the performing center: The first dose varies between 22, 30, and 45 mL, with the second dose usually being half of these quantities [29]. Multiple investigations have confirmed the utmost importance of NaP application in CCE. In a study conducted by Sieg et al. [35] the median colon transit time decreased from 8.25 to 4.5 h when NaP was used. Later, Spada et al. [34] showed that NaP boosters achieved 100% CCE excretion rates when used compared with only 75% when not used, in less than 10 h. Additional formulations, such as magnesium citrate, are also available. However, their use is generally restricted to patients prone to acute nephropathy by NaP, as its efficacy is not well established [29].

Small bowel capsule endoscopy cleansing quality evaluation

Prokinetic drugs CCE follows the same principles as SBCE until reaching the colon. Therefore there must be a similar concern in avoiding prolonged gastric transit time, so the capsule can preserve relevant battery time to allow visualization of all colonic segments. The role of prokinetic drugs in CCE is the same as for SBCE—it must be selectively applied to patients with increased risk of an incomplete SBCE study and those with CE remaining in the stomach for more than 1 h as confirmed by real-time viewing.

Small bowel capsule endoscopy cleansing quality evaluation As previously mentioned, the quality of visualization and thus the diagnostic yield of SBCE rely on the presence/absence of air bubbles, bile, and intestinal debris [2]. The evaluation of small bowel preparation quality is mandatory to assess the reliability of SBCE findings. Several scores assessing small bowel cleanliness have been proposed, which can be divided into operator-dependent and automated scores. Nevertheless, a consensus has not been reached on which scale is the ideal. A recent ESGE position states that the development of a single, universally accepted, validated scale must be sought to allow standardized evaluation and monitoring of proposed performance measures [36].

Automated scores Aside from having an objective, reliable, and reproducible scoring system, performing the evaluation in a swift manner is also important. Thus computer-generated scores could have a role in fulfilling all these criteria [37]. Van Weyenberg et al. [38] created a proof of concept, computed assessment of cleansing (CAC) score based on the objective evaluation of color intensities of the red over green (R/G) channels of the tissue color bar of the rapid reader in the PillCam CE system, containing the summary of all CE images. This was then converted to red-green-blue mode (RGB). The relation between the mean intensity of red and green areas was used to evaluate small bowel cleanliness— properly visible mucosa associated with red signals as opposed to debriscontaining lumen associated with green signals. Abou Ali et al. [39] used an adapted approach based on the R/G pixel ratio of still frame images. A small bowel-CAC score cut-off of 1.6 demonstrated a sensitivity of 91.3% and a specificity of 94.7%, defining an adequate SB visualization. This principle was further adapted to OMOM and MiroCam CE systems. The MiroCam software has a function called Map View, a bar containing a representation of all the available images recorded, which was also submitted for the

187

188

CHAPTER 11 Small bowel and colon cleansing in capsule endoscopy

determination of the red and green channel intensities. The adapted CAC score for MiroCam achieved great reproducibility, with a moderate-to-good agreement in three subjective evaluations—quantitative index, qualitative evaluation, and overall adequacy assessment [40]. Klein et al. [41] created and validated a computer algorithm based on the pixels in the tissue color bar of the PillCam system, with each pixel being independently categorized as adequate or inadequate. The algorithm then calculated and summarized the total number of “inadequate” pixels, their locations, the “adequate” to “inadequate” pixel ratio, and the longest duration of consecutive “inadequate” pixels in the color bar. This algorithm had an overall sensitivity of 95%, specificity 82%, and accuracy 90%. Recently, Oumrani et al. [42] projected a score relying on three electronic characteristics—colorimetry, abundance of bubbles, and brightness. These parameters were compared with the Brotz score, an operator-dependent score, as assessed by different experts, with a score of at least 7/10, being compatible with adequate mucosal visualization. Pietri et al. [43] developed a computer-based algorithm to evaluate the abundance of bubbles in SBCE. For instance, they divided still frames in terms of the percentage of air bubbles (,10% vs $ 10%). A gray-level of co-occurrence matrix algorithm achieved perfect reproducibility, with a sensibility of 95.79%, specificity of 95.19%, and a calculation time of 0.037 s per frame, showing its utility for the evaluation of a specific parameter of evaluation of preparation quality. More recently, the development of deep learning-based tools for the evaluation of the quality of small bowel visualization is intensely growing. Nam et al. [44] developed a deep learning model to evaluate SBCE cleansing. The training dataset consisted of a five-step scoring system based on mucosal visibility (0% 25%, 25% 50%, 50% 75%, 75% 90%, over 90%). The performance of the cleansing score is evaluated with a validation dataset. The 5-step cleansing score [1 5] was compared with a validated small bowel clinical preparation scale (A to C). The cleansing score correlated with the clinical score (with decreased cleansing scores as clinical grading worsened). The ROC curve analysis obtained an optimal cleansing score cut-off of 3.25 with an AUC of 0.977, suggesting this value as a standard criterion for adequate bowel preparation. Even though numerous automated scores have been proposed, practically, no readily available score is integrated on the CE reading software to this day. The development of deep learning tools for bowel cleansing scores could be the way to standardize and integrate accurate scores in CE reading software.

Operator-dependent scores Several studies evaluate the cleanliness of the small bowel through operatordependent scores [37]. These investigations and subsequent cleansing scores differ significantly, from bowel preparation regimens to the application of different

Small bowel capsule endoscopy cleansing quality evaluation

descriptive methods—quantitative and/or qualitative. Typically, quantitative evaluations apply a numerical score (e.g., from 1 to 10), and qualitative scores rely on descriptive terms such as adequate versus inadequate or a scale of poor fair good excellent. Disadvantages of these scores are that they are operatordependent (eventually leading to variability) and time-consuming compared with automated scores. In the absence of a universally accepted score, we are focusing on the two most commonly used scores endorsed by ESGE recommendations on performance measures for SB endoscopy, the validated Park and Brotz scales, in this chapter. Brotz et al. [45] conducted a prospective study, randomized, single-center validation study, including 40 CE videos (PillCam) visualized by five CE experienced readers, who proceeded to classify the SB cleanliness based on three distinctive scoring systems previously defined by the authors (Table 11.1). One month after the initial scoring, the same 40 CE videos were randomly reassigned to the same readers, who reevaluated SB quality of preparation according to the three different scores. A clear liquid diet with overnight fasting was the only preparation applied previously to SBCE. The three evaluated scales were: A quantitative index (QI 0 10, with higher scores corresponding to better cleansing); a qualitative evaluation (QE poor, fair, good, or excellent), and an overall adequacy assessment (OAA adequate or inadequate). Despite presenting a strong and highly significant association among all three scales, the reported interobserver agreement was rather disappointing (QI and OAA moderate k 5 0.47 0.52 and 0.41, Table 11.1 Brotz score. QI Points

Percentage of visualized mucosa

Fluids and debris abundance

Bubble abundance

Bile/ chyme staining

Brightness reduction

0 1 2

,80% 80% 89% 90% or more

Severe Moderate Minimal/mild

Severe Moderate Minimal/ mild

Severe Moderate Minimal/ mild

Severe Moderate Minimal/ mild

Absent/ minimal Mild Moderate Severe

Absent/ minimal Mild Moderate Severe

Absent/ minimal Mild Moderate Severe

QE Excellent

90% or more

Absent/minimal

Good Fair Poor

90% or more ,90% ,80%

Mild Moderate Severe

OAA Adequate Inadequate OAA, Overall adequacy assessment; QE, qualitative evaluation; QI, quantitative index.

189

190

CHAPTER 11 Small bowel and colon cleansing in capsule endoscopy

Table 11.2 Park score. Score

0

1

2

3

Percentage of visualized mucosa Obscuration

# 25% $ 50%

25% 50% 25% 50%

50% 75% 5% 25%

.75% ,5%

respectively, and slight to fair for the QE k 5 0.20 0.24). This scale is hampered by a few limitations. First, it implies three different gradings considering all SBCE frames, an evaluation that may become time-consuming and eventually affect its feasibility. Moreover, it only evaluates the small bowel as a whole, which may misrepresent the actual cleansing, as different segments might present significantly different preparation quality levels. Park et al. [46] developed a different cleansing quantitative score based on the proportion of visualized mucosa and the degree of obscuration (Table 11.2). In contrast to the former score, these patients were given a 4 L PEG solution as bowel preparation. The two visual parameters were scored based on two fourdegree scales—the proportion of visualized mucosa (0 3) and the degree of obscuration by bubbles, debris, and bile (also 0 3)—by evaluating images from the entire small bowel selected at 5-min intervals. The overall classification was obtained by adding the scores of all selected images and then dividing them by the number of frames examined for each parameter. The final score is the average of the two mean partial scores. The author proposed a cut-off value of 2.25 to definite adequate small bowel cleanliness. Despite presenting good intra- and interobserver agreements, this cleansing scale evaluates images from the entire small bowel selected at 5-min intervals, representing a crucial limitation of the scale, as the majority of frames collected by SBCE will not be considered in the final cleansing score, which may therefore not reflect the actual quality of the visualization of the mucosa. Currently, a validated scale universally accepted for the classification of small bowel quality of preparation in SBCE is still missing. There are numerous grading systems with different technical characteristics, such as the parameters and the portion of the CE video that are analyzed, the objectivity of the analysis, the lesser or greater dependency on the operator, and the validation of the score [47]. In this setting, despite being time-consuming, the operator-dependent scores of Brotz and Park should be used during CE interpretation until further improved scales are developed [4].

Colon capsule endoscopy cleansing quality evaluation Adequate colon cleansing is paramount to reliable colonoscopy reported findings interpretation, a premise that is equally applicable to CCE. Sulz et al. [48] reported a decreased adenoma detection rate of 47% for inadequate bowel

Colon capsule endoscopy cleansing quality evaluation

cleansing classifications, with an early interval examination recommended in this setting. A CCE with suboptimal preparation may be associated with greater inconvenience to both patients (need for repeated procedure, a new bowel preparation, delayed diagnosis) and caretakers (limited resources, increasing waiting lists, and higher costs). Therefore quality assessment criteria are crucial to optimizing bowel cleansing protocols and improving the quality of the procedure as an instrument for clinical decisions and research reliability and validity. Qualitative terms such as adequate, inadequate, good, excellent, fair, or poor have been classically yet unevenly applied by physicians, leading to the lack of clear, consistent, and standard quality of preparation assessment [49]. In this setting, and considering the recent growth in CE expansion, the search for a reliable bowel preparation assessment scale for CCE has emerged. The most commonly used cleansing scale for CCE was conceived in 2011 by Leighton and Rex [50], which is a qualitative scale dividing the colon into five segments (cecum, right-sided colon, transverse, left-sided colon, and rectum), classifying each segment based on the degree of obscuration of the mucosa by fecal residues into poor (largely obscured), fair (partial obscuration of the mucosa, preventing the detection of polyps over 5 mm), good (partial obscuration, yet allowing detection of polyps . 5 mm), and excellent (clear view). Overall, this grading system mainly relies on the physician’s subjective impression for qualitative assessment. In addition, the Leighton Rex scale considers a subclassification, the bubbles effect scale, which grades the visualization specifically considering bubbles, further relying on the physician’s subjective impression [51]. Due to these limitations, Sousa Magalha˜es et al. [52] developed a novel scale called the Colon Capsule Cleansing Assessment and Report (CC-CLEAR) (Fig. 11.1). This is a quantitative scale developed by an analysis of 58 CCE videos in patients undergoing PillCam COLON2. In this score, the colon was divided into three segments: right-sided colon, transverse, and left-sided colon. Each segment was then scored according to an estimation of the percentage of

FIGURE 11.1 CC-CLEAR. CC-CLEAR, Colon capsule cleansing assessment and report.

191

192

CHAPTER 11 Small bowel and colon cleansing in capsule endoscopy

visualized mucosa (0% , 50%; 1% 50% 75%; 2 . 75%; 3 . 90%), with the overall cleansing classification being a sum of each segment score, grading between Excellent [8,9], Good [6,7], and Inadequate (0 5). Any segment scoring 1 or less resulted in overall inadequate classification. The colon videos were reviewed by two experienced readers, blinded to each other, and were scored by the two available operator-dependent scores: CC-CLEAR and the Leighton Rex score scale. The CC-CLEAR interobserver agreement was superior to the one previously described for Leighton Rex scale (Kendall’s W 0.911 vs 0.806, P , .01). This might be due to the subjective clinical judgment on which the Leighton Rex scale relies instead of the more objective, quantitative assessment of the percentage of visualized mucosa needed for CC-CLEAR. Furthermore, the Leighton Rex scale is more cumbersome to apply in clinical practice as there is a need to divide the colon in five different segments, which is a challenging feature (especially in CCE) (Fig. 11.2).

FIGURE 11.2 Examples of small bowel capsule endoscopy frames of segments considered to have inadequate (A), fair (B), good (C), and excellent (D) cleansing.

Final remarks

Endoscopists might find the CC-CLEAR a more user-friendly score as the colon needs to be divided into just three segments: Right colon, transverse and left colon, based on the hepatic and splenic flexure landmarks and therefore is expected to become increasingly applied to assess bowel preparation quality in CCE. Moreover, this classification has recently been validated in a multicenter study, where it was shown to impact major outcomes in CCE: Lesion detection, surveillance recommendations, and post-CC endoscopic treatment [53].

Final remarks Bowel preparation is key to achieving effective small bowel as well as colon capsule endoscopies. Improved bowel preparation protocols and reliable scales to assess the quality of mucosal visualization are pivotal in further boosting the technique

FIGURE 11.3 Examples of colon capsule endoscopy frames of segments considered to have inadequate (A), fair (B), good (C), and excellent (D) cleansing.

193

194

CHAPTER 11 Small bowel and colon cleansing in capsule endoscopy

and its quality indicators. Standardization of bowel preparation protocols and cleansing evaluation scales implemented in CE devices is pivotal in the uniformization of the procedure and for clinical research purposes (Fig. 11.3).

References [1] Koulaouzidis A, Marlicz W, Wenzek H, Koulaouzidis G, Eliakim R, Toth E. Returning to digestive endoscopy normality will be slow and must include novelty and telemedicine. Dig Liver Dis 2020;52(10):1099 101. [2] Viazis N, Sgouros S, Papaxoinis K, Vlachogiannakos J, Bergele C, Sklavos P, et al. Bowel preparation increases the diagnostic yield of capsule endoscopy: a prospective, randomized, controlled study. Gastrointest Endosc 2004;60(4):534 8. [3] Yung DE, Rondonotti E, Sykes C, Pennazio M, Plevris JN, Koulaouzidis A. Systematic review and meta-analysis: is bowel preparation still necessary in small bowel capsule endoscopy? Expert Rev Gastroenterol Hepatol 2017;11(10):979 93. [4] Rondonotti E, Spada C, Adler S, May A, Despott EJ, Koulaouzidis A, et al. Smallbowel capsule endoscopy and device-assisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Technical Review. Endoscopy. 2018;50(4):423 46. [5] Chen HB, Lian-Xiang P, Yue H, Chun H, Shu-Ping X, Rong-Pang L, et al. Randomized controlled trial of 3 days fasting and oral senna, combined with mannitol and simethicone, before capsule endoscopy. Med (Baltim) 2017;96(43):e8322. [6] Squirell E, Ricci M, Hookey L. Preparation, timing, prokinetics, and surface agents in video capsule endoscopy. Gastrointest Endosc Clin N Am 2021;31(2):251 65. [7] Mergener K, Ponchon T, Gralnek I, Pennazio M, Gay G, Selby W, et al. Literature review and recommendations for clinical application of small-bowel capsule endoscopy, based on a panel discussion by international experts. Consensus statements for small-bowel capsule endoscopy, 2006/2007. Endoscopy 2007;39(10):895 909. [8] Rokkas T, Papaxoinis K, Triantafyllou K, Pistiolas D, Ladas SD. Does purgative preparation influence the diagnostic yield of small bowel video capsule endoscopy?: A meta-analysis. Am J Gastroenterol 2009;104(1):219 27. [9] Wu S, Gao YJ, Ge ZZ. Optimal use of polyethylene glycol for preparation of small bowel video capsule endoscopy: a network meta-analysis. Curr Med Res Opin 2017;33(6):1149 54. [10] Mascarenhas-Saraiva MJ, Mascarenhas-Saraiva M. A case-control study demonstrates an improved visualization when capsule endoscopy is performed after preparation with polyethylene glycol and ascorbic acid. Rev Esp Enferm Dig 2021;113 (4):261 8. [11] Pons Beltran V, Gonzalez Suarez B, Gonzalez Asanza C, Perez-Cuadrado E, Fernandez Diez S, Fernandez-Urien I, et al. Evaluation of different bowel preparations for small bowel capsule endoscopy: a prospective, randomized, controlled study. Dig Dis Sci 2011;56(10):2900 5. [12] Hansel SL, Murray JA, Alexander JA, Bruining DH, Larson MV, Mangan TF, et al. Evaluating a combined bowel preparation for small-bowel capsule endoscopy: a prospective randomized-controlled study. Gastroenterol Rep (Oxf) 2020;8(1):31 5.

References

[13] Stein DJ, Copland A, McDaniel D, Hays RA. Single-dose linaclotide is equal in efficacy to polyethylene glycol for bowel preparation prior to capsule endoscopy. Dig Dis 2019;37(4):297 302. [14] Matsuura M, Inamori M, Inou Y, Kanoshima K, Higurashi T, Ohkubo H, et al. Lubiprostone improves visualization of small bowel for capsule endoscopy: a double-blind, placebo-controlled 2-way crossover study. Endosc Int Open 2017;5(6): E424 9. [15] Xavier S, Rosa B, Monteiro S, Arieira C, Magalhaes R, Curdia Goncalves T, et al. Bowel preparation for small bowel capsule endoscopy the later, the better!. Dig Liver Dis 2019;51(10):1388 91. [16] Mascarenhas-Saraiva MJ, Oliveira E, Mascarenhas-Saraiva MN. The use of a PEG/ ascorbate booster following standard bowel preparation improves visualization for capsule endoscopy in a randomized, controlled study. Turk J Gastroenterol 2021;32 (5):437 42. [17] Koulaouzidis A, Giannakou A, Yung DE, Dabos KJ, Plevris JN. Do prokinetics influence the completion rate in small-bowel capsule endoscopy? A systematic review and meta-analysis. Curr Med Res Opin 2013;29(9):1171 85. [18] Kotwal VS, Attar BM, Gupta S, Agarwal R. Should bowel preparation, antifoaming agents, or prokinetics be used before video capsule endoscopy? A systematic review and meta-analysis. Eur J Gastroenterol Hepatol 2014;26(2):137 45. [19] Cotter J, de Castro FD, Magalhaes J, Moreira MJ, Rosa B. Finding the solution for incomplete small bowel capsule endoscopy. World J Gastrointest Endosc 2013;5(12): 595 9. [20] Albert J, Gobel CM, Lesske J, Lotterer E, Nietsch H, Fleig WE. Simethicone for small bowel preparation for capsule endoscopy: a systematic, single-blinded, controlled study. Gastrointest Endosc 2004;59(4):487 91. [21] Rosa BJ, Barbosa M, Magalhaes J, Rebelo A, Moreira MJ, Cotter J. Oral purgative and simethicone before small bowel capsule endoscopy. World J Gastrointest Endosc 2013;5(2):67 73. [22] Papamichael K, Karatzas P, Theodoropoulos I, Kyriakos N, Archavlis E, Mantzaris GJ. Simethicone adjunct to polyethylene glycol improves small bowel capsule endoscopy imaging in non-Crohn’s disease patients. Ann Gastroenterol 2015;28(4):464 8. [23] Wu L, Cao Y, Liao C, Huang J, Gao F. Systematic review and meta-analysis of randomized controlled trials of Simethicone for gastrointestinal endoscopic visibility. Scand J Gastroenterol 2011;46(2):227 35. [24] Sey M, Yan B, McDonald C, Segal D, Friedland J, Puka K, et al. A randomized controlled trial of high volume simethicone to improve visualization during capsule endoscopy. PLoS One 2021;16(4):e0249490. [25] Zeng X, Ye L, Liu J, Yuan X, Jiang S, Huang M, et al. Value of the diving method for capsule endoscopy in the examination of small-intestinal disease: a prospective randomized controlled trial. Gastrointest Endosc 2021;94(4):795 802 e1. [26] Spada C, Hassan C, Galmiche JP, Neuhaus H, Dumonceau JM, Adler S, et al. Colon capsule endoscopy: European Society of Gastrointestinal Endoscopy (ESGE) Guideline. Endoscopy. 2012;44(5):527 36. [27] Wang YC, Pan J, Liu YW, Sun FY, Qian YY, Jiang X, et al. Adverse events of video capsule endoscopy over the past two decades: a systematic review and proportion meta-analysis. BMC Gastroenterol 2020;20(1):364.

195

196

CHAPTER 11 Small bowel and colon cleansing in capsule endoscopy

[28] Sieg A. Capsule endoscopy compared with conventional colonoscopy for detection of colorectal neoplasms. World J Gastrointest Endosc 2011;3(5):81 5. [29] Singhal S, Nigar S, Paleti V, Lane D, Duddempudi S. Bowel preparation regimens for colon capsule endoscopy: a review. Ther Adv Gastroenterol 2014;7(3):115 22. [30] Tan JJ, Tjandra JJ. Which is the optimal bowel preparation for colonoscopy a meta-analysis. Colorectal Dis 2006;8(4):247 58. [31] Van Gossum A, Munoz-Navas M, Fernandez-Urien I, Carretero C, Gay G, Delvaux M, et al. Capsule endoscopy versus colonoscopy for the detection of polyps and cancer. N Engl J Med 2009;361(3):264 70. [32] Spada C, Hassan C, Ingrosso M, Repici A, Riccioni ME, Pennazio M, et al. A new regimen of bowel preparation for PillCam colon capsule endoscopy: a pilot study. Dig Liver Dis 2011;43(4):300 4. [33] Hartmann D, Keuchel M, Philipper M, Gralnek IM, Jakobs R, Hagenmuller F, et al. A pilot study evaluating a new low-volume colon cleansing procedure for capsule colonoscopy. Endoscopy. 2012;44(5):482 6. [34] Spada C, Riccioni ME, Hassan C, Petruzziello L, Cesaro P, Costamagna G. PillCam colon capsule endoscopy: a prospective, randomized trial comparing two regimens of preparation. J Clin Gastroenterol 2011;45(2):119 24. [35] Sieg A, Friedrich K, Sieg U. Is PillCam COLON capsule endoscopy ready for colorectal cancer screening? A prospective feasibility study in a community gastroenterology practice. Am J Gastroenterol 2009;104(4):848 54. [36] Spada C, McNamara D, Despott EJ, Adler S, Cash BD, Fernandez-Urien I, et al. Performance measures for small-bowel endoscopy: a European Society of Gastrointestinal Endoscopy (ESGE) Quality Improvement Initiative. Endoscopy. 2019;51(6):574 98. [37] Rosa B, Margalit-Yehuda R, Gatt K, Sciberras M, Girelli C, Saurin JC, et al. Scoring systems in clinical small-bowel capsule endoscopy: all you need to know!. Endosc Int Open 2021;9(6) E802-e23. [38] Van Weyenberg SJ, De Leest HT, Mulder CJ. Description of a novel grading system to assess the quality of bowel preparation in video capsule endoscopy. Endoscopy. 2011;43(5):406 11. [39] Abou Ali E, Histace A, Camus M, Gerometta R, Becq A, Pietri O, et al. Development and validation of a computed assessment of cleansing score for evaluation of quality of small-bowel visualization in capsule endoscopy. Endosc Int Open 2018;6(6):E646 51. [40] Ponte A, Pinho R, Rodrigues A, Silva J, Rodrigues J, Carvalho J. Validation of the computed assessment of cleansing score with the Mirocam® system. Rev Esp Enferm Dig 2016;108(11):709 15. [41] Klein A, Gizbar M, Bourke MJ, Ahlenstiel G. Validated computed cleansing score for video capsule endoscopy. Dig Endosc 2016;28(5):564 9. [42] Oumrani S, Histace A, Abou Ali E, Pietri O, Becq A, Houist G, et al. Multicriterion, automated, high-performance, rapid tool for assessing mucosal visualization quality of still images in small bowel capsule endoscopy. Endosc Int Open 2019;7 (8):E944 8. [43] Pietri O, Rezgui G, Histace A, Camus M, Nion-Larmurier I, Li C, et al. Development and validation of an automated algorithm to evaluate the abundance of bubbles in small bowel capsule endoscopy. Endosc Int Open 2018;6(4):E462 9.

References

[44] Nam JH, Hwang Y, Oh DJ, Park J, Kim KB, Jung MK, et al. Development of a deep learning-based software for calculating cleansing score in small bowel capsule endoscopy. Sci Rep 2021;11(1):4417. [45] Brotz C, Nandi N, Conn M, Daskalakis C, DiMarino M, Infantolino A, et al. A validation study of 3 grading systems to evaluate small-bowel cleansing for wireless capsule endoscopy: a quantitative index, a qualitative evaluation, and an overall adequacy assessment. Gastrointest Endosc 2009;69(2):262 70 70.e1. [46] Park SC, Keum B, Hyun JJ, Seo YS, Kim YS, Jeen YT, et al. A novel cleansing score system for capsule endoscopy. World J Gastroenterol 2010;16(7):875 80. [47] Ponte A, Pinho R, Rodrigues A, Carvalho J. Review of small-bowel cleansing scales in capsule endoscopy: a panoply of choices. World J Gastrointest Endosc 2016; 8(17):600 9. [48] Sulz MC, Kroger A, Prakash M, Manser CN, Heinrich H, Misselwitz B. Metaanalysis of the effect of bowel preparation on adenoma detection: early adenomas affected stronger than advanced adenomas. PLoS One 2016;11(6):e0154149. [49] Sharma RS, Rossos PG. A review on the quality of colonoscopy reporting. Can J Gastroenterol Hepatol 2016;2016:9423142. [50] Leighton JA, Rex DK. A grading scale to evaluate colon cleansing for the PillCam COLON capsule: a reliability study. Endoscopy. 2011;43(2):123 7. [51] Tabone T, Koulaouzidis A, Ellul P. Scoring systems for clinical colon capsule endoscopy—all you need to know. J Clin Med 2021;10(11). [52] de Sousa Magalhaes R, Arieira C, Boal Carvalho P, Rosa B, Moreira MJ, Cotter J. Colon Capsule CLEansing Assessment and Report (CC-CLEAR): a new approach for evaluation of the quality of bowel preparation in capsule colonoscopy. Gastrointest Endosc 2021;93(1):212 23. [53] de Sousa Magalha˜es R, Cha´lim Rebelo C, Sousa-Pinto B, Pereira J, Boal Carvalho P, Rosa B, et al. CC-CLEAR (Colon Capsule Cleansing Assessment and Report): the novel scale to evaluate the clinical impact of bowel preparation in capsule colonoscopy a multicentric validation study. Scand J Gastroenterol 2022;1 8.

197

This page intentionally left blank

CHAPTER

Introducing blockchain technology in data storage to foster big data and artificial intelligence applications in healthcare systems

12

Miguel Mascarenhas1,2, Andre´ Santos3 and Guilherme Macedo1,2,4 1

Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, 2 Faculty of Medicine, University of Porto, Porto, 3 Centro Hospitalar do Baixo Vouga, Aveiro, 4 World Gastroenterology Organization Porto Training Center, Porto,

Portugal Portugal Portugal Portugal

Introduction We are currently immersed in a new digital era in which data has become one of the world’s most valuable commodities. In the world of healthcare, digitalization offers clear benefits both in terms of access to and quantification of different types of medical data. Indeed, the gradual shift toward the implementation of electronic health records (EHRs) worldwide bears witness to the acknowledgment of the value of digitized medical data. However, these systems remain relatively local and often with limited access, negating some of their potential benefits. In addition to EHRs, there are other clinical areas where digital transformation has driven important and beneficial changes. One such area is medical imaging with the advancement in the acquisition and handling of medical images that enhance the possibilities for their quantification and the capacity to more accurately compare images. As a result, imaging techniques are becoming increasingly reliable in generating not only diagnostic information but also in establishing prognostic disease biomarkers and monitoring disease progression. For example, it is now more feasible to identify lesions and monitor their size or histological changes in cancer [1] or neurological diseases (e.g., multiple sclerosis) [2], even when the affected tissues may not be readily accessible. However, this expansion in the use of clinical imaging opens up two issues that must be addressed. The first involves the storage and accessibility of the vast amounts of data generated this way. The second is related to our understanding and interpretation of the Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00011-3 © 2023 Elsevier Inc. All rights reserved.

199

200

CHAPTER 12 Introducing blockchain technology in data storage

images obtained and their relationship to disease status and progression. Both these issues have been the subject of recent advances related to the management of big data and the application of artificial intelligence (AI) to the datasets with a view to enhancing the information that can be extracted from them. The concerns regarding the storage and use of digitized clinical data are security, traceability, and reliability. Blockchain technology offers the possibility of, at least in part, addressing these problems, as a blockchain is a linked cryptographic list of records (or blocks) [3,4] containing a cryptographic hash of the previous block, a timestamp, and transaction data (Fig. 12.1). The timestamp proves the prior existence of the transaction when the block was generated or published, and thus as each block contains information about the previous one from which it is derived, they form a chain. Consequently, blockchains are resistant to data modification since the data in any given block cannot be altered retroactively without altering all the ensuing blocks—characteristics that have led to proposals for the implementation of this technology in different aspects of the healthcare system. In this review, we will discuss the solutions that blockchain technology can offer in the storage of medical data, allowing its safe, traceable, and efficient handling. Such changes could open the way for establishing decentralized digital platforms in which the data compiled is readily accessible at any time and from any location. Here we will focus on how this technology, particularly the incorporation of big data and AI applications, can help in the introduction of other novel and important advances in modern healthcare systems (Fig. 12.2).

A brief picture of present-day medical challenges Modern sociocultural changes brought about by increasing urbanization have led to sedentary lifestyles, unhealthy dietary patterns, and a rise in obesity, leading to

FIGURE 12.1 Scheme showing the concept underlying blockchain technology and the benefits it has to offer in terms of security and maintaining the traceability of information. Each record or block contains a cryptographic hash that encodes information regarding the previous block in the chain, linking the two. This way, information cannot be erased. When sent to the user, the digital signature and timestamping of the information in a block provide undisputable evidence of the provenance of a record at any point in time.

A brief picture of present-day medical challenges

FIGURE 12.2 Scheme showing the incorporation of AI algorithms and blockchain technology in healthcare. The blockchain platform is fed by medical data and information that can be collected from EHRs, anamnesis or from other physical examinations, biochemical or other more specific analyses (e.g., biomarkers), and medical imaging data (e.g., capsule endoscopy images: see the left-hand side of the figure). This data can all be handled through diverse informatics platforms in a network (e.g., desktop or laptop computers and even mobile devices). This data can then be stored directly in a blockchain platform within a defined network, in a safe, secure, and traceable manner, and/or be analyzed using different AI algorithms to extract features of interest. Likewise, the results of these analyses with AI algorithms can also be stored in the blockchain platform, and the data in the blockchain can be used to train the algorithms (see the feedback loop in the center of the figure). At each step, the owner of the data must approve the treatment of their personal and health data, giving them control over its use. Similarly, they can control the use of this data in more universal global environments, with the potential to monetize their data. AI, Artificial intelligence; EHRs, electronic health records.

an increase in chronic diseases, such as heart disease, diabetes, and even cancer [5], in the aging population. Technological developments are expected to help address the healthcare challenges associated with these changes. However, the implementation of technological developments themselves poses new challenges for society. We are unarguably witnessing a profound shaping of society by technological advances. While the current generation is accustomed to the instant retrieval of information (from anywhere and anytime), the health sector has fallen far short of expectations in this regard. For healthcare to embrace technological

201

202

CHAPTER 12 Introducing blockchain technology in data storage

advances that can address the challenges faced, medical database management must evolve to favor access to information in a secure and reliable way. The introduction of EHRs has improved access to medical records for healthcare professionals and patients alike [6]. However, current healthcare systems are not designed for the interinstitutional sharing of medical information, partly because of their compliance with the measures of data protection and security. Medical institutions store their patients’ medical records often only accessible through their computer system, leading to fragmentation [7]. This suboptimal model of data storage is a major obstacle when contemplating the possibility of improving the quality and efficiency of medical care. The prominence that digitalized medical imaging has achieved in the diagnosis and monitoring of a wide variety of diseases and clinical research is partly due to the advancement in technology that have enhanced the capability to store and classify data. A wide variety of imaging techniques are currently being used as diagnostic procedures for disease monitoring and in clinical research as endpoints and biomarkers, and they are even being combined with therapeutic procedures. The data generated by these approaches is often complex and requires considerable computing power to be processed and handled. The advent of big data and AI applications is generating much expectation, both in healthcare in general and in relation to medical imaging. These new approaches appear to offer solutions regarding the most appropriate way in which vast amounts of medical imaging data being generated can be managed and used.

Emergence of blockchain in healthcare It is perhaps not surprising that the digitalization of healthcare has awoken increasing interest in the possible application of blockchain technologies in this sphere, not least as many analysts regard this technology as an innovation with true disruptive potential [8]. Indeed, forecasts of the World Economic Forum indicate that 10% of the global GDP will involve distributed ledger technology by 2025 [8], and national governments and public institutions have also appreciated the potential of blockchain technology to provide new solutions to public issues [8]. While medicine is often slow to adopt new technologies, falling behind other industries, there are clear reasons why the healthcare industry should embrace blockchain technology without delay [9]. Significantly, there has been a large increase in healthcare executives putting in place commercial blockchain solutions [10]. One area in which blockchain could have several advantages is in the management of EHRs [9] or in the management of large datasets. Indeed, as early as 2017, the health authorities in Estonia implemented a blockchain solution to handle the storage of a million health records [11]. The uniqueness of blockchain technology is that it permits the establishment of a massive, secure, and decentralized public store of ordered records or events

Emergence of blockchain in healthcare

[11]. It involves a distributed database that contains a growing list of transactions/ records, each organized into blocks, and uses algorithms that ensure the data included in the ledger is resistant to manipulation. This way, information cannot be erased. The valid transactions that are stored in a blockchain are timestamped and signed digitally by the sender, providing an irrefutable testimony of both the source and the existence of a record at any given time [12]. Importantly, the blockchain dataset is controlled by its users rather than being governed by any centralized regulatory body [11]. The local storage of medical information is a barrier to sharing this information, potentially compromising its security. Blockchain technology enables data to be carefully protected and safely stored, assuring its immutability. Thus blockchain technology could help overcome the fragmentation of patients’ medical records, potentially benefiting patients and healthcare professionals alike. This approach would provide a more complete picture of patients’ health, reduce the time healthcare professionals spend harvesting these records, promote communication between healthcare professionals, and radically reduce costs associated with sharing medical data.

Blockchain and its utility for big data and artificial intelligence in healthcare The next phase in the digitalization of healthcare appears to have begun with the revolutionary implementation of big data and AI technologies. However, their full integration into healthcare systems can be further facilitated by their complementation with blockchain technologies, with their integration overcoming, to some extent, the weaknesses of each of these technologies [13]. AI applications can benefit from different features of blockchain, offering trustworthiness and enhanced privacy and traceability to AI, while AI can serve to construct better machine learning systems based on blockchain, which have tighter security, enhanced scalability, and improved personalization and handling. Hence we can consider these two disrupting technologies to be synergistic. AI algorithms rely on data to be trained and on information to interpret the data and reach conclusions. When this data is acquired from a reliable, secure, and trusted platform, AI algorithms perform better. Data on the blockchain is stored robustly and with high integrity, such that when this data is used by machine learning algorithms to take decisions and perform an analysis, the results are trustworthy and reliable. Thus the combining of AI with blockchain establishes a safe, irreversible, and decentralized environment to manage sensitive health-related data and apply AI-driven techniques [14]. In the healthcare context, AI applications with access to the blockchain of medical data can employ various algorithms, enhancing their decision-making capabilities by accessing large amounts of robust data. Thus converging blockchain and next-generation AI healthcare technologies will accelerate the benefits of these technologies in areas where the power of deep learning and AI applications has already been seen, such

203

204

CHAPTER 12 Introducing blockchain technology in data storage

as in medical diagnostics, image analysis, drug synthesis and drug classification, etc. Indeed, the implementation of AI systems can help alleviate the demands produced when human resources are limited and by the intense workloads in clinical practice. As blockchain provides a unique capacity to trace records, known variations (genetic, geographic, or demographic) can be more readily taken into account, and this improved partitioning of data is fundamental for AI applications and precision medicine [15]. However, in real-life clinical applications inherent bias cannot be ignored and must be considered before validating AI solutions. For example, for the technique of capsule endoscopy (CE) [16] and in the field of cardiovascular medicine [17], spectrum bias is a potential pitfall for AI applications. Spectrum bias occurs when a diagnostic test is studied in individuals who differ from the population for which the test was intended. AI systems are tailor-made and designed based on a training dataset, with the risk that they may be presented with an underrepresentative population [16,18]. As a result the overfitting of these models should not be ignored. Hence AI learning models might not always be fully valid and applicable to new datasets. The integration of blockchain-enabled data from other healthcare platforms could serve to augment the number of what would otherwise be underrepresented cases in a dataset, thereby improving the training of the AI application and, ultimately, its successful implementation. Several AI applications that integrate blockchains have emerged. For example, the American Hospital Association (AHA) has partnered with Open Health Network to develop AI and blockchain products like PatientSphere, a blockchainbased HIPAA (Health Insurance Portability and Accountability Act)-compliant data-sharing platform that uses AI to deliver treatment plans and exercise tips [19]. Much interest has been shown in three additional areas that could drive important advancements in personalized cardiovascular medicine [15]: (1) using blockchain for data storage to develop AI models that predict events like acute myocardial infarction by analyzing demographics, cardiac imaging data, and other inputs [20]; (2) decentralizing AI algorithms and data, like that seen in the partnership between MedStar Health Research Institute and ObEN, to monitor patients with heart failure and offer incentives to patients who engage in healthrewarding behaviors; (3) using blockchain as the backbone for AI-linked network of sensors to predict cardiovascular diseases, such as the initiatives by Farasha Labs or Health2Sync diabetes data trust [15]. The combined technology would greatly enhance data availability to develop and train AI applications, help share proprietary AI algorithms for generalization, and promote database decentralization. However, such applications are still in the early stages of development and concerns have been raised regarding their implementation (see below). Blockchain and AI can be used to disseminate locally learned models for big data analytics, such as lung cancer CT scans, together building a global model to extract patterns for lung cancer patients [21]. Significantly, blockchain and AI can provide solutions to manage the coronavirus epidemic [22]. In fact, blockchain and AI have been used together to manage other infectious epidemics, such

Emergence of blockchain in healthcare

as Ebola [23], to conduct real-time contact tracing, surveillance of transmission patterns, and vaccine delivery. AI can provide computer-aided solutions to analyze medical images and symptoms to ensure efficient patient management and treatment, which can also be used to predict future outbreak prediction. For example, Infravision is a company that uses deep learning medical imaging platforms to speed up the detection of COVID-19 cases by recognizing complex features in the lung [24]. A decentralized framework integrating blockchain and AI technologies has also been proposed for a patient-centered healthcare system, wherein blockchain reduces the impact of siloed patient datasets using AI models to predict diagnosis and improve treatments. In addition, blockchain facilitates federated learning techniques, whereby AI models train COVID-19 data with patients’ permission while preserving privacy, enabling knowledge to be aggregated from the nodes in AI models to inform a global model [25]. Thus deep learning models can be trained with a distributed network while maintaining data privacy [25].

Blockchain and use of big data and artificial intelligence in imaging Medical imaging has become a vital part of modern medicine, contributing significantly to patient diagnosis, monitoring, and treatment decisions. However, medical imaging data is currently stored and transferred with the aid of trusted third-party intermediaries, an inefficient system that carries the risk of possible data tampering. As already indicated, blockchain-based frameworks favor secure sharing and decentralized access to data, and although such systems have been described, the feasibility of their large-scale implementation remains to be demonstrated [26,27]. The use of blockchain systems facilitates the exchange of data between clinicians and radiologists, enhancing the transmission of information obtained through radiology studies. Blockchains also allow individual contributions to be tracked and annotated by different specialists as individual contributions, preserving the order in which each analysis was performed. Radiology reports may also include AI-generated content, and the contribution of this information to medical imaging reports is likely to increase in the future. Again, blockchain differentiates whether the contribution is provided by a radiologist directly or with assistance from AI, identifying the version of the AI algorithms used [28]. The integration of blockchain into EHRs in conjunction with AI algorithms allow radiologists to find relevant information in EHRs easily, leading to efficient and better interpretation of the imaging data. Moreover, follow-up recommendations, reporting of incidental findings, and indications for management can be stored, validated, and tracked on the blockchain, thereby improving patient healthcare. Although AI has possible applications in many distinct medical specialties, there is particular interest in implementing this technology in those disciplines

205

206

CHAPTER 12 Introducing blockchain technology in data storage

with strong imaging and diagnostic components [29]. Creating and training efficient AI algorithms require vast amounts of high-quality data, usually from different institutions, validated to build efficient learning models. If there is insufficient data to train deep learning models, rare events will not always be detected, leading to selection biases affecting the generalization of the AI approach, which may have pernicious effects on decision-making [30]. For AI applications to be effective, AI algorithms need to constantly evolve, after their initial training, using real-world data. This is where blockchain can help enable institutions to share datasets, offering improved traceability of imaging data and radiologists’ annotations to train AI algorithms. Blockchain technology supports the continuous assessment and improvement of learning models, which is an important element of AI. When associated with blockchain, AI can not only learn from the shared data, but engineers can also track and evaluate its learning by reviewing the chain, gaining greater insight into AI decision-making [28]. Positive and negative values (both true and false) validated by radiologists can generate large datasets suitable for training complex systems [31]. Indeed, all the annotations by experts, including the conditions in which a deep learning model is trained and the adjustments are made to it, provide information on the quality of the model.

Growing field of artificial intelligence applied to capsule endoscopy The benefits of combining blockchain technology with AI applications are best seen in areas where AI application is already proving to be useful. Gastrointestinal (GI) conditions are on the rise, making these an interesting area to consider in the recent advances in the application of AI in improving the detection of the condition. Moreover, like many other medical conditions, early diagnosis of GI tract diseases improves their prognosis, which can be achieved by using techniques that are more sensitive and faster (particularly in terms of processing and analysis by specialists). In gastroenterology, CE is one such area in which AI is showing considerable promise, and this will become more relevant with the technique being implemented more widely for screening, diagnosis, and disease monitoring purposes. CE has become a common minimally invasive procedure to study the small intestine. However, several challenges remain in this field, including the lengthy and tedious reading process where image interpretation is strongly dependent on the human reader’s ability. A single CE video examination normally involves assessing 50,000 60,000 frames on average, requiring 30 120 min to be read by gastroenterologists. Furthermore, manual reading has an inherent risk of oversight given the limitations of human concentration, particularly as small-bowel abnormalities may only be present in a single or few frames of the video. Thus CE is an ideal target for AI systems to assist gastroenterologists in the identification of areas of interest and suspicious abnormalities [32].

Growing field of artificial intelligence applied to capsule endoscopy

There is growing evidence that AI will be fundamental to the evolution of digestive endoscopy. The detection of GI hemorrhage due to ulcers and vascular lesions has witnessed major advances through the expansion of automated videocapsule diagnostics. Over a decade ago a computer-based procedure was developed to detect hemorrhages using color coding, although it was limited by poor-quality video images [33]. Subsequently, a support vector machine (SVM) model [34] as well as a multilayer perceptron (MLP) [35] and a convolutional neural network (CNN) of hemorrhage were developed to detect bleeding [36], although the latter has a limited value in clinical practice as it only detects relatively severe cases [37]. An SVM-based computer-aided design (CAD) method provides novel image analysis by studying superpixels (grouped sets of pixels in each frame with similar characteristics), reducing computational demands while improving detection capacity [38]. Since then the sensitivity and accuracy of these automated approaches have increased: First with a new local texture descriptor with less computational demands that makes it suitable for real-time implementation [39], and more recently, with a CNN capable of detecting smallbowel angiectasias [40]. The CNN represents an excellent launch pad for future automated diagnostic software [40], particularly as angiectasias are the most common lesions detected by video CE in patients with medium GI bleeding. Further progress has since been made in this area [41 44], and the performance and effectiveness of these networks and models are expected to improve. Videocapsule images are also useful in detecting abnormal structures in the small intestine mucosa and elsewhere in the GI tract. An MLP AI method [45] and a CAD system have been developed to detect GI tumors through CE [46], which benefited from the analysis of consecutive frames [47]. The MLP method appeared to perform better than an earlier SVM AI method [48], and further advances have since led to the development of a CNN method capable of identifying lesions as polyps, nodules, epithelial tumors, submucosal tumors, and venous structures [49,50]. This AI approach is better adapted to clinical practice, as multiple pathological events may be detected in a single study, and they must each be classified correctly [51]. The evaluation of inflammatory bowel disease with CE, particularly Crohn’s disease (CD), is enhanced by assessing the entire small intestine mucosa, aiding in CD diagnosis. AI approaches that take advantage of specific scales (e.g., the Lewis score) can assist in the assessment of disease activity and responses to therapy [52]. An SVM-based similarity learning method [53] and deep learning algorithms have been employed in CD patients to predict small-bowel pathologies through video CE [54]. Recently, a neural network has been developed that appears to be capable of assessing the severity of the ulcers in CD patients [55]. Celiac disease is a chronic autoimmune disorder in which the ingestion of gluten by a genetically susceptible individual triggers damage to the small intestine mucosa [56]. Currently, this condition has a global prevalence of around 1%, although this is believed to be on the increase. Celiac disease is diagnosed when duodenal villous atrophy is evident in endoscopic biopsies. The high cost and

207

208

CHAPTER 12 Introducing blockchain technology in data storage

invasive nature of this procedure suggest CE would represent a better diagnostic tool with fewer associated risks [56]. Accordingly, computer models to aid in the diagnosis of this disorder by CE have been developed, involving the use of classifiers, a CNN that measures the damage to the intestinal mucosa quantitatively, and a deep learning approach applied to a CNN system [57]. Together, these studies highlight the scope for introducing AI into the assessment and diagnosis of celiac disease. When performing CE, AI may also help locate the capsule in the intestine and manage artifacts that potentially compromise the clinical assessment of the mucosa. For example, a CNN method has been used to develop an algorithm that accurately detects bubbles and food debris [58], whereas other AI models can classify luminal content to prevent it from interfering with the results obtained [59]. Consequently, AI tools can help optimize the evaluation of videocapsule images, reducing the time required for reading as well as the bias and interpretation errors [16]. Despite the growing evidence that AI can offer important benefits in digestive endoscopy, most studies have been performed retrospectively, often with data from individual centers or from small multicenter studies, which may introduce selection and spectrum bias, limiting the possible generalizations of the AI systems for CE [32]. In addition, the size of training and test datasets vary widely. Thus large controlled trials in real-time clinical settings will be necessary to obtain robust evidence for the performance of AI systems with CE [60]. This is where blockchain can help by facilitating the sharing of datasets by multiple institutions to train AI algorithms, supporting the continuous assessment and improvement of learning models, and making AI a reliable tool for CE studies of the GI tract. As the use of video CE becomes more widespread in the study and diagnosis of GI diseases, the potential of AI applications to learn from and improve through the large amounts of data available will only grow. Accordingly, the opportunities for blockchain technology to offer enhanced security and traceability of this data can only benefit this field.

Limitations and challenges to applications of blockchain in healthcare There are several challenges to the implementation of blockchain technology in healthcare. Ethical issues related to privacy are a concern when sharing health information, although when considering private blockchains this is similar to when medical research is carried out, as data must be accessible without revealing the identity of the patients (usually achieved by anonymization). Even though medical data in a blockchain ledger may be anonymous, it might not always be desirable to allow public access to data to minimize the risk of data being linked to an individual by analyzing the blockchain transactions, perhaps in conjunction with other information. Medical data should not generally be publicly available

Limitations and challenges to applications of blockchain in healthcare

but encrypted transactions or blocks in the chain should require permission to access the data stored elsewhere to maintain privacy and ensure security [15]. Data privacy may also be achieved by newer approaches, such as homomorphic and attribute-based encryption, secure multiparty computation, zero-knowledge proof, obfuscation, or format-preserving encryption [25]. Security can be reinforced at different levels by using hybrid privacy methods and security-enhancing technologies like homomorphic signatures [61], which work better than public key certificates [25]. The feasibility of setting up a blockchain-based network to share medical data must also be looked at in light of current national/regional/local regulations. For example, it is unclear if using random public keys and hashed study identifiers constitute sufficient de-identification to be exempt of the standard disclosure restrictions under data protection laws (e.g., HIPPA and GDPR). If not, it may be necessary to evaluate whether a patient’s digital signature on the blockchain transaction serves as an authorization to release personal health information [26]. In addition, how will blockchain reconcile data subject rights under GDPR to rectify and erase inaccurate personal data, given that blockchains are purposefully designed to render such modifications of data onerous to ensure data integrity and to enhance trust in the network? Although off-chain storage in appropriate databases would enable the rectification and erasure of personal data, the status of the on-chain hashtag when transactional data is stored off-chain and subsequently erased needs to be examined [62]. Additional national and local regulations are likely to apply, which may hinder the adoption of this unconventional sharing framework by healthcare institutions. Aside from the technical, ethical, and regulatory limitations, end-user experience will probably determine the acceptance of the blockchain framework proposed, which involves concepts that are generally unfamiliar to most. Such complexities could be overcome by designing user-friendly interfaces to optimize user engagement, probably best deployed as web and/or mobile applications [26]. Healthcare providers must also be considered in this context, as there are more wide-reaching problems related to the sharing of health information that are not resolved by the blockchain. For example, in the absence of the clinical context, clinical data may be shared but not usable [26]. Indeed, the ever-increasing volume of clinical data itself represents a challenge, as it might augment the time required to verify and approve transactions in the blockchain environment. This may in turn affect scalability, depending on the consensus algorithm, although such scalability issues can be circumvented by using a permissioned blockchain better suited to process large volumes of transactions, overriding the need for time-consuming validation [25]. Novel solutions to achieve network-wide scalability are being developed, such as sharding [63], dividing rapidly growing blockchain networks into groups or shards. The issue of scalability can also be overcome by designing hierarchical blockchain systems. Finally, legal or regulatory issues are probably the biggest barrier to adopting frameworks based on AI and blockchain, as each healthcare organization is

209

210

CHAPTER 12 Introducing blockchain technology in data storage

legally liable and responsible to the central healthcare system. However, it may be more difficult to resolve legal disputes or discrepancies in decentralized, public blockchain structures. Copyright infringement and defamation problems could arise when personal information is run on converged AI and blockchain platforms. As such, administrative procedures and new legal frameworks may need to be developed to oversee the implementation of this novel technology [25,64].

