173 40 37MB
English Pages 233 Year 2015
Image-Guided Surgery Fundamentals and Clinical Applications in Otolaryngology
Image-Guided Surgery Fundamentals and Clinical Applications in Otolaryngology
Robert F. Labadie, MD, PhD J. Michael Fitzpatrick, PhD
5521 Ruffin Road San Diego, CA 92123 e-mail: [email protected] Web site: http://www.pluralpublishing.com Copyright 2016 © by Plural Publishing, Inc. Typeset in 10½/13 Palatino by Flanagan’s Publishing Services, Inc. Printed in Korea by Four Colour Print Group All rights, including that of translation, reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, recording, or otherwise, including photocopying, recording, taping, Web distribution, or information storage and retrieval systems without the prior written consent of the publisher. For permission to use material from this text, contact us by Telephone: (866) 758-7251 Fax: (888) 758-7255 e-mail: [email protected] Every attempt has been made to contact the copyright holders for material originally printed in another source. If any have been inadvertently overlooked, the publishers will gladly make the necessary arrangements at the first opportunity. NOTICE TO THE READER Care has been taken to confirm the accuracy of the indications, procedures, drug dosages, and diagnosis and remediation protocols presented in this book and to ensure that they conform to the practices of the general medical and health services communities. However, the authors, editors, and publisher are not responsible for errors or omissions or for any consequences from application of the information in this book and make no warranty, expressed or implied, with respect to the currency, completeness, or accuracy of the contents of the publication. The diagnostic and remediation protocols and the medications described do not necessarily have specific approval by the Food and Drug administration for use in the disorders and/or diseases and dosages for which they are recommended. Application of this information in a particular situation remains the professional responsibility of the practitioner. Because standards of practice and usage change, it is the responsibility of the practitioner to keep abreast of revised recommendations, dosages, and procedures.
Library of Congress Cataloging-in-Publication Data Names: Labadie, Robert F., author. | Fitzpatrick, J. Michael, author. Title: Image-guided surgery : fundamentals and clinical applications in otolaryngology / Robert F. Labadie, J. Michael Fitzpatrick. Description: San Diego, CA : Plural Publishing, [2016] | Includes bibliographical references and index. Identifiers: LCCN 2015042351| ISBN 9781597567190 (alk. paper) | ISBN 1597567191 (alk. paper) Subjects: | MESH: Otorhinolaryngologic Surgical Procedures. | Surgery, Computer-Assisted — methods. Classification: LCC RF46.5 | NLM WV 168 | DDC 617.5/1 — dc23 LC record available at http://lccn.loc.gov/2015042351
Contents Introduction ix Acknowledgments xi
1 Brief History of Image-Guided Surgery
1
Overview of How Image-Guided Surgery Works 1 The Evolution of IGS 3 Images 3 Putting It All Together: CT and MRI in IGS 12 References 21
2 CT and MRI
25
3 Tracking Systems
75
4 Registration
99
How CT Works 25 Intraoperative CT Scanners 30 Stationary CT Scanners 30 Portable CT Scanners 31 MRI 40 How MRI Works 40 Why Is Understanding Imaging Important in Using IGS? 70 Inaccuracy in Images 70 2D Presentation of 3D Images 71 References 73
Overview 75 Optical Tracking 78 Electromagnetic (EM) Tracking 90 Summary 97 References 97
Fiducial Markers and Points 100 Skin-Affixed, Fiducial Markers and Points 100 Bone-Affixed, Fiducial Markers and Points 102 Rigid Point Registration 107 Surface Registration 109 Accuracies for Various Types of Fiducials 112 Fiducials During Navigation 113 Fusion 113 References 114 v
vi
Image-Guided Surgery
5 Error Analysis
117
6 Best Practices for Use of IGS
141
7 Surgical Systems
151
8 What Does the Future Hold?
185
Extrinsic Versus Intrinsic Error 118 Fiducial Localization Error (FLE) 119 Fiducial Registration Error (FRE) 120 Target Registration Error (TRE) 121 Error Relationships 122 Relating TRE to FLE 122 Relationships Involving FRE 127 Probe Localization Error (PLE) 128 Stubborn Myths 129 Myth 1: FRE Is an Indicator of TRE 130 Myth 2: Dropping a Fiducial Will Increase Accuracy 134 Myth 3: Planar Fiducial Configurations Are Bad 137 Summary 137 References 138
Who Is Using IGS? 141 Is IGS Safer for Patients? 142 Does IGS Help Make Better Surgeons? 144 Professional Society Position Statements 144 Is IGS Financially Sustainable? 146 Less Litigation? 147 Overview 147 References 148
Current FDA-Cleared IGS for ENT 153 Medtronic 153 Brainlab 162 Clinical Accuracy of Electromagnetic (EM) Tracking Systems 171 Stryker 172 Smaller IGS Companies 180 Fiagon 180 ClaroNav 181 Summary 181 References 183
Computer-Assisted Navigation Augmented Reality Visual Augmented Reality Nonvisual Augmented Reality
185 187 189 190
Contents
Robots 193 History 194 History of Autonomous Robots 194 Current FDA-Cleared Autonomous Robots 197 What Autonomous Robots Does Otolaryngology Need? 197 Conclusions 202 References 203
Index 205
vii
Introduction ings, and scientific publications in this field, we have written this book to explain the theory behind the technology and to present the current state of the art in the context of clinical applications. While some clinicians may at first be put off by the inclusion of theory in this work, we have found that it is vital to understanding both the power and the limits of this emerging technology, and we have worked hard to make it accessible. Leonardo da Vinci was unaware of the rapid deterioration of the egg-based tempura he used to paint the Last Supper, and as result that masterpiece is no longer with us in its full glory. Likewise, the skull-base surgeon who is unaware of the limitations of image-guided surgery may either not use the technology to its full capacity or, worse, use it in a way that is dangerous. To make it feasible for the busy surgeon to learn the basic foundations of image guidance without devoting an inordinate amount of time to it, we have done our best to trim all technical descriptions to their bare essentials. To make it easier to understand these descriptions, we have augmented them with clarifying explanations, analogies, and examples, such that all clinicians should be able to understand the technology. In addition, scattered through the book we have included boxed text highlighted in light blue-gray to provide additional details for the interested reader. These details tell “the rest of the story” (quoting the late Paul Harvey) but can be skipped without interrupting the flow and content of the text. Finally, we have included plenty
A tool is only as good as its human operator. Perhaps this truth is most evident in art, where a masterpiece may be created by a talented operator using very simple, basic tools. While Michelangelo produced the masterpiece that graces the ceiling of the Sistine Chapel with mere brushes, paint, and plaster, the authors of this work would be hard pressed to use these same tools to create anything considered art! Image guidance for surgical interventions is a tool, and it too is only as good as its human operator. Proper use of an image-guidance system is vital to both safety and efficacy in the operating room, and because these systems are becoming so widespread and at the same time are becoming so sophisticated, there is a real danger that their human operators may become captive to the technology instead of mastering it. The authors have watched this situation develop over a period of some 25 years, slowly at first but much more rapidly in the last 5 or 10 years. And, in that quarter century, the following theme has emerged: running through all this complexity is a common theoretical thread that, once grasped, will subdue the technology and make these marvelous tools far easier to master. Furthermore, understanding the theory of image guidance does not require an advanced degree in engineering or physics. It is our contention that these ideas can be understood by anyone who is willing to learn. To prove our claim and to help surgeons navigate the bewildering array of features, manuals, guidelines, warnix
x
Image-Guided Surgery
of figures to help illustrate our points and make them simpler to absorb. But, at the same time, in keeping with our underlying goal of providing the necessary information to help clinicians understand the technology such that they will be able to better utilize it, we have abided by Albert Einstein’s admonition that Everything should be made as simple as possible, but not simpler. Thus, little of this book would qualify as light reading, and some may wish to skip some parts of it. In fact, it was written to be accessible in whole or in part, and each chapter can stand alone. For those who want a more general overview absent the underlying technology, other references exist, but we are confident that those who are willing to read all or most of this book will find it well worth the effort. While each chapter can be read in isolation from the rest of the book, we have included cross-references between chapters to help the reader integrate the various components into the whole of image-guided surgery (IGS). Chapter 1, “Brief History of Image-Guided Surgery”, traces the development of IGS from Roentgen’s discovery of x-rays through development of computed tomography (CT) and magnetic resonance imaging (MRI) and on to the current level incorporating tracking systems allowing navigation on CT and MRI images while operating. Chapter 2, “CT and MRI”, covers the basics of CT and MRI, including explicit presentations of the limits of imaging technology. Chapter 3, “Tracking Systems”, investigates optical (infrared and visual)
and electromagnetic tracking of objects — a necessary requirement for IGS. Chapter 4, “Registration”, explains how a CT or MRI image is superimposed onto intraoperative anatomy. Chapter 5, “Error Analysis”, treats the ubiquitous errors that exist in all IGS systems, and it provides explicit recommendations of ways in which a surgeon can minimize those errors. Chapter 6, “Best Practices for Use of IGS”, debates the evidence supporting the use of IGS in clinical settings. Chapter 7, “Surgical Systems”, presents currently approved IGS systems and expected accuracies based on laboratory and clinical studies. Chapter 8, “What Does the Future Hold?”, discusses likely short- and long-term uses of IGS, including augmented reality and robotic applications. The 21st century is still young but already promises to be a century noted for technological progress in medicine. Image guidance in surgery will certainly continue to be a big part of that progress, and new systems with new capabilities will steadily appear. As they do, the field may seem to grow more and more daunting, but the foundations of this field are in fact likely to remain the same. And, the practitioner who has mastered those foundations should be able to keep abreast of the changing landscape. We hope that this book will help guide you through that landscape, and we look forward to hearing from you, and about you, as you master these marvelous tools, as you use them in surgical interventions, and as you improve them to advance the future of medicine.
Rob Labadie [email protected] [email protected]
[email protected] [email protected]
Mike Fitzpatrick
Acknowledgments Koscielak, Megan Carter, and Nicole Bowman — for taking a chance on a book of this topic and for editorial and publishing expertise. I am sure this project would never have reached its conclusion without the continual prodding, critiquing, and encouraging of my coauthor, Mike Fitzpatrick. When I arrived at Vanderbilt in 2001 and met Mike (Bob Galloway, now Professor Emeritus of Biomedical Engineering, introduced us), little did I know how entwined our careers would be although on different ends of the career spectrum as I was but a naïve assistant professor and he was a sage full professor. Over the 15 years we have worked together, I have learned more from him than anyone else at Vanderbilt both about image-guided technology and about navigating academia. Our relationship has thrived based on our mutual respect for each other’s expertise yet the freedom to propose any idea. (The fact that we have a lot of fun working together only adds to the experience!) I am honored he choose to share the byline with me on this textbook. Another benefit of working with Mike was that I became integrated into the School of Engineering at Vanderbilt University through which I have had numerous fruitful collaborations many of which have had a direct impact on this book including those with Michael Goldfarb, PhD, Ramya Balachandran, PhD, Benoit Dawant, PhD, Jack Noble PhD, and Bob Webster, PhD. Collaborations such as these between surgeons and engineers may seem obvious but
Like most first-time authors (and despite admonitions from my veteran coauthor!) I grossly underestimated the amount of effort required to write a book. Having survived the journey, I am deeply indebted to the many who helped me along the way. First and foremost is my immediate family, especially my wife Karyn, who dealt with my fluctuating moods during the highs and lows of the project. Our four boys were supportive but thought it was just another one of dad’s crazy ideas. I thoroughly enjoyed bouncing ideas off our oldest son, an undergraduate physics major, over lunches during his summer internship at Vanderbilt in 2015. This project could not have transpired without a sabbatical — rare for an academic surgeon — which my chair, Ron Eavey, granted me for six weeks in January and February of 2015. During this time, my clinical colleagues, especially the neurotology service consisting of David Haynes, Marc Bennett, George Wanna, Alejandro Rivas, and Ken Watford, DNP; my nurse, Georgette Smiley, RN; and multiple other nurses, residents, and fellows who covered patient emergencies and urgencies, allowing me the privilege of protected time dedicated to the project. My administrative assistant, Maria Ashby, retired in the midst of the project but substantially started the enormous task of obtaining figure permissions. This task was carried forth by our newly hired lab manager, Jody Peters, who both finished that task and also proofread the entire manuscript. Thanks also to Plural Publishing — namely, Valerie Johns, Kalie xi
xii
Image-Guided Surgery
often are impeded by the lack of shared resources and/or misaligned incentives within academia. Lucky for us at Van derbilt, such collaborations have been encouraged and supported through the Vanderbilt Institute of Surgery and Engineering which receives generous internal and external funding to facilitate advances in health care outcomes based on implementation of engineering technology (https://www4.vander bilt.edu/vise/). Finally, I must state that this book is a huge “badge of honor” to me perhaps
born out of my heritage, most especially my aunt and godmother, Bernadine Meyer, a retired business law professor who wrote Legal Systems in 1978 and incorporated all members of our extended family into case studies (I was the president of a labor union and witness to an industrial accident) and my uncle and namesake, Fr. Earl (Robert) Meyer, who had Homilies of Father Earl Meyer published in 2013, 2014, and 2015. It is with great pride that I add ImageGuided Surgery: Fundamentals and Clinical Applications in Otolaryngology to this list! — Robert F. Labadie
Many people helped me along the path toward this book. Professor Edward Desloge was first. He had a profound influence on my writing and teaching as my PhD advisor in physics at Florida State University. I defended my dissertation in 1972 only a few months after Godfrey Hounsfield announced the invention of computer-assisted tomography. Desloge sagely advised me to enter the burgeoning field of medical physics, but I was young and foolish and ignored that advice — for nine years. In 1981, while I was on a sabbatical from teaching undergraduate physics, I met Professor Stephen Pizer of the Department of Computer Science at the University of North Carolina. His inspiring enthusiasm for medical image processing, which combined physics, computers, and medicine, finally won me over. I quit my tenured position, earned a master’s degree in computer science, and with Steve’s help landed an assistant professorship in computer science at Vanderbilt University.
Just five years later, Dr Robert Maciunas, a recently hired surgeon at Vanderbilt, walked into my office, introduced himself, began divulging some exciting ideas for improving brain surgery, and suggested that we might work together to make them happen. I took the plunge, and after 12 years, on the basis of that work, I was awarded a full professorship at Vanderbilt and he was awarded a chairmanship in New York. It was 1999, and I was now prepared for a relaxed glide path into retirement in the new century. Instead, the curtain was about to rise on major new phase of my life. Just two years after Bob Maciunas left, a new person appeared, bringing with him a hefty dose of déjà vu. Dr Robert Labadie, a recently hired surgeon at Vanderbilt, walked into my office, introduced himself, began divulging some exciting ideas for improving ear surgery, and suggested that we might work together to make them happen. Rob Labadie had a tougher sell than
Acknowledgments
Bob Maciunas. My position was secure, I was only nine years from retirement, and I was reluctant to tackle yet another region of anatomy. However, one need experience the Labadie personality but once to understand why I was swept along. I have never met a more persuasive, enthusiastic, optimistic, amiable, kind, and brilliant person. And there was another factor — Rob Labadie is a surgeon who understands physics and mathematics! So I took another plunge. We formed a team, and, with this book, we are completing our 15th year of a collaboration that has been the most successful of my career — and the most fun. All four of these people shared important ideas with me and encouraged me, and as a result I owe them a huge debt of gratitude for any success that I may have had, but others have shared ideas with me as well. First, there are the many graduate students that I have had the pleasure to advise over the last 30 years, including my dear friend Dr Ramya Balachandran, who, upon receiving her PhD in computer science in 2008, worked with Rob and me
as a research assistant professor. I am indebted to many other colleagues as well, both inside and outside Vanderbilt with whom I have collaborated for over 25 years, most notably Professor Benoit Dawant of Electrical Engineering, Professor Emeritus George Allen of Neurological Surgery, and Professor Emeritus Robert Galloway of Biomedical Engineering. I thank them all. Finally, I wish to thank my dear wife, Dr Patricia Robinson, who put up with many lonely evenings and weekends while Rob and I worked on this book and who, despite a busy pediatric practice, has supported my career every inch of the way and has given me two wonderful children. Pat continually amazes me both with her deep and abiding concern for her patients and with her extraordinary deductive powers when the data are so sparse and the diagnosis is so crucial. The way she practices medicine reminds me daily that while the marvelous technological breakthroughs described in this book represent major advances in health care, the most important tool will always be the human mind. — J. Michael Fitzpatrick
xiii
1 Brief History of Image-Guided Surgery when to trust anatomical knowledge. Although the comparison between IGS and GPS is useful in conveying what each technology can do, there are some fundamental differences between the two (see boxed text Chapter 3). One useful similarity, however, is how— despite their obvious benefits — they can get users into trouble — for example, the naïve driver who does not understand the limits of GPS while operating at the limits of his or her knowledge of the terrain and trusts it when it recommends a shorter route over a mountain pass in inclement weather or the naïve surgeon who does not understand the limits of IGS while operating at the limits of his or her surgical skills and trusts it when it puts the crosshairs inaccurately on the surgical target, erroneously guiding the surgeon to remove vital tissue. Image-guided surgery (IGS) involves linking a preoperative image, most commonly computed tomography (CT) or magnetic resonance imaging (MRI),i
Overview of How Image-Guided Surgery Works Anyone who has learned to drive within the past decade probably thinks of a paper map as a museum artifact and is unlikely to navigate anywhere without using a global positioning system (GPS). Similarly, current surgical trainees are unlikely to practice without imageguided surgical (IGS) systems, which are often compared to GPS albeit on a smaller scale in the operating room. And who wouldn’t want to have this amazing technology available to see things inside the human body and know precisely where those things are? Who doesn’t want Superman’s x-ray vision? But, as the mythical Superman understood, with great abilities come great responsibilities, and in the surgical arena, this means understanding how IGS works so that surgeons know the limits of the technology — when to trust IGS and i
In this book, we will consider only CT and MRI because these are the imaging modalities overwhelmingly used for IGS in otolaryngology.
1
2
Image-Guided Surgery
to a patient’s intraoperative anatomy, allowing one to navigate using the image as a guide or map. Current IGS systems are more similar than different among various vendors (Chapter 7), and all use tracking both to identify points that will be used to register the preoperative image (Chapter 2) to the patient’s location in the operating room and to navigate during surgery (Figure 1–1). Tracking (Chapter 3) may take the form of opticalii tracking, which localizes via triangulation, in which dimensions of virtual triangles connecting known points with an unknown point are solved using geometry, or the form of electromagnetic tracking, in which a probe disrupts an electromagnetic field (EMF) and the disruption can be correlated
to position. Some tracked points are denoted as fiducials (Chapter 4), which may consist of unique patient anatomy or of markers affixed to the patient. The locations of fiducials are specified in the preoperative image (Chapter 2) and, after they are localized in the operating room using the tracking system (Chapter 3), the two sets of locations are overlaid onto each other in a process known as registration (Chapter 4). After registration is performed, a tracked probe can be used to navigate in the surgical space, which is registered to the image space, and it is this registration plus tracking that makes IGS possible. Sounds simple — right? Well . . . the underlying concepts are sound, but Le
Figure 1–1. A generic optical tracking IGS system is shown at the left and a generic EMF IGS system is shown at the right. The surgeon stands opposite a video monitor that shows the position of the tracked probe on a preoperative image such as a CT or MRI. For the optical system, an infrared camera system sends out pulses of infrared light that reflect off markers attached to the probe, held by the surgeon, and a coordinate reference frame (CRF), affixed to the patient to allow tracking of the head; depicted is the more common “passive system”, which does not require hardwiring between the tracked devices (the probe and the CRF) and the computer. For the EMF system, the probe is usually hardwired to the EMF generation unit, shown as a gray cube. Tracking systems are discussed in more detail in Chapter 3. ii
In this book, unless indicated otherwise, “optical” denotes the portion of the electromagnetic spectrum that can be directed and focused by means of conventional lenses and mirrors (eg, visible and infrared light).
Chapter 1
Brief History of Image-Guided Surgery
n
bon Dieu est dans le detail!iii Be you an optimist or a pessimist, the details of IGS are vital in minimizing error (Chapter 5) — which never goes away — and ignoring these details accounts for the vast majority of misuses of IGS. But, we’re getting ahead of ourselves because to appreciate current IGS systems, we need to learn how we have arrived at a world where IGS systems are all but ubiquitous in modern operating theaters.
The Evolution of IGS Images Without images, there would be no image-guided surgery, so the history of IGS is intimately linked to the history of radiology, which had its seminal event in 1895 on November 8 when Wilhelm Conrad Röntgen discovered x-rays (called “x” to designate an unknown) at the University of Würzburg, Germany. Two weeks after his initial discovery (and without institutional review board approval!), he captured a now famous image of his wife’s hand (Figure 1–2). Although he knew that the commercial potential was huge, Röntgen decided against patent protection because he felt that his discovery belonged to humankind. His findings were published December 28, 1895, and he was awarded the first Nobel Prize in Physics in 1901. He donated the prize money to the University of Würzburg. Röntgen’s discovery had a surprisingly quick bench-to-bedside transition being used early the next year, 1896, for multiple applications. J. H. Clayton of iii
Figure 1–2. The first known x-ray image produced by Röntgen of his wife’s hand at the University of Würzburg, Germany (albeit without ethical board approval!).
Birmingham, England, is given credit for the first IGS intervention, which occurred a mere eight days after the publication of Röntgen’s discovery. Clayton used an x-ray machine to identify an industrial sewing needle embedded in a factory worker’s hand and to assist him in its removal.1 IGS crossed the Atlantic one month later when John Cox at McGill University in Montreal, Canada, used an x-ray image to remove a bullet from a limb,2 and in America, in February 1896, a New York surgeon, Dr Bull, asked physicist Michael Pupin to obtain an x-ray image of the hand of a patient with embedded buckshot to assist in its removal3 (Figure 1–3). Otolaryngology’s ties to this history began early when the first Department
Gustave Flaubert, 19th century, God is in the detail. This original quote is the origin of today’s The devil is in the details.
3
4
Image-Guided Surgery
the discovery of computed tomography and magnetic resonance imaging. Discovery of CT
Figure 1–3. An early x-ray image taken by physicist Michael Pupin circa 1896, which was used by Dr Bull to extract buckshot from the patient’s hand. This greatly facilitated the intervention that was performed quicker than anticipated.
of Radiology at the Glasgow Royal Infirmary (Glasgow, Scotland) was established in 1896 by laryngologist John Macintyre, who imaged, among other things, a coin in a child’s throat.4 Military applications followed quickly, when x-rays were used to find and treat both fractures and embedded shrapnel, first during the Italo-Abyssinian War and subsequently the Boer War and World War I.5 Although multiple other uses were conceived of during the early part of the 20th century, IGS was hampered by the need for three-dimensional depictions instead of the two-dimensional shadows produced by x-ray projection. The third dimension would come in the third quarter of the 20th century with
In 1967, while the Beatles were working on their groundbreaking album, Sgt. Pepper’s Lonely Hearts Club Band, for release by Electric and Musical Industries (EMI), Ltd., another groundbreaking development was taking place inside the same company. While EMI was prospering from the profits generated by the Beatles phenomenon,6,7 an EMI engineer, Godfrey Newbold Hounsfield, was working stealthily on the first CT scanner. By 1968, he had completed a working prototype that proved the concept and had submitted a patent application that would be granted in 1971. In 1973, EMI announced the world’s first working clinical model.8 Its images were crude and required two days to compute, but it revolutionized diagnostic medicine and surgery and led ultimately to today’s remarkable instruments. Early clinicians were amazed by the EMI image (Figure 1–4). Mike Glasscock, founder of the Otology Group in Nashville, Tennessee, recalls, for example, that it allowed him — for the first time ever — to be able to see how big a vestibular schwannoma was prior to beginning the surgery. Thus, he could plan his surgical approach and estimate time of intervention before cutting skin. Today, continual improvements in CT seem nowhere near their end. But neither was Hounsfield’s work the beginning. The idea of combining radiographic information of a patient acquired from multiple directions to produce a picture of a single slice (“tomos” =
Chapter 1
Brief History of Image-Guided Surgery
n
Figure 1–4. An early EMI scan circa 1975 from Dr Michael Glasscock’s practice in Nashville, Tennessee, showing a large leftsided vestibular schwannoma. Such images, although considered “rough” by today’s standards, allowed surgeons to predict how long a surgical intervention might take.
slice; “graphy” = producing a picture) through the body was born a half century earlier. In 1917, just over 20 years after Röntgen’s discovery of x-rays, André-Edmund-Marie Bocage conceived of x-ray tomography.8 In 1921, he applied for a French patent on the idea, which was issued in 1923, and by 1940, nineteen additional tomography patents had been awarded. More were to follow, and by the time Hounsfield began his work on his prototype at EMI, over 3000 articles had been published on it and over 50 commercial tomographic imagers had been introduced. At that time, in the late 1960s, over 60% of the radiologists in the United States had one, but they used them in only 1% of their cases!8 Why? — because the images of these slices through the body were badly blurred by shadows of remote parts of the body. Tomography from 1917 to the 1960s was noth-
ing like the tomography of today. Those largely unused imagers produced their slice images by directing x-rays roughly perpendicularly to the slice, similarly to plain-film radiography, and as a result, the tissue overlying and underlying that slice added confusing shadows that confounded the desired tissue slice with irrelevant anatomy. This corruption of the image is an inherent problem with any approach to x-ray tomography because the signal produced in each sensor, whether it is a grain of x-ray film, a phosphor, or a silicon chip, is produced by the cumulative effect of everything the x-ray encounters along its path through the body. It is up to the imaging device to tease out the true individual intensities at each of the points within a slice from these integrated signals. Early researchers in the field of x-ray tomography were well aware of the problem, and in 1940, a major step toward solving it was invented and patented by Gabriel Frank.8 Frank approached slice imaging from a different angle — literally! His rays were projected not perpendicularly to the slice but sideways, into the edge of the slice, and they traveled entirely within the slice, and — most important — never passed through any anatomy outside the slice. Frank’s patent describes an ingenious photographic process not only for acquiring such projections from multiple angles (by rotating the patient) but also for combining those projections via a second photographic projection (by rotating a cylinder of film) to produce a tomogram. His method was not implemented, but if it had been, while it was a giant step beyond the prevailing approaches, the image would still
5
6
Image-Guided Surgery
have suffered from some residual blurring because of a flaw in his method for combining the projections. To produce a clear image from Frank’s projections, a sophisticated mathematical computation is required. A method for doing it had been published in a paper unrelated to imaging in 1917,9 which by a remarkable coincidence was the same year that Bocage invented tomography, but there was no practical way for Bocage or Frank or anyone else to implement that method or any method for avoiding the blur until a breakthrough appeared in another field. The other field was numerical computation, the breakthrough was the digital computer, and Hounsfield was an expert in both. In fact, his first job at EMI was to lead the development of a transistorized computer, the EMIDEC 1100, which he completed in 1959. Computers like the EMIDEC provided the first opportunity to implement in a practical way a method to solve the blurring problem of tomography. Hounsfield had one handy, and he knew how to use it. Hounsfield used the EMIDEC to show that CT was clinically practical, but he was in fact not the first to produce a scanner. Indeed, he shared the Nobel Prize with Alan MacLeod Cormack, who had built a working prototype in 1963 that was not very different from Hounsfield’s,10 and a complete system for doing it had been described in Russia in the 1950s.11 But Hounsfield was the first to suggest the idea of digital reconstruction and was the first to show that it could be practical with a high-speed computer. In short, he
iv
added the term “computed” to tomography, and that was the groundbreaking development at EMI. Hounsfield’s first tomograms were still a bit blurry, but they represented an obvious, major leap forward. Thanks to digital reconstruction with a high-speed computer, the tomographic fog had begun to lift, and the world of medicine would never be the same. Discovery of MRI Our history of magnetic resonance imaging (MRI) begins with the discovery in 1933 that a proton is not only a charged particle but also a magnet.12,13,iv Its magnetism means that the nucleus of hydrogen, which is itself a single proton, will — like the needle on a compass — tend to line itself up with the field of any nearby magnet. Many atomic nuclei are magnetic, but the ubiquity of water in the human body makes the proton’s magnetism very important medically. It is important not because of any physical effect that magnetism has on the body itself but because its magnetism makes MRI possible. When a patient slides into the maw of an MR imager’s cylindrical electromagnet and is exposed to its magnetic field, although there is virtually no sensation or visible sign of it, hydrogen nuclei within every cubic nanometer of the patient’s body will immediately experience a torque toward the axis of the magnetic field. Their alignment with the field is always imperfect, and, analogous to a child’s top when it is not aligned with the vertical, they begin to precess around that axis.
Otto Stern was awarded the Nobel Prize in 1943 for this discovery.
Chapter 1
Brief History of Image-Guided Surgery
n
It is perhaps amazing that exposure of the entire body to a strong magnetic field appears to have no effect on physiologic states. The lone exception is transient eye motion, nystagmus, which occurs in scanners with field strengths of 3T and higher.14 Although the alignment of nuclei would not be expected to alter chemical reactivity, the alignment of the protons (electron alignment is negligble) might be thought to change it in some way. It does not. The real concern with regard to physiologic effects within an MR scanner arises not from the magnetic field directly but from electric current in tissue induced by rapid changes in the field. These changes are necessary to make imaging possible, but they are today kept far below the danger level. Nevertheless, what follows below attests to either the bravery or blind ambition of the physicists who discovered MRI. They knew the potential danger of induced currents, and yet they experimented on themselves, colleagues, and family members. The magnetism of hydrogen nuclei would be of little consequence in medicine except for a phenomenon discovered in 1938 called nuclear magnetic resonance (NMR).15 NMR is the absorption of energy by magnetic nuclei that are precessing around a static magnetic field when they are exposed to a second magnetic field that is oriented at right angles to the first and is rotatingv v
at, or very near, the precession frequency. The energy change imparted is similar to that which occurs when one pushes a child on a swing. If the pushes are timed correctly (the resonance frequency of the swing), each push imparts increased energy to the child, who swings with a larger amplitude, but it does not change the frequency of the swing’s oscillations. When the rotating field is turned off, the nuclei emit some of their energy as electromagnetic radiation, which oscillates at the same frequency. MRI relies on this effect (see the detailed explanation of how MRI works in Chapter 2), but turning it into an image requires much more ingenuity than CT. The difficulty is that the wavelength of radiation produced by these magnetic nuclei is very long — about 2 m or more, depending on the strength of the MR magnetvi — compared with 1 × 10−11 m for the typical x-ray and 5 × 10−7 m for visible light. So, the detector cannot determine the direction from which NMR radiation comes other than that it comes from somewhere inside the scanner. Attempting to use MR radiation directly to form an image, in the manner that Röntgen did when he cast a shadow of his wife’s hand on photographic film with beams of x-rays, would be like trying to paint a life-sized picture depicting anatomical detail with a brush that is bigger than the human body! Although the long wavelength of MRI is a disadvantage relative to the short wavelength of CT, there is a big advantage as well. The advantage is that there is no danger of ionization with MR, because the
Or oscillating. 2.3 m for a scanner whose magnetic field strength is 3T and varies inversely with field strength.
vi
7
8
Image-Guided Surgery
energy of a photon is inversely proportional to the wavelength and thus the energy of a photon of MR radiation is tiny — less than one millionth of the energy required to ionize even one atom. By comparison, one typical x-ray photon used in CT is powerful enough to ionize many thousands of atoms on its way through the body. The impossibility of using MR “beams” to produce a picture directly means that indirect methods are required. Such methods were first developed in the 1970s, and they remain in use today. The first one is today called “slice selection”. Slice selection is a technique for exciting protons in just one thin section, or “slice” of the body, so that the radiated energy is proportional to the sum of the hydrogen atoms in just that slice. Slice selection can be repeated multiple times to obtain signals from multiple slices. Mapping the relative number of hydrogen atoms in slices of the body by spatially selective excitation is a first step toward mapping the number of protons at each point in the body, and the idea of spatially selective excitation for the purpose of imaging occurred to at least three people independently during the beginning of the 1970s. During the period from 1971 to 1973, Raymond V. Damadian (Figure 1–5) at the State University of New York (SUNY) Downstate Medical Center in Brooklyn16; Paul C. Lauterbur, just 50 miles away at SUNY’s Stony Brook campus16,17; and Peter Mansfield at the University of Nottingham, England,18,19 all conceived of methods for doing it, and in 2003, Lauterbur and Mansfield ivii
were awarded the Nobel Prize in Physiology or Medicine for MRI. The reason for the omission of Raymond Damadian by the Nobel committee probably had to do with practicality. His method never led to a useful scanner. Lauterbur’s approach was itself only marginally better, but he was the first to publish the idea of slice selection,vii and variations on that theme have been used ever since. Lauterbur’s approach was to acquire a series of signals from slices selected at multiple positions and angles and then to reconstruct an image from those signals. His method was remarkably similar to Hounsfield’s highly successful method for producing an image from multiple x-ray projections, including employing Hounsfield’s method for calculating the image from the projections.20–22 However, when applied to MRI, Lauterbur’s approach suffers from two major problems: (a) Because of limitations inherent in NMR relaxation times, the rate at which signals can be obtained with his technique is orders of magnitude below that of x-ray acquisition, so Lauterbur’s method was extremely slow. (b) Unlike an x-ray projection, which can be counted on to give information from a straight, narrow path, the information received from NMR slice selection is not so well defined. Slight spatial variations in the static field caused by magnetization of the patient’s body itself viii cause the slice to be warped in an unknown way, resulting in a blurred image. (We will revisit this problem in Chapter 2 at the beginning of the section named “Phase Encod-
Lauterbur did not call it that, and one needs to read his first paper carefully even to recognize it. Patient magnetization comes predominantly from the electrons of the atoms — not from their nuclei.
viii
Chapter 1
Brief History of Image-Guided Surgery
n
Figure 1–5. Raymond Damadian, from the State University of New York (SUNY) Downstate Medical Center in Brooklyn, in an early MRI scanner. The man on the left is a cardiologist present to monitor his EKG and blood pressure. The device on his chest consists of the antennae for detecting the emitted signal. Image provided courtesy of FONAR Corporation.
ing”.) Lauterbur’s method also suffered from a minor problem: as described first in 1973, it worked only for objects that varied in only two dimensions.
In 1974, Mansfield and his students, Allen Garroway and Peter Grannell, published a paper that described and demonstrated images acquired with
9
10
Image-Guided Surgery
an approach to MRI that reduced the problem of slow acquisition, solved the problem of blurring, and also allowed for three-dimensional objects.18 While Lauterbur’s approach measured the strength of the signal emitted from an entire selected slice, Mansfield caused the nuclei within it to precess at different frequencies across the slice, so that frequency analysis of the signal could reveal their positions within the slice. This “frequency encoding” of position is now standard in virtually all MRI image acquisitions. It eliminated the need for Lauterbur’s CT-like imageconstruction step, reduced the imaging time by a factor of 30 or so, and eliminated the blurring problem caused by spatial variations in the static field. That same paper describes a feasible approach to true three-dimensional imaging as well, and in 1976, Mansfield showed the first MR image of a living human — the cross section of a finger of one of his studentsix— at a special meeting of the London Medical Research Council on MRI (Figure 1–6). Scanning one cross section of one finger required 20 minutes in 1976, and it would take many hours to produce an MR image approaching the resolution and size of CT images available at the time. However, in 1974, the same year that Mansfield’s group published frequency encoding, while listening to Lauterbur explain his method at a conference in Raleigh, North Carolina, Richard Ernst of ETH Zürich conceived of a second way to speed things up.23 His method exploited an additional piece of information available in the NMR signal — phase. Phase, which is ix
often measured in degrees, is the portion of an oscillation that has been completed by an oscillator at a given moment in time, and when the oscillator is a precessing nucleus, the phase also measures the portion of the trip around the axis that it has completed — much like longitude measures a trip around the Earth’s equator. Ernst realized that the phase of the MR signal, which could be measured with available hardware, could be exploited during the imageformation process. Ernst’s method was subject to rather severe geometric distortion, which makes it unsatisfactory for IGS, and was superseded in 1980 by a superior method called “spinwarp imaging” that was introduced by William Edelstein at The University of Aberdeen, Scotland24 and is appropriate for IGS. Spin-warp imaging largely rendered Ernst’s approach obsolete, but the speedup that he introduced with phase encoding was the final innovation that gave MRI the potential to become a clinically useful modality. Meanwhile, Mansfield had realized that the speed of MR acquisition would be the major hurdle between the current state of MRI and routine medical imaging. In his quest to speed things up, he discovered the benefit of phase independently about two years after Ernst. According to Mansfield’s recent memoir on the history of MRI,25 he learned of Ernst’s method at an MRI meeting in Nottingham in 1976 where he himself presented a different phaseencoding method. In 1977, he published that method and gave it the now well-known name, “echo-planar imaging” (EPI). EPI combines slice selection
he one member of his research group with a finger small enough to fit into the 1.5-cm bore T of their NMR spectrometer!
Chapter 1
Brief History of Image-Guided Surgery
n
A
B Figure 1–6. First magnetic resonance image of live human tissue, specifically the finger of one of Peter Mansfield’s graduate students, whose finger could fit inside the 1.5-cm bore of the machine! Although “rough” by today’s standards, it was remarkable in that soft tissue could be differentiated, including blood vessels, muscles, and bone marrow. A. Cross-sectional images of a finger obtained in vivo by NMR (see text for full details). Left panel: TR (see the “Relaxtion Times Section of Chapter 2) = 0.5 seconds. Right panel: TR delay = 0.3 seconds. B. Key. Republished with permission of British Institute of Radiology, from Mansfield P, Maudsley AA. Medical imaging by NMR. Br J Radiol. 1977;50:188–194.
with phase encoding, and it produced a dramatic improvement in acquisition speed with the potential to acquire a two-dimensional image of one slice in
40 to 50 ms.26 EPI is unfortunately subject to even more severe geometrical distortion than Ernst’s method and is likewise ill-suited to IGS, but neverthe-
11
12
Image-Guided Surgery
less, because of its incredible speed, it is a widely used acquisition method. It is in fact the standard MR protocol for “functional” MRI, in which physiologic changes in the brain are measured as a function of time and that require acquisition of sets of dozens of high-resolution slices of the brain at the rate of a few seconds per set. In 1978, after designing and overseeing the manufacture of a full-sized body scanner, Mansfield stepped inside it, closed the door, and remained as motionless as he could in absolute darknessx while it acquired signals from his body for 50 minutes. The resulting image of his abdomen was presented at the Experimental NMR Conference in Blacksburg, Virginia, in April 1978 (Figure 1–7). The acquisition and presentation of this image proved that clinical imaging based on NMR was a reality,
and it marked the end of the beginning of MRI.
Putting It All Together: CT and MRI in IGS Although various individuals have attempted to take credit for the concept of IGS, it appears that the concept has been universal among surgeons since Röntgen discovered x-rays in 1895, as is perhaps most evident from the almost immediate utilization of x-rays to help guide surgical interventions (see Figure 1–3). Neurosurgery was the first surgical discipline to embrace IGS, and efforts at intracranial navigation date back to Horsley and Clark’s stereotactic frame circa 1908,27 which begat numerous other stereotactic devices as nicely summarized by others.28 Why neurosurgery led the charge over otolaryn-
Figure 1–7. First whole-body magnetic resonance image of Peter Mansfield’s abdomen presented at the Experimental NMR Conference in Blacksburg, Virginia, in April 1978. For this and other work, Peter Mansfield shared the 2003 Nobel Prize in Physiology or Medicine with another MRI pioneer, Paul C. Lauterbur. Republished with permission of British Institute of Radiology, from Mansfield P, Maudsley AA. Medical imaging by NMR. Br J Radiol. 1977;50:188–194. x
The scanner’s cavity was vertical, so Mansfield was standing.
Chapter 1
Brief History of Image-Guided Surgery
n
The first scans required a measure of bravery because the participants did not know whether exposure to the magnetic field would cause any physiologic changes in either the short or long term. One week before Mansfield’s historic first whole-body scan, a colleague, Thomas Budinger, who 7000 miles away in San Francisco had heard of what Mansfield planned to do, sent him a warning that, according to his calculations, changing magnetic fields during scanning could induce currents in his body that would trigger cardiac fibrillation! Mansfield, relying on his own calculations, nevertheless subjected himself to the scan, after a brief test pulse, without incident. He then traveled to Virginia to give his
gology, which traditionally has been one of the most progressive surgical subspecialties, is open to debate, but it appears to have been motivated at least in part by the early recognition that limiting removal of vital neural tissue was an imperative in neurosurgery and that IGS could help meet this imperative. Early ear, nose, and throat (ENT) IGS systems were essentially modified neurosurgical systems, and even today, there is wide overlap between the two. As noted above, you can’t have IGS without the I — imaging. Perhaps not as obvious, but very important, is that the imaging must accurately depict the tissue that it images. When CT scanners were introduced in 1973, those early scanners had poor axial consistency. As detailed in Chapter 2, CT scanners take two-dimensional (2D) “slices” of a patient and “stack” them into a threedimensional (3D) volume. When the
talk and afterward became so ill that he canceled other talks in the United States and returned to England. Upon presenting to his primary care doctor, Mansfield was told that he was the expert on the interaction of magnetic fields on the human body and that he was the only one who could determine whether his illness was related to his scan! Mansfield suffered with insomnia for years afterward, worried that he had overlooked some aspect that would lead scanners to harm human beings. To date, MRI has never been shown to lead to any permanent deleterious effects other than those caused by the induction of currents in surgically implanted objects with large conducting loops.29
slices are geometrically misaligned with each other, as was common with early CT scanners, and/or there is movement of the tissue between the acquisitions of slices, which was common because of patient motion during long acquisition times, the slices do not properly align, and, when they are stacked together, the result is a geometrically skewed volume. Patient motion was overcome with the stereotactic frame by using it to fix the patient’s head to the scanner bed. The frame was rigidly attached to the patient’s skull, and the frame was rigidly attached to the bed, and as a result, the skull and its contents remained motionless relative to the CT bed as it advanced through the gantry. The remaining problem of the geometrical misalignment of the scanner’s slices with each other was overcome in 1979 by the invention of the N-bar reference
13
14
Image-Guided Surgery
system by Russell Brown.30,31,xi This ingenious device consists of a set of three N-shaped bars that are rigidly attached to the frame. These N-bars surround the patient’s head, and cross sections of their individual members appear in each slice (Figure 1–8). The two-dimensional coordinates of the surgical target in the slice in which it appears and of the centroids of the cross sections in that same slice are enough to allow the determination of the target position relative to the frame. That position is then used for targeting in the operating room. This arrangement provided a precise geometric reference in each
slice that could be linked to the patient’s intraoperative anatomy independently of all the other slices, whether or not those slices were aligned with each other. The stereotactic frame in combination with the ingenious N-bar is of historical significance in IGS because it provided reliable geometry in the face of geometrically skewed images. This combination is still widely used today (eg, the BRW [Brown-Roberts-Wells], CRW [CosmanRoberts-Wells], and Leksell frames), and it is considered by many neurosurgeons as the gold standard for IGS. However, in the last 10 years, the N-bar has become largely unnecessary
A
B
Figure 1–8. The N-bar system (three rigid “N”-shaped bars), invented by Russell Brown in 1979 and still used today. This fiducial system solved the problem of geometric misalignment (skewing) between CT scan slices by providing reference marks in each slice that allow determination of the position relative to a target in one slice without reference to other slices. A. Schematic of the N-bar system attached to a base ring that is rigidly attached to the patient’s skull prior to imaging. The tilted rectangle represents a CT slice, and the seven ellipses show where the slice intersects the seven members of the N-bar system. B. Corresponding CT slice showing a head cross section and the seven ellipses. The two-dimensional slice coordinates of the target and of the centroids of the ellipses are sufficient to calculate the target position relative to the base ring. In the operating room, the N-bar system is replaced with an adjustable probe attached to the base ring and aimed at the calculated target position in reference to the ring. Figure A republished with permission of the Institution of Mechanical Engineers from J M Fitzpatrick, The role of registration in accurate surgical guidance, Proc. IMechE Vol 224 Part H: J. Engineering in Medicine, 2010. xi
Russell Brown, MD, PhD, was both a physician and a computer scientist.
Chapter 1
Brief History of Image-Guided Surgery
n
because of breakthroughs in CT technology. First, in the late 1980s and early 1990s, the “slip-ring” was introduced into CT.xii This technology consists of a set of electrified metal rings on the stationary part of the machine and a set of contacts on the rotating part. The contacts touch the rings to allow current flow for powering the x-ray generator and detectors and for communicating signals between the rotating and stationary parts. This feature allows for continuous rotation in one direction. Before slip rings, a back-andforth movement was required to allow connecting wires to wrap and unwrap between slice acquisitions. This repetitive reversing required that brakes be applied to bring the gantry to a complete stop, after which it had to be accelerated in the opposite direction to full speed, limiting the rate at which successive slices could be acquired to roughly one slice per 3 seconds.32 This jerking motion often led to angular misalignments between slices. After slip rings, CT scanning technology advanced to the point of producing accurately aligned slices at the rate of one slice per second. Continuous rotation via the slip ring along with advances in image-reconstruction algorithms made possible the introduction of “helical scanning”xiii in commercial systems in the early 1990s. Before helical scanning, each slice was acquired while the x-ray source and the detectors executed a 180-degree circular path about a stationary patient, after which the bed and patient were advanced and xii
then halted in preparation for the next slice. With helical scanning, the bed moves continuously throughout the acquisition of all slices, so that, relative to the patient, the path of the rotating x-ray source describes a helix. In the late 1990s, multislice helical scanners were introduced. With these systems, multiple slices are acquired simultaneously. At about the same time (late 1990s), rotational speed increased to the point that two rotations of a fourslice CT were being acquired in just one second.32 This remarkable trend toward faster and faster CT acquisition has continued, and in 2007, Toshiba introduced a scannerxiv that could acquire 160 slices of 2.0 mm thickness, covering the entire brain in one rotation in less than one second! With this development, the goal of geometrically accurate CT with negligible patient motion had been achieved, and today 640 slices of 0.5-mm thickness can be acquired in less than one-third of a second. Thanks to these remarkable strides in CT technology, today, geometrically faithful CT image volumes of the head are readily available without the need for frame fixation or N-bar localization. Thus, we consider this achievement of geometrically accurate, rapid volume CT during the two decades from roughly 1990 through 2010 to be as important a milestone in IGS as the discovery of x-rays (1895), the development of the first stereotactic frame (1908), the discovery of CT (1972), the discovery of MRI (1973), and the invention of the N-bar (1979) (Figure 1–9).
First appearing in the SOMATOM PLUS (Siemens Medical Systems) and the TCT 900S (Toshiba Medical Systems). xiii Also called “spiral scanning”. xiv The Aquillon ONE (Toshiba Medical Systems).
15
16
Figure 1–9. Time line of major events that have made IGS possible. Although IGS was clinically utilized within mere months after the discovery of x-rays, it took over 100 years for further technological advances to be developed in order to capitalize on the true power of IGS, namely real-time, three-dimensional navigation on the basis of a geometrically accurate three-dimensional image of the patient’s anatomy.
Chapter 1
Brief History of Image-Guided Surgery
n
Meanwhile, MRI technology was not standing still. MR scanners of the late 1980s and early 1990s, which suffered from thick slices, long imaging times, and geometric distortion, were replaced in the late 1990s and early 2000s by scanners with higher resolution, faster imaging, and more geometric fidelity. MRI is treated in detail in the next chapter, but here we will point out that there was another important breakthrough that made MRI and CT both more useful to IGS. In 1995, a new computer algorithm for aligning, or “fusing”, a patient’s MR and CT scans was published,33,34 and it was shown in a blinded validation35,36 to be capable of submillimetric accuracy when MR distortion was corrected. The advent of submillimetric MRI-CT fusion meant that fiducial registration employed in one of these image modalities was sufficient to register both modalities to physical space with fiducials required in only one modality. This method of fusion, which is based on a measure called “mutual information”, remains today the standard for MR-CT registration and also deserves to be recognized as a critical milestone in IGS development. Fusion is discussed further in Chapter 4. With the advent of geometrically accurate, rapid, volume CT scanning, point-based registration, in which a set of small, discrete fiducial markers is used to register (Chapter 4) the preoperative images to the intraoperative anatomy, IGS became capable of achieving navigational accuracy comparable to that with an N-bar frame. xv
Upon its introduction, IGS was known as “frameless stereotaxy” because it eliminated the frame. The analogy between frameless stereotaxy and the horseless carriage was nicely developed by the late Robert Maciunas, MD, in his introductory chapter to the book Interactive Image-Guided Neurosurgery.37 In that work, he sagely pointed out that, as advances are made, they are referenced to existing technology perhaps to allow some comfort to adoptees through its comparison to the prior standard. However, as the technology evolves, referencing the prior standard becomes passé, as it eventually did with the “horseless carriage”, which became known, more aptly, as the “automobile”, As IGS becomes the new standard in surgical navigation, the stereotactic frame can be expected to take its place in history alongside the horse. But this reference to old technology has now been dropped by most practitioners in favor of the more apt phrase “image-guided surgery”, which appears to have been coined and first published by Robert L. Galloway in 1991.38 A consequence of these advances was that, in principle, IGS without the stereotactic frame became a possibility. Today, IGSxv without a frame has become commonplace throughout the surgical world with a great variety of commercial options available, but specialized technology had to be developed
IGS is a very broad field. From this point on, we limit our discussion of IGS to cranial applications in which the anatomy (eg, sinus and/or mastoid cavities and their contents) is enclosed by rigid anatomy, namely bone.
17
18
Image-Guided Surgery
in order to allow it to happen. Systems for navigation (Chapter 3) and fiducial markers for registration (Chapter 4) had to be developed to allow tracking of surgical instruments relative to the anatomy with submillimetric accuracy. Tracking began with mechanical linkages that employed posable, articulated arms with kinematic encoders that allowed the calculation of the tip location with respect to the base.39 Probably because of the awkwardness of these systems in the surgeon’s hands, they were soon replaced by systems that track without the need for physical linkage. These armless options (“armless” is another adjective that, like “frameless”, has fallen out of use) include sound localization,40,41 optical localization,42–46 and electromagnetic localization.47,48 Tracking by sound localization was short-lived because the speed of sound in air is dependent on the temperature of the air along the sound’s path, and as a result, operating room (OR) lights can cause significant tracking errors. A better technology was infrared (IR) tracking via triangulation, which was modified for clinical use in the early 1990s and is today the most common tracking method used in commercial IGS systems. Another alternative, and one that was commercially successful in ENT before IR, is electromagnetic tracking, in which the location within a generated EMF is sensed and can be correlated to positional changes. Fiducial marker technology needed to evolve as well and progressed from the N-bar frame to “frameless systems” with a minimum of three boneimplanted markers.45,49–51 Although bone-implanted markers are still considered the gold standard for fiducial markers, for ENT applications, in which
interventions are often undertaken to treat non-life-threatening diseases (eg, chronic sinusitis), less-invasive fiducials were adopted despite the decrease in navigation accuracy associated with them. This step backward in accuracy to reduce invasiveness has included skin-affixed fiducials,52 use of anatomical structures,53 and anatomical surface matching.54 The first detailed report of IGS in otolaryngology was by a group from Aachen Technical University in Aachen, Germany, which utilized bone-affixed fiducials for registration and a mechanical arm for tracking. Their work was first published in German49 and subsequently in English.55,56 These reports described use of their IGS system for endoscopic sinus surgery (n = 232, Figure 1–10) as well as skull base approaches (n = 37) at multiple institutions (5 hospitals) by numerous surgeons. Although their system does not appear to have been a commercial success, they are the first group to have routinely used and reported their experience with IGS in otolaryngology. The Aachen group’s work was followed by numerous reports of the use of ISG Technologies’ “Viewing Wand” (Mississauga, Ontario, Canada) for IGS in otolaryngology (Figure 1–11). It was initially used in Europe57 and England,58 following which it received Food and Drug Administration (FDA) approval in 1994 and was quickly embraced by centers on the forefront of IGS.50 Like the Aachen system, the ISG system used an articulated mechanical arm (FARO Medical Technologies, Lake Mary, Florida) to reference intraoperative position to preoperative CT or MRI data displayed on an Allegro computer. With the advent of less obtrusive tech-
Chapter 1
Brief History of Image-Guided Surgery
n
Figure 1–10. A view of the display of the IGS system used by the group from Aachen Technical University in Aachen, Germany, who were the first to routinely use IGS for ENT surgery. Their systems utilized bone-implanted markers and a mechanical arm for registration as well as navigation. Although they used the systems on hundreds of patients, the system never achieved commercial success. Republished with permission of the American Association of Neurosurgeons, from Schlöndorff G, Mösges R, Klimek L. An articulated localizing arm for otolaryngology. Chapter 13 of Interactive Image-Guided Neurosurgery, edited by Robert J. Maciunas, American Association of Neurological Surgeons, 1993.
nology for tracking (eg, infrared tracking), the use of mechanical arms for tracking was short-lived. However, as presented in Chapter 8, there are some applications for which they offer advantages (ie, robotic interventions). Infrared (IR) camera technology allows tracking without the need for physical linkage. Although commonly used in today’s systems, IR technology was initially too large and expensive for integration into ENT IGS. As a result, efforts turned to EMF tracking, and perhaps the most successful of these first systems was the InstaTrak, which was developed by Visualization Technologies (Boston, Massachusetts).48,59,xvi This device used EMF to track metallic instruments and used as fiducials a proprietary headset that patients wore in the preoperative scanner as well as at the time of surgery (Figure 1–12). This xvi
Sold to General Electric in 2002.
highly utilitarian device dominated the market, capturing 46.6% of the ENT IGS market in 2003.60 Fast forward to the present: current systems are discussed in detail in Chapter 7, and we will see that the majority of the current market consists of devices that use either IR or EMF tracking and is dominated by three big players in the following order of current marker share, beginning with largest: Medtronic, Brainlab, and Stryker. The use of IGS in ENT is still predominantly limited to sinus surgery but is increasingly finding applicability in skull-base surgery. Although many who use IGS systems routinely would consider them ubiquitous, given that they are present in so many community hospitals, they are still not regarded as routine or “standard of care”, as discussed in Chapter 6. The future is never clear, but the path
19
Figure 1–11. ISG Technologies’ “Viewing Wand” was FDA cleared in 1994 and utilized by many groups on the cutting edge of IGS in both Europe and America. The wand precisely related distal tip position to the base unit via motion encoders in each joint. Because the wand’s base had to be rigidly fixed to the patient during active navigation, the head had to be “pinned” via the use of a Mayfield head holder as shown in this photograph. Such rigid linkage systems were replaced by optical and EMF tracking, which did not require rigid linkage and could track both the tool and patient, allowing the head to be mobile during surgical interventions. Republished with permission of Cambridge University Press, from Carney AS, Patel N, Baldwin DL, Coakham HB, Sandeman DR. Intra-operative image guidance in otolaryngology — the use of the ISG viewing wand. J Laryngol Otol. 1996;110(4):322-327. Figure 1.
20
Chapter 1
Brief History of Image-Guided Surgery
n
A
B
Figure 1–12. InstaTrak Headframe from the first widely commercially successful IGS system for ENT. Headsets were made of plastic and were disposable. They reliably and repeatably repositioned on patients via slight compression with earbuds (A, lower arrows) and pads that rested on the nasal prominences. They were worn both in the preoperative CT scan and during the surgical intervention with seven fiducial markers embedded in the rectangular piece (A, upper arrows).
that surgery has taken from the first x-ray in 1895 to the imaging and imageguidance systems available today seems aimed directly toward IGS as the standard of tomorrow. The chapters that follow have been written to aid those who embrace this trend as they attempt to navigate the field of IGS.
References 1. Burrow E. Pioneers and Early Years: A History of British Radiology. Alderney: Colophon Press; 1986. 2. Cox J, Kirkpatrick RC. The new photography with report of a case in which a bullet was photographed in the leg. Montreal Med J. 1896;24:661–665. 3. Pupin M. From Immigrant to Inventor.
New York, NY: Charles Scribner’s Sons; 1925:165. 4. John Macintyre. Wikipedia. http://en .wikipedia.org/John_Macintyre. Accessed November 20, 2015. 5. Thomas AMK, Banerjee AK. Military radiology. The History of Radiology. Oxford, UK: Oxford University Press; 2013:37–58. 6. Filler AG. The history, development and impact of computed imaging in neurological diagnosis and neurosurgery: CT, MRI, and DTI. Nat Proc. 2009;7(1):1–69. 7. Goodman LR. The Beatles, the Nobel Prize, and CT scanning of the chest. Radiol Clin North Am. 2010;48(1):1–7. 8. Webb S. From the Watching of Shadows: The Origins of Radiological Tomography. Bristol, UK: Adam Hilger; 1990. 9. Radon J. On Determination of Functions by Their Integral Values along Certain Multiplicities.Vol. 69. Leipzig, Germany:
21
22
Image-Guided Surgery
Ber der Sachische Akademie der Wissenschaften Leipzig; 1917:262–277. 10. Cierniak R. X-Ray Computed Tomography in Biomedical Engineering. London, UK: Springer-Verlag London; 2011. 11. Barrett HH, Hawkins WG, Joy ML. Historical note on computed tomography. Radiology. 1983;147(1):172. 12. Estermann I, Stern O. Magnetic deviation of hydrogen molecules and the magnetic moment of the proton. II [in German]. Zeitschrift für Physik. 1933; 85(1):17–24. 13. Frisch R, Stern O. Magnetic deviation of hydrogen molecules and the magnetic moment of the proton. I [in German]. Zeitschrift für Physik. 1933;85(1):4–16. 14. Roberts DC, Marcelli V, Gillen JS, Carey JP, Della Santina CC, Zee DS. MRI magnetic field stimulates rotational sensors of the brain. Curr Biol. 2011;21(19):1635–1640. 15. Rabi II, Zacharias JR, Millman S, Kusch P. A new method of measuring nuclear magnetic moment. Phys Rev. 1938;53:318. 16. Mattson J, Simon M. The Pioneers of NMR and Magnetic Resonance in Medicine: The Story of MRI. Jericho, NY: Bar-Ilan University Press and Dean Books Co; 1996. 17. Lauterbur PC. Image formation by induced local interactions: examples employing nuclear magnetic resonance. Nature. 1973;242:190–191. 18. Garroway AN, Grannell PK, Mansfield P. Image formation in NMR by a selective irradiative process. J Phys C Solid State Phys. 1974;7(24):L457–L462. 19. Mansfield P. Sir Peter Mansfield — Biographical. http://www.nobelprize.org/ nobel_prize.org/nobel-prizes/medic ine/laureates/2003/mansfield-bio. html. Accessed January 24, 2015. 20. Gordon R, Bender R, Herman GT. Algebraic Reconstruction Techniques (ART) for three-dimensional electron microscopy and X-ray photography. J Theor Biol. 1970;29(3):471–481. 21. Gordon R, Herman GT. Reconstruction of pictures from their projections. Commun ACM. 1971;14(12):759–768.
22. Hounsfield GN. A Method and Apparatus for Examination of a Body by Radiation such as X or Gamma Radiation. London, UK: British Patent No 1283915; 1972. 23. Kumar A, Welti D, Ernst RR. NMR Fourier zeugmatography. J Magnetic Resonance. 1975;18(1):69–83. 24. Edelstein WA, Hutchinson JMS, Johnson G, Redpath T. Spin warp NMR imaging and applications to human whole-body imaging. Phys Med Biol. 1980;25(4):751–756. 25. Mansfield P. The Long Road to Stockholm: The Story of Magnetic Resonance Imaging. Oxford, UK: Oxford University Press; 2013:120–121. 26. Mansfield P. Multi-planar image formation using NMR spin echoes. J Phys C Solid State Phys. 1977;10(3):L55–L58. 27. Horsley V, Clarke RH. The structure and functions of the cerebellum examined by a new method. Brain. 1908;31:45–124. 28. Galloway R, Peters T. Overview and history of image-guided interventions. In: Peters T, Cleary K, eds. Image-Guided Interventions: Technology and Applications. New York, NY: Springer Science and Business Media, LLC; 2008:1–21. 29. The Long Road to Stockholm: The story of MRI by Peter Mansfield, Oxford University Press, Oxford, UK, 2013. 30. Brown RA. A stereotactic head frame for use with CT body scanners. Invest Radiol. 1979;14(4):300–304. 31. Brown RA. System Using Computed Tomography as for Selective Body Treatment. Alexandria, VA: US patent 4608977; 1986. 32. Kalender WA. Computed Tomography: Fundamentals, System Technology, Image Quality, Applications. 3rd ed. Hoboken, NJ: John Wiley & Sons; 2011. 33. Collignon A, Maes F, Delaere D, Vandermeulen D, Suetens P, Marchal G. Automated multi-modality image registration based on information theory. Inf Process Med Imaging. 1995;3(6):263–274. 34. Viola P, Wells WM III. Alignment by maximization of mutual information. Intl J Comput Vision. 1997;24(2):137–154.
Chapter 1
Brief History of Image-Guided Surgery
n
35. West JB, Fitzpatrick JM, Wang MY, et al. Comparison and evaluation of retrospective intermodality image registration techniques. Proc SPIE 2710, Medical Imaging. 1996;2710:332–347. 36. West JB, Fitzpatrick JM, Wang MY, et al. Comparison and evaluation of retrospective intermodality brain image registration techniques. J Comput Assist Tomogr. 1997;21(4):554–568. 37. Maciunas RJ. Yesterday’s tomorrow: the clinical relevance of interactive imageguided surgery. In: Maciunas RJ, ed. Interactive Image-Guided Neurosurgery. New York, NY: Thieme Publishing Group; 1994:3–8. 38. Galloway RL Jr, Edwards CA II, Thomas JG, Schreiner S, Maciunas RJ. New device for interactive image-guided surgery. Proc SPIE 1444, Medical Imaging V. 1991;1444:9–18. 39. Watanabe E, Watanabe T, Manaka S, Mayanagi Y, Takakura K. Three-dimensional digitizer (neuronavigator): new equipment for computed tomographyguided stereotaxic surgery. Surg Neurol. 1987;27(6):543–547. 40. Roberts DW, Strohbehn JW, Hatch JF, Murray W, Kettenberger H. A frameless stereotaxic integration of computerized tomographic imaging and the operating microscope. J Neurosurg. 1986;65(4):545–549. 41. Friets EM, Strohbehn JW, Hatch JF, Roberts DW. A frameless stereotaxic operating microscope for neurosurgery. IEEE Trans Biomed Eng. 1989;36(6):608–617. 42. Zamorano LJ, Nolte LP, Kadi AM, Jiang Z. Interactive intraoperative localization using an infrared-based system. Stereotact Funct Neurosurg. 1994;63(1–4):84–88. 43. Nolte LP, Zamorano LJ, Visarius H, et al. Clinical evaluation of a system for precision enhancement in spine surgery. Clin Biomech. 1995;10(6):293–303. 44. Rohling R, Munger P, Hollerbach JM, Peters T. Comparison of relative accuracy between a mechanical and an optical position tracker for image-guided
neurosurgery. J Image Guid Surg. 1995; 1(1):30–34. 45. Maurer CR Jr, Fitzpatrick JM, Wang MY, Galloway RL Jr, Maciunas RJ, Allen GS. Registration of head volume images using implantable fiducial markers. IEEE Trans Med Imaging. 1997;16(4): 447–462. 46. Balachandran R, Fitzpatrick JM, Labadie RF. Accuracy of image-guided surgical systems at the lateral skull base as clinically assessed using boneanchored hearing aid posts as surgical targets. Otol Neurotol. 2008;29(8):1050– 1055. 47. Tan KK, Grzeszczuk R, Levin DN, et al. A frameless stereotactic approach to neurosurgical planning based on retrospective patient-image registration: technical note. J Neurosurg. 1993;79(2): 296–303. 48. Fried MP, Kleefield J, Gopal H, Reardon E, Ho BT, Kuhn FA. Image-guided endoscopic surgery: results of accuracy and performance in a multicenter clinical study using an electromagnetic tracking system. Laryngoscope. 1997;107(5): 594–601. 49. Schlöndorff G, Mösges R, MeyerEbrecht D, Krybus W, Adams L. CAS (computer assisted surgery): a new procedure in head and neck surgery [in German]. HNO. 1989;37(5):187–190. 50. Carrau RL, Snyderman CH, Curtin HD, Janecka IP, Stechison M, Weissman JL. Computer-assisted intraoperative navigation during skull base surgery. Am J Otolaryngol. 1996;17(2):95–101. 51. Wang MY, Maurer CR Jr, Fitzpatrick JM, Maciunas RJ. An automatic technique for finding and localizing externally attached markers in CT and MR volume images of the head. IEEE Trans Biomed Eng. 1996;43(6):627–637. 52. Grunert P, Müller-Forell W, Darabi K, et al. Basic principles and clinical applications of neuronavigation and intraoperative computed tomography. Comput Aided Surg. 1998;3(4):166–173.
23
24
Image-Guided Surgery
53. Hill DL, Hawkes DJ, Crossman JE, et al. Registration of MR and CT images for skull base surgery using point-like anatomical features. Br J Radiol. 1991;64 (767):1030–1035. 54. Pelizzari CA, Chen GTY, Spelbring DR, Weichselbaum RR, Chen CT. Accurate three-dimensional registration of CT, PET, and/or MR images of the brain. J Comput Assist Tomogr. 1989;13(1):20–26. 55. Adams L, Krybus W, Meyerebrecht D, et al. Computer-assisted surgery. IEEE Comput Graph Appl. 1990;10(3):43–51. 56. Schlöndorff G, Moesges R, Klimek L. An articulated localizing arm for otolaryngology. In: Maciunas RJ, ed. Interactive Image-Guided Neurosurgery. New York, NY: Thieme Publishing Group: 1994:149–158.
57. Freysinger W, Gunkel AR, Thumfart WF. Image-guided endoscopic ENT surgery. Eur Arch Otorhinolaryngol. 1997; 254(7):343–346. 58. Carney AS, Patel N, Baldwin DL, Coak ham HB, Sandeman DR. Intra-operative image guidance in otolaryngology — the use of the ISG viewing wand. J Laryngol Otol. 1996;110(4):322–327. 59. Fried MP, Kleefield J, Taylor R. New armless image-guidance system for endoscopic sinus surgery. Otolaryngol Head Neck Surg. 1998;119(5):528–532. 60. Group MR. US Markets for Image Guided Surgical Systems 2004. Technical Report, Millennium Research Group Inc., Toronto, Canada, 2004.
2 CT AND MRI and scatter reveal the presence of tissue by contributing to the attenuation of the beam as it travels through the body, and it is that attenuation that is measured by the scanner by comparing the x-ray intensity entering the body with that transmitted through the body to the detector. When an x-ray photon is absorbed, it ceases to exist, and its energy is transferred to an electron causing ionization. This absorption of the x-ray attenuates the intensity of the detected beam, thereby indicating the presence of tissue somewhere along the x-ray path. When an x-ray photon is scattered, only part of its energy is transferred to an electron, with the rest being scattered as a photon deflected into a different direction. Scatter also contributes to the attenuation of the intensity of the detected beam and indicates the presence of tissue. However, it has the potential to cause errors in the calculation of attenuation because the scattered beam can still strike the detector. Fortunately, the angle of deflection of the scattered photon will typically be large, and this large angle makes it possible to eliminate almost completely the resulting error by means of directional
Despite its complex history, CT is a simpler technology than MRI, and yet its discovery depended on a perfect storm of events — extensive experimentation with tomography and development of modern computers — and a brilliant and ambitious engineer who capitalized upon that perfect storm — Godfrey Hounsfield. Although the description below is intended to simplify CT scanning such that it can be understood by practitioners, we also hope that it allows the reader to appreciate the incredible contributions that Hounsfield and others made to provide us with modern CT scanners, which are used daily and are vital diagnostic tools in every subspecialty of otolaryngology.
How CT Works CT scanning uses a paired x-ray source and detector that rotate around the patient (Figure 2–1). The matter between the x-ray source and the detector disrupts the travel of the x-rays on their way to the detector by either absorption or scatter (Figure 2–2). Both absorption 25
Figure 2–1. CT scan with cover removed showing the x-ray source (T) as well the detectors (D). The fan beam travels along path X, and rotation of the emitter and detectors is in the direction of R. Republished with a Creative Commons license. https://upload.wikimedia.org/wikipedia/commons/7/76/ Ct-internals.jpg
Figure 2–2. As x-rays pass through tissue, they may (a) not interact with the tissue, (b) hit an electron and scatter, or (c) be absorbed by the tissue, which results in the emission of an electron (ionization).
26
Chapter 2
filtering. A directional filter comprising a set of thin fins of x-ray absorbent material, such as lead, each of which is aligned with the direction from x-ray source directly to fin, is placed in front of the detector such that those x-rays that arrive directly from the source can pass between the fins, but those that do not will be absorbed by them and will not reach the detector. In Figure 2–3, a simple CT scanner is depicted with the matter being subdivided into three rows and three columns of squares, each of which is labeled with its x-ray attenuation coefficient, µij. The attenuation coefficient is the fraction of x-ray radiation absorbed or scattered per centimeter of travel. One square represents the smallest resolvable two-dimensional (2D) unit, also known as a pixel, which is a con-
ct and mri
n
traction of picture element, or, when the third dimension is included, known as a voxel, which is a contraction of volume element. The goal of CT is to determine the x-ray attenuation coefficient for each voxel, and it is done indirectly via observation of the x-rays that pass through these voxels in multiple directions (ie, those that are neither absorbed nor scattered). The intensity of each beam emitted from the x-ray source is known, and, as it passes through the matter, the x-rays are absorbed or scattered as they pass through regions with different attenuation coefficients. What is measured at the detector is the resulting residual intensity. The energy decreases in an exponential fashion as it is absorbed and scattered, meaning that, if an x-ray of intensity I0 passes through a voxel of width x with an
Figure 2–3. CT scanning consists of an x-ray emitter and detector pair rotating around the subject being scanned. At discrete intervals during the rotation, information about attenuation of the beam (arrows) is determined. These series of equations are then solved to determine the attenuation coefficients that create the image.
27
28
Image-Guided Surgery
attenuation coefficient of μ, then the residual intensity that emerges from that subdivision is I = I0 e−µx. In Figure 2–3, each arrow represents an x-ray beam that passes through the patient as the gantry rotates. For the first horizontal beam, the intensity starts out as I0 , then passes through first voxel,i and, assuming that the width of each voxel is x, is diminished to I0 e−µ1,1x. As the beam goes through the second voxel, the intensity starts out as I0e−µ1,1x and is then diminished to I0e−µ1,1xe−µ1,2x. And, going through the third voxel, it starts as I0e−µ1,1xe−µ1,2x and is then diminished to I = I0e−µ1,1xe−µ1,2xe−µ1,3 x, which can be written using the rules of exponentiation, I = I0e−(µ1,1+µ1,2+µ1,3)x. Taking the natural log (ln) and rearranging gives this simple result: µ1,1 + µ1,2 + µ1,3 = (lnI0 – lnI)/x. The right side of the equation can be calculated from the observed beam intensity I, the known initial intensity I0, and the known voxel width x. The attenuation coefficients on the left are the unknowns, whose values are desired. A similar equation can be formed for each of the 12 arrows in Figure 2–3, representing all the beams of radiation during the rotation. A fundamental principal of mathematics is that in order to find the values of N unknowns, it is necessary to have at least N equations. For the three-bythree grid in Figure 2–3, we have nine variables, so we need at least nine equations. In fact, because of the particular form of equations that we get from the three-by-three arrangement, we need i
at least 10 equations, so beams from all four orientations are required. We can then solve any subset set of 10 or 11 or even all 12 equations using methods developed for linear algebra in order to determine the nine attenuation coefficients. These coefficients can be scaled to indicate those types of tissue that absorb much x-ray energy — large positive values, which are depicted as white in typical CTs, those that absorb less x-ray energy — smaller positive values, which are depicted as gray in typical CTs, and those that absorb no x-ray energy — the smallest value on the chosen scale, which are depicted as black. Now imagine that we have a grid, more appropriately called a matrix, that isn’t 3 × 3 as shown in Figure 2–3 but 512 × 512, which is typical for a modern CT scanner. And, now we take, say, 750 projections at each of 1000 angles of rotation. We would have 750 000 equations to solve for 512 × 512 = 262 144 attenuation coefficients. The number of computations for this problem is enormous, and it explains why the evolution of computers had to catch up with tomography before Hounsfield, or anyone else, could invent computerized tomography. And for this feat he received not only fame, wealth, and a mantel groaning with awards but also the lasting designation of naming of the scale used to indicate the relative absorption of x-ray energy, the Houns field unit (HU),ii which is used to specify each separate μ in both the simple 3 × 3 CT scanner of Figure 2–3 and today’s modern 512 × 512 CT scanners.
Although Figure 2–3 is 2D and the individual squares are pixels, actual CT scanners have 3D voxels and thus we will use the term “voxel” throughout this discussion. ii Technically, the HU is a measure of radiodensity, which is a linear transformation of the attenuation coefficient to a measure in which the radiodensity of distilled water at standard temperature and pressure (STP) is set to 0 HU and that of air at STP is set to −1000 HU. Bone, depending on density, has an HU of multiple hundreds to multiple thousands.
Chapter 2
Although the above is how the first CT scanners worked, there soon followed major advances in both hardware (eg, the use of fan-beam x-ray sources which cover the specimen more efficiently) and software (eg, computational advances including more efficient solution methods, such as filtered back-projection). So how do we get more than one slice? We can move the x-ray emitter and detector a small amount, say 0.5 mm in the axial or z direction, rescan, and then lay that information on top of the previous scan like one sheet of paper laid on another, continuing along z until, like a stack of such paper sheets, we have a set of individual two-dimensional images that together form a threedimensional whole. This approach of course depends on the patient not moving between scans or else the sheets will become askew much like what happens when you open a ream of paper and push its edge: the individual sheets no longer align correctly. In CT-speak, this skewing is known as a motion artifact. Alternatively, to minimize this motion artifact, we can have multiple pairs of x-ray emitters and detectors that rotate around the patient simultaneously. Such multislice CT (MSCT) scanners cover a larger volume, also known as a field of view (FOV), in each rotation around the patient.iii For example, a 64-slice CT scanner with a 0.5-mm slice thickness can scan a FOV with a 32-mm axial z range in one rotation of the scanner. The patient is then moved 32 mm in the z direction, and the scan is repeated, which is where motion artifacts for iii
ct and mri
n
MSCT scanners can come into play. Or, better yet, we can slide the patient very precisely at a known speed and simultaneously rotate the scanner so that, as pointed out in Chapter 1, each ray follows a helical path through the patient, like the cutting blade through a spiralcut ham. The relationship between the voxels and the detected beams in spiral CT (also known as “helical-scan” CT) is clearly more complicated than that of the rotate-then-advance approach, but the basic mathematical problem remains the same. A set of equations is solved for a set of attenuation coefficients. The size of a CT voxel depends not only on the resolution of the individual slice matrix (3 × 3 in Figure 2–3 and 512 × 512 in most modern CT scanners by which the FOV is divided, to determine the spatial resolution) but also on the slice thickness. Regarding individual slice resolution, the dimension of one side of the FOV within a slice must be divided by the number of detection elements to determine the in-slice resolution. For example, for a 30-cm × 30-cm FOV, for the 3 × 3 matrix of Figure 2–3, the resolution would be 30 cm/3 = 10 cm, while for a 512 × 512 CT scanner, the resolution would be 30 cm/512 = 0.6 mm. A typical voxel for a modern CT scanner could be 0.6 × 0.6 × 0.8 mm (assuming slice thickness of 0.8 mm). And, perhaps most important, the HU calculated for each voxel has only one value, instead of varying with the true change in attenuation across the 0.6 × 0.6 × 0.8-mm chunk of tissue that the voxel represents. This forced constant intensity across each voxel means that even high-resolution CT scanners are “pixelated” (more appro-
Here, we are referring to a three-dimensional FOV, but FOV can also refer to a two-dimensional region. The meaning will be clear from the context.
29
30
Image-Guided Surgery
priately “voxelated”). To make prettier pictures and potentially aid clinicians in diagnosis, these pixelated images are often replaced by higher resolution images with smoother changes in brightness from one point to the next. For example, a 512 × 512 array of voxel intensities may be replaced by a 1024 × 1024 array of subvoxels in which a 2 × 2 set of “subvoxels” represents one original voxel. The intensity of these subvoxels will be varied within the set to “ramp up” the intensity from one edge of the set to the next to reduce the jump in intensity between any neighboring pairs and thereby reduce the effect of artificial edges that can be perceived in the unsmoothed version. This smoothing can result in dramatic differences between actual data and processed data as shown in Figure 2–4, where the lower right panel is a raw CT image — also known as a Digital Imaging and Com-
munications in Medicine (DICOM) image — and the lower left panel of Figure 2–4 shows the same image after processing, which is typically done using proprietary software that varies according to the manufacturer.
Intraoperative CT Scanners Perhaps most pertinent for IGS in this discussion of imaging are CT scanners that can be used in the operating room (OR). These scanners fall broadly into two categories — stationary scanners that reside in the OR and portable CT scanners.
Stationary CT Scanners The availability of high-quality, true MSCT imaging for surgical guidance is a clear advantage of having a scanner that resides in the OR. The disadvan-
Figure 2–4. The top panel shows a coronal, temporal bone CT scan with screw-in fiducial marker. This image is enlarged in the lower panels, which show on the right the raw DICOM data and, on the left, the processed image.
Chapter 2
tage of having a scanner that resides in the OR is that it resides in the OR! Because it is in the OR, the scanner cannot be used for other purposes while the OR is occupied. Although not typically thought of as CT, fluoroscopy units, which are used in interventional cardiology, vascular, and neurovascular suites, are typically capable of functioning as CT scanners. Such units are usually placed in highly specialized OR suites with unique OR tables and surrounding equipment. This arrangement makes use of such ORs and intraoperative CT scanners suboptimal for IGS in otolaryngology because the specialized OR suite is typically reserved for emergent interventions, and the equipment is designed for other procedures. Nonetheless, there are reports of such suites having been used early on during the incorporation of IGS in otolaryngology when intraoperative imaging was required.1 Another commercially available option is the Siemens Artis Zee line of C-arms (Erlanger, Germany), including the Zeego, which is a twoaxis C-arm mounted on a robotic drive system that allows high flexibility in positioning. At least one research group is working on the refinement of such systems for ENT IGS applications, specifically those involving the skull base.2 Perhaps a more efficient setup for an intraoperative CT scanner is that offered by Innovative Magnetic Resonance Imaging Systems (IMRIS, Inc, Minnetonka, Minnesota) in which a MSCT scanner is mounted on ceiling rails that extend between two rooms (Figure 2–5). Termed the VISIUS iCT, it also offers a similar design for its intraoperative MRI (VISIUS iMRI). Although both rooms could be ORs, it is likely more efficient to have one of
ct and mri
n
these rooms be an imaging suite such that the CT scanner can be used for routine scanning while the OR is being used for components of the case that do not require CT scanning (the vast majority of the time!). This setup does require colocalization of the OR and the radiology suites and, in addition, an understanding of priority of the CT scanner for intraoperative cases as well as agreements for the sharing of operating costs and/or revenue between surgical and radiologic departments. At present, IMRIS offers both 64-slice and 128-slice VISIUS iCT units with pricing comparable to stationary units in radiographic suites, and, at the time of this writing, the cost of a VISIUS iCT with installation exceeds $1 million.
Portable CT Scanners Portable CT scanners are likely to be the approach that is most amenable to IGS in the future because the technology is back-compatible with current ORs and can be used in multiple locations, including the OR, intensive care units (ICUs), and radiology suites. At the time of this writing, there are three portable CT scanners commercially available with a fourth previously available and poised to reenter the market soon. These devices are the O-arm (Medtronic, Inc, Minneapolis, Minnesota), the BodyTom and CereTom (Neurologica Corp, Danvers, Massachusetts), which are two sizes of the same technology, and the Airo Mobile (Brainlab Corp, Munich, Germany), with the fourth being the xCAT (Xoran Technologies, Ann Arbor, Michigan). Of these, the BodyTom/CereTom and Airo are MSCT, while the O-arm and xCAT
31
A
B Figure 2–5. The IMRIS intraoperative CT scanner (A) is mounted on ceiling rails, which traverse between two rooms (B). The scanner resides in a mid-position, allowing the rooms to be used primarily for other clinical activities (eg, operating). The scanner can travel at upward of 1 foot/ second. Images provided courtesy of IMRIS by Deerfield Imaging © 2015.
32
Chapter 2
are flat-panel volumetric computerized tomography (fpVCT) machines. Portable MSCT Scanners Neurologica, which is now a subsidiary of Samsung Electronics Co, produced the first truly portable MSCT scanner, obtaining FDA approval in 2005. The design was motivated by the desire to make CT scanners available to anyone in any location after a family member of the founder died in a rural hospital secondary to head trauma for which CT scanning may have guided life-saving intervention (https://www.youtube .com/watch?v=hsAe72kbKro. Accessed January 4, 2016). The original device, still available commercially as of this writing, consisted of an eight-slice MSCT scanner that had two drive systems, the first system being a set of large caster wheels for transport from location to location and the second being a tank-tread, continuous-track drive system that allowed the scanner to precisely navigate from one FOV to the next during scanning (ie, the eight-slice, 1.25-mm thickness scanner would take an image over a 10-mm axial FOV and then move 10 mm on the tank-tread drive to take the next FOV). Unfortunately, the tank-tread drive system can create problems with geometric fidelity, which is a requirement for IGS. The problem arises in most ORs, where the floor on which the scanner traverses is not perfectly planar. As the scanner traverses between each set of slice acquisitions, the scanner experiences pitching and yawing over small bumps and depressions in the floor, and thus one set of slices is
ct and mri
n
not aligned with the next, creating geometric discrepancies between them.3 (Referring back to our analogy of a CT scan matrix being like a ream of paper, when groups of sheets are misaligned, the volume is no longer geometrically true; Figure 2–6.) Neurologica offers several “fixes” for this issue, including laser scanning of the floor where the device will be operated to allow back correction and/or motion detectors on the tank-tread drive system to correct for the amount of pitching and yawing continuously during imaging. Until these “fixes” have been shown to repeatedly overcome this issue in multiple settings, the CereTom and BodyTom would seem inappropriate for IGS applications that require a high degree of geometric accuracy. Nonetheless, for non-IGS applications such as exploratory scanning in the emergency room and ICU, the device offers high utility. Having received FDA clearance in 2013 (citing the BodyTom as its predicate device!), the Airo portable CT scanner is manufactured by Mobius Imaging LLC (Ayer, Massachusetts) for Brainlab (Figure 2–7). It is a 32-slice MSCT with 1-mm slice thickness, allowing a 32-mm axial FOV. Coupled with a linearly translatable patient gurney/ bed, which is tracked in reference to the scanner, it accurately scans over relatively large axial volumes. Because Brainlab also manufactures the most widely used IGS system in otolaryngology (see Chapter 6), their intent appears to be able to offer seamless imaging and navigation, but their system may be limited in applications in the lateral skull base because of the relatively large slice thickness of 1 mm. Nonetheless, as of the date of this writing, over a dozen are in use in the United States.
33
Figure 2–6. Portable CT scanners, such as Neurologica’s BodyTom and CereTom, which move along the floor while scanning, are subject to geometric distortions of the floor, including pitching and yawing, which creates misaligned stacks of CT slices.
Figure 2–7. The BrainLab Airo portable CT scanner consists of a true MSCT, which docks to a radiolucent patient positioning table or can be used with a separate gurney/bed. In transport mode, the gantry is rotated around a vertical axis 90 degrees from the scanning position shown here, allowing it to traverse hallways and doors. Image provided courtesy of BrainLab, Inc.
34
Chapter 2
Portable fpVCT Scanners fpVCT (aka cone-beam) scanners are different from MSCT in that the x-ray source is akin to a flashlight from which a cone of radiation illuminates all the tissue in the FOV simultaneously (Figure 2–8), leading to more complex mathematical calculations in determining the distribution of x-ray attenuation than that of conventional CT, which uses a planar beam in the shape of a fan. In addition to more complex mathematics, fpVCT is more sensitive to beam hardening, which occurs when x-rays pass through large volumes of similar tissue and their beam becomes “harder” as it reaches the center of the
ct and mri
n
tissue. Beam hardening means that as the beam progresses farther through the tissue, it is composed of more highenergy protons because lower energy protons are absorbed or scattered by tissue at the periphery. The result is that the attenuation coefficients in the center of the tissue volume are slightly miscalculated as artificially lower than their true values.iv This creates both an altered visual projection (ie, the center of the object appears darker than the periphery) because the HUs tend to be lower in that region despite that fact that the tissue is the same. Another disadvantage is that the soft-tissue contrast tends to be lower and noise tends to be higher for fpVCT than for
Figure 2–8. Flat-panel-volume CT (fpVCT) (a) versus conventional CT (b). Most x-rays originate at the source, but scattered rays (red arrows) originate within the patient. This schematic shows that scatter is more likely to strike the detector (red splashes) in (a) than in (b) because most scatter travels outside the fan-shaped beam of (b). Scatter that strikes the detector reduces contrast and increases noise in fpVCT relative to conventional CT. iv
Beam hardening is sometimes referred as “cupping artifact” because of the dip in the attenuation coefficients at the center of the tissue giving a plot relating Hounsfield units to depth of tissue penetration a central dip which makes the curve appear somewhat like a cup.
35
36
Image-Guided Surgery
conventional CT. This problem is present because the cone shape of the radiation makes it more likely that scattered x-rays will strike the detector (see Figure 2–8), and those scattered rays tend to slightly “fog” the image. An advantage of fpVCT over MSCT is that the concept of slice is absent, being replaced by a volume. This allows “slices” and their composite voxels to be defined after the acquisition in ways beneficial for image processing. The benefits are that the voxels tend to be smaller than MSCT voxels, and are usually isotropic, which means that the height, width, and depth of the voxel are the same. The isotropy of fpVCT voxels is advantageous because slices can be displayed at any angle without the low-resolution “blocky” artifact in the slice direction seen with conventional CT in any view other than axial. However, perhaps the greatest long-term benefit of fpVCT is that in practice, it achieves a lower radiation dose than MSCT, although in theory, MSCT could achieve similar low radiation doses if soft-tissue delineation is not required. Technically, virtually any rotation fluoroscopy machine could be converted to an fpVCT unit using appropriate reconstruction software, but such units are typically used for subjective assessment of placement of catheters or stents in vascular procedures. Perhaps the first group to take a rotation fluoroscopy machine and repackage it into a machine for which IGS would be feasible is Breakaway Imaging, LLC (Littleton, Massachusetts), which received FDA clearance for its O-arm
v
in 2005. Licensed and available exclusively from Medtronic, it has a unique “breakaway” design wherein the track for the rotating x-ray emitter and flatpanel detector retracts via a telescoping sleeve (Figure 2–9). This design was created to facilitate its use during spine surgery, allowing the track to extend itself around the OR table without violating sterility. As with the BrainLab-Mobius partnership, their intent appears to be to offer seamless imaging and navigation. Hundreds of O-arms have been sold and are operational in the United States with a new model predicted for 2016. Designed specifically for otolaryngology, the Xoran xCAT (Figure 2–10) is a portable fpVCT that received FDA clearance in 2006. Clearance occurred at a time when upright, in-office fpVCT scanners for sinus applications were in vogue with Xoran’s offering in that realm being the miniCAT. Xoran Technologies created a highly utilitarian horizontal fpVCT in the xCAT, but their timing could not have been much worse because it entered the market just prior to the global economic recession of 2008. After a failed business distribution arrangement with Medtronic and a voluntary FDA halt in 2011v when the insulation in the wiring harness of the rotating gantry melted, Xoran suspended production of the xCAT but continues to service existing units. A limited number of these units continue to be used in the United States, including one at the home institution of the authors, and Xoran intends to reenter the market in 2016 with an updated unit.
http://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfMAUDE/detail.cfm?mdrfoi__id= 2062570 (Accessed January 4, 2016).
A
Figure 2–9. Medtronic’s O-arm has a unique telescoping gantry that “builds” itself around the patient as shown in the open position (A) and closed position (B). C. Shows it as used in the operating room, where it is predominantly used in spine surgery for pedicle screw placement.
B
C 37
A
B Figure 2–10. The Xoran xCAT is a portable, flat-panel volumetric computerized tomography machine used to image the head and neck region as seen at a trade show (A) and in use in the operating room with a sterile drape covering the patient (B). Images provided courtesy of Xoran Technologies.
38
Chapter 2
Accuracy of Portable Scanners To test the accuracy of such scanners, Dillon et al4 have reported on scanning phantoms with the xCAT, BodyTom, Airo, and O-Arm and compared them to a clinically used, conventional, fixedbase, multislice CT scanner (Siemens Corp, Munich, Germany). Each phantom consisted of 64 spherical titanium fiducials, 4 mm in diameter, arranged in a three-dimensional 4 × 4 × 4 rectangular grid embedded in Plexiglas. The fiducials were placed by means of a computer numeric controlled (CNC) milling machine whose accuracy was 0.0002 inch absolute with 0.0001-inch repeatability. The phantom was then imaged on each scanner in multiple orientations, and a three-dimensional (3D) fiducial-localization algorithm was employed to determine the locations of the centers of the spheres in each image volume. Then, the set of localized centers was compared to the set of positions at which the fiducials’ centers were placed by the machine. Distortion was then reported as the maximum percentage change in the distance between any two points in the image
caused by expansion, compression, or skewing. Their data are replicated in Table 2–1. As can be seen from these results, each of the scanners was remarkably accurate. It should be emphasized that these errors are percentages, so, for example, 0.234% denotes an error of only 0.00234! The corresponding scale factors, which equal 1 for a perfect scanner, are remarkably close to perfection, ranging from 0.99766(= 1 − 0.1249/100) to 1.00234 (= 1 + 0.234/100). These accuracies are far better than any level needed for visual inspection of CT images for diagnosis or interventional planning. However, the question for us is whether they are good enough for applications in which IGS is intended for submillimetric targeting. Taking a FOV approximately the size of a human head (sphere with radius 100 mm), then the error between a location in the middle of the head and on the surface of the sphere would have a 1-mm error for each 1% of error. Using the values in Table 2–1 shows that, in the worst case, each of these scanners would produce a geometrical targeting error of less than 0.249 mm, which — in our opinion — is
Table 2–1. Accuracy of Portable CT Scanners Scanner
Voxel Size (x,y,z in mm)
% Error
Siemens
0.74 × 0.74 × 0.75
0.182
0.4 × 0.4 × 0.4
0.074
1.16 × 1.16 × 1.25
0.129
0.59 × 0.59 × 1
0.186
0.42 × 0.42 × 0.83
0.249
xCAT BodyTom Airo O-Arm
ct and mri
n
Note. The Siemens scanner, which is not portable, is included for comparison.
39
40
Image-Guided Surgery
more than adequate for most current IGS applications. Note: The BodyTom, despite its excellent performance in laboratory conditions, as indicated by the table, with its propensity to pitch and yaw during translation between FOVs (see Figure 2–6), is not recommended for IGS applications until a “fix” has been undertaken such as those mentioned above.
MRI As we mentioned in our history of MRI, it is far more difficult to produce an image via nuclear magnetic resonance than to produce one via x-ray interactions. It should not be surprising therefore to find that understanding the problem and its solution is also more challenging. Consequentially, our explanation of MRI is considerably more involved than that for CT. So bear with us; the understanding is worth the journey!
How MRI Works MRI comprises the use of nuclear magnetic resonance (NMR) to produce electromagnetic signals from the nuclei of hydrogen atoms in the body, the recording of those signals, and the construction of an image from those recordings. This process is just like CT, except that CT uses x-rays to produce electromagnetic signals from all atoms in the body. The difficulty presented by MRI is not its limitation to hydrogen atoms, however. Most of the atoms in the body are in fact hydrogen atoms. The difficulty is
caused by the wavelength of the electromagnetic radiation produced by NMR. Whereas CT wavelengths are much smaller than even the tiniest anatomic feature, MRI wavelengths are roughly the length of the entire body. It is impossible to use such long wavelengths to produce an image directly, so for MRI, indirect methods are required. There are three parts to the method employed by all MRI scanners: A. Apply magnetic fields that vary spatially. B. Excite hydrogen nuclei via NMR in a manner determined by the spatial variation. C. Analyze the electromagnetic signal from the nuclei to infer their spatial distribution. This ABC of MRI is a good starting point, but it hides a great deal of complexity. There is an enormous variety in the combinations of both the applied magnetic fields (A) and the excitations (B). Even the overall strength of the magnetic field, which is measured in tesla (T; mT = millitesla = 0.001 tesla), varies from 0.2 to 1 T for some so-called open-bore units (Figure 2–11), varies from 1.5 to 3 T for the more common closed-bore units (Figure 2–12), and ranges even to 7 T for the relatively rare “high-field” closed-bore units. Most scanners are closed bore, with a field strength of 1.5 T being the most common, and in our description of MRI, we will use that model. Compared to the 0.00005 T strength of the Earth’s magnetic field, these field strengths are extremely high, and to achieve them, an MR scanner uses superconducting magnets bathed in liquid helium. Note, however, that one
Figure 2–11. Example of an open-bore MRI scanner (Philips Panorama HFO, field strength = 1.0 T). “Open bore” means that the patient is afforded a wide view of the room. Two sets of magnetic coils with vertical axes are placed one above the other with an opening between them that is wider than the typical human body. Image provided courtesy of Philips Healthcare, Inc.
Figure 2–12. Example of a closed-bore MRI scanner (GE Healthcare Discovery MR750, field strength = 3.0 T). “Closed bore” means that the patient is afforded only a very limited view of the room. Magnetic coils with a horizontal axis surround a hollow horizontal cylindrical cavity whose width is greater than the typical human body. Image provided courtesy of GE Healthcare, Inc.
41
42
Image-Guided Surgery
can easily purchase a small (1 cubic inch) neodymium permanent magnet with a 1.5 T field. The important difference between this magnet and an MRI magnet is the spatial uniformity of their fields. The field strength within an MRI scanner is exquisitely uniform over a large volume, varying by no more than 0.5 parts per million (ppm), which equals 50 millionths of 1%, over a radius of 175 mm from the center of the bore!5,vi By contrast, the neodymium magnet varies by approximately 10% over 1 mm. The engineering required to produce 0.5 ppm uniformity is one major reason for the high cost of an MR scanner, but there is a lot of other very expensive engