Computational Methods and Deep Learning for Ophthalmology 0323954146, 9780323954143

Citation preview

1

Computational Methods and Deep Learning for Ophthalmology  Editor  D. Jude Hemanth Professor, Karunya University, Tamil Nadu, India

2

Table of Contents  Cover image  Title page  Copyright  Contributors  1. Classification of ocular diseases using transfer learning approaches and glaucoma severity grading 1.1. Introduction 

1.2. Literature review 

1.3. Proposed methodology 

1.4. Results and discussion 

1.5. Conclusion 

2. Early diagnosis of diabetic retinopathy using deep learning techniques 2.1. Introduction 

2.2. Related background 

2.3. Experimental methodology 

2.4. Proposed flow 

2.5. Results and discussion 

3

2.6. Conclusion and future direction    3. Comparison of deep CNNs in the identification of DME structural  changes in retinal OCT scans    3.1. Introduction    3.2. Structural changes of DME    3.3. Convolutional neural networks    3.4. Results and discussion    3.5. Conclusion    4. Epidemiological surveillance of blindness using deep learning ap-  proaches    4.1. Conceptualizing surveillance systems in ophthalmic epidemiology    4.2. Deep learning in ophthalmic epidemiological surveillance    4.3. Limitations    4.4. Conclusion    5. Transfer learning-based detection of retina damage from optical coher-  ence tomography images    5.1. Introduction    5.2. Experimental methodology    5.3. Proposed model    4

5.4. Experimental results and observations    5.5. Conclusion    6. An improved approach for classification of glaucoma stages from color  fundus images using Efficientnet-b0 convolutional neural network and  recurrent neural network    6.1. Introduction    6.2. Related work    6.3. Methodology    6.4. Experimental findings    6.5. Conclusion    7. Diagnosis of ophthalmic retinoblastoma tumors using 2.75D CNN seg-  mentation technique    7.1. Introduction    7.2. Literature review    7.3. Materials    7.4. Methodology    7.5. Experimental analysis    7.6. Discussions    7.7. Conclusion 

5

8. Fast bilateral filter with unsharp masking for the preprocessing of optical  coherence tomography images—an aid for segmentation and classification    8.1. Introduction    8.2. Methodology    8.3. Results and discussion    8.4. Conclusion    9. Deep learning approaches for the retinal vasculature segmentation in fun-  dus images    9.1. Introduction    9.2. Significance of deep learning    9.3. Convolutional neural network    9.4. Fully convolved neural network    9.5. Retinal blood vessel extraction    9.6. Artery/vein classification    9.7. Summary    10. Grading of diabetic retinopathy using deep learning techniques    10.1. Introduction    10.2. Materials and methods    10.3. Methodology  6

10.4. Results and discussion    10.5. Conclusion    11. Segmentation of blood vessels and identification of lesion in fundus image by using fractional derivative in fuzzy domain    11.1. Introduction    11.2. Preliminary ideas    11.3. Proposed method of blood vessel extraction    11.4. Proposed method of lesion extraction    11.5. Experimental analysis    11.6. Conclusion    12. U-net autoencoder architectures for retinal blood vessels segmentation    12.1. Introduction    12.2. Related works    12.3. Proposed works    12.4. Experiment    12.5. Conclusion    13. Detection and diagnosis of diseases by feature extraction and analysis on  fundus images using deep learning techniques    13.1. Introduction  7

13.2. Fundus image analysis    13.3. Eye diseases with retinal manifestation    13.4. Diagnosis of glaucoma    13.5. Diagnosis of diabetic retinopathy    13.6. Conclusion    Index   

8

Copyright  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 King-  dom    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 writ-  ing 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 Copy-  right

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 expe-  rience and knowledge in evaluating and using any information, meth-  ods, 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 pro-  fessional 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 

9

methods, products, instructions, or ideas contained in the material  herein.    ISBN: 978-0-323-95415-0    For information on all Academic Press publications visit our website  at https://www.elsevier.com/books-and-journals   

  Publisher: Mara Conner  Acquisitions Editor: Chris Katsaropoulos  Editorial Project Manager: Tom Mearns  Production Project Manager: Prem Kumar Kaliamoorthi  Cover Designer: Christian Bilbow    Typeset by TNQ Technologies   

10

Contributors   R. Adarsh,     National Institute of Technology, Tiruchirappalli, Tamil  Nadu, India  Gadipudi Amarnageswarao,     National Institute of Technology, Tiruchi-  rappalli, Tamil Nadu, India  V.P. Ananthi,     Department of Mathematics, Gobi Arts & Science Col-  lege, Gobichettipalayam, Tamil Nadu, India  K. Balakrishnan,     Department of Computer Science and Engineering,  Indian Institute of Information Technology, Tiruchirappalli, Tamil Nadu,  India  Ebenezer Daniel,     Diagnostic Radiology, City of Hope National Medical  Center, Duarte, CA, United States  S. Deivalakshmi,     National Institute of Technology, Tiruchirappalli,  Tamil Nadu, India  R. Dhanalakshmi,     Department of Computer Science and Engineering,  Indian Institute of Information Technology, Tiruchirappalli, Tamil Nadu,  India  Poonguzhali Elangovan,     Department of ECE, National Institute of Technology Puducherry, Thiruvettakudy, Karaikal, Puducherry, India  Kurubaran Ganasegeran,     Clinical Research Center, Seberang Jaya Hos-  pital, Ministry of Health Malaysia Seberang Perai, Penang, Malaysia  G. Indumathi,     Mepco Schlenk Engineering College, Sivakasi, Tamil  Nadu, India  Anitha J,     Department of Electronics and Communication Engineering,  Karunya Institute of Technology and Sciences, Coimbatore, Tamil Nadu,  India  D. Jasmine David,     Department of Electronics and Communication Engineering, Karunya Institute of Technology and Sciences, Coimbatore,  Tamil Nadu, India  T. Jemima Jebaseeli,     Department of Computer Science and Engi-  neering, Karunya Institute of Technology and Sciences, Coimbatore, Tamil  Nadu, India  Mohd Kamarulariffin Kamarudin,     Department of Social and Preventive  Medicine, Faculty of Medicine, University of Malaya Kuala Lumpur,  Malaysia  S.N. Kumar,     Department of EEE, Amal Jyothi College of Engineering,  11

Kottayam, Kerala, India  Malaya Kumar Nath,     Department of ECE, National Institute of Tech-  nology Puducherry, Thiruvettakudy, Karaikal, Puducherry, India  N. Padmasini,     Department of Biomedical Engineering, Rajalakshmi  Engineering College, Chennai, Tamil Nadu, India  Asha Gnana Priya H,     Department of Electronics and Communication  Engineering, Karunya Institute of Technology and Sciences, Coimbatore,  Tamil Nadu, India  Ranjitha Rajan  Lincoln University College, Kota Bharu, Malaysia  LUC MRC, Kuttikanam, Kerala, India  G. Santhiya,     Department of Mathematics, Gobi Arts & Science Col-  lege, Gobichettipalayam, Tamil Nadu, India  V. Sathananthavathi,     Mepco Schlenk Engineering College, Sivakasi,  Tamil Nadu, India  D. Selvathi,     Senior Professor and Head, Biomedical Engineering Pro-  gramme, Department of ECE, Mepco Schlenk Engineering College, Sivakasi,  Tamilnadu, India  Mohamed Yacin Sikkandar,     CAMS, Majmaah University, Al Majmaah,  Saudi Arabia  Manavi D. Sindal,     Head Vitreo-Retina Services, Aravind Eye Hospital,  Pondicherry, India  Bam Bahadur Sinha,     Department of Computer Science and Engi-  neering, Indian Institute of Information Technology, Ranchi, Jharkhand,  India  J. Sudaroli Sandana,     National Institute of Technology, Tiruchirappalli,  Tamil Nadu, India  R. Umamaheswari,     Department of Electrical and Electronics Engi-  neering, Velammal Engineering College, Chennai, Tamil Nadu, India  Ajantha Devi Vairamani,     AP3 Solutions, Chennai, Tamil Nadu, India  D. Vijayalakshmi,     Department of ECE, National Institute of Tech-  nology Puducherry, Thiruvettakudy, Karaikal, Puducherry, India  Alongbar Wary,     School of Computer Science and Engineering, Vellore  Institute of Technology—AP University, Amaravati, Andhra Pradesh, India 

12

1: Classification of ocular diseases using transfer learning ap-  proaches and glaucoma severity grading  D. Selvathi     Senior Professor and Head, Biomedical Engineering Pro-  gramme, Department of ECE, Mepco Schlenk Engineering College, Sivakasi,  Tamilnadu, India   

Abstract  The recent advancement in medical imaging technologies resulted in  an abundant availability of retinal images to analyze ocular pathologies.  The ever-growing medical database in the form of images and the com-  plexity of human analysis of these medical images made a way for the  intrusion of Computer Aided Diagnosis in the medical field. Ocular dis-  eases are analyzed using retinal fundus images by extracting optimal  features for the particular disease. Successful identification and grading  of ocular diseases demand rich human expertise. In addition, the  process is time consuming and hard for nonexperts to identify the rele-  vant features within the image. The ophthalmologist find the region of  interest for the pathology analysis and grading, which is now eased by  the automation of medical image analysis using machine learning tech-  niques. This prompted the automation of ocular disease examination  using image processing. Automatic image processing paved way for  the experts as well as nonexperts to do the task. In this proposed  method, deep convolutional neural network (DCNN) models classify  the ocular pathologies into eight classes such as age-related macular  degeneration, cataract, glaucoma, diabetic retinopathy, hypertensive  retinopathy, myopia, normal, and others. The input sources are trained  by DCNN and then validation is done by the trained DCNN for evalu-  ating the accuracy. Among the DCNN networks ResNet50, AlexNet,  VGGNet, and GoogLeNet, ResNet50 yielded a promising result on the  classification of ocular diseases with a classification accuracy of  89.9%. The classification process is tailed by the grading of glaucoma  realized by segmenting the region of interest that is, cup and disk.  Based on the cup-to-disc ratio, the severity of glaucoma is graded.  Severity grading is done to four labels, that is, no glaucoma, mild glau-  coma, moderate glaucoma, and severe glaucoma. 

13

Keywords  CNN; Glaucoma grading; Ocular diseases; Transfer learning 

14

1.1. Introduction  The primary requisite of medical diagnosis is accuracy, which has a signif-  icant effect on mitigation and treatment. Medical diagnosis is the process of  identification of the presence and severity of the pathology based on the  symptoms and signs. Medical diagnosis is further assisted by medical imaging techniques that are drastically improved in the past few decades.  These medical imaging techniques are predominantly used for creating the  visual depiction of the organs or tissues that are hidden from the human  eyes by the skin. Several imaging modalities are present in the medical field.  These techniques portray the structure of the retinal fundus. Vision is the  prominent sense for all living beings. Therefore, eye is considered as the  most precious of all sensory organs because 80% of human perception is  based on what is seen. So, visual health is given much importance nowa-  days. Retinal examination is often done regularly. This in turn adds a re-  sponsibility of proper monitoring, analysis, and diagnosis of ocular fundus  images. If any careless is done during diagnosis, it may result in delayed  treatment and sometimes even loss of vision.  Glaucoma is one such dread ocular disease that mostly affects adults  above 40 having a serious risk of irreversible blindness. It is the second  most ocular disease to cause blindness to a huge number. WHO surveys  say that it affects nearly 12 million people annually and causes 1.2 million vi-  sion losses. Glaucoma analysis demands a rich human expertise. But  human analysis has an issue of intra- and interobserver variability. Thus,  automatic ocular disease identification and grading are emerging nowadays.  But ophthalmic imaging is a challenging job. There are several imaging tech-  nologies for capturing eye images. Among that, color fundus imaging using  fundus cameras is used for imaging the rear portion of the eye. So medical  diagnoses using color fundus imaging modalities are having more advan-  tages and less risk comparing other modalities. Color fundus photography  is the best method for visualizing and analyzing ocular pathologies com-  pared to other imaging techniques.  Ocular diseases are analyzed by fundus images (rear portion of the eye)  rather than lens images because the retina, optic disc, fovea, posterior pole,  and macula can be visualized better in the fundus images. Medical struc-  tures like hemorrhages, optical nerve heads, blood vessels, and their abnor-  malities can be conveniently extracted from the fundus image. Thus, the 

15

presence and progress of most deadly ocular diseases can be analyzed from  the fundus image. But human diagnosis will take considerable time and it  demands expert knowledge. Therefore, automation of image analysis is  needed. The automation of medical image analysis has been a topic of inter-  est over the past 2decades, which induced many researchers to work in this  field. 

16

1.2. Literature review  Over the past 3decades, the automation of medical image analysis paved the  attention of everyone, which led to several useful improvements in medical  image processing. Several works are done by researchers in this domain.  Ophthalmic imaging is a subdomain in medical imaging which is auto-  mated by the CAD process. The following are such works that automated  the ocular disease diagnosis and severity grading.  The work focuses on classifying the data into two classes—diabetic mac-  ular edema and age-related macular degeneration—and also improves the  network's adaptability to datasets. Duke University and Noor Eye Hospital  in Tehran SD-OCT imaging are used. The reusability of the networks is im-  proved by using transfer learning based on CliqueNet, DPN92,  DenseNet121, ResNet50, and ResNext101. CliqueNet's precision, recall, and  accuracy scores are comparatively higher than other networks. It represents  that CliqueNet has the highest adaptability to datasets [1].  The work proposed in Ref. [2] is an automatic diabetic retinopathy classi-  fication and grading system referred as “Deep DR”. The dataset used in this  work is retinal fundus images from the Sichuan Academy of Medical Sci-  ences. The first stage was a transfer learning network ResNet50 followed by  the classifier Standard Deep Neural Network, which is a custom classifier.  Then, the grading system works as a four class classifier (normal, NPDR,  NPDR2PDR, and PDR), which utilizes ensemble learning. This strategy con-  sistency and reproducibility for several diagnostics yielded promising re-  sults.  In [3], three Deep Neural Networks such as AlexNet, GoogLeNet, and  ResNet50 performance are analyzed for image classification. Multiple  datasets such as ImageNet, CIFAR10, CIFAR100, and MNIST are evaluated  to prove the performance capability of the model. It is observed that increas-  ing the training data increases the performance accuracy but also increases  the complexity of the network.  In this work, the diabetic retinopathy fundus image classification using  convolutional neural networks (CNNs)-based transfer learning is imple-  mented. The publicly available retinal fundus images in DR1 and MESSIDOR  datasets are used. The pretrained networks such as AlexNet, GoogLeNet,  and VGGNet are evaluated for grading diabetic retinopathy (DR) in which  VGGNet achieved a better performance in five classes such as DR, mild DR, 

17

moderate DR, severe DR, and proliferate DR). Two types of fine-tunings,  layer-wise fine-tuning and all-layer tuning, are tested. The pretrained CNN's  layer-wise will reduce the risk of overfitting and also obtain better results for  small datasets [4].  This study explores the classification process using three CNNs, Cifar-  Net, AlexNet, and GoogLeNet. The CNNs are examined for two different dis-  eases Thoraco-abdominal lymph node and interstitial lung disease classi-  fication. This study shows that limited datasets can cause bottleneck prob-  lems. Therefore, large-scale annotated datasets can be beneficially classified  using transfer learning models [5].  This work took a deep inspection on different requisites of deep learning  in medical imaging. It suggested that deep convolutional neural networks  (DCNNs) can auto-extract the mid and high-level features from the images.  Increasing the number of iterations or epochs optimized the network  parameters. When a medium-sized dataset is not available, then pretrained  CNN is suggested and also the fine-tuning of the pretrained CNNs achieved  better results. The challenges in medical imaging are need of a huge dataset,  need of expensive medical expertise for high-quality annotation, and privacy  issues in sharing the medical dataset [6].  This paper explained the processing operations to perform disease recog-  nition using different approaches such as support vector machine (SVM),  discrete cosine transform (DCT), hidden Markov model (HMM), and prin-  cipal component analysis (PCA). The first step is the image acquisition fol-  lowed by segmentation where the boundary of the iris is taken as circles,  and they need not be cocentric. Then normalization of image is done to  eliminate nonuniform illumination. Circular symmetric filter and grabber fil-  ter are used to extract features. Finally, a matching process is done by using  encoding followed by hamming distance method [7].  This paper suggested a method that has a flow of region of interest (ROI)  segmentation, image scaling, disc diameter calculation, cup diameter calcu-  lation, and cup-to-disc ratio (CDR) calculation from spectral domain OCT  images of the Armed Forces Institute of Ophthalmology. This system uses  the green channel of the preprocessed image for feature extraction. ROI ex-  traction has two steps first extracting a circle with the center of the retina,  and the next bilinear interpolation is done to improve spatial resolution and  also for improving accuracy. This paper also indicates that CDR