Advantages of using blockchain in capsule endoscopy: how it can be enhanced with artificial intelligence tools If we consider the specific case of CE and the incorporation of AI tools to enhance the procedure, there are clear advantages to be gained by implementing blockchain technology in conjunction with these approaches. As indicated above, the benefits of AI applications are not only that they can handle large numbers of images but that these applications are trained by this data. The incorporation of blockchain technology will not only help manage data, but it will also help to confirm the veracity of the data used and reported. This has been demonstrated by incorporating blockchain into a CNN model used to recognize stomach abnormalities, making the model more resistant to data tampering due to the internal validation inherent in the blockchain [65]. The clear advantages provided by blockchain when managing large amounts of verifiable and reliable data will therefore favor the accuracy and improvement of these tools. Moreover, the secure sharing of data through blockchain technology will favor the extended use and distribution of these tools and the incorporation of data from multiple centers to train these tools. The capacity to handle data from different verified sources may open the way to use AI tools to detect local tendencies in diagnosis trends that are difficult to detect on a manual basis. Indeed, the implementation of AI tools in CE will generate large amounts of data and benefit from being able to draw on this, which will require tools and protocols to ensure the veracity and reliability of the data. Moreover, throughout these processes, patients can remain assured that their personal and medical data will remain private and confidential despite its extended use.

Concluding remarks It is clear that healthcare is evolving rapidly to address the changes faced by society, both in terms of epidemiological and demographic changes as well as due to technological advancements. The incorporation of new technologies raises new issues that must be addressed while opening the door for improvements over the existing systems. Here we have focused on the recent interest in the incorporation of big data and AI-based applications into today’s health systems. While these

References

technologies offer great opportunities to improve healthcare, they also raise questions regarding the clinical systems currently employed. One possibility of addressing some of these questions is by introducing blockchain technology to enhance the management, storage, and accessibility of clinical data. The benefits that such technology can offer are particularly relevant to AI-based applications that rely on verifiable and reliable data delivered in a safe and secure manner.

Acknowledgments The authors would like to thank all those colleagues whose helpful discussions have helped shape this review article. No external funding was received in regard to the preparation of this manuscript.

Conflicts of interest The authors have no conflicts of interest to declare in relation to this work.

References [1] Fass L. Imaging and cancer: a review. Mol Oncol 2008;2:115 52. Available from: https://doi.org/10.1016/j.molonc.2008.04.001. [2] Young PNE, Estarellas M, Coomans E, Srikrishna M, Beaumont H, Maass A, et al. Imaging biomarkers in neurodegeneration: current and future practices. Alzheimer’s Res Ther 2020;12:49. Available from: https://doi.org/10.1186/s13195-020-00612-7. [3] Puthal D, Malik N, Mohanty SP, Kougianos E, Das G. Everything you wanted to know about the blockchain. IEEE Consum Electron Mag CEM 2018;7:6 14. [4] Ahmad SS, Khan S, Kamal MA. What is blockchain technology and its significance in the current healthcare system? A brief insight. Curr Pharm Des 2019;25:1402 8. Available from: https://doi.org/10.2174/1381612825666190620150302. [5] Cooper T, Allen S. Global health care outlook: the evolution of smart health care. Deloitte analysis; 2018. ,https://www2.deloitte.com/content/dam/Deloitte/global/Documents/LifeSciences-Health-Care/gx-lshc-hc-outlook-2018.pdf. [accessed 25.11.21]. [6] Mamoshina P, Ojomoko L, Yanovich Y, Ostrovski A, Botezatu A, Prikhodko P, et al. Converging blockchain and next-generation artificial intelligence technologies to decentralize and accelerate biomedical research and healthcare. Oncotarget 2018;9:5665 90. Available from: https://doi.org/10.18632/oncotarget.22345. [7] Roehrs A, da Costa CA, da Rosa Righi R. OmniPHR: a distributed architecture model to integrate personal health records. J Biomed Inf 2017;71:70 81. Available from: https://doi.org/10.1016/j.jbi.2017.05.012. [8] Gaggioli A. Blockchain technology: living in a decentralized everything. Cyberpsychol Behav Soc Netw 2018;21:65 6. Available from: https://doi.org/ 10.1089/cyber.2017.29097.csi.

211

212

CHAPTER 12 Introducing blockchain technology in data storage

[9] Hoy MB. An introduction to the blockchain and its implications for libraries and medicine. Med Ref Serv Q 2017;36:273 9. Available from: https://doi.org/10.1080/ 02763869.2017.1332261. [10] IBM Institute for Business Value. Healthcare rallies for blockchains: keeping patients at the center. IBM Institute for Business Value; 2016. [11] Gammon K. Experimenting with blockchain: can one technology boost both data integrity and patients’ pocketbooks? Nat Med 2018;24:378 81. Available from: https://doi.org/10.1038/nm0418-378. [12] Nugent T, Upton D, Cimpoesu M. Improving data transparency in clinical trials using blockchain smart contracts. F1000Research 2016;5:2541. Available from: https://doi.org/10.12688/f1000research.9756.1. [13] Campbell D. Combining AI and blockchain to push frontiers in healthcare. Macadamian; 2018. ,https://www.macadamian.com/learn/combining-ai-and-blockchain-in-healthcare/. [accessed 25.11. 21]. [14] Xing B, Marwala T. The synergy of blockchain and artificial intelligence. SSRN Electron J 2018. Available from? https://doi.org/10.2139/ssrn.3225357. [15] Krittanawong C, Rogers AJ, Aydar M, Choi E, Johnson KW, Wang Z, et al. Integrating blockchain technology with artificial intelligence for cardiovascular medicine. Nat Rev Cardiol 2020;17:1 3. Available from: https://doi.org/10.1038/s41569019-0294-y. [16] Mascarenhas M. Artificial intelligence and capsule endoscopy: unravelling the future. Ann Gastroenterol 2021. Available from? https://doi.org/10.20524/aog.2021.0606. [17] Minchole´ A, Rodriguez B. Artificial intelligence for the electrocardiogram. Nat Med 2019;25:22 3. Available from: https://doi.org/10.1038/s41591-018-0306-1. [18] Kim SH, Lim YJ. Artificial intelligence in capsule endoscopy: a practical guide to its past and future challenges. Diagnostics 2021;11:1722. Available from: https://doi. org/10.3390/diagnostics11091722. [19] Wiggers K. PatientSphere uses AI and blockchain to personalize treatment plans. VentureBeat; 2018. ,https://venturebeat.com/2018/10/25/patientsphere-uses-ai-andblockchain-to-personalize-treatment-plans/. [accessed 25.11.21]. [20] Popov G. The future of artificial intelligence in healthcare! 2019. [21] Kumar R, Wang W, Kumar J, Yang T, Khan A, Ali W, et al. An integration of blockchain and AI for secure data sharing and detection of CT images for the hospitals. Comput Med Imaging Graph 2021;87:101812. Available from: https://doi.org/ 10.1016/j.compmedimag.2020.101812. [22] Chamola V, Hassija V, Gupta V, Guizani M. A comprehensive review of the COVID-19 pandemic and the role of IoT, drones, AI, blockchain, and 5G in managing its impact. IEEE Access 2020;8:90225 65. Available from: https://doi.org/ 10.1109/ACCESS.2020.2992341. [23] Kangbai JB, Jame PB, MS, Fofanah AB, George A, Briama A, et al. Tracking Ebola through cellphone, internet of things and blockchain technology. Curr Res Integr Med 2018;1(2):19 21. Available from: https://doi.org/10.4172/2529-797X.1000035. [24] Chinese Critical Care Ultrasound Study Group (CCUSG)Peng Q-Y, Wang X-T, Zhang L-N. Findings of lung ultrasonography of novel corona virus pneumonia during the 2019 2020 epidemic. Intensive Care Med 2020;46:849 50. Available from: https://doi.org/10.1007/s00134-020-05996-6.

References

[25] Jabarulla MY, Lee H-N. A Blockchain and artificial intelligence-based, patient-centric healthcare system for combating the COVID-19 pandemic: opportunities and applications. Healthcare 2021;9:1019. Available from: https://doi.org/10.3390/healthcare9081019. [26] Patel V. A framework for secure and decentralized sharing of medical imaging data via blockchain consensus. Health Inform J, 25. SAGE Publications Ltd; 2019, p. 1398 411. Available from: https://doi.org/10.1177/1460458218769699. [27] Tang H, Tong N, Ouyang J. Medical images sharing system based on blockchain and smart contract of credit scores. In: 2018 first IEEE international conference on hot information-centric networking. Shenzhen: 2018, p. 240 1. [28] European Society of Radiology (ESR)Kotter E, Marti-Bonmati L, Brady AP, Desouza NM. ESR white paper: blockchain and medical imaging. Insights Imaging 2021;12:82. Available from: https://doi.org/10.1186/s13244-021-01029-y. [29] Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts HJWL. Artificial intelligence in radiology. Nat Rev Cancer 2018;18:500 10. Available from: https://doi. org/10.1038/s41568-018-0016-5. [30] Geis JR, Brady A, Wu CC, Spencer J, Ranschaert E, Jaremko JL, et al. Ethics of artificial intelligence in radiology: summary of the joint European and North American multisociety statement. Insights Imaging 2019;10:101. Available from: https://doi. org/10.1186/s13244-019-0785-8. [31] Raman B, Chandrasekaran K. Blockchain for radiology. Health Manage Org 2019;19:38 41. [32] Yang YJ. The future of capsule endoscopy: the role of artificial intelligence and other technical advancements. Clin Endosc 2020;53:387 94. Available from: https:// doi.org/10.5946/ce.2020.133. [33] Lau PY, Correia PL. Detection of bleeding patterns in WCE video using multiple features. In: 2007 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). Lyon, France: IEEE; 2007, p. 5601 4. https://doi.org/10.1109/IEMBS.2007.4353616. [34] Giritharan B, Yuan, X, Liu, J, Buckles B, Oh JH, Tang SJ. Bleeding detection from capsule endoscopy videos. In: 2008 30th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). Vancouver, BC: IEEE; 2008, p. 4780 3. https://doi.org/10.1109/IEMBS.2008.4650282. [35] Li B, Meng MQ-H. Computer-aided detection of bleeding regions for capsule endoscopy images. IEEE Trans Biomed Eng 2009;56:1032 9. Available from: https://doi. org/10.1109/TBME.2008.2010526. [36] Pan G, Yan G, Song X, Qiu X. BP neural network classification for bleeding detection in wireless capsule endoscopy. J Med Eng Technol 2009;33:575 81. Available from: https://doi.org/10.1080/03091900903111974. [37] Charisis V, Hadjileontiadis LJ, Liatsos CN, Mavrogiannis CC, Sergiadis GD. Abnormal pattern detection in wireless capsule endoscopy images using nonlinear analysis in RGB color space. Annu Int Conf IEEE Eng Med Biol Soc IEEE Eng Med Biol Soc Annu Int Conf 2010;2010:3674 7. Available from: https://doi.org/ 10.1109/IEMBS.2010.5627648. [38] Fu Y, Zhang W, Mandal M, Meng MQ-H. Computer-aided bleeding detection in WCE video. IEEE J Biomed Health Inf 2014;18:636 42. Available from: https://doi. org/10.1109/JBHI.2013.2257819.

213

214

CHAPTER 12 Introducing blockchain technology in data storage

[39] Hassan AR, Haque MA. Computer-aided gastrointestinal hemorrhage detection in wireless capsule endoscopy videos. Comput Methods Prog Biomed 2015;122:341 53. Available from: https://doi.org/10.1016/j.cmpb.2015.09.005. [40] Leenhardt R, Vasseur P, Li C, Saurin JC, Rahmi G, Cholet F, et al. A neural network algorithm for detection of GI angiectasia during small-bowel capsule endoscopy. Gastrointest Endosc 2019;89:189 94. Available from: https://doi.org/10.1016/j.gie.2018.06.036. [41] Pogorelov K, Suman S, Azmadi Hussin F, Saeed Malik A, Ostroukhova O, Riegler M, et al. Bleeding detection in wireless capsule endoscopy videos—color vs texture features. J Appl Clin Med Phys 2019;20:141 54. Available from: https://doi.org/ 10.1002/acm2.12662. [42] Aoki T, Yamada A, Aoyama K, Saito H, Fujisawa G, Odawara N, et al. Clinical usefulness of a deep learning-based system as the first screening on small-bowel capsule endoscopy reading. Dig Endosc 2020;32:585 91. Available from: https://doi.org/ 10.1111/den.13517. [43] Tsuboi A, Oka S, Aoyama K, Saito H, Aoki T, Yamada A, et al. Artificial intelligence using a convolutional neural network for automatic detection of small-bowel angioectasia in capsule endoscopy images. Dig Endosc 2020;32:382 90. Available from: https://doi.org/10.1111/den.13507. [44] Aoki T, Yamada A, Kato Y, Saito H, Tsuboi A, Nakada A, et al. Automatic detection of blood content in capsule endoscopy images based on a deep convolutional neural network. J Gastroenterol Hepatol 2020;35:1196 200. Available from: https://doi.org/ 10.1111/jgh.14941. [45] Barbosa DJC, Ramos J, Lima CS. Detection of small bowel tumors in capsule endoscopy frames using texture analysis based on the discrete wavelet transform. In: 2008 Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). Vancouver, BC: IEEE; 2008, p. 3012 5. https://doi.org/10.1109/ IEMBS.2008.4649837. [46] Li B-P, Meng MQ-H. Comparison of several texture features for tumor detection in CE images. J Med Syst 2012;36:2463 9. Available from: https://doi.org/10.1007/ s10916-011-9713-2. [47] Zhao Q, Dassopoulos T, Mullin G, Hager G, Meng MQ-H, Kumar R. Towards integrating temporal information in capsule endoscopy image analysis. Annu Int Conf IEEE Eng Med Biol Soc IEEE Eng Med Biol Soc Annu Int Conf 2011;2011:6627 30. Available from: https://doi.org/10.1109/IEMBS.2011.6091634. [48] Vieira PM, Ramos J, Lima CS. Automatic detection of small bowel tumors in endoscopic capsule images by ROI selection based on discarded lightness information. In: 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). EMBC, Milan: IEEE; 2015, p. 3025 8. https://doi. org/10.1109/EMBC.2015.7319029. [49] Yuan Y, Meng MQ-H. Deep learning for polyp recognition in wireless capsule endoscopy images. Med Phys 2017;44:1379 89. Available from: https://doi.org/ 10.1002/mp.12147. [50] Blanes-Vidal V, Baatrup G, Nadimi ES. Addressing priority challenges in the detection and assessment of colorectal polyps from capsule endoscopy and colonoscopy in colorectal cancer screening using machine learning. Acta Oncol 2019;58:S29 36. Available from: https://doi.org/10.1080/0284186X.2019.1584404.

References

[51] Saito H, Aoki T, Aoyama K, Kato Y, Tsuboi A, Yamada A, et al. Automatic detection and classification of protruding lesions in wireless capsule endoscopy images based on a deep convolutional neural network. Gastrointest Endosc 2020;92:144 51. Available from: https://doi.org/10.1016/j.gie.2020.01.054 e1. [52] Kalla R, McAlindon ME, Drew K, Sidhu R. Clinical utility of capsule endoscopy in patients with Crohn’s disease and inflammatory bowel disease unclassified. Eur J Gastroenterol Hepatol 2013;25:706 13. Available from: https://doi.org/10.1097/ MEG.0b013e32835ddb85. [53] Seshamani S, Kumar R, Dassopoulos T, Mullin G, Hager G. Augmenting capsule endoscopy diagnosis: a similarity learning approach. Med Image Comput ComputAssist Interv 2010;13:454 62. Available from: https://doi.org/10.1007/978-3-64215745-5_56. [54] Klang E, Barash Y, Margalit RY, Soffer S, Shimon O, Albshesh A, et al. Deep learning algorithms for automated detection of Crohn’s disease ulcers by video capsule endoscopy. Gastrointest Endosc 2020;91:606 13. Available from: https://doi.org/ 10.1016/j.gie.2019.11.012 e2. [55] Barash Y, Azaria L, Soffer S, Margalit Yehuda R, Shlomi O, Ben-Horin S, et al. Ulcer severity grading in video capsule images of patients with Crohn’s disease: an ordinal neural network solution. Gastrointest Endosc 2021;93:187 92. Available from: https://doi.org/10.1016/j.gie.2020.05.066. [56] Rubio-Tapia A, Ludvigsson JF, Brantner TL, Murray JA, Everhart JE. The prevalence of celiac disease in the United States. Am J Gastroenterol 2012;107:1538 44. Available from: https://doi.org/10.1038/ajg.2012.219. [57] Wang X, Qian H, Ciaccio EJ, Lewis SK, Bhagat G, Green PH, et al. Celiac disease diagnosis from videocapsule endoscopy images with residual learning and deep feature extraction. Comput Methods Prog Biomed 2020;187:105236. Available from: https://doi.org/10.1016/j.cmpb.2019.105236. [58] Ionescu M, Tudor A, Vatamanu O, Apostol S, Vere C. Detection of lumen and intestinal juices in wireless capsule endoscopy. Comput Sci Ser 2013;11:61 5. [59] Segui S, Drozdzal M, Vilarino F, Malagelada C, Azpiroz F, Radeva P, et al. Categorization and segmentation of intestinal content frames for wireless capsule endoscopy. IEEE Trans Inf Technol Biomed 2012;16:1341 52. Available from: https://doi.org/10.1109/TITB.2012.2221472. [60] Pecere S, Milluzzo SM, Esposito G, Dilaghi E, Telese A, Eusebi LH. Applications of artificial intelligence for the diagnosis of gastrointestinal diseases. Diagnostics 2021;11:1575. Available from: https://doi.org/10.3390/diagnostics11091575. [61] Lin Q, Yan H, Huang Z, Chen W, Shen J, Tang Y. An ID-based linearly homomorphic signature scheme and its application in blockchain. IEEE Access 2018;6:20632 40. Available from: https://doi.org/10.1109/ACCESS.2018.2809426. [62] Panel for the Future of Science and Technology. Scientific Foresight Unit (STOA). Blockchain and the General Data Protection Regulation: can distributed ledgers be squared with European data protection law? PE 634.445. Brussels: European Parliamentary Research Service; 2019. Retrieved from https://www.europarl.europa. eu/RegData/etudes/STUD/2019/634445/EPRS_STU(2019)634445_EN.pdf. [63] Yu G, Wang X, Yu K, Ni W, Zhang JA, Liu RP. Survey: sharding in blockchains. IEEE Access 2020;8:14155 81. Available from: https://doi.org/10.1109/ACCESS.2020.2965147.

215

216

CHAPTER 12 Introducing blockchain technology in data storage

[64] Kakavand H, Kost De Sevres N, Chilton B. The blockchain revolution: an analysis of regulation and technology related to distributed ledger technologies. Rochester, NY: Social Science Research Network; 2017. Available from: https://doi.org/ 10.2139/ssrn.2849251. [65] Attique Khan M, Mashood Nasir I, Sharif M, Alhaisoni M, Kadry S, Ahmad Chan Bukhari S, et al. A blockchain based framework for stomach abnormalities recognition. Comput Mater Contin 2021;67:141 58. Available from: https://doi.org/ 10.32604/cmc.2021.013217.

CHAPTER

Magnetic capsule endoscopy: concept and application of artificial intelligence

13

Chen He1, Qiwen Wang1, Xi Jiang2, Bin Jiang2, Yang-Yang Qian1, Jun Pan3 and Zhuan Liao1 1

Department of Gastroenterology, Changhai Hospital, Naval Medical University, Shanghai, P.R. China 2 Department of Gastroenterology, The First Naval Hospital of Southern Theater Command, Guangdong, P.R. China 3 Department of Endoscopy, Hospital de Especialidades, Centro Me´dico Nacional Siglo XXI, Instituto Mexicano del Seguro Social, Mexico City, Mexico

Magnetic capsule endoscopy (CE) (MCE) has gained significant attention since the concept of magnetic control was initially put forward by Carpi in 2006 [1]. After efforts of technological optimization and development, current research mainly focuses on external magnetic field maneuvering [2]. Hand-held MCE, magnetic resonance imaging (MRI)-based MCE, and robotic MCE are the three main ways of achieving external active control of CE.

Types of magnetic capsule endoscopy and their feasibility Hand-held magnetic capsule endoscopy The first hand-held MCE used was developed by Given Imaging Ltd., Israel, in 2010 [3]. The capsule was a modified version of PillCam Colon (11 mm 3 31 mm) with one of its cameras replaced by a magnetic disk. It was positioned and oriented by a large hand-held permanent magnet (100 3 100 3 30 mm2), with a maximum magnetic force of 256 g/cm2. Keller et al. [4] conducted a pilot study in ten healthy volunteers using hand-held MCE, which revealed that seven subjects had good mucosal visibility (75% 90%) and three had moderate visibility (50% 60%). MiroCam Navi (Intromedic, Korea) developed a similar system with an external hand-held magnet based on the modifications of MiroCam small bowel CE [5,6] (Fig. 13.1). Rahman et al. [7] performed a pilot study on 26 volunteers using MiroCam Navi and successfully visualized approximately 88% 100% of the major landmarks of the stomach during gastric examination (esophagogastric junction: 92%, gastric cardia: 88%, Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00009-5 © 2023 Elsevier Inc. All rights reserved.

217

218

CHAPTER 13 Magnetic capsule endoscopy

FIGURE 13.1 MiroCam-Navi system. (A) Real-time viewer; (B) receiver and sensor; (C) hand-held magnet; (D) MiroCam-Navi capsule [9].

fundus: 96%, body: 100%, incisura: 96%, antrum: 96%, and pylorus: 100%). Besides, the OMOM (Jinshan, China) system with a hand-held MCE has demonstrated its maneuverability in examining the upper gastrointestinal (UGI) tract in human models [8]. However, it has not been used in clinical practice. Here, we list the studies on the clinical application of hand-held MCE (Table 13.1), although any well-designed diagnostic study with a large sample size has not been carried out. The hand-held MCE systems mentioned above have not been widely used in clinical practice.

Magnetic resonance imaging-based magnetic capsule endoscopy Olympus Medical Systems Corporation and Siemens Healthcare collaborated to develop an MRI-based MCE in 2010 (Fig. 13.2). The magnetic force used to move the capsule is not more than 0.1 T, which is much smaller than the conventional MRI field (1.5 T) [13]. This technology made MRI-based MCE quieter during working condition, and a cooling system was no longer needed. To prove its feasibility, Rey et al. [14] performed an investigation on 29 volunteers and 24 patients in 2010, and the overall success rate was 98%, with the full visualization of anatomic marks listed as follows: antrum: 98%, body: 96%, fundus: 73%, and cardia: 75%. However, the orientation control of the magnetic capsule was not possible since the base magnetic field B0 direction is fixed. Ko´sa et al. [15] tried to solve this problem by adding flexible tails holding miniature coils, which could

Table 13.1 Overview of studies on diagnostic accuracy of hand-held magnetic capsule endoscopy (MCE). Handheld MCE MirocamNavi

Number of centers

Sample size

Author

Year

Journal

Patients

Location of comparison

Keller et al. [3]

2010

Healthy volunteers

Esophagus

1

10

Rahman et al. [7]

2016

Healthy volunteers

Stomach

1

26

Ching et al. [10]

2018

Gastrointestinal Endoscopy Gastrointestinal Endoscopy Endoscopy

Having recurrent and refractory iron deficiency

1

49

Ching et al. [11]

2019

Gastrointestinal Endoscopy

Having suspected acute upper gastrointestinal bleeding

1

33

Beg et al. [12]

2020

Gastrointestinal Endoscopy

Healthy people and patients having known esophageal disease

Upper gastrointestinal tract Upper gastrointestinal tract Esophagus

1

50

220

CHAPTER 13 Magnetic capsule endoscopy

FIGURE 13.2 General overview of guidance equipment. (A) Magnetic guidance system and control panel; (B) capsule gastroscopy; (C) details of the control panel [9].

propel the magnetic capsule (23 29 mm 3 10 mm) by moving tails according to the magnetic field generated from the alternating currents in the coils. Fringe field navigation (FFN) [16] is a new strategy proposed to make good use of MRI. The fringe fields around an MRI scanner cause large field gradients (2 4 T/m), which can navigate the magnetic object in the entire body area. Changes in patients’ positions are needed to make full use of the immobile fringe field. Dipole field navigation (DFN) [17], which is a modified version of FFN, can generate large fields and gradients without having to move patients. In 2015, a total of 189 symptomatic patients (105 male; mean age 53 years) from two French centers subsequently and blindly received capsule and conventional gastroscopy by nine and six examiners, respectively [18]. The final gold standard was unblinded conventional gastroscopy with biopsy under propofol sedation. Capsule accuracy was 90.5%, with a specificity of 94.1% and a sensitivity of 61.9%. No study on the evaluation of the diagnostic accuracy of MRI-based MCE has been made since 2015. However, the sensitivity achieved in that study was relatively low (61.9%) (Table 13.2), which limited the further clinical application of MRI-based MCE.

Robotic magnetic capsule endoscopy The first robotic platform steering CE performed was developed by the CRIM Lab in 2009 [19]. The robotic arm with a permanent magnet working as an endeffector could move in six degrees of freedom (DOF) and control the movement

Table 13.2 Overview of studies on diagnostic accuracy of magnetic resonance imaging (MRI)-based magnetic capsule endoscopy (MCE). MRI-based MCE Endocapusle MCE

Author

Year

Journal

Patients

Rey et al. [14]

2010

Endoscopy

Rey et al. [13]

2012

Denzer et al. [18]

2015

Gastrointestinal Endoscopy Journal of Clinical Gastroenterology

Healthy people and patients with epigastric pain and/or reflux symptoms Having indications for upper gastrointestinal examination Having upper abdominal complaints

Location of comparison

Number of centers

Sample size

Stomach

1

53

Stomach

1

61

Stomach

2

189

222

CHAPTER 13 Magnetic capsule endoscopy

of the magnet-incorporated capsule swallowed by patients. A camera was incorporated with the capsule to observe the condition inside the GI tract. Ciuti et al. [19] conducted a comparative experiment in vivo and found that compared with hand-held MCE, robotic MCE was better at reaching targets (87% 6 13% vs. 37% 6 14%). The duration of the hand-held MCE examination was shorter, while the robotic system was more accurate and had a higher precision. Another robotic platform was reported by the University of Utah in 2013 [20], in which the 6 DOF robotic arm was upgraded to have a motor-rotatable permanent magnet as an end-effector. The CE inside the GI tract was controlled in 5 DOF. However, this system was still in its investigation stage without being experimented on animals or undergoing clinical trials. In China, research teams such as Ankon Technology Co. Ltd., JIFU Medical Technologies Co. Ltd., and Chongqing Science and Technology Co. Ltd. have made a lot of effort in developing robotic MCE. In 2012, the first robotic MCE was developed by Ankon Technology Co. Ltd. (Fig. 13.3). The guidance magnet robot is a C-arm type robot with two rotational and three translational degrees of freedom. The magnetic field generated by the magnetic robot can reach a maximum of 200 mT and is adjustable. This system was clinically investigated by Liao’s group [21] from Changhai Hospital. Then a multicenter prospective study was performed with 350 patients [22]. MCE detected gastric focal lesions with a sensitivity of 90.4%, specificity of 94.7%, positive predictive value of 87.9%, and

FIGURE 13.3 NaviCam magnetic control system: (A) Guidance magnet robot and computer workstation; (B) Magnetic capsule endoscope; (C) Capsule locator; (D) ESNavi software; and (E) Data recorder [24].

Operation procedure, indications, and contradictions

FIGURE 13.4 Parts of FAMCE. (A) Capsule endoscopy, magnetic-controlled capsule endoscopy, and image recorder; (B) Capsule endoscope controller; (C) Operations Console of FAMCE. FAMCE, Fully automated magnetic capsule endoscopy.

accuracy of 93.4%. This system obtained European Union CE certification and was clinically approved by the State Food and Drug Administration of China (CFDA) and the US FDA. Standing-type MCE (SMCE) (JIFU, China) [23] allowed patients to undergo examination without lying on the bed, which was convenient for some of them. The positive compliance rate, negative compliance rate, and overall compliance rate among SMCE and gastroscopy were 92.0%, 95.5%, and 94.41%, respectively. A third-generation of OMOM system could guide the fully automated MCE (FAMCE) system automatically without manual control by an operator [25] (Fig. 13.4). A study of 114 participants was carried out to compare the safety and efficacy of FAMCE and conventional gastroscopy in detecting gastric lesions. FAMCE performed a good concordance with conventional gastroscopy (99.61%) in screening lesions, suggesting that FAMCE would be an effective method for screening of GI tract. Here we list studies on the clinical application of robotic MCE (Table 13.3).

Operation procedure, indications, and contradictions of magnetic capsule endoscopy A series of clinical studies have explored gastric preparation, operation procedures, and indications based on NaviCam MCE. It has been used in over a hundred medical centers worldwide [29]. Here we are going to introduce the

223

Table 13.3 Overview of studies on diagnostic accuracy of robotic magnetic capsule endoscopy (MCE). Robotic MCE NaviCam MCCG

Location of comparison

Number of centers

Sample size

Author

Year

Journal

Patients

Liao et al. [22]

2016

Having upper abdominal complaints

Stomach

7

350

Zou et al. [26]

2015

Clinical Gastroenterology and Hepatology Endoscopy

Stomach

2

68

Qian et al. [27]

2018

Stomach

1

10

Chen et al. [28]

2019

Having upper abdominal complaints Having superficial gastric neoplasia Healthy people and patients having suspected esophageal disease

Esophagus

1

25

Digestive and Liver Disease Endoscopy

Operation procedure, indications, and contradictions

operation procedure, the indications, and contradictions of MCE based on the robotic MCE developed by Ankon Technology Co. Ltd. An MCE system consists of a guidance magnet robot, a computer workstation with ESNavi software, an endoscopic capsule, a capsule locator, and a data recorder. A robotic system consists of a 5 DOF arm containing two rotational and three translational degrees of freedom. The field strength was up to 0.2 T in a 50 cm3 working area, and the magnet-embedded video capsule (12 mm 3 9 mm 3 28 mm) was maneuvered by 5 30 mT field strengths. The capsule measures 26.8 mm 3 11.6 mm, weighs 4.8 g, and has a permanent magnet inside its dome. Images are captured with a resolution of 480 3 480 pixels. The view angle of the capsule is 140 degree, and the view depth is 0 mm to 60 mm. The battery life of the capsule is more than 8 hours. The live video captured pictures at 0.5 6 frames per second (fps) from a single CMOS sensor (Fig. 13.5). Previous studies have confirmed the safety and feasibility of MCE [26] and a great agreement of diagnosis with conventional gastroscopy [22]. MCE was reported to have a better tolerance and patients’ preference than conventional gastroscopy. Besides, the technology of NaviCam MCE has undergone a great revolution recently, and the observing spectrum and parameters of MCE have improved a lot. Detachable string MCE (DS-MCE) [28] was developed to overcome the limitations of passive control of CE when examining the esophagus, which allowed the examination of the esophagus and the stomach at one time. It has shown high diagnostic accuracy in detecting esophageal diseases in both healthy people and

FIGURE 13.5 Workstation of magnetic capsule endoscopy. (A) Guidance magnet robot and workstation; (B) Endoscopist rotated the capsule until the camera end faced the pylorus; (C) In the duodenal bulb; the 360-degree automatic scanning model was used during the procedure [30].

225

226

CHAPTER 13 Magnetic capsule endoscopy

patients with advanced chronic liver disease (cACLD) [28,31,32]. A secondgeneration MCE (MCE-2) is an upgradation of MCE with a wider field of view, adaptable imaging frequency, better image resolution, and longer battery life [33 36], which performed better diagnosis in UGI examination. In addition, MCE could reduce gastric transit time by assisting the transpyloric passage of the capsule to improve the examination of the small bowel, which added further support to MCE as an approach for small bowel examination [37].

Operation procedure of gastric examination For gastric preparation, patients are required to consume soft foods the day before the examination and fast overnight ( . 8 h). Patients are instructed to ingest 400 mg of simethicone suspension dissolved in 100 mL of water 40 min before the examination [8,10]. In addition, patients are instructed to drink 1000 mL of water 10 min before the examination to optimize gastric distension. After completing the gastric preparation, the operator puts on the data recorder with the help of an assistant. The assistant activates the capsule with the capsule locator. The patient is instructed to assume the supine or left lateral decubitus position and to swallow the capsule with a small amount of water to effectively observe the esophagus and dentate line. Once the capsule enters the stomach, it is lifted away from the posterior wall, rotated, and advanced to the fundus and cardiac regions, followed by the gastric body, angulus, antrum, and the pylorus (Fig. 13.6). During this procedure, position changes such as

FIGURE 13.6 Standardized gastric examination procedure. When the capsule reaches the stomach, the capsule is lifted away from the posterior wall, rotated, and advanced to the fundus (A), long shots (B), and close-ups (C) of the cardiac regions, and then the posterior wall (D), the lesser and greater curvature (E) and anterior wall (F) of the gastric body, followed by the angulus (G) and antrum (H), and finally the pylorus (I) [38].

Operation procedure, indications, and contradictions

FIGURE 13.7 Standard landmarks of upper GI tract: (A) Z-line; (B) Gastric cardia; (C) Fundus; (D) Body; (E) Angulus; (F) Antrum; (G) Pylorus; (H) Duodenal papilla [24]. GI, Gastrointestinal.

supine, prone, left, and right lateral are also helpful in achieving clear observation and smooth transition [38]. For a better understanding of the procedure, we present the standard landmarks of the upper GI tract (Fig. 13.7) and typical GI lesions detected under MCE (Fig. 13.8). If the patient needs a further examination of the small bowel after the capsule has passed through the duodenum, the small intestine mode button under the realtime view interface is pressed. The patient is allowed to leave the hospital along with the data recorder for further collection of images of the small intestine.

Indications and contradictions Conventional gastroscopy (including painless gastroscopy) has more risks of complications for patients suffering from cardiopulmonary malfunction, liver cirrhosis, and other diseases [39]. MCE, which is a noninvasive procedure, has shown its safety and feasibility for patients with a high risk of undergoing conventional gastroscopy [40,41]. MCE can be performed in patients with suspected gastric diseases (Table 13.4). MCE has shown a good agreement with esophagogastroduodenoscopy (EGD) in gastric diseases examination (including gastric cancer), especially in those who were asymptomatic [38,42]. When compared with pathology, the per-patient sensitivity of MCE for superficial gastric neoplasia detection was 100%, and the per-lesion sensitivity was 91.7% [27]. Besides adults, MCE can be used in both pediatric patients above 6 years old [43] and elderly patients up to 94 years old [38,44].

227

228

CHAPTER 13 Magnetic capsule endoscopy

FIGURE 13.8 Typical GI lesions of upper GI tract observed on magnetically controlled capsule endoscopy in different examinees: (A) Polyp; (B) Ulcer; (C) Gastric tumor; (D) Varices of gastric fundus; (E) Erosive gastritis; (F) Vasodilatation; (G) Diverticulum; (H) Reflux esophagitis [24]. GI, Gastrointestinal.

Contraindications for MCE include the same contradictions for CE and MRI examination. Absolute contraindications include patients (1) who are unable or unwilling to tolerate any abdominal surgery, including endoscopic operation; (2) have implanted electronic devices or magnetic metal foreign bodies; (3) have pacemakers that are not MRI compatible; or (4) who are pregnant. Relative contraindications include patients: (1) with known or suspected GI obstruction, stenosis, or fistula; or (2) with dysphagia.

Overview of artificial intelligence and its integration into gastrointestinal practice With the upgradation of MCE, flourishing technologies have shown great potential in combining with MCE. Artificial intelligence (AI), especially represented by the deep learning (DL) models with the use of the convolutional neural network (CNN), has emerged as an efficient and accurate automated image recognition method in various medical fields, showing remarkable performances in lesion identification or differentiation with high sensitivity and specificity. Meanwhile, AI has a vast space for development in versatile clinical practice: multilesion classification, disease grading, prognosis prediction, automatic report generation, quality monitoring, virtual training, and multiscenario applications and

Table 13.4 Indications and contraindications of magnetic capsule endoscopy [45]. Indications

Patients undergoing EGD examination with or without complaints

Best indications

Relative indications

Contraindications

Patients with contraindications for small bowel capsule endoscopy or MRI

Absolute contraindications

Relative contraindications EGD, Esophagogastroduodenoscopy; MRI, magnetic resonance imaging.

Unwilling or unable to tolerate EGD (including EGD with sedation) and at a high risk of EGD Health management for gastric examination (health examination) Preliminary screening for gastric cancer (superficial tumors, etc.) Follow-up of gastric lesions, such as ulcers, polyps, varices, and erosive and atrophic inflammation Drug-related gastrointestinal mucosal injury Noncontact endoscopy examination (including remote control) Acute upper gastrointestinal bleeding with stable hemodynamics Esophageal diseases, such as varices, reflux inflammation, and Barrett’s esophagus Follow-up of partial gastrectomy and minimally invasive endoscopic treatment Duodenal diseases, such as ulcers and polyps Small bowel inspection after gastric examination Unable or unwilling to tolerate any abdominal surgery, including endoscopic operation Electronic devices and magnetic metal objects in the body, such as pacemakers, cochlear implants, drug infusion pumps, and nerve stimulators; but MRI-compatible products are excluded Pregnancy Suspected or known gastrointestinal obstruction, stenosis, and fistula Dysphagia

230

CHAPTER 13 Magnetic capsule endoscopy

interactions. Here, summaries of the latest research, translational applications, and prospects of AI in MCE are presented as follows.

Artificial intelligence: Definition and role in technology enhancement AI, the science of creating machines or software that mimics human intelligence to perform tasks, was first proposed in 1956 and has become a hotspot under intense research and industrial focus since then. With the evolution of huge data diversity, large storage availability, and faster and more powerful computers, AI algorithms have developed from machine learning (ML), such as support vector machines built on given data to make decisions, to DL models with the most popular algorithm of CNN. A CNN constitutes convolutional and mutually fully connected pooling layers. It is characterized by the efficiency of classification and strong adaptability, showing outstanding performance, especially in the visual area. AI has been portrayed to play a pivotal role in product development and service enhancement and as a solution provider over healthcare services.

Development and validation of artificial intelligence systems in gastrointestinal practice Undoubtly, AI will be of widespread use in the medical examination field. However, whether AI implementation can be realistically translated into a trustworthy clinical reality is a critical question when expectations for AI techniques are shifting from preclinical research to commercial products. AI-assisted clinical diagnostic strategy is generally conducted as a four-stage implementing methodology to ensure generalizability. Preclinical trials are first deployed to construct an algorithm model with a given training set, followed by verification of accuracy through a validation set. In the second stage, comparison of accuracy among the different algorithm methods is measured. Then randomized controlled trials (RCT) are conducted to evaluate the gains and risks of novel automatic detection systems compared with existing approaches. In the final stage, large cohort studies have to be performed to evaluate its clinical effectiveness [46]. Another inherent property of AI refers to the "black box", the internal working of which is difficult to be clearly understood. From the end-user perspective, since AI applications largely depend on the training dataset, data transparency will be an essential requirement to prevent the inappropriate use of AI applications. Instead of mastery of in-depth knowledge of AI, familiarity with the training data and the testing scenarios will be a convenient and practical way of using a specific AI system. However, additional efforts should still be expended to guarantee the generalizability such as setting up benchmarking datasets as a gold standard even though AI is put into service.

Current artificial intelligence applications in magnetic capsule

At present, in the field of gastroenterology, AI has become relatively mature in the analysis of images captured by endoscopic devices, including EGD, colonoscopy, and CE. In traditional EGD, clinical AI studies mainly focused on Barrett dysplasia assisted detection, which is already available in some areas of the world, detection of gastric cancers and precursors, evaluation of invasion depth, diagnosis of Helicobacter pylori infection, stomach anatomical classification, and blind spots monitoring [47,48]. In colonoscopy, AI has yielded considerable strides in the automatic detection and classification of colonic polyps. AI-assisted computer-aided detection (CADe) or diagnosis (CADx) for characterization of colorectal polyps under colonoscopy has been approved in the United States, Europe, Japan, and some other counties [49 51]. In CE, state-of-the-art CNN-based systems have been developed to automatically identify various GI diseases such as bleeding, infection, ulcers, polyps, and tumors and monitor capsule localization during CE reading, reducing both oversight and reading time of physicians [52,53].

Current artificial intelligence applications in magnetic capsule endoscopy Various CNN-based methodologies have been published in the classification of digestive lesions in CE with a satisfactory accuracy. Yet few studies have investigated the role of AI in MCE procedure as it has recently become a novel modality mainly used to noninvasively examine the whole stomach with an actively magnetic-controlled system since its first introduction in 2010 (Table 13.5) [3].

Artificial intelligence-assisted magnetic capsule endoscopy localization strategy Since MCE is controlled through magnetic actuation, it allows physicians to visualize the GI tract. Pose tracking of MCE plays a critical role in optimizing operation and maintaining effective magnetic actuation. There are two main localization approaches: external magnetic sensors outside the body and internal sensors inside the capsule. Regardless of sensor approaches, methods have been reported to determine the capsule’s position and orientation. In 2017, Turan et al. [54] proposed the first multisensor fusion-based sequence-to-sequence DL approach, which fused hand-eye calibrated and synchronized RGB camera information. The system was made up of optical flow estimation, CNN-based feature vector extraction and multirate long short-term memory) based sensor fusion. The dataset was documented on five different real pig stomachs and four different commercial endoscopic cameras were employed. The study achieved a reliable and real-time robot localization for both translational and rotational movements during video recording.

231

Table 13.5 Summary of clinical studies using artificial intelligence (AI) in magnetic capsule endoscopy (MCE) field. Study design

Type of AI

Retrospective

ML

Ex vivo pilot

CNN

Evaluation of AIaided detection ability for MCE images

Retrospective

CNN

Demonstration of AI-based real-time diagnostic system in MCE

Prospective pilot

CNN

Reference

Year

Aim

Mewes et al.

2011

Turan et al.

2017

Recognition of gastritis and GI bleeding in stomach Achievement of precise magnetic sensor localization

Xia et al.

2020

Pan et al.

2022

CNN, Convolutional neural network; GI, gastrointestinal; ML, machine learning.

Training set

Validation set

Outcome

100 healthy and 100 diseased MCE images 40,000 frames from 5 pig stomachs with 4 cameras 822,590 images from 697 patients undergoing MCE 34,062 MCE images from 856 patients

300 MCE images from 44 patients (100 gastritis, 100 bleeding, and 100 normal) 20,000 frames from 5 pig stomachs with 4 cameras

Accuracy: Gastritis: 86% Bleeding: 86% Normal: 92%

201,365 images from 100 patients undergoing MCE

Sensitivity: 96.2% Specificity: 76.2% Accuracy: 77.1%

50 patients referred for MCE

Sensitivity for gastric lesions: 98.9% Accuracy for gastric landmarks: 94.2%

Translational and rotational movements: submillimeter precision

Current artificial intelligence applications in magnetic capsule

Artificial intelligence-assisted magnetic capsule endoscopy diagnostic procedure With the revolutionized development of CE, it has been increasingly deployed in the clinical practice of SB diseases. Despite not requiring manipulation by endoscopists, the major limitation of CE, however, is evident, i.e., is long reading time and extremely large amount of video data, usually consuming 1 2 h per case and easily leading to omission of lesions. In 2019, Ding et al. [55] first developed a CNN-based DL model to help in the diagnosis of multiple SB lesions using Ankon CE equipment with the validation of a large multicenter dataset, achieving outstanding performance compared with other AI algorithms over the same period with a high sensitivity (99.90%), a high detection rate (70.91%), and a high time efficiency (5.9 min per case). In the same year, Luo et al. [56] carried out the largest, multicentre, case-control, diagnostic study on AI application in upper GI cancers in China, the most common malignancies worldwide with high mortality rate and overall poor prognosis, to overcome challenges regarding the fluctuated missing rate of gastric cancers highly dependent on endoscopist qualifications and experience. The Gastrointestinal AI diagnostic System (GRAIDS), a DL learning semantic segmentation model capable of automated real-time detecting upper GI cancers, was developed and validated by 1036 496 endoscopy images from 84 424 individuals, obtaining high diagnostic accuracy (0.913 0.918) in upper GI cancer detection with a sensitivity similar to that of expert endoscopists (0.942 vs. 0.945, P 5 .692). Later in 2021, Wu’s team designed the first RCT tandem trial to evaluate the effect of their ENDOANGEL-LD AI system on gastric neoplasm detection using white light endoscopy. The miss rates of gastric neoplasms were notably reduced in the AI-assisted group with an RR of 0.224, indicating AI’s great potential in neoplasm detection and miss rate reduction in clinical practice [57]. The MCE procedure, which is an alternative tool for gastric cancer screening, encounters similar inherent challenges as in GI endoscope, mainly regarding the reliability of MCE, which is wildly influenced by endoscopist operation with variable levels of experience and a lengthy period of reporting time. This faces a critical juncture in the widespread promotion of the procedure, necessitating diagnostic effectiveness with intrinsic robustness regardless of operators and environment. The first attempt of integrating AI techniques into MCE was made by Mewes et al. [58], who presented the two-step computer-assisted diagnostic-procedure (CAPD) algorithm for detecting gastritis and GI bleeding during MCE examination in 2011. The training dataset contained 100 healthy and 100 diseased images. The two-step algorithm was: (1) a region-of-interest segmentation to separate medically relevant sections of the image from parts containing bubbles; (2) a contrast-normalized filtering to identify and localize possible lesions of pathologies and a feature vector for classifying pathologies in an ML approach. The research achieved sensitivity and specificity results well over 80% for 100 healthy images and 100 images for each of the two pathologies from 44 patients. This

233

234

CHAPTER 13 Magnetic capsule endoscopy

initial step offered implications for the future development of AI applications on this imaging modality. In 2021, Liao’s group developed the first CNN-based auxiliary diagnostic system to assist physicians in the diagnosis of gastric lesions of seven categories (erosions, polyps, ulcers, submucosal tumors, xanthomas, normal mucosa, and invalid images) using NaviCam MCE systems [59]. The model was designed with three CNNs to form a pipeline and was illustrated in Fig. 13.9. The first CNN using Resnet-34 architecture categorized images by three types: images with lesions, normal images, and invalid images, The second CNN also used Resnet34 architecture categorizing 5 types of lesions (erosions, polyps, ulcers, submucosal tumors, and xanthomas) and an additional type, which is normal. The former two architectures were skilled at large lesion detection. The third, the fasterregion-based CNN (RCNN) was designed to deal with multiple small lesions, and the output samples are shown in Fig. 13.10. The system trained and tested a total of 1,023,955 MCE images from 797 patients treated at Changhai Hospital, showing good performance with 96.2% sensitivity and 76.2% specificity in gastric focal lesions detection. The receiver operating characteristic curve for all positive images was 0.84. Besides, one image processing time was 44 milliseconds for AI system and 0.38 6 0.29 s for clinicians (P , .001). The AI-assisted diagnostic system showed potential for stable clinical application by greatly reducing physicians’ workload and variations without neglecting important lesions. Furthermore, the novel CNN faster-RCNN-based diagnostic performance exhibited a perfect

FIGURE 13.9 Schematic illustration of overall architecture of AI system consisting of three convolutional neural networks. AI, Artificial intelligence; RCNN, region-based convolutional neural network.

Current artificial intelligence applications in magnetic capsule

FIGURE 13.10 Output of faster-CNN. Faster-RCNN delt with multiple small lesions in an image. (A) and (B) Polyp; (C) and (D) Erosion/Ulcer. CNN, Convolutional neural network; RCNN, regionbased convolutional neural network.

overall detection rate and accuracy, laying the foundation for mass screening application. On the basis of the above-mentioned first-step investigation of AI-based auxiliary diagnostic system in MCE still images, later in 2022, Liao’s group further demonstrated the real-time performance of the smart data service system AI (SDSS-AI)-based diagnostic system on 50 patients referred for MCE at Changhai Hospital, which had been trained by a total of 34062 MCE images from 856 patients (Fig. 13.11) [60]. The overall sensitivity for detecting gastric lesions and accuracy for identifying gastric anatomical landmarks of SDSS-AI was 98.9% and 94.2%, respectively (Fig. 13.12). Image processing time was 94 Ms per

235

236

CHAPTER 13 Magnetic capsule endoscopy

FIGURE 13.11 Monitor interface of AI-based real-time diagnostic system in MCE. AI, Artificial intelligence; MCE, magnetic capsule endoscopy.

FIGURE 13.12 Output samples of real-time identification of gastric lesions marked with blue frames and heat map. (A), (B) Erosion/Bleeding/Ulcer; (C), (D) Polyp/Submucosal tumor.

image. It was the first study that demonstrated the real-time use of AI in MCE, and evaluated the identification ability of gastric anatomical landmarks, bringing lots of excitement for standardizing MCE procedure in the future.

Prospects of artificial intelligence in magnetic capsule endoscopy

Prospects of artificial intelligence in magnetic capsule endoscopy AI has been portrayed as a silver bullet for the revolution of diagnosis of GI endoscopy, especially of CE. The threat of COVID-19 pandemic has led to a greater expectation of noninvasive and contactless MCE being a promising alternative tool for GI cancer screening. Advancement of CNN-based systems in MCE practice will simplify the real-time diagnostic procedures by automatedly analyzing every single frame, preventing missing lesions that may be overlooked otherwise, minimizing more than 90% reading time while reserving an accuracy comparable with experts. Furthermore, AI has been shown to achieve promising innovations in simultaneous localization of capsule trajectory, real-time identification of anatomical landmarks, and classification of various lesions. However, AI techniques still have some limitations before it can be applied clinically. For example, the current training set is mainly based on still images, which cannot be correlated with dynamic changes of the target image, affecting identification accuracy of the location and nature of the lesion; at the same time, most studies are generally based on single-center retrospective trials with limitedscale and usually duplicated dataset, more prospective, case-controlled, and largescale clinical studies are appealed to test the stability and generalizability of system performance; in addition, current research is largely restricted to the limited field of detection regions or lesion types. In the future, well-tailored and realistic preclinical AI-based MCE systems awaiting quick incorporation into multiple clinical scenarios are expected to meet the following improvements: (1) one-step AI-assisted diagnostic systems applying to the whole GI tract (not only confining to stomach, but also includes esophagus, SB, and colorectum); (2) AI-assisted classification of lesions extending to all GI abnormalities (such as bleeding, inflammation, polyps, cancers, etc.) to ensure accurate and general diagnosis; (3) AI-assisted identification extending from still images to real-time videos with automated comprehensive information documentation, including lesion type, number, nature, location, size, depth, suspected pathology, activity and grade; (4) AI-assisted endoscopic treatment such as automated biopsy operation and topical administration of medication; (5) AI-assisted surveillance guidance based on integrated overall information extracted automatedly from MCE reports and electronic medical records including patients’ race, age, gender, medical history, clinical symptoms, laboratory indices, and medication history; (6) AI-assisted quality-control indicators to evaluate the process efficiency and quality involving integrity of mucosa observation, filling degree of GI cavity, cleansing level of GI tract, exposure of anatomical landmarks, and image resolution; (7) AI-assisted telemedicine services such as remote AI manipulation of MCE with autonomous report generation; (8) AI-assisted virtual training on endoscopists, patient autonomous management, and other medical service supports.

237

238

CHAPTER 13 Magnetic capsule endoscopy

References [1] Carpi F, Galbiati S, Carpi A. Magnetic shells for gastrointestinal endoscopic capsules as a means to control their motion. Biomed Pharmacother 2006;60:370 4. [2] Carpi F, Pappone C. Magnetic maneuvering of endoscopic capsules by means of a robotic navigation system. IEEE Trans Biomed Eng 2009;56:1482 90. [3] Keller J, Fibbe C, Volke F, et al. Remote magnetic control of a wireless capsule endoscope in the esophagus is safe and feasible: results of a randomized, clinical trial in healthy volunteers. Gastrointest Endosc 2010;72:941 6. [4] Keller J, Fibbe C, Volke F, et al. Inspection of the human stomach using remotecontrolled capsule endoscopy: a feasibility study in healthy volunteers (with videos). Gastrointest Endosc 2011;73:22 8. [5] Rahman I, Kay M, Bryant T, et al. Optimizing the performance of magnetic-assisted capsule endoscopy of the upper GI tract using multiplanar CT modelling. Eur J Gastroenterol Hepatol 2015;27:460 6. [6] Hale MF, Rahman I, Drew K, et al. Magnetically steerable gastric capsule endoscopy is equivalent to flexible endoscopy in the detection of markers in an excised porcine stomach model: results of a randomized trial. Endoscopy 2015;47:650 3. [7] Rahman I, Pioche M, Shim CS, et al. Magnetic-assisted capsule endoscopy in the upper GI tract by using a novel navigation system (with video). Gastrointest Endosc 2016;83:889 95 .e1. [8] Li ZS, Deng XM, Sun T, et al. China expert consensus on sedative and anesthesia of digestive endoscope diagnosis and treatment. Chin J Pract Int Med 2014;34:756 64. [9] Nam S-J, Lee HS, Lim YJ. Evaluation of gastric disease with capsule endoscopy. Clin Endosc 2018;51:323 8. [10] Ching HL, Hale MF, Kurien M, et al. Diagnostic yield of magnetically assisted capsule endoscopy vs gastroscopy in recurrent and refractory iron deficiency anemia. Endoscopy 2019;51:409 18. [11] Ching HL, Hale MF, Sidhu R, et al. Magnetically assisted capsule endoscopy in suspected acute upper GI bleeding vs esophagogastroduodenoscopy in detecting focal lesions. Gastrointest Endosc 2019;90:430 9. [12] Beg S, Card T, Warburton S, et al. Diagnosis of Barrett’s esophagus and esophageal varices using a magnetically assisted capsule endoscopy system. Gastrointest Endosc 2020;91:773 81. [13] Rey JF, Ogata H, Hosoe N, et al. Blinded nonrandomized comparative study of gastric examination with a magnetically guided capsule endoscope and standard videoendoscope. Gastrointest Endosc 2012;75:373 81. [14] Rey JF, Ogata H, Hosoe N, et al. Feasibility of stomach exploration with a guided capsule endoscope. Endoscopy 2010;42:541 5. [15] Ko´sa G, Jakab P, Sze´kely G, et al. MRI driven magnetic microswimmers. Biomed Microdevices 2012;14:165 78. [16] Tremblay C.C., Martel S., Conan B., et al. Fringe field navigation for catheterization. In: Sixth European conference of the international federation for medical and biological engineering; 2014. [17] Latulippe M, Martel S. Dipole field navigation: theory and proof of concept. IEEE Trans Robot 2015;31:1 11.

References

[18] Denzer UW, Ro¨sch T, Hoytat B, et al. Magnetically guided capsule vs conventional gastroscopy for upper abdominal complaints: a prospective blinded study. J Clin Gastroenterol 2015;49:101 7. [19] Ciuti G, Donlin R, Valdastri P, et al. Robotic vs manual control in magnetic steering of an endoscopic capsule. Endoscopy 2010;42:148 52. [20] Mahoney AW, Abbott JJ. Five-degree-of-freedom manipulation of an untethered magnetic device in fluid using a single permanent magnet with application in stomach capsule endoscopy. Int J Robot Res 2016;35(1 3):129 47. [21] Liao Z, Duan XD, Xin L, et al. Feasibility and safety of magnetic-controlled capsule endoscopy system in examination of human stomach: a pilot study in healthy volunteers. J Interv Gastroenterol 2012;2:155 60. [22] Liao Z, Hou X, Lin-Hu EQ, et al. Accuracy of magnetically controlled capsule endoscopy, compared with conventional gastroscopy, in detection of gastric diseases. Clin Gastroenterol Hepatol 2016;14:1266 73. [23] Lai HS, Wang XK, Cai JQ, et al. Standing-type magnetically guided capsule endoscopy vs gastroscopy for gastric examination: multicenter blinded comparative trial. Dig Endosc 2020;32:557 64. [24] Jiang X, Pan J, Li ZS, et al. Standardized examination procedure of magnetically controlled capsule endoscopy. VideoGIE 2019;4:239 43. [25] Xiao YF, Wu ZX, He S, et al. Fully automated magnetically controlled capsule endoscopy for examination of the stomach and small bowel: a prospective, feasibility, two-centre study. Lancet Gastroenterol Hepatol 2021;6:914 21. [26] Zou WB, Hou XH, Xin L, et al. Magnetic-controlled capsule endoscopy vs. gastroscopy for gastric diseases: a two-center self-controlled comparative trial. Endoscopy 2015;47:525 8. [27] Qian YY, Zhu SG, Hou X, et al. Preliminary study of magnetically controlled capsule gastroscopy for diagnosing superficial gastric neoplasia. Dig Liver Dis 2018;50:1041 6. [28] Chen YZ, Pan J, Luo YY, et al. Detachable string magnetically controlled capsule endoscopy for complete viewing of the esophagus and stomach. Endoscopy 2019;51:360 4. [29] Li ZCD, Eliakim R, et al. Handbook of capsule endoscopy. Dordrecht, The Netherlands: Springer; 2014. [30] Jiang X, Qian YY, Liu X, et al. Impact of magnetic steering on gastric transit time of a capsule endoscopy (with video). Gastrointest Endosc 2018;88:746 54. [31] Qi X, Berzigotti A, Cardenas A, et al. Emerging non-invasive approaches for diagnosis and monitoring of portal hypertension. Lancet Gastroenterol Hepatol 2018;3:708 19. [32] Garcia-Tsao G, Abraldes JG, Berzigotti A, et al. Portal hypertensive bleeding in cirrhosis: risk stratification, diagnosis, and management: 2016 practice guidance by the American Association for the study of liver diseases. Hepatology 2017;65:310 35. [33] Tontini GE, Wiedbrauck F, Cavallaro F, et al. Small-bowel capsule endoscopy with panoramic view: results of the first multicenter, observational study (with videos). Gastrointest Endosc 2017;85:401 8. [34] Monteiro S, de Castro FD, Carvalho PB, et al. PillCam® SB3 capsule: does the increased frame rate eliminate the risk of missing lesions? World J Gastroenterol 2016;22:3066 8.

239

240

CHAPTER 13 Magnetic capsule endoscopy

[35] Takamaru H, Yamada M, Sakamoto T, et al. Dual camera colon capsule endoscopy increases detection of colorectal lesions. Scand J Gastroenterol 2016;51:1532 3. [36] Ou G, Shahidi N, Galorport C, et al. Effect of longer battery life on small bowel capsule endoscopy. World J Gastroenterol 2015;21:2677 82. [37] Luo YY, Pan J, Chen YZ, et al. Magnetic steering of capsule endoscopy improves small bowel capsule endoscopy completion rate. Dig Dis Sci 2019;64:1908 15. [38] Zhao AJ, Qian YY, Sun H, et al. Screening for gastric cancer with magnetically controlled capsule gastroscopy in asymptomatic individuals. Gastrointest Endosc 2018;88:466 74 .e1. [39] Levy I, Gralnek IM. Complications of diagnostic colonoscopy, upper endoscopy, and enteroscopy. Best Pract Res Clin Gastroenterol 2016;30:705 18. [40] Liao Z, Zou W, Li ZS. Clinical application of magnetically controlled capsule gastroscopy in gastric disease diagnosis: recent advances. Sci China Life Sci 2018;61:1304 9. [41] Hu J, Wang S, Ma W, et al. Magnetically controlled capsule endoscopy as the firstline examination for high-risk patients for the standard gastroscopy: a preliminary study. Scand J Gastroenterol 2019;54:934 7. [42] Li J, Ren M, Yang J, et al. Screening value for gastrointestinal lesions of magneticcontrolled capsule endoscopy in asymptomatic individuals. J Gastroenterol Hepatol 2021;36:1267 75. [43] Gu Z, Wang Y, Lin K, et al. Magnetically controlled capsule endoscopy in children: a single-center, retrospective cohort study. J Pediatr Gastroenterol Nutr 2019;69:13 17. [44] Zhang S, Sun T, Xie Y, et al. Clinical efficiency and safety of magnetic-controlled capsule endoscopy for gastric diseases in aging patients: our preliminary experience. Dig Dis Sci 2019;64:2911 22. [45] National Clinical Research Center for Digestive System Diseases (Shanghai) NCfQCoDE. Capsule Endoscopy Collaborative Group of Digestive Endoscopy Branch of Chinese Medical Association, etc. [Chinese Guidelines for Clinical Application of Magnetic Control Capsule Gastroscopy (2021, Shanghai)]. Chin J Digestive Endoscopy 2021;38:949 63. [46] Colli A, Fraquelli M, Casazza G, et al. The architecture of diagnostic research: from bench to bedside research guidelines using liver stiffness as an example. Hepatology 2014;60:408 18. [47] NEC NEC’s AI supports doctors to detect neoplasia in Barrett’s esophagus during endoscopic procedures. NEC; 2021. Available from: http://www.nec.com/en/press/ 202105/global_20210528_01.html. [48] Pannala R, Krishnan K, Melson J, et al. Artificial intelligence in gastrointestinal endoscopy. VideoGIE 2020;5:598 613. [49] Administration UFaD. FDA authorizes marketing of first device that uses artificial intelligence to help detect potential signs of colon cancer. 2021. Available from: http://www.fda.gov/news-events/press-announcements/fdaauthorizes-marketing-firstdevice-uses-artificialintelligence-help-detect-potential-signs-colon. [50] Weigt J, Repici A, Antonelli G, et al. Performance of a new integrated computerassisted system (CADe/CADx) for detection and characterization of colorectal neoplasia. Endoscopy 2022;54:180 4.

References

[51] Mori Y, Kudo SE, Misawa M, et al. Artificial intelligence-assisted colonic endocytoscopy for cancer recognition: a multicenter study. Endosc Int Open 2021;9: E1004 11. [52] Yang YJ. The future of capsule endoscopy: the role of artificial intelligence and other technical advancements. Clin Endosc 2020;53:387 94. [53] Kim S.H., Lim Y.J. Artificial intelligence in capsule endoscopy: a practical guide to its past and future challenges. Diagnostics (Basel) 2021;11(9):1722. [54] Turan M, Shabbir J, Araujo H, et al. A deep learning based fusion of RGB camera information and magnetic localization information for endoscopic capsule robots. Int J Intell Robot Appl 2017;1:442 50. [55] Ding Z, Shi H, Zhang H, et al. Gastroenterologist-level identification of small-bowel diseases and normal variants by capsule endoscopy using a deep-learning model. Gastroenterology 2019;157:1044 54. [56] Luo H, Xu G, Li C, et al. Real-time artificial intelligence for detection of upper gastrointestinal cancer by endoscopy: a multicentre, case-control, diagnostic study. Lancet Oncol 2019;20:1645 54. [57] Wu L, Xu M, Jiang X, et al. Real-time artificial intelligence for detecting focal lesions and diagnosing neoplasms of the stomach by white-light endoscopy (with videos). Gastrointest Endosc 2022;95:269 80. [58] Mewes PW, Neumann D, Licegevic O, et al. Automatic region-of-interest segmentation and pathology detection in magnetically guided capsule endoscopy. Med Image Comput Comput Assist Interv 2011;14:141 8. [59] Xia J, Xia T, Pan J, et al. Use of artificial intelligence for detection of gastric lesions by magnetically controlled capsule endoscopy. Gastrointest Endosc 2021;93(133139):e4. [60] Pan J., Xia J., Jiang B., et al. Real-time identification of gastric lesions and anatomical landmarks by artificial intelligence during magnetically controlled capsule endoscopy. Endoscopy 2022.

241

This page intentionally left blank

CHAPTER

Nonwhite light endoscopy in capsule endoscopy: Fujinon Intelligent Chromo Endoscopy and blue mode

14

Catarina Gomes1, Emanuel Dias2 and Rolando Pinho1 1

Gastroenterology Department, Centro Hospitalar Vila Nova de Gaia/Espinho, Vila Nova de Gaia, Porto, Portugal 2 Gastroenterology Department, Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal

Background Even though capsule endoscopy is a revolutionary method to access small-bowel mucosa, it is not without limitations. The limitations are mainly associated with the low-resolution images compared to conventional endoscopy, as miniaturization of several image components are required, and also a significant proportion of missing lesions result from gut peristalsis and enteric content [1 3]. Moreover, some lesions are captured only in one frame, which makes their assessment difficult and could lead to underestimation of their clinical significance [4]. Image enhancement technologies were developed to overcome these problems to optimize the overall diagnostic yield, which is around 70% with white-light endoscopy (WLE) according to the current literature [5].

White light White light can be visibly split into seven colors, with each color corresponding to a different wavelength. The depth of mucosa penetration depends on the wavelength [6] violet represents the shortest wavelength (400 nm) while red represents the largest (700 nm), with blue, green, and yellow representing the wavelengths in-between. Hemoglobin is responsible for the peak absorption in the blue parcel, which makes the blood vessels look dark brown in contrast to the surrounding mucosa reflecting the blue light. Vascular areas, such as malignancy and inflammation, could also enhance this pattern [7] (Fig. 14.1A, F, K, P).

Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00002-2 © 2023 Elsevier Inc. All rights reserved.

243

244

CHAPTER 14 Nonwhite light endoscopy in capsule endoscopy

FIGURE 14.1 Small-bowel pathological images in capsule endoscopy. (A) Angioectasia WLE, (B) FICE 1, (C) FICE 2, (D) FICE 3, (E) BM; Erosion (F) WLI, (G) FICE 1, (H) FICE 2, (I) FICE 3, (J) BM; celiac disease: (K) WLI, (L) FICE 1, (M) FICE 2, (N) FICE 3, (O) BM; Tumor (P) WLI, (Q) FICE 1, (R) FICE 2, (S) FICE 3, (T) BM. BM, Blue mode; FICE, Fujinon Intelligent Chromo Endoscopy.

Virtual chromoendoscopy in capsule endoscopy Fujinon Intelligent Chromo Endoscopy system Flexible spectral imaging color enhancement [or Fujinon Intelligent Chromo Endoscopy (FICE) Fujinon, Tokyo, Japan] was developed in 2005 and is a digital technology that takes WLE images and mathematically reconstructs the images by enhancing the mucosa according to certain ranges of wavelengths [6]. Unlike Narrow Band Imaging (NBI, Olympus Medical Systems, Tokyo, Japan), which is a real-time system that incorporates optical filters that emit blue (415 nm) and green (540 nm) wavelengths, the FICE system uses an image processing algorithm based on spectral emission methods. In this sense, FICE is a postprocessing technology in opposition to technologies like NBI, which are preprocessing technologies. The chromoendoscopy method has been largely evaluated in the

Virtual chromoendoscopy in capsule endoscopy

assessment of lesions in the upper and lower gastrointestinal (GI) tract as an alternative and/or a complement to dye spray, showing satisfactory results [7]. In flexible endoscopy, the FICE processor generates a vast number of wavelength variations to create different optical images, and this is influenced by several factors, such as the spectrum of the light source, the optical device, and the spectral sensitivity of the sensing element [8]. However, the use of FICE in capsule endoscopy does not require any reengineering of the capsule; it only requires the integration of the software in the workstation. Three spectral wavelengths of the FICE settings for capsule endoscopy can be selected to display a composite color-enhanced image: setting 1 (red 595 mm, green 540 mm, blue 535 mm), setting 2 (red 420 mm, green 520 mm, blue 530 mm), and setting 3: red 595 mm, green 570 mm, blue 415 mm [8,9]. From the several types of FICE settings, setting 1 (Fig. 14.1B, G, L, Q) achieves the preferred appearance of the vascular and mucosal contrast aiming to reduce bile interference, and the overall color appears pale but more natural compared with the other settings [10]. FICE 2 (Fig. 14.1C, H, M, R ) colors bile pigments with cyanogen, emphasizing blood and differences between normal mucosa and lesions, and FICE 3 (Fig. 14.1D, I, N, S) colors blood with magenta and bile pigments with yellow and should be used when bile pigments are mixed with blood [11]. The FICE technology has been incorporated in the reading software (RAPID) of the PillCam capsule endoscopy workstation of Given Imaging (Medtronic, Yokneam, Israel), and enables the reader to select images under these three different settings through a specific menu in the Rapid Reader software (Fig. 14.2).

FIGURE 14.2 Rapid reader software menu in PillCam capsule endoscopy workstation of Given Imaging (Medtronic).

245

246

CHAPTER 14 Nonwhite light endoscopy in capsule endoscopy

Blue mode Blue mode (BM) was introduced by Given Imaging (Medtronic, Yokneam, Israel) in 2007, and it was the first virtual chromoendoscopy (VC) to be added to capsule endoscopy, although its utility in clinical practice has been less discussed [12,13]. The BM enables light in the short-wavelength range of 490 430 nm to be picked up from white-light images (Fig. 14.1E, J, O, T).

Narrow band imaging NBI is an advanced imaging system that applies optical-digital methods to enhance endoscopic images. It is based on the penetration properties of light: shorter wavelengths penetrate only superficially into the mucosa, whereas longer wavelengths can penetrate more deeply. NBI utilizes red, green, and blue filters to modify WLE: the blue light filter (400 430 nm) highlights the capillaries in the superficial mucosa through mean peak absorption of hemoglobin (415 nm), while the green light filter (525 555 nm) penetrates deeper into the mucosa. This results in greater clarity of mucosal surface structures due to the increased contrast between mucosa and superficial vessels, which appear brown/black, improving the visualization of the mucosal surface architecture and the microvascular patterns, which are useful in predicting the histological structure of tissues [14]. In the esophagus, NBI provides better detail in mucosal evaluation compared with WLE, particularly for mucosal and vascular patterns associated with Barrett’s esophagus, dysplasia, and malignancy. Normal intraepithelial capillaries are observed as brown loops originating from a branching vessel, running perpendicularly in the lamina propria and finally reaching the intraepithelial papillae [15]. In the stomach, NBI has also been demonstrated to improve correlation with histological findings, with several patterns useful in the identification of different pathologies, such as intestinal metaplasia, dysplasia, intramucosal cancer, or invasive cancer. Importantly, the normal mucosal pattern on NBI varies according to the gastric location: the normal gastric body shows a regular arrangement of small round pits, surrounded by a regular capillary network with a honeycomb appearance, while normal antral mucosa has a coil-shaped appearance of a subepithelial capillary network [16]. Normal colonic mucosa presents a circular and regular gland and vessel pattern on NBI. Colon inflammation maintains the same pattern but with thicker vessels and variable vascular density, conferring a reddish appearance of the mucosa. Additionally, NBI’s contribution is also valuable in the detection and characterization of colorectal polyps, predicting histology and estimating the depth of invasion of a colorectal cancer [17]. Although NBI technology has made considerable contributions to the detection of pathological changes in the human GI tract, its usage in capsule endoscopy is still very limited.

Evidence of virtual chromoendoscopy in capsule endoscopy

Evidence of virtual chromoendoscopy in capsule endoscopy Fujinon Intelligent Chromo Endoscopy A few years ago, Yung et al. [3] performed a metaanalysis that evaluated the impact of FICE in improving lesion delineation and detection in capsule endoscopy. According to the authors [3], lesion delineation is defined as improved visualization by aiding lesion characterization and enhanced delineation of the lesion surface and/or borders. On the other hand, lesion detection was relative to the average number of lesions detected during the capsule endoscopy recording. In the three studies [13,18,19] concerning lesion delineation, only FICE 1 improved both delineation of angioectasias and erosions/ulcers in 89% and 45% of images, respectively. The FICE 2 mode improved the delineation of less than 50% (43%) of angioectasias images. Conversely, FICE mode 2 for erosions/ulcers and FICE 3 for all images achieved close to no improvement in visualization compared with WLE. However, almost all pooled analyses were associated with high heterogeneity of studies (I2 . 70%). Data from Sato et al. [20] was not included in the metaanalysis since they graded lesion visibility according to a visual analogue scale (VAS; 0 100), and their data concluded that FICE 1 and 2 performed better than WLE for delineation of angioectasias and erosions/ulcers. Rinco´n et al. [21] corroborated some of these data, as they found advantages of FICE 1 for delineation of vascular and inflammatory lesions and of FICE 2 for delineation of vascular lesions. For the lesion detection analysis, the authors included five studies [20,22 25] in which the average number of lesions detected per video was considered by a single reader in a specific mode. The repeated-measure analysis of variance for the detection of angioectasias and erosions/ulcers revealed no differences between WLE and the three FICE modes. Nevertheless, in a study-level interpretation, FICE 1, and essentially FICE 2, seemed to perform better than WLE in the detection of those lesions. Furthermore, there were four other studies identified in the systematic review but not included in the metaanalysis, because in three studies [11,26,27] the average number of lesions was obtained by a consensus of multiple readers and, in one study [28], there was no differentiation by types of lesions. Among those studies, the data from Gupta et al. [28] was the most contradictory, as the authors concluded that FICE modes were not better than WLE for diagnosing significant lesions for obscure GI bleeding. Dias de Castro et al. [29] found that the application of FICE 1 mode in a previous negative procedure enabled the identification of P1 and P2 lesions in 62% and 21% of the patients, respectively. They also demonstrated that among patients who rebled, 81% had P1 lesions and 6% had P2 lesions detected with the FICE 1 mode. As expected, none of the FICE modes seemed to perform better than WLE for tumor/polyp recognition [19 22,26,27] since the FICE system operates with color enhancement and not with pattern enhancement. Moreover, FICE modes could be associated with a higher detection rate of nonsignificant findings and/or false positive lesions [24,25,28,30,31].

247

248

CHAPTER 14 Nonwhite light endoscopy in capsule endoscopy

Blue mode In the study by Krystallis et al. [18], utilization of the BM technology improved lesion delineation for all types of images, whereas the FICE 1 mode only improved the delineation of vascular lesions. Likewise, Abdelaal et al. [32] found that vascular or inflammatory lesion detection by BM was higher than with WLE, and BM conferred an additional advantage with Quick View (QV) at a high reading speed (20 fps) for the detection of vascular lesions, with an overall missing rate of only 4%. Nonetheless, this data was not supported in a previous series [33] evaluating the usefulness of QV BM. As was the case with FICE, studies concerning BM also presented some contradictory data. Cotter et al. [13] showed that BM was only able to improve lesion delineation in around 60% of erosions/ulcers but not with angioectasias or edema/ atrophy. Conversely, in the delineation analysis using a VAS by Sato et al. [20], BM was able to improve visualization for angioectasias but not for erosions/ ulcers. However, in the lesion detection analysis by the same group [20], BM was not better than WLE for all types of lesions. In a study evaluating the role of BM in the small-bowel assessment of Crohn’s disease patients, the authors indicated that BM did not perform better than WLE in calculating the Lewis score [34].

Fujinon Intelligent Chromo Endoscopy and blue mode Rimbas et al. [35] conducted an analysis that evaluated the application of VC (FICE 1, 2, 3 modes and BM) in categorizing difficult-to-interpret suspected ulcerative images and concluded that there was a 16.5% global improvement with its application. FICE 1 and 2 settings were rated as more useful for the recognition of true ulcerative lesions, although both settings were similar to BM for the recognition of image artifacts. Zammit et al. [36] recently studied the role of VC in celiac disease (CD) and found that both FICE settings and BM were not superior to conventional WLE in the delineation of macroscopic CD-related changes. Concerning colon capsule endoscopy, a recent study from Nakazawa et al. [37] aimed to investigate whether a specific color space method (CIELab) could differentiate colorectal polyps using WLE and VC (FICE and BM) by calculating the color difference. The color differences were higher among adenomatous polyps for both FICE and BM but not for hyperplastic polyps, which may play a role in polyp categorization with VC and require further studies to confirm its efficacy.

Narrow band imaging To date, there has been only a few studies attempting to employ NBI in capsule endoscopy, and, in general, the resolution is also not high enough to allow for easy disease diagnoses by clinicians [38,39]. Recently, a study demonstrated that a capsule endoscope with an NBI lens design of 415 and 540 nm outperformed traditional endoscopes and rivaled chromoendoscopes in terms of the rate of accuracy in detecting and screening neoplastic and nonneoplastic intestinal lesions [38].

Evidence of virtual chromoendoscopy in capsule endoscopy

Another NBI capsule endoscope profited from a dedicated dual-mode complementary metal-oxide semiconductor sensor and demonstrated efficacy in assisting effective diagnoses of early GI cancers. Also, experimental results on backside mucosa of a human tongue and pig’s small intestine showed that it can significantly improve the image quality compared with a commercial-of-the-shelf capsule endoscope for GI tract diagnosis [40]. However, a wireless capsule endoscope with the NBI technology has not been presented in the market up till now. Further studies regarding its implementation in capsule endoscopy are needed.

Other virtual chromoendoscopy methods Although the scope of this review is mainly focused on the application of FICE, BM, and NBI, other VC software in MiroCam (IntroMedic, Korea) (Fig. 14.3) and EndoCapsule (EC) type 1 (Olympus, Japan) are there with limited data. A study group evaluating the performance of settings 1 10 of the Augmented Live-Body Image Color-Spectrum Enhancement (ALICE) system from MiroCam (IntroMedic, Korea) revealed that most settings were able to enhance lesion visibility for depressed and flat lesions (setting 2, 3, 6, and 10 for depressed lesions; setting 3, 5, and 10 for flat lesions) (Fig. 14.4) [41]. On the other hand, another chromoendoscopy system from MiroCam, with image enhancement in three color modes (CM1, CM2, and CM3), did not prove to be useful in improving the evaluation and characterization of any elementary lesions (Fig. 14.5) [42].

FIGURE 14.3 Image Enhancement menu in MiroCam capsule endoscopy workstation (IntroMedic).

249

250

CHAPTER 14 Nonwhite light endoscopy in capsule endoscopy

FIGURE 14.4 Angioectasia in capsule endoscopy. (A) WLE, (B) ALICE 1, (C) ALICE 2, (D) ALICE 3, (E) ALICE 4, (F) ALICE 5, (G) ALICE 6, (H) ALICE 7, (I) ALICE 8, (J) ALICE 9, (K) ALICE 10. ALICE, Augmented Live-Body Image Color-Spectrum Enhancement.

FIGURE 14.5 Angioectasia in capsule endoscopy: (A) WLE, (B) CM1, (C) CM2, (D) CM3. CM, Color mode.

References

The EC type 1 is a contrast capsule that is equipped with a white-light lightemitting diode (WL-LED) called contrast image capsule endoscope (CICE), an optical-digital method that increases the intensity in the blue light wavelength range, making it appropriate for hemoglobin visualization, and consequently, increases the contrast between vascular, inflammatory, or polypoid lesions and the surrounding mucosa, which appears brownish in color. Ogata et al. [43] proved that the use of CICE appeared to facilitate the identification of small-bowel erosions, ulcers, and areas of angioectasia compared with WLE. Aihara et al. [44] reported similar findings with CICE to those of NBI in conventional endoscopy. Similarly, Hatogai et al. [45] showed that CICE was effective in enhancing polyp visibility in patients with polyposis syndrome (familial adenomatous polyposis, Cowden syndrome, and Cronkhite-Canada syndrome), although the authors studied only six patients, and the polyp detection rate was not different with this contrast capsule.

Conclusion Only a few observational series have evaluated the benefits of VC in capsule endoscopy. Moreover, these studies varied substantially in terms of study design, selected population, image/video acquisition, capsule endoscope versions, and measured outcomes. Globally, the studies yielded scant results and, at the same time, discordant data, which make it difficult to evaluate the potential role of FICE, BM, or NBI in clinical practice [3,12,13,46]. Yung et al. [3] hypothesized that these miscellaneous results could also be influenced by other factors, such as the absence of a color blindness test for the capsule readers and the fact that most studies did not specify the size and clinical significance of the lesions. Furthermore, lesion delineation and detection improvement are important, but the ultimate goal is to understand if this achievement truly increases the diagnostic yield and could potentially change patient management. Currently, several groups are looking to develop novel tools to overcome the detection methods of static lesions, and data seems to be promising concerning their overall performance [47,48]. Other alternative imaging techniques that rely on the detection of submucosal or transmural pathology, such as ultrasonography, fluorescence imaging, optical coherence tomography, and ionizing radiation, are still in the early stage of development and require further clinical testing [2]. It is expected that in the future, with technical advancement, the capsule endoscopy image quality can be improved by increasing resolution and enhancing VC technology systems.

References [1] Pennazio M, Spada C, Eliakim R, Keuchel M, May A, Mulder CJ, et al. Small-bowel capsule endoscopy and device-assisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Clinical Guideline. Endoscopy. 2015;47(4):352 76.

251

252

CHAPTER 14 Nonwhite light endoscopy in capsule endoscopy

[2] Cummins G, Cox BF, Ciuti G, Anbarasan T, Desmulliez MPY, Cochran S, et al. Gastrointestinal diagnosis using non-white light imaging capsule endoscopy. Nat Rev Gastroenterol Hepatol 2019;16(7):429 47. [3] Yung DE, Boal Carvalho P, Giannakou A, Kopylov U, Rosa B, Rondonotti E, et al. Clinical validity of flexible spectral imaging color enhancement (FICE) in small-bowel capsule endoscopy: a systematic review and meta-analysis. Endoscopy. 2017;49(3):258 69. [4] Koulaouzidis A, Iakovidis DK, Karargyris A, Plevris JN. Optimizing lesion detection in small-bowel capsule endoscopy: from present problems to future solutions. Expert Rev Gastroenterol Hepatol 2015;9(2):217 35. [5] Spada C, McNamara D, Despott EJ, Adler S, Cash BD, Ferna´ndez-Urie´n I, et al. Performance measures for small-bowel endoscopy: a European Society of Gastrointestinal Endoscopy (ESGE) Quality Improvement Initiative. U Eur Gastroenterol J 2019;7 (5):614 41. [6] Pohl J, May A, Rabenstein T, Pech O, Ell C. Computed virtual chromoendoscopy: a new tool for enhancing tissue surface structures. Endoscopy. 2007;39(1):80 3. [7] Negreanu L, Preda CM, Ionescu D, Ferechide D. Progress in digestive endoscopy: Flexible Spectral Imaging Colour Enhancement (FICE)-technical review. J Med Life 2015;8(4):416 22. [8] Van Gossum A. Image-enhanced capsule endoscopy for characterization of small bowel lesions. Best Pract Res Clin Gastroenterol 2015;29(4):525 31. [9] Manfredi MA, Abu Dayyeh BK, Bhat YM, Chauhan SS, Gottlieb KT, Hwang JH, et al. Electronic chromoendoscopy. Gastrointest Endosc 2015;81(2):249 61. [10] Pohl J, Aschmoneit I, Schuhmann S, Ell C. Computed image modification for enhancement of small-bowel surface structures at video capsule endoscopy. Endoscopy. 2010;42(6):490 2. [11] Konishi M, Shibuya T, Mori H, Kurashita E, Takeda T, Nomura O, et al. Usefulness of flexible spectral imaging color enhancement for the detection and diagnosis of small intestinal lesions found by capsule endoscopy. Scand J Gastroenterol 2014;49(4):501 5. [12] Ibrahim M, Van Gossum A. Novel imaging enhancements in capsule endoscopy. Gastroenterol Res Pract 2013;2013:304723. [13] Cotter J, Magalha˜es J, de Castro FD, Barbosa M, Carvalho PB, Leite S, et al. Virtual chromoendoscopy in small bowel capsule endoscopy: new light or a cast of shadow? World J Gastrointest Endosc 2014;6(8):359 65. [14] East JE, Vleugels JL, Roelandt P, Bhandari P, Bisschops R, Dekker E, et al. Advanced endoscopic imaging: European Society of Gastrointestinal Endoscopy (ESGE) Technology Review. Endoscopy 2016;48:1029 45. [15] ASGE Technology Committee, Manfredi MA, Abu Dayyeh BK, Bhat YM, Chauhan SS, Gottlieb KT, et al. Electronic chromoendoscopy. Gastrointest Endosc 2015;81:249 61. [16] Pimentel-Nunes P, Libaˆnio D, Lage J, Abrantes D, Coimbra M, Esposito G, et al. A multicenter prospective study of the real-time use of narrow-band imaging in the diagnosis of premalignant gastric conditions and lesions. Endoscopy 2016;48:723 30. [17] Utsumi T, Iwatate M, Sano W, Sunakawa H, Hattori S, Hasuike N, et al. Polyp detection, characterization, and management using narrow-band imaging with/without magnification. Clin Endosc 2015;48:491 7. [18] Krystallis C, Koulaouzidis A, Douglas S, Plevris JN. Chromoendoscopy in small bowel capsule endoscopy: blue mode or Fuji Intelligent Colour Enhancement? Dig Liver Dis 2011;43(12):953 7.

References

[19] Imagawa H, Oka S, Tanaka S, Noda I, Higashiyama M, Sanomura Y, et al. Improved visibility of lesions of the small intestine via capsule endoscopy with computed virtual chromoendoscopy. Gastrointest Endosc 2011;73(2):299 306. [20] Sato Y, Sagawa T, Hirakawa M, Ohnuma H, Osuga T, Okagawa Y, et al. Clinical utility of capsule endoscopy with flexible spectral imaging color enhancement for diagnosis of small bowel lesions. Endosc Int Open 2014;2(2):E80 7. [21] Rinco´n ON, Merino Rodrı´guez B, Gonza´lez Asanza C, Ferna´ndez-Pacheco PM. [Utility of capsule endoscopy with flexible spectral imaging color enhancement in the diagnosis of small bowel lesions]. Gastroenterol Hepatol 2013;36(2):63 8. [22] Imagawa H, Oka S, Tanaka S, Noda I, Higashiyama M, Sanomura Y, et al. Improved detectability of small-bowel lesions via capsule endoscopy with computed virtual chromoendoscopy: a pilot study. Scand J Gastroenterol 2011;46(9):1133 7. [23] Duque G, Almeida N, Figueiredo P, Monsanto P, Lopes S, Freire P, et al. Virtual chromoendoscopy can be a useful software tool in capsule endoscopy. Rev Espan˜ola de Enfermedades Digestivas 2012;104:231 6. [24] Sakai E, Endo H, Kato S, Matsuura T, Tomeno W, Taniguchi L, et al. Capsule endoscopy with flexible spectral imaging color enhancement reduces the bile pigment effect and improves the detectability of small bowel lesions. BMC Gastroenterol 2012;12(1):83. [25] Boal Carvalho P, Magalha˜es J, Dias de Castro F, Gonc¸alves TC, Rosa B, Moreira MJ, et al. Virtual chromoendoscopy improves the diagnostic yield of small bowel capsule endoscopy in obscure gastrointestinal bleeding. Dig Liver Dis 2016;48(2):172 5. [26] Kobayashi Y, Watabe H, Yamada A, Hirata Y, Yamaji Y, Yoshida H, et al. Efficacy of flexible spectral imaging color enhancement on the detection of small intestinal diseases by capsule endoscopy. J Dig Dis 2012;13(12):614 20. [27] Matsumura T, Arai M, Sato T, Nakagawa T, Maruoka D, Tsuboi M, et al. Efficacy of computed image modification of capsule endoscopy in patients with obscure gastrointestinal bleeding. World J Gastrointest Endosc 2012;4(9):421 8. [28] Gupta T, Ibrahim M, Deviere J, Van Gossum A. Evaluation of Fujinon intelligent chromo endoscopy-assisted capsule endoscopy in patients with obscure gastroenterology bleeding. World J Gastroenterol 2011;17(41):4590 5. [29] Dias de Castro F, Magalha˜es J, Boal Carvalho P, Cu´rdia Gonc¸alves T, Rosa B, Moreira MJ, et al. Improving diagnostic yield in obscure gastrointestinal bleeding— how virtual chromoendoscopy may be the answer. Eur J Gastroenterol Hepatol 2015;27(6):735 40. [30] Minami-Kobayashi Y, Yamada A, Watabe H, Suzuki H, Hirata Y, Yamaji Y, et al. Efficacy of repeat review with flexible spectral imaging color enhancement in patients with no findings by capsule endoscopy. Saudi J Gastroenterol 2016;22(5):385 90. [31] Nakamura M, Ohmiya N, Miyahara R, Ando T, Watanabe O, Kawashima H, et al. Usefulness of flexible spectral imaging color enhancement (FICE) for the detection of angiodysplasia in the preview of capsule endoscopy. Hepato-gastroenterology 2012;59(117):1474 7. [32] Abdelaal UM, Morita E, Nouda S, Kuramoto T, Miyaji K, Fukui H, et al. Blue mode imaging may improve the detection and visualization of small-bowel lesions: a capsule endoscopy study. Saudi J Gastroenterol 2015;21(6):418 22. [33] Koulaouzidis A, Smirnidis A, Douglas S, Plevris JN. QuickView in small-bowel capsule endoscopy is useful in certain clinical settings, but QuickView with Blue Mode is of no additional benefit. Eur J Gastroenterol Hepatol 2012;24(9):1099 104.

253

254

CHAPTER 14 Nonwhite light endoscopy in capsule endoscopy

[34] Koulaouzidis A, Douglas S, Plevris JN. Blue mode does not offer any benefit over white light when calculating Lewis score in small-bowel capsule endoscopy. World J Gastrointest Endosc 2012;4(2):33 7. [35] Rimba¸s M, Negreanu L, Ciobanu L, Bengu¸s A, Spada C, B˘aicu¸s CR, et al. Is virtual chromoendoscopy useful in the evaluation of subtle ulcerative small-bowel lesions detected by video capsule endoscopy? Endoscopy Int Open 2015;3(6):E615 20. [36] Chetcuti Zammit S, McAlindon ME, Ellul P, Rondonotti E, Carretero C, Sanders DS, et al. Improving diagnostic yield of capsule endoscopy in coeliac disease: can flexible spectral imaging colour enhancement play a role? Digestion. 2020;101(4):347 54. [37] Nakazawa K, Nouda S, Kakimoto K, Kinoshita N, Tanaka Y, Tawa H, et al. The differential diagnosis of colorectal polyps using colon capsule endoscopy. Intern Med (Tokyo, Jpn) 2021;60(12):1805 12. [38] Yen C, Lai Z, Lin Y, Cheng H. Optical design with narrow-band imaging for a capsule endoscope. J Healthc Eng 2018;2018 Jan. [39] Tang L., Hu C., Xie K., Cheng C., Liu Z. Optimized design of capsule endoscopy lens based on ZEMAX. In: 2011 IEEE international conference on information and automation. China: Shenzhen; 2011. p. 57 62. [40] Dung L, Wu Y. A wireless narrowband imaging chip for capsule endoscope. IEEE Trans Biomed Circuits Syst 2010;4(6):462 8. Available from: https://doi.org/ 10.1109/TBCAS.2010.2079932. [41] Ryu CB, Song J-Y, Lee MS, Shim CS. Mo1670 does capsule endoscopy with alice improves visibility of small bowel lesions? Gastrointest Endosc 2013;77(5):AB466. [42] Ribeiro da Silva J, Pinho R, Rodrigues A, Ponte A, Rodrigues J, Sousa M, et al. Evaluation of the usefulness of virtual chromoendoscopy with different color modes in the MiroCam® system for characterization of small bowel lesions. GE Portuguese J Gastroenterol 2018;25(5):222 9. [43] Ogata N, Ohtsuka K, Sasanuma S, Ogawa M, Maeda Y, Ichimasa K, et al. White light-emitting contrast image capsule endoscopy for visualization of small intestine lesions: a pilot study. Endosc Int Open 2018;6(3) E315-e21. [44] Aihara H, Ikeda K, Tajiri H. Image-enhanced capsule endoscopy based on the diagnosis of vascularity when using a new type of capsule. Gastrointest Endosc 2011;73 (6):1274 9. [45] Hatogai K, Hosoe N, Imaeda H, Rey JF, Okada S, Ishibashi Y, et al. Role of enhanced visibility in evaluating polyposis syndromes using a newly developed contrast image capsule endoscope. Gut Liver 2012;6(2):218 22. [46] Spada C, Hassan C, Costamagna G. Virtual chromoendoscopy: will it play a role in capsule endoscopy? Dig Liver Dis 2011;43(12):927 8. [47] Iakovidis DK, Koulaouzidis A. Software for enhanced video capsule endoscopy: challenges for essential progress. Nat Rev Gastroenterol Hepatol 2015;12(3):172 86. [48] Costa D, Vieira P, Pinto C, Arroja B, Leal T, Mendes S, et al. Clinical performance of new software to automatically detect angioectasias in small bowel capsule endoscopy. GE Portuguese J Gastroenterol 2021;28(2):87 96.

CHAPTER

Colon capsule endoscopy and artificial intelligence: a perfect match for panendoscopy

15

Tiago Ribeiro1, Ignacio Ferna´ndez-Urien2 and He´lder Cardoso1,3,4 1

Centro Hospitalar Universita´rio de Sa˜o Joa˜o, Porto, Portugal 2 Navarra Hospital Complex, Pamplona, Spain 3 Faculty of Medicine, University of Porto, Porto, Portugal 4 World Gastroenterology Organization Porto Training Center, Porto, Portugal

Introduction Principles of colon capsule endoscopy A colonic disease is a major health issue in several parts of the world, particularly in high-income countries. This is embodied by colorectal cancer (CRC), a leading cause of new cancer cases as well as cancer-related mortality [1], whose screening programs in high-resource settings are mainly focused on conventional colonoscopy (CC) [2]. However, evaluation of the colon by CC has significant drawbacks spanning across the full length of the chain of care: from the patient’s perspective: discomfort, need for sedation, risk of perforation, risk of bleeding; from the practitioner’s perspective: risk of iatrogenic injury with potential medical liability, long procedures and physical strain with the risk of musculoskeletal injuries; from the institutional perspective: high personnel-related and equipment-related costs, increasing waiting lists. Over the course of the new millennium, intense research has been dedicated to minimally invasive alternatives for the endoscopic assessment of the gastrointestinal tract. The ultimate example is the advent of capsule endoscopy (CE) for the study of suspected small bowel diseases. CE is a first-line tool for the assessment of patients with obscure gastrointestinal bleeding, small bowel evaluation in polyposis syndromes, and assessment of inflammatory activity in patients with Crohn’s disease (CD). The evaluation of other topographies of the gastrointestinal tract by a CE solution has started to be explored more recently. Indeed, the first generation of colon CE, CCE-1 (PillCam Colon, Given Imaging, Yokneam, Israel), was introduced in 2006 by Eliakim et al. [3]. The CCE-1 system was the first to combine two cameras in the same device, capturing images at a combined rate of 4 frames per second, Artificial Intelligence in Capsule Endoscopy. DOI: https://doi.org/10.1016/B978-0-323-99647-1.00007-1 © 2023 Elsevier Inc. All rights reserved.

255

256

CHAPTER 15 Colon capsule endoscopy and artificial intelligence

with an overall field of view of approximately 312 degrees. The CCE introduced an ingenious system to overcome the limited battery life, including an initial delay mode of two hours. Later, in 2009, the first generation was upgraded into a second generation CCE-2 (PillCam Colon 2, Given Imaging, Yokneam, Israel) [4]. CCE-2 provided a field of view of 344 degrees between both cameras, and, in addition, this system included an adaptive frame rate providing automatic control of the capture frame rate according to the speed of bowel transit, ranging from 4 to 35 frames per second, with the aim of optimizing the battery energy and mucosal observation. More recently, in 2017, a later-generation panendoscopy device was released. This device was designated PillCam Crohn’s Capsule (Medtronic, Dublin, Ireland) as it was designed for the assessment of the full length of the gastrointestinal disease in patients with CD, therefore contributing to a more accurate staging [5]. This system is based on the PillCam Colon 2 suite, with a similar adaptive frame rate and field of view and a minimum operating time of at least 10 h [5]. Additionally, this novel system included modifications in the reading software, namely renewed applications for capsule localization and evaluation of disease extent and severity [6].

Indications for colon capsule endoscopy/panendoscopy Endoscopic assessment of the colonic mucosa by CE has been proposed as an alternative for patients who are not considered the best candidates for diagnostic CC. More recently, the development of commercially available CE systems for panendoscopy has enabled the possibility of a one-time full-length inspection of the gastrointestinal tract. A recent real-life prospective study from 14 centers in France included data derived from a total of 689 CCE examinations performed over a time span of 5 years [7]. CCE was mainly indicated for colonic exploration of elderly patients (median age, 70 years old) for whom colonoscopy or anesthesia was contraindicated (n 5 307, 44.6%), of patients with previous incomplete colonoscopy (n 5 217, 31.5%), or of patients refusing colonoscopy (n 5 144, 20.9%). The main clinical situation for colon investigation was the presence of bowel symptoms (n 5 208, 30.2%), abdominal pain, diarrhea or constipation, colorectal neoplasia screening in people with personal or family history of CRC or with a previous positive screening test (n 5 178, 25.8%), irondeficiency anemia (n 5 148, 21.5%), and overt bleeding (n 5 103, 15.0%). The role of CCE and panendoscopy in the routine clinical management of patients with suspected gastrointestinal and colonic diseases has yet to be completely defined. Nevertheless, international guidelines have come to endorse its use in specific clinical situations, particularly for CRC screening in average-risk populations [2,8].

Colorectal cancer screening CCE, and, in particular, devices allowing the endoscopic assessment of the entire gastrointestinal tract, constitute a minimally invasive alternative to conventional

Indications for colon capsule endoscopy/panendoscopy

endoscopic modalities, namely colonoscopy. The role of CCE has been most extensively studied in the setting of CRC screening. CC, enabling tissue sampling and lesion resection, is the gold standard for the diagnosis of colorectal neoplasia. Nevertheless, screening protocols centered on CC may present with significant obstacles to patient adherence, a critical factor in population-based screening programs. These hindrances include the fear of pain or discomfort by the patient, concerns regarding the use of sedation and possible procedure complications, unease with the high degree of invasiveness, and loss of working days [9]. These limitations may contribute to the suboptimal uptake rates of screening colonoscopy [10]. In the existing literature, nonadherence rates to screening colonoscopy vary between 20.3% and 44.5% [11,12]. Moreover, although this data can be influenced by socioeconomic status, gender, comorbidities, and healthcare coverage [13 15], some evidence suggests compliance with CC as an entry screening test may be lower compared with surveillance CC [13,16]. CCE was originally designed to provide a minimally invasive, ambulatory alternative to CC, allowing painless observation of the colonic mucosa without the need for sedation or air insufflation, and to mitigate patient discomfort by allowing the simultaneous execution of compatible work or leisure tasks. Validation studies of CCE-1 focused on the detection of polyps in a pool of 84 patients. The sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for the detection of polyps of any size by the expert panel were 56%, 69%, 57%, and 67%, respectively [3]. The development of CCE-2 led to a significant improvement in the performance features. The sensitivity and specificity for the detection of any polyps $ 6 mm increased to 89% and 76%, respectively [4]. Additionally, this second generation of CCE had a sensitivity of 88% and a specificity of 89% for the detection of more significant lesions ($10 mm) [4]. The combined results have highlighted the potential of CCE for CRC screening. Nevertheless, the pilot studies acknowledged some limitations, particularly suboptimal completion rates (78% in CCE-1 study vs 81% in CCE-2 study) [3,4]. Moreover, the pilot study on CCE-2 evaluated the burden of reading these examinations and concluded that more than 25% of videos required more than 50 min to complete the reading [4]. Over the last few years, several countries, particularly in Europe, have been developing organized CRC screening programs. The participation rate is a key factor for the efficiency of such programs, and data suggests that the uptake rate for CRC screening is suboptimal and has decreased in recent years [17,18]. Some authors have suggested that this trend may be associated with the anticipation of discomfort or invasiveness associated with screening CC [19,20]. Several studies have assessed the performance of CCE for CRC screening in different population subgroups, namely average-risk individuals, people with a positive fecal immunological test (FIT), and first-degree relatives of CRC cases or patients. Holleran and coworkers [21] developed a prospective study to assess the screening performance of CCE compared with CC in an FIT-positive population. In their study, polyps were found in 69% of patients (n 5 43) and significant findings (.3

257

258

CHAPTER 15 Colon capsule endoscopy and artificial intelligence

polyps or any polyp .10 mm) were found in 18 patients (29%), compared with 58% and 29% detected by CC, respectively. A single case of CRC was diagnosed and detected both by CCE and CC. The sensitivity, specificity, PPV, and NPV of CCE (vs CC) were 95%, 65%, 79%, and 90%, respectively. The specificity of CCE increased only when significant lesions were considered (96%). The analysis of these studies must take into account that the reference standard (i.e., CC) has a pooled adenoma miss rate of 26%. Therefore false positive results may be overestimated. Moreover, this trend was also seen when the correlation between CCE and CC was assessed (r 5 0.62 for any polyp and r 5 0.84 for significant lesions). As the likelihood of malignancy is low for smaller polyps, the use of metrics for significant lesions may be reasonable and help avoid unnecessary invasive procedures, as in CC [21]. More recently, Parodi et al. [22] conducted a study on first-degree relatives (n 5 177) comparing the performance of CCE against CC, which was segmentally unblinded for the results of CCE. CCE-2 had a per-patient sensitivity and specificity of 91% and 88%, respectively, for polyps . 6 mm. These values modified to 89% and 95%, respectively, for polyps . 10 mm. When the authors restricted the analysis to adenomas the sensitivity and specificity were 95% and 80% for adenomas $ 6 mm and 91% and 92% for adenomas $ 10 mm, respectively. The sensitivity for adenomas with high-degree dysplasia was 82%. A recent comprehensive systematic revision and a metaanalysis, including 12 studies on CCE-2 (n 5 1898 patients), assessed for the accuracy of CCE for the detection of colorectal polyps [17]. Significant heterogeneity between the studies resulted in the restriction of metaanalysis for three outcomes: polyps of any size and polyps $ 6 mm and $ 10 mm. Overall, CCE-2 had a sensitivity and specificity of 85% and a diagnostic odds ratio of 31. Specifically, polyps $ 6 and $ 10 mm were diagnosed by CCE-2 with a sensitivity of 87% and specificity of 88% and 95%, respectively. The diagnostic odds ratio for these lesions were 51 and 136, respectively. Another recent systematic review with metaanalysis included 13 studies (n 5 2328) measuring the performance of CCE in different populations: averagerisk population (n 5 2), positive FIT (n 5 4), first-degree relatives (n 5 2), family history of CRC (n 5 1), and studies that included mixed populations (n 5 4). Performance metrics were calculated for the diagnosis of polyps $ 6 and $ 10 mm. The pooled sensitivities and specificities were concordant with those found in previous studies (sensitivities of 87% and specificities of 87% and 95%, for $ 6 and $ 10 mm, respectively). These authors calculated the pooled sensitivities and specificities for the different population segments and demonstrated that the performance of CCE was similar for different groups of patients, with values of sensitivity and specificity of over 80% [23]. One of the main concerns regarding the adoption of CCE as a screening method for CRC is the relatively high rate of incomplete examinations. The rate of complete examinations reported in the literature varies greatly (54% 100%) [17,24,25]. The variability in these completeness rates appears to be significantly linked to the lack of standardization of bowel preparation and booster regimens.

Indications for colon capsule endoscopy/panendoscopy

Holleran et al. reported in their prospective study an overall completion rate of 73%, with a significantly higher rate for patients undergoing NaP booster (88%) versus picosulfate (70%) [21]. In addition to an adequate screening performance, wider adoption of CCE as a valid screening method will depend on the demonstration that CCE can improve screening uptake. Parodi et al. reported that 41% of their cohort preferred CCE over CC, which contrasted with the 23% who preferred CC over CCE [22]. Nevertheless, studies evaluating patients’ preferences do not allow practitioners to draw definite conclusions. In a trial on first-degree relatives that randomly offered CCE or CC for CRC screening, where participants could switch across to the other arm, more patients chose CC instead of CCE [26]. Nevertheless, these results may not reflect the overall reality as this population may have a higher awareness of the risk of colorectal neoplasia and are, therefore, favor standard optical colonoscopy that allows for the detection and resection of colorectal polyps. Moreover, the need to undergo a second bowel preparation in case of a positive CCE may limit the acceptability of CCE [17]. Indeed, the need for rigorous bowel preparation seemed to be the main limitation of CCE [27]. Although these studies aim to assess patient perspectives regarding screening with CCE, they do not allow to assess the impact of CCE in the screening uptake of an opportunistic or population-based program. Groth and coworkers assessed CCE as a primary screening tool in an opportunistic screening setting in Germany [20]. The authors assessed screening uptake in a geographical area with low participation rates (1% annually). This study showed a fourfold increase in the uptake rate when CCE was offered as the primary screening tool [20].

Inflammatory bowel disease CE has a recognized role in the management of patients with suspected or established CD. The advent of CCE, allowing colonic exploration and panendoscopy, has generated the hope that these new systems would allow a more accurate evaluation of gastrointestinal involvement by CD [28]. The concept of treat-to-target underlying IBD management frequently requires endoscopic evaluations to assess mucosal healing. In this sense, a single-time panenteric evaluation is appealing. A feasibility study using PillCam Colon 2, involving a cohort of patients under corticosteroid-free clinical remission (n 5 12), reported that this system allowed panendoscopy in 83% of patients (n 5 10). Importantly, the authors found significant inflammatory activity in 75% of patients [29]. More recently, the clinical impact of the panendoscopic evaluation in CD patients was evaluated in a multicenter study, which included 93 patients, 71 with established and 22 with suspected CD, submitted to PillCam Crohn’s Capsule (PCC) [6]. Active disease was found in 71% of symptomatic patients with established CD. Significantly, active disease was found in 58% of patients with clinically remitting CD. PCC led to the upstaging of CD in a third of those with established CD, particularly by the detection of a higher prevalence of upper gastrointestinal involvement, with an increase

259

260

CHAPTER 15 Colon capsule endoscopy and artificial intelligence

in the notification of L4 disease (Montreal classification) from 7.0% to 19.7%. Management was ultimately modified after PCC in 39% of patients. This data highlights the potential of panenteric systems for the simplification of the evaluation of patients with known or suspected CD, as it allows the inspection of the upper gastrointestinal tract, small bowel, and colon in a single test. Several scores have been developed for the classification of CD activity. The Capsule Endoscopy Crohn’s Disease Activity Index (CECDAI) or the Niv score is one of the most commonly used scores that allows for estimating the activity in patients with isolated small bowel disease. More recently, a score has been developed, Eliakim’s score, for panenteric characterization of CD activity in PCC, which integrates the most common lesions, the most severe lesions, the extent of the disease, and the presence of strictures. This score has shown to have an excellent interobserver correlation and correlated better with other markers of mucosal inflammation (e.g., fecal calprotectin) [30]. This is particularly relevant since, to this day, patients are submitted to separate colonic and small bowel evaluations, the latter usually by radiology examinations (either computed tomography or magnetic resonance enterography). Hence, panenteric systems may also allow the avoidance of repeated exposure to radiation and overcome the limited sensitivity of radiology in recognizing inflammatory activity in the proximal segments of the small bowel [31]. The utilization of CCE for the evaluation of mucosal healing in patients with ulcerative colitis has been explored since its first steps. Existing evidence is consensual in that CCE cannot be recommended for the initial diagnosis of ulcerative colitis as it is intrinsically dependent on histopathology. Hence, most studies on the application of CCE to ulcerative colitis have focused on its ability to assess disease activity, which is useful as a potential tool to guide therapies [32]. In a multicenter study by Sung et al. the first generation of CCE was able to detect ulcerative colitis activity with a sensitivity of 89% and a specificity of 75% [32]. These findings are in line with those reported by Ye and coworkers, which documented a good correlation between disease activity (r 5 0.75) and extension (r 5 0.52) evaluated by CCE-1 and CC [33]. Studies on the second generation of CCE agree with those of the first generation. Hosoe et al. evaluated the correlation between CCE-2 and CC in the evaluation of inflammatory activity based on Matts score, which was good for each segment of the colon (r between 0.67 and 0.91), although suboptimal bowel preparation was acknowledged as a limitation by the authors [34]. These results were confirmed by the same group using a different bowel preparation protocol [35]. More recently, the role of panendoscopy was evaluated in the setting of ulcerative colitis. Adler et al. used the PillCam Crohn’s Capsule to determine ulcerative colitis activity (based on Mayo’s endoscopic subscore) and extent and its correlation to CC [36]. PCC had an excellent correlation with CC regarding disease activity (95.7%, k 5 0.86) and a good overall agreement estimating the extent of mucosal inflammation (56.5%, k 5 0.42). When proctitis and left colitis were grouped together the agreement rate for extent was 78.3% (k 5 0.61). This panendoscopy system allowed the detection of small

Limitations of colon capsule endoscopy

bowel findings in 13% of patients (3 out of 23 enrolled): 2 had inflammatory activity in the distal ileum and 1 had jejunal ulceration, leading to a modification of the diagnosis of the latter from ulcerative colitis to CD [36].

Gastrointestinal bleeding and anemia Obscure gastrointestinal bleeding, either overt or occult, is the most common indication for conventional CE [37,38]. Indeed, the routine use of CE in patients with gastrointestinal bleeding with negative esophagogastroduodenoscopy and colonoscopy led to the adoption of the nomenclature mid-gastrointestinal bleeding, as most originate from the small bowel [39]. The role of panendoscopy systems in this setting has not been established and evidence remains at early stages. Mussetto et al. enrolled patients with melena and negative esophagogastroduodenoscopy (n 5 12) and offered them investigation by PillCam Colon 2 previous to routine colonoscopy [40]. Complete evaluation of the full length of the gastrointestinal tract was possible in 11 cases. The diagnostic yield of CCE was 83.3%. Most findings were located in the small bowel (41.7%) and colon (33.3%), and device-assisted enteroscopy was performed in half of the cases. Therefore the authors concluded that minimally invasive panendoscopy devices have a high diagnostic yield and may contribute to avoiding unnecessary colonoscopies [40]. However, more data is required before recommendations on the use of this tool in this setting can be made. Considering the advantages of CCE, one can speculate that in patients with iron-deficiency anemia who are poor candidates for conventional endoscopic procedures or who are reluctant to undergo these procedures, CCE may constitute to be a more patient-friendly first endoscopic approach.

Limitations of colon capsule endoscopy CCE and, particularly, panendoscopy systems appear to be safer and minimally invasive alternatives to conventional endoscopic examinations. Nevertheless, the existing evidence suggests that these systems have limitations that have not been completely addressed, thus limiting their applicability. Several requisites are yet to be optimized at preprocedure, intraprocedure, or postprocedure stages. The performance of CCE and panenteric CE examinations require extensive bowel preparation since it is not capable of cleaning existing residues. A recent metaanalysis concluded that the rates of adequate bowel cleansing fall short of those of CC [41]. Bowel preparation is dependent on a large number of variables, including diet prior to the CCE examination, laxative type, volume and timing of administration, and the type and number of boosters. Bowel preparations based on polyethylene glycol (PEG) solutions combined with NaP boosters are the most common strategy for CCE. Although there is a trend toward higher completion

261

262

CHAPTER 15 Colon capsule endoscopy and artificial intelligence

rates and adequate cleansing with PEG 1 NaP, a recent metaanalysis did not show statistically significant differences [41]. These results must be interpreted with caution, as significant heterogeneity exists between studies. Additionally, although the rate of complications during CCE examinations is low, mild complications (e.g., nausea) are most frequently associated with bowel preparation [42]. Another significant limitation of these capsule systems is the inability to acquire specimens for histologic assessment as well as to perform therapeutic procedures. One of the most commonly cited limitations of panendoscopy examinations is regarding the time required for reading and reporting these examinations. Indeed, reading CCE and panendoscopy examinations is time-consuming as each examination may produce over 50,000 frames, requiring approximately one hour for reading and reporting a single examination [4,27,43]. The long reading times limit the applicability of this technique to routine clinical practice. Moreover, the reader’s fatigue when reviewing multiple examinations is a known factor that may lead to an increase in the miss lesion rate of CE examinations [44]. The final hurdles to overcome to generalize the use of the panendoscopy system are the significant cost associated with these techniques [45] and the lack of centers with expertise in the performance of these examinations. The introduction of new techniques to simplify the reading process is expected to potentiate the use of panendoscopy systems in low-volume centers. Finally, the increasing use of these techniques is expected to lower the per-examination cost.

Impact of artificial intelligence The development of artificial intelligence (AI) algorithms has been the focus of intense research across different medical specialties, particularly those intrinsically dependent on image analysis [46 48]. Convolutional neural networks (CNN) are a subtype of AI architecture that resembles the animal visual cortex and is particularly designed for automatic image analysis (Chapter 1). These systems have been studied in the setting of conventional small bowel CE with promising results for the detection of several types of lesions (Chapters 7 11), and commercial solutions, including deep learning models, are now available (NaviCam SB system, AnX Robotica Corp., Plano, Texas, USA), although not yet cleared for commercialization by the Food and Drug Administration (FDA). The evidence regarding the application of deep learning modules to CCE or other panendoscopy systems is still in its early stages, and, to this date, this stems mainly from retrospective studies including a small number of patients and with limited datasets (mostly using still frames instead of full-length videos). Most studies evaluate the performance of deep neural networks for the detection of colorectal neoplasia. Blanes-Vidal and coworkers were the first to evaluate the performance of a CNN for the automatic detection of polyps in a CCE system (PillCam Colon 2). The authors collected images from polyps detected by CCE

Impact of artificial intelligence

(n 5 375) and confronted the images with results from optical colonoscopy [49], and integrated them into an adaptation of a preexistent CNN (AlexNet), which demonstrated a high accuracy (96%), sensitivity (97%), and specificity (93%) for the detection of colorectal polyps. Their work was followed by that of Yamada et al., who developed a CNN-based model based on the CNN Single Shot MultiBox Detector for the diagnosis of colorectal neoplasia. The authors used data from a pool of 184 patients, including a total of 20,717 frames divided for the constitution of training and validation datasets (n 5 15,933 and n 5 4784, respectively). The authors reported a sensitivity, specificity, and area under the curve (AUC) of 79%, 87%, and 0.90, respectively, for the detection of colorectal polyps and cancers [50]. The Saraiva et al. group conducted a proof-of-concept study where they developed a deep learning algorithm that was able to detect colonic protruding lesions, including a total of 24 CCE-2 examinations, from which 3640 frames (860 showing protruding lesions and 2780 showing normal mucosa or other mucosal lesions) [51]. The authors included polyps, epithelial tumors, and subepithelial lesions under the label “protruding lesions” and reported a sensitivity of 91%, a specificity of 93%, and an AUC of 0.97 for the detection of these lesions. They further evaluated the potential role of AI in increasing the diagnostic yield of CCE and in the setting of gastrointestinal bleeding, which may be of particular importance in clinical settings where the performance of CC is unwanted or unfeasible. Saraiva and coworkers developed a pilot CNN for automatic detection of blood or hematic residues in CCE images. Their algorithm detected blood and hematic traces with a sensitivity of almost 100%, a specificity of 93%, and an overall accuracy of 97% [52]. The study was followed by a multicentric study in which the authors designed a trinary network aiming to detect and differentiate normal colonic mucosa, blood, and hematic residues as well as mucosal lesions that included ulcers and erosions, vascular lesions (red spots, angiectasia, and varices), and protruding lesions (as defined previously) [53]. For the development of this network, the authors included 9,005 frames from 124 patients from two highvolume CE centers. Overall, the CNN had a sensitivity and specificity of 96% and 98%, respectively. The network provided accurate predictions in 98%. Specifically, the network was able to detect and differentiate each of the categories with sensitivities and specificities of over 90%. In addition to potential applications for automatic detection of colorectal neoplasia and in the setting of gastrointestinal bleeding, studies have emerged evaluating the potential of deep learning for automatic detection and classification of mucosal inflammation in the setting of IBD. These studies have applied AI algorithms to the novel PillCam Crohn’s Capsule. Ferreira et al. provided the framework of a CNN for automatic detection of both small bowel and colonic ulcers and erosions using PCC images [54]. This proof-of-concept study involved 59 PCC examinations from two centers in Portugal, including a total of 24,675 frames showing enteric or colonic mucosa (5,300 showing ulcers or erosions). The total pool was divided for the constitution of the training and validation

263

264

CHAPTER 15 Colon capsule endoscopy and artificial intelligence

dataset. Overall, this CNN had a sensitivity, specificity, and accuracy of 98%, 99%, and 99%, respectively. The AUC for the detection of ulcers and erosions was approximately 1.00. Recently, a Danish group evaluated the performance of an automated algorithm for the automatic detection and classification of the severity of lesions associated with CD (normal mucosa, nonulcerated inflammation, aphthous ulceration, ulcer, and extensive ulceration), using small bowel and colon images from the PCC system [55]. The authors included 7744 (2772 from the colon) images of 38 patients. A total of 2748 frames included at least one mucosal break. Their algorithm was able to predict the presence of mucosal breaks compatible with CD with a sensitivity of 96%, a specificity of 100%, and an accuracy of 99%. These figures were similar for the detection of nonerosive mucosal inflammation and did not differ significantly between small bowel or colonic lesions. Their model was able to differentiate the small bowel from colonic lesions with an overall good agreement. Moreover, this deep learning tool showed substantial agreement for the classification of the severity of mucosal lesions compared with manual classification (k 5 0.72). The introduction of AI systems for assistance in the revision of panendoscopy examinations may help to improve some of the limitations that limit the generalizability of these endoscopic systems. The most evident limitation to be tackled by the implementation of AI software is the time required for the revision of each procedure, which is expected to drop significantly. The optimization of the reading and reporting process will contribute significantly to a more widespread utilization of these techniques. These improvements may even allow for consecutive therapeutic CC with the same bowel preparation of CCE. Consequently the adoption of AI-assisted capsule panendoscopy may be a true game changer as the potential increase in the use of CCE due to implementation may ultimately tackle its financial costs by decreasing CCE system unit price. Moreover, capsule panendoscopy is not widely available, and most endoscopists are not familiar with reviewing CCE images. Acquiring expertise in reviewing and reporting panendoscopy examinations is time-consuming and requires intensive training [56]. The introduction of AI-assisted CCE image review may enhance the acquisition of competences in CCE reading, thus shortening the learning curve for unexperienced gastroenterologists. This is particularly important in centers with a low volume of CE examinations, increasing the accessibility to underserved or outreach regions.

Future directions AI is proposed to expand the clinical role of CCE in routine practice, provided it becomes more accessible and affordable. Indeed, deep neural networks are expected to increase the diagnostic yield of CCE and mitigate the drawbacks

References

associated with this technique, thus enhancing its use for the minimally invasive investigation of the GI tract. Additionally, the implementation of these techniques will be helpful in providing support to clinicians with less expertise in reading panendoscopy examinations. These hypotheses should be tested in head-to-head comparison studies to evaluate the impact of the introduction of automated technologies in this field. Nevertheless, one of the most anticipated impacts of the application of deep learning tools to CCE refers to its use as a tool for CRC screening. Indeed, current guidelines already endorse the use of CCE as an alternative for screening in average-risk populations. The use of AI may help to democratize access to this minimally invasive technique in screening settings. This is particularly helpful in the postpandemic state in which access to healthcare facilities is limited, generating a significant delay in the performance of screening colonoscopies. CCE will be potentially helpful in recovering from long waiting lists for CRC screening examinations. This is consummated by the pilot project of the National Health Service (United Kingdom), which launched a pilot project to screen 11,000 patients for CRC with CCE [57]. AI may potentiate these programs by streamlining the time-consuming process of capsule reading and reporting. The introduction of AI to these endoscopic systems should be accompanied by studies assessing its impact on the diagnostic yield, the time required for the performance of each examination, patient compliance and satisfaction, and the impact on endoscopy services.

Conclusion The introduction of AI in clinical practice is expected to be a significant evolution in routine clinical practice. Although practitioners often regard the introduction of disruptive tools with excitement, this technologic innovation must be integrated within ethical and regulatory standards in a planned, stepwise, controlled, and supervised program [58], which must simultaneously improve the diagnostic process and maintain the trust of patients and professionals alike. Scientific societies are expected to publish position papers or guidelines regarding AI implementation in clinical practice. However, with the improvement in panendoscopy enabled by AI, its use can now be significantly expanded as a minimally invasive procedure for the assessment of the entire GI tract for the benefit of patients at large.

References [1] Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2021;71(3):209 49. [2] Shaukat A, Kahi CJ, Burke CA, Rabeneck L, Sauer BG, Rex DK. ACG clinical guidelines: colorectal cancer screening 2021. Am J Gastroenterol 2021;116(3).

265

266

CHAPTER 15 Colon capsule endoscopy and artificial intelligence

[3] Eliakim R, Fireman Z, Gralnek IM, Yassin K, Waterman M, Kopelman Y, et al. Evaluation of the PillCam Colon capsule in the detection of colonic pathology: results of the first multicenter, prospective, comparative study. Endoscopy. 2006;38(10):963 70. [4] Eliakim R, Yassin K, Niv Y, Metzger Y, Lachter J, Gal E, et al. Prospective multicenter performance evaluation of the second-generation colon capsule compared with colonoscopy. Endoscopy. 2009;41(12):1026 31. [5] Tontini GE, Rizzello F, Cavallaro F, Bonitta G, Gelli D, Pastorelli L, et al. Usefulness of panoramic 344 -viewing in Crohn’s disease capsule endoscopy: a proof of concept pilot study with the novel PillCamt Crohn’s system. BMC Gastroenterol 2020;20(1):97. [6] Tai FWD, Ellul P, Elosua A, Fernandez-Urien I, Tontini GE, Elli L, et al. Panenteric capsule endoscopy identifies proximal small bowel disease guiding upstaging and treatment intensification in Crohn’s disease: a European multicentre observational cohort study. United European Gastroenterol J 2021;9(2):248 55. [7] Benech N, Vinet O, Gaudin JL, Benamouzig R, Dray X, Ponchon T, et al. Colon capsule endoscopy in clinical practice: lessons from a national 5-year observational prospective cohort. Endosc Int Open 2021;9(10) E1542 E8. [8] Spada C, Hassan C, Galmiche JP, Neuhaus H, Dumonceau JM, Adler S, et al. Colon capsule endoscopy: European Society of Gastrointestinal Endoscopy (ESGE) Guideline. Endoscopy. 2012;44(5):527 36. [9] Eliakim R, Adler SN. Colon PillCam: why not just take a pill? Dig Dis Sci 2015;60 (3):660 3. [10] Kanth P, Inadomi JM. Screening and prevention of colorectal cancer. BMJ. 2021;374:n1855. [11] Badurdeen DS, Umar NA, Begum R, Sanderson 2nd AK, Jack M, Mekasha G, et al. Timing of procedure and compliance with outpatient endoscopy among an underserved population in an inner-city tertiary institution. Ann Epidemiol 2012;22 (7):531 5. [12] Kazarian ES, Carreira FS, Toribara NW, Denberg TD. Colonoscopy completion in a large safety net health care system. Clin Gastroenterol Hepatol 2008;6(4):438 42. [13] Shuja A, Harris C, Aldridge P, Malespin M, de Melo Jr. SW. Predictors of no-show rate in the GI endoscopy suite at a safety net academic medical center. J Clin Gastroenterol 2019;53(1):29 33. [14] O’Neil J, Winter E, Hemond C, Fass R. Will they show? Predictors of nonattendance for scheduled screening colonoscopies at a safety net hospital. J Clin Gastroenterol 2021;55(1):52 8. [15] Weiss JM, Smith MA, Pickhardt PJ, Kraft SA, Flood GE, Kim DH, et al. Predictors of colorectal cancer screening variation among primary-care providers and clinics. Am J Gastroenterol 2013;108(7):1159 67. [16] Greenspan M, Chehl N, Shawron K, Barnes L, Li H, Avery E, et al. Patient nonadherence and cancellations are higher for screening colonoscopy compared with surveillance colonoscopy. Dig Dis Sci 2015;60(10):2930 6. [17] Kjølhede T, Ølholm AM, Kaalby L, Kidholm K, Qvist N, Baatrup G. Diagnostic accuracy of capsule endoscopy compared with colonoscopy for polyp detection: systematic review and meta-analyses. Endoscopy. 2021;53(7):713 21. [18] Brenner H, Altenhofen L, Stock C, Hoffmeister M. Prevention, early detection, and overdiagnosis of colorectal cancer within 10 years of screening colonoscopy in Germany. Clin Gastroenterol Hepatol 2015;13(4):717 23.

References

[19] Thygesen MK, Baatrup G, Petersen C, Qvist N, Kroijer R, Kobaek-Larsen M. Screening individuals’ experiences of colonoscopy and colon capsule endoscopy; a mixed methods study. Acta Oncol 2019;58(sup1) S71 S6. [20] Groth S, Krause H, Behrendt R, Hill H, Bo¨rner M, Bastu¨rk M, et al. Capsule colonoscopy increases uptake of colorectal cancer screening. BMC Gastroenterol 2012;12:80. [21] Holleran G, Leen R, O’Morain C, McNamara D. Colon capsule endoscopy as possible filter test for colonoscopy selection in a screening population with positive fecal immunology. Endoscopy. 2014;46(6):473 8. [22] Parodi A, Vanbiervliet G, Hassan C, Hebuterne X, De Ceglie A, Filiberti RA, et al. Colon capsule endoscopy to screen for colorectal neoplasia in those with family histories of colorectal cancer. Gastrointest Endosc 2018;87(3):695 704. [23] Mo¨llers T, Schwab M, Gildein L, Hoffmeister M, Albert J, Brenner H, et al. Secondgeneration colon capsule endoscopy for detection of colorectal polyps: systematic review and meta-analysis of clinical trials. Endosc Int Open 2021;9(4) E562 E71. [24] Vuik FER, Nieuwenburg SAV, Moen S, Spada C, Senore C, Hassan C, et al. Colon capsule endoscopy in colorectal cancer screening: a systematic review. Endoscopy. 2021;53(8):815 24. [25] Kobaek-Larsen M, Kroijer R, Dyrvig AK, Buijs MM, Steele RJC, Qvist N, et al. Back-to-back colon capsule endoscopy and optical colonoscopy in colorectal cancer screening individuals. Colorectal Dis 2018;20(6):479 85. [26] Adria´n-de-Ganzo Z, Alarco´n-Ferna´ndez O, Ramos L, Gimeno-Garcı´a A, AlonsoAbreu I, Carrillo M, et al. Uptake of colon capsule endoscopy vs colonoscopy for screening relatives of patients with colorectal cancer. Clin Gastroenterol Hepatol 2015;13(13):2293 301 e1. [27] Rondonotti E, Borghi C, Mandelli G, Radaelli F, Paggi S, Amato A, et al. Accuracy of capsule colonoscopy and computed tomographic colonography in individuals with positive results from the fecal occult blood test. Clin Gastroenterol Hepatol 2014;12(8):1303 10. [28] Eliakim R. The impact of panenteric capsule endoscopy on the management of Crohn’s disease. Therap Adv Gastroenterol 2017;10(9):737 44. [29] Boal Carvalho P, Rosa B, Dias de Castro F, Moreira MJ, Cotter J. PillCam COLON 2 in Crohn’s disease: A new concept of pan-enteric mucosal healing assessment. World J Gastroenterol 2015;21(23):7233 41. [30] Eliakim R, Yablecovitch D, Lahat A, Ungar B, Shachar E, Carter D, et al. A novel PillCam Crohn’s capsule score (Eliakim score) for quantification of mucosal inflammation in Crohn’s disease. United European Gastroenterol J 2020;8(5):544 51. [31] Greener T, Klang E, Yablecovitch D, Lahat A, Neuman S, Levhar N, et al. The impact of magnetic resonance enterography and capsule endoscopy on the reclassification of disease in patients with known Crohn’s disease: a prospective Israeli IBD Research Nucleus (IIRN) study. J Crohns Colitis 2016;10(5):525 31. [32] Sung J, Ho KY, Chiu HM, Ching J, Travis S, Peled R. The use of Pillcam Colon in assessing mucosal inflammation in ulcerative colitis: a multicenter study. Endoscopy. 2012;44(8):754 8. [33] Ye CA, Gao YJ, Ge ZZ, Dai J, Li XB, Xue HB, et al. PillCam colon capsule endoscopy vs conventional colonoscopy for the detection of severity and extent of ulcerative colitis. J Dig Dis 2013;14(3):117 24. [34] Hosoe N, Matsuoka K, Naganuma M, Ida Y, Ishibashi Y, Kimura K, et al. Applicability of second-generation colon capsule endoscope to ulcerative colitis: a clinical feasibility study. J Gastroenterol Hepatol 2013;28(7):1174 9.

267

268

CHAPTER 15 Colon capsule endoscopy and artificial intelligence

[35] Usui S, Hosoe N, Matsuoka K, Kobayashi T, Nakano M, Naganuma M, et al. Modified bowel preparation regimen for use in second-generation colon capsule endoscopy in patients with ulcerative colitis. Dig Endosc 2014;26(5):665 72. [36] Adler SN, Gonza´lez Lama Y, Matallana Royo V, Sua´rez Ferrer C, Schwartz A, BarGil Shitrit A. Comparison of small-bowel colon capsule endoscopy system to conventional colonoscopy for the evaluation of ulcerative colitis activity. Endosc Int Open 2019;7(10) E1253 E61. [37] Nennstiel S, Machanek A, von Delius S, Neu B, Haller B, Abdelhafez M, et al. Predictors and characteristics of angioectasias in patients with obscure gastrointestinal bleeding identified by video capsule endoscopy. United European Gastroenterol J 2017;5(8):1129 35. [38] Liao Z, Gao R, Xu C, Li ZS. Indications and detection, completion, and retention rates of small-bowel capsule endoscopy: a systematic review. Gastrointest Endosc 2010;71(2):280 6. [39] Pennazio M, Spada C, Eliakim R, Keuchel M, May A, Mulder CJ, et al. Small-bowel capsule endoscopy and device-assisted enteroscopy for diagnosis and treatment of small-bowel disorders: European Society of Gastrointestinal Endoscopy (ESGE) Clinical Guideline. Endoscopy. 2015;47(4):352 76. [40] Mussetto A, Arena R, Fuccio L, Trebbi M, Tina Garribba A, Gasperoni S, et al. A new panenteric capsule endoscopy-based strategy in patients with melena and a negative upper gastrointestinal endoscopy: a prospective feasibility study. Eur J Gastroenterol Hepatol 2021;33(5):686 90. [41] Bjoersum-Meyer T, Skonieczna-Zydecka K, Cortegoso Valdivia P, Stenfors I, Lyutakov I, Rondonotti E, et al. Efficacy of bowel preparation regimens for colon capsule endoscopy: a systematic review and meta-analysis. Endosc Int Open 2021;9 (11) E1658 E73. [42] Tal AO, Vermehren J, Albert JG. Colon capsule endoscopy: current status and future directions. World J Gastroenterol 2014;20(44):16596 602. [43] Saurin JC, Lapalus MG, Cholet F, D’Halluin PN, Filoche B, Gaudric M, et al. Can we shorten the small-bowel capsule reading time with the “Quick-view” image detection system? Dig Liver Dis 2012;44(6):477 81. [44] Beg S, Card T, Sidhu R, Wronska E, Ragunath K. The impact of reader fatigue on the accuracy of capsule endoscopy interpretation. Dig Liver Dis 2021;53 (8):1028 33. [45] Triantafyllou K, Beintaris I, Dimitriadis GD. Is there a role for colon capsule endoscopy beyond colorectal cancer screening? A literature review. World J Gastroenterol 2014;20(36):13006 14. [46] Yasaka K, Akai H, Abe O, Kiryu S. Deep learning with convolutional neural network for differentiation of liver masses at dynamic contrast-enhanced CT: a preliminary study. Radiology. 2018;286(3):887 96. [47] Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, et al. Dermatologistlevel classification of skin cancer with deep neural networks. Nature. 2017;542 (7639):115 18. [48] Gargeya R, Leng T. Automated identification of diabetic retinopathy using deep learning. Ophthalmology 2017;124(7):962 9. [49] Blanes-Vidal V, Baatrup G, Nadimi ES. Addressing priority challenges in the detection and assessment of colorectal polyps from capsule endoscopy and colonoscopy in

References

[50]

[51]

[52]

[53]

[54]

[55]

[56]

[57]

[58]

colorectal cancer screening using machine learning. Acta Oncol 2019;58(sup1) S29 S36. Yamada A, Niikura R, Otani K, Aoki T, Koike K. Automatic detection of colorectal neoplasia in wireless colon capsule endoscopic images using a deep convolutional neural network. Endoscopy. 2020;. Saraiva MM, Ferreira JPS, Cardoso H, Afonso J, Ribeiro T, Andrade P, et al. Artificial intelligence and colon capsule endoscopy: development of an automated diagnostic system of protruding lesions in colon capsule endoscopy. Tech Coloproctol 2021;25(11):1243 8. Saraiva MM, Ferreira JPS, Cardoso H, Afonso J, Ribeiro T, Andrade P, et al. Artificial intelligence and colon capsule endoscopy: automatic detection of blood in colon capsule endoscopy using a convolutional neural network. Endosc Int Open 2021;9(8) E1264 E8. Mascarenhas M, Ribeiro T, Afonso J, Ferreira JPS, Cardoso H, Andrade P, et al. Deep learning and colon capsule endoscopy: automatic detection of blood and colonic mucosal lesions using a convolutional neural network. Endosc Int Open 2022;10(2) E171 E7. Ferreira JPS, de Mascarenhas Saraiva M, Afonso JPL, Ribeiro TFC, Cardoso HMC, Ribeiro Andrade AP, et al. Identification of ulcers and erosions by the novel Pillcamt Crohn’s capsule using a convolutional neural network: a multicentre pilot study. J Crohns Colitis 2022;16(1):169 72. Majtner T, Brodersen JB, Herp J, Kjeldsen J, Halling ML, Jensen MD. A deep learning framework for autonomous detection and classification of Crohn’s disease lesions in the small bowel and colon with capsule endoscopy. Endosc Int Open 2021;9(9) E1361 E70. Watabe H, Nakamura T, Yamada A, Kakugawa Y, Nouda S, Terano A. Assessment of an electronic learning system for colon capsule endoscopy: a pilot study. J Gastroenterol 2016;51(6):579 85. National Health Service 2021 [11th March 2021]. NHS rolls out capsule cameras to test for cancer, ,https://www.england.nhs.uk/2021/03/nhs-rolls-out-capsule-camerasto-test-for-cancer/.; 2022 [accessed on 30.06.22]. Robertson AR, Segui S, Wenzek H, Koulaouzidis A. Artificial intelligence for the detection of polyps or cancer with colon capsule endoscopy. Ther Adv Gastrointest Endosc 2021;14 26317745211020277.

269

This page intentionally left blank

Index Note: Page numbers followed by “f” and “t” refer to figures and tables, respectively.

A AD. See Angiodysplasia (AD) Agile Patency capsule, 30 AHA. See American Hospital Association (AHA) AHSCT. See Allogeneic hematopoietic stem cell transplantation (AHSCT) AlbuNet-34 architecture, 156, 157f AlexNET CNN, 112 ALICE system. See Augmented Live-Body Image Color-Spectrum Enhancement (ALICE) system Allogeneic hematopoietic stem cell transplantation (AHSCT), 33 34 American Hospital Association (AHA), 204 Ancylostoma spp., 172 Anemia, gastrointestinal bleeding and, 261 Angiectasia, 150 Angiodysplasia (AD), 150 ANNs. See Artificial neural networks (ANNs) Applicability, artificial intelligence, 3 Artificial neural networks (ANNs), 92 94, 108 Augmented Live-Body Image Color-Spectrum Enhancement (ALICE) system, 107, 249 Automatic hookworm detection, 172

B Barrett’s esophagus, 125 Barrett’s metaplastic tissue, 6 BF. See Blue filter (BF) Blockchain advantages of using, 210 growing field of AI applied to CE, 206 208 and its utility for big data and AI in healthcare, 203 205 limitations and challenges to applications of, 208 210 present-day medical challenges, 200 202 and use of big data and AI in imaging, 205 206 Blood and hematic residues, AI for automatic detection of, 92 94 active bleeding and hematic residues, AI in detecting, 97 artificial neural network (ANN), 94 convolutional neural network (CNN), 94 97 ESNavi, 96 Inception-Resnet-V2, 96 97 SSD 1 ResNet50, 96 support vector machines, 94

Blue filter (BF), 107 Blue mode (BM), 246, 248 BM. See Blue mode (BM) Bowel cleansing, 41 42 Brotz score, 188, 189t

C CAC score. See Computed assessment of cleansing (CAC) score CAD-CAP 2020. See Computer-Assisted Diagnosis for CAPsule Endoscopy (CADCAP) 2020 CADe. See Computer-aided detection (CADe) CAD systems. See Computer-aided diagnosis (CAD) systems Cameron’s erosions, 58 CAPD algorithm. See Computer-assisted diagnostic-procedure (CAPD) algorithm CapsoCam Plus, 102 103 CapsoCam SV-1, 61 Capsule endoscopy (CE), 7 in patients with established Crohn’s disease, 75 76 in reclassification of inflammatory bowel disease, 77 78 in suspected Crohn’s disease, 72 74 Capsule Endoscopy Crohn’s Disease Activity Index (CECDAI), 27 28, 29t, 72 73, 74t, 80 81, 259 260 Capsule endoscopy models, 13t. See also Wireless capsule endoscopy (WCE) Capsule retention, 42 46 Capsules, advanced technologies in, 61 Capsule Scoring of Ulcerative Colitis (CSUC) score, 82 CC. See Conventional colonoscopy (CC) CC-CLEAR. See Colon Capsule Cleansing Assessment and Report (CC-CLEAR) CCE. See Colon capsule endoscopy (CCE) CD. See Crohn’s disease (CD) CE. See Capsule endoscopy (CE) CECDAI. See Capsule Endoscopy Crohn’s Disease Activity Index (CECDAI) CECDAIic score, 81 Celiac disease, 15, 207 208 Challenges in CE, 103 104 CICE. See Contrast image capsule endoscope (CICE)

271

272

Index

CLBP. See Color local binary patterns (CLBP) CNN. See Computational neural networks (CNNs) Convolutional neural network (CNN) Colon, 141 142 Colon Capsule Cleansing Assessment and Report (CC-CLEAR), 189t, 191 193 Colon capsule endoscopy (CCE), 12, 69, 78, 122, 141 142 cleansing quality evaluation, 190 193 cost-effectiveness of, in inflammatory bowel disease, 83 84 in Crohn’s disease, 79 81 future directions, 264 265 impact of AI, 262 264 indications for, 256 261 colorectal cancer screening, 256 259 gastrointestinal bleeding and anemia, 261 inflammatory bowel disease, 259 261 limitations of, 261 262 preparation, 185 187 boosters, 186 diet and fasting, 185 oral purgatives, 186 prokinetic drugs, 187 principles of, 255 256 in ulcerative colitis, 81 83 Colon examination, 15 16 Colonoscopy, 58 Colon pathology, 12 Colorectal cancer (CRC), 122, 142, 255 259 Color local binary patterns (CLBP), 170 Complications of CE, 59 60, 84 85 Computational neural networks (CNNs), 22 23, 30 Computed assessment of cleansing (CAC) score, 187 Computer-aided detection (CADe), 109 Computer-aided diagnosis (CAD) systems, 109, 123, 135 136, 141, 163 164, 207 Computer-Assisted Diagnosis for CAPsule Endoscopy (CAD-CAP) 2020, 152 153 GIANA—MICCAI 2017, 153 GIANA—MICCAI 2018, 153 Computer-assisted diagnostic-procedure (CAPD) algorithm, 233 234 Contraindications of CE, 59 60 Contrast image capsule endoscope (CICE), 251 Conventional colonoscopy (CC), 255 Conventional reading (CR), 136 Convolutional neural network (CNN), 2 3, 5 6, 94 97, 103, 108, 110t, 125, 129, 135 136, 142, 164, 207, 228 230, 262 architecture, 157 158 ESNavi, 96

Inception-Resnet-V2, 96 97 SSD 1 ResNet50, 96 COVID-19 pandemic, 126 129, 204 205, 237 CR. See Conventional reading (CR) CRC. See Colorectal cancer (CRC) C-reactive protein, 75 76 Crohn’s disease (CD), 15, 25 30, 69 70, 207 capsule endoscopy in, 72 76 colon capsule endoscopy in, 79 81 endoscopic appearance of, 70f extent of, 27 monitoring activity of, 27 30 suspected, 26 27 ulcers, 12 14 CSUC score. See Capsule Scoring of Ulcerative Colitis (CSUC) score

D DAE. See Device-assisted enteroscopy (DAE) Data augmentation, 108 DBE. See Double-balloon enteroscopy (DBE) Deep hookworm detection framework (DHDF), 174 Deep learning (DL), 2 3, 104 105, 108, 228 230 Definition of artificial intelligence, 1 2 DenseNet-161, 158 Detachable string MCE (DS-MCE), 225 226 Device-assisted deep enteroscopy, 55 Device-assisted enteroscopy (DAE), 12 14, 130 DHDF. See Deep hookworm detection framework (DHDF) Direct oral anticoagulants (DOACs), 54 55 DL. See Deep learning (DL) DOACs. See Direct oral anticoagulants (DOACs) Double-balloon enteroscopy (DBE), 56 DS-MCE. See Detachable string MCE (DS-MCE)

E EHRs. See Electronic health records (EHRs) Electronic health records (EHRs), 199 EM algorithm. See Expectation maximization (EM) algorithm Endocapsule, 102 ESGE. See European Society of Gastrointestinal Endoscopy (ESGE) ESNavi, 96 Esophagogastroduodenoscopy (EGD), 124 126, 227 Esophagogastroscopy, 12 14 Esophagus, 123 125

Index

European Society of Gastrointestinal Endoscopy (ESGE), 25, 31 32 Expectation maximization (EM) algorithm, 155 156

F Facilitating technology, artificial intelligence in healthcare as, 5 Fake news detection and cybersecurity, 5 FAMCE system. See Fully automated MCE (FAMCE) system Familial adenomatous polyposis (FAP), 14, 31 32 FAP. See Familial adenomatous polyposis (FAP) Fecal calprotectin, 75 76 Fecal immunological test (FIT), 257 258 FFN. See Fringe field navigation (FFN) FICE. See Flexible spectral imaging color enhancement (FICE)Fujinon Intelligent Chromo Endoscopy (FICE) system FIT. See Fecal immunological test (FIT) Flexible spectral imaging color enhancement (FICE), 102t, 141 Fringe field navigation (FFN), 220 Fujinon Intelligent Chromo Endoscopy (FICE) system, 244 245, 247 248 Fully automated MCE (FAMCE) system, 223

G Gastrointestinal (GI) bleeding, 53, 55f, 56f, 57f, 60f, 206 advanced technologies in capsules, 61 artificial intelligence in CE, 61 contraindications and complications of CE, 59 60 suspected small bowel bleeding, 54 58 timing of CE, 58 59 Gastrointestinal AI diagnostic System (GRAIDS), 233 Gastrointestinal bleeding and anemia, 261 Gastrointestinal Image ANAalysis (GIANA) Challenge 2017, 156 Gastrointestinal tract, vascular lesions in, 150 151 Gauss Laguerre transform (GLT), 169 GIANA Challenge 2017. See Gastrointestinal Image ANAalysis (GIANA) Challenge 2017 GIANA—MICCAI 2017, 153 GIANA—MICCAI 2018, 153 GI bleeding. See Gastrointestinal (GI) bleeding Given Imaging Ltd., 11 12 GLT. See Gauss Laguerre transform (GLT) Graft versus host disease (GVHD), 33 36

GRAIDS. See Gastrointestinal AI diagnostic System (GRAIDS) GVHD. See Graft versus host disease (GVHD)

H Hand-held magnetic capsule endoscopy, 217 218 HCN frames. See Highly contaminated nonbubbled (HCN) frames Healthcare, artificial intelligence in, 5 Helicobacter pylori, 126 Hereditary polyposis syndromes, 14 Highly contaminated nonbubbled (HCN) frames, 169 Hookworms and foreign bodies, 172 175

I IBD. See Inflammatory bowel disease (IBD) IEE. See Image enhanced endoscopy (IEE) Ileocolonoscopy, 70 Image enhanced endoscopy (IEE), 104, 106 107 Inception-Resnet-V2, 96 97 Indications in capsule endoscopy, 22 36, 37t Crohn’s disease (CD), 25 30 extent of, 27 monitoring activity of, 27 30 suspected, 26 27 graft versus host disease, 33 36 polyposis syndromes, screening of, 31 33 familial adenomatous polyposis, 31 32 Peutz Jeghers syndrome, 33 refractory celiac disease, evaluation of, 30 31 suspected middle digestive hemorrhage, 22 25 suspected small intestine tumors, 33 Inflammatory bowel disease (IBD), 12, 25, 38 39, 103, 259 261 capsule endoscopy (CE), 77 78 in patients with established Crohn’s disease, 75 76 in suspected Crohn’s disease, 72 74 colon capsule endoscopy (CCE), 78 in Crohn’s disease, 79 81 in ulcerative colitis, 81 83 complications of capsule endoscopy, 84 85 cost-effectiveness of colon capsule endoscopy in, 83 84 Crohn’s disease, 70 IBD unclassified (IBDU), 77 new research areas for future, 85 86 postoperative recurrence, assessment of, 76 ulcerative colitis, 70 71 Inflammatory lesions, CE software enhancements to improve detection of, 106 107

273

274

Index

K KIDs dataset, 152, 155 K-nearest neighbor (KNN), 170 KNN. See K-nearest neighbor (KNN) Kvasir Capsule dataset, 154, 158

L Lewis score, 24, 29t, 72 74, 74t, 105 Limitations of endoscopic capsule, 40 46 bowel cleansing, 41 42 capsule retention, 42 46 pediatrics, challenges in, 40 41 swallowing the capsule, 40 41 Lubiprostone, 183 Luminal content analysis hookworms and foreign bodies, 172 175 lymphangiectasia, 170 172 small bowel preparation and luminal content, 164 170 Lymphangiectasia, 170 172 Lynch syndrome, 137f

M MACE. See Magnetically assisted CE (MACE) Machine learning (ML), 2, 104 105, 108, 230 Magnetically assisted CE (MACE), 125 129 Magnetic capsule endoscopy (MCE) AI-assisted magnetic capsule endoscopy diagnostic procedure, 233 236 localization strategy, 231 232 definition and role of AI in technology enhancement, 230 development and validation of AI systems in gastrointestinal practice, 230 231 hand-held magnetic capsule endoscopy, 217 218 indications and contradictions, 227 228 operation procedure of gastric examination, 226 227 prospects of AI in magnetic capsule endoscopy, 237 robotic magnetic capsule endoscopy, 220 223 Magnetic resonance enterography (MRE), 14 Magnetic resonance imaging-based magnetic capsule endoscopy, 218 220 Markov random fields, 155 156 MCE. See Magnetic capsule endoscopy (MCE) MCL. See Most common lesion (MCL) Meckel’s diverticulum, 12 14 Medicine, artificial intelligence in, 5 7 MiroCam, 102, 107, 249 ML. See Machine learning (ML) MLP. See Multilayer perceptron (MLP)

Most common lesion (MCL), 103 Most severe lesion (MSL), 103 MRE. See Magnetic resonance enterography (MRE) MSL. See Most severe lesion (MSL) Multilayer perceptron (MLP), 207

N Narrow band imaging (NBI), 246, 248 249 NaviCam MCE systems, 223 226, 234 235 NaviCam SBCE system, 114 115 NaviCam SB system, 103 NaviCam system, 129 NBI. See Narrow band imaging (NBI) Necator americanus, 172 Next-generation capsule endoscopes, AI in, 114 115 Niv score, 259 260 Nonwhite light endoscopy in capsule endoscopy background, 243 virtual chromoendoscopy, evidence of, 247 251 blue mode (BM), 248 Fujinon Intelligent Chromo Endoscopy (FICE), 247 Fujinon intelligent chromoendoscopy and blue mode, 248 narrow band imaging (NBI), 248 249 other virtual chromoendoscopy methods, 249 251 virtual chromoendoscopy in capsule endoscopy, 244 246 blue mode (BM), 246 Fujinon Intelligent Chromo Endoscopy (FICE) system, 244 245 narrow band imaging (NBI), 246 white light, 243

O Obscure gastrointestinal bleeding (OGIB), 22, 24, 24t, 131 Occult gastrointestinal bleeding (OGIB), 37 38, 150 151 OGIB. See Obscure gastrointestinal bleeding (OGIB) OMOM, 102 Online experience, 4

P PCC. See PillCam Crohn’s Capsule (PCC) PCCE2. See Second-generation PillCam Colon Capsule (PCCE2) Pediatrics, CE clinical scope in, 36 40 indications, 37 inflammatory bowel disease (IBD), 38 39

Index

occult gastrointestinal bleeding, 37 38 polyposis syndromes, 39 40 PEG solution. See Polyethylene-glycol (PEG) solution Peutz Jeghers syndrome (PJS), 14, 33 PillCam capsule endoscopy workstation, 245, 245f PillCam Crohn’s Capsule (PCC), 12, 78, 113, 259 261 PillCam ESO3, 125 PillCam SB3, 102, 165 166 PJS. See Peutz Jeghers syndrome (PJS) Polyethylene-glycol (PEG) solution, 182, 261 262 Polyp detection, 123 Polyposis syndromes, 39 40 screening of, 31 33 familial adenomatous polyposis, 31 32 Peutz Jeghers syndrome, 33 Postoperative recurrence, assessment of, 76 PRISMA methodology, 135 136 Protruding lesions, AI for perspectives on challenges and developments, 142 143 state-of-the-art clinical aspects, 123 142 colon, 141 142 esophagus, 123 125 small bowel, 130 140 stomach, 125 129 state-of-the-art technological aspects, 123

R Randomized controlled trials (RCT), 230 RAPID (Medtronic) software, 105 RCT. See Randomized controlled trials (RCT) Red lesion endoscopy dataset, 152 Refractory celiac disease, evaluation of, 30 31 ResNet-152, 158 Resnet-34 architecture, 234 235 Robot-guided magnetic capsules, 61 Robotic magnetic capsule endoscopy, 220 223 Robotics, 4

S SBB. See Small bowel bleeding (SBB) SBCE. See Small bowel capsule endoscopy (SBCE) SBI. See Suspected blood indicator (SBI) SB mucosa. See Small bowel (SB) mucosa SB tumors, 12 14 SDSS-AI-based diagnostic system. See Smart data service system AI (SDSS-AI)-based diagnostic system Second-generation PillCam Colon Capsule (PCCE2), 103 SegNet network, 157 159

Self-driving cars, 5 Sessile serrated lesions (SSLs), 141 ShotMultiBox Detector, 158 Simethicone, 184 185 Single Shot MultiBox Detector (SSD), 112 113 Small bowel (SB) mucosa, 101, 130 140 Small bowel bleeding (SBB), 53 58, 91 Small bowel capsule endoscopy (SBCE), 12 15, 131, 135 136, 142 143 cleansing quality evaluation, 187 190 automated scores, 187 188 operator-dependent scores, 188 190 preparation, 181 185 antifoaming agents, 184 185 diet and fasting, 182 oral purgatives, 182 184 prokinetic drugs, 184 water ingestion, 185 Small bowel inflammation, CE scoring systems for, 104 106 CE Crohn’s disease activity index, 106 Lewis score, 105 Small bowel preparation and luminal content, 164 170 Small bowel tumors, 14 Small bowel ulcerations and erosions, AI for detection of, 109 110 Small bowel xanthoma, 171f Small intestine tumors, 33 Smart data service system AI (SDSS-AI)-based diagnostic system, 235 236 SmartScan, 140 SPICE (smooth, protruding lesions index in CE), 170 171 Spice index, 35t SSD 1 ResNet50, 96 SSD. See Single Shot MultiBox Detector (SSD) SSLs. See Sessile serrated lesions (SSLs) SSSB. See Suspected small bowel bleeding (SSSB) State-of-the-art clinical aspects, 123 142 colon, 141 142 esophagus, 123 125 small bowel, 130 140 stomach, 125 129 State-of-the-art technological aspects, 123 Stomach, 125 129 Stomach, protruding lesions of, 126, 127t Support vector machines (SVMs), 92 94, 109, 207 Suspected blood indicator (SBI), 91 Suspected middle digestive hemorrhage, 22 25 Suspected small bowel bleeding (SSSB), 12 14, 53 58

275

276

Index

Suspected small intestine tumors, 33 SVMs. See Support vector machines (SVMs)

T Timing of CE, 58 59 Triple ANet, 158

U Ulcerative colitis, 69 71 colon capsule endoscopy in, 81 83 endoscopic appearance of, 71f Ulcers and erosions, detection of AI in next-generation capsule endoscopes, 114 115 artificial intelligence and its application in CE, 108 109 automatic detection, 110 114 capsule endoscopes, 102 103 current challenges in CE, 103 104 grading of ulcers and erosions severity, 114 inflammatory lesions, CE software enhancements to improve detection of, 106 107 image enhanced endoscopy (IEE), 106 107 small bowel inflammation, CE scoring systems for, 104 106 CE Crohn’s disease activity index, 106 Lewis score, 105 small bowel ulcerations and erosions, AI for detection of, 109 110 U-Net network architecture, 156 Upper gastrointestinal endoscopy, 70

V Vascular lesions, AI for, 154 159 CAD CAP 2020, 152 153 GIANA—MICCAI 2017, 153 GIANA—MICCAI 2018, 153 KIDs dataset, 152, 155 Kvasir Capsule, 154 red lesion endoscopy dataset, 152 wireless capsule endoscopy (WCE) and AI, 149 151

vascular lesions in gastrointestinal tract, 150 151 Vehicles, 5 Virtual chromoendoscopy (VC), 244 246 blue mode (BM), 246 evidence of, 247 251 blue mode (BM), 248 Fujinon Intelligent Chromo Endoscopy (FICE), 247 Fujinon intelligent chromoendoscopy and blue mode, 248 narrow band imaging (NBI), 248 249 Fujinon Intelligent Chromo Endoscopy (FICE) system, 244 245 narrow band imaging (NBI), 246

W WCE. See Wireless capsule endoscopy (WCE) White light, 243 White light endoscopy (WLE) images, 106 107 White-light light-emitting diode (WL-LED), 251 Wireless capsule endoscopy (WCE), 149 152, 155 background, 11 future perspectives, 16 gastrointestinal tract, vascular lesions in, 150 151 indications, 12 16 celiac disease, 15 colon examination, 15 16 Crohn’s disease, 15 hereditary polyposis syndromes, 14 small bowel tumors, 14 suspected small bowel bleeding (SSSB), 12 14 pros and cons of, 12t types of capsules, 11 12 WLE images. See White light endoscopy (WLE) images WL-LED. See White-light light-emitting diode (WL-LED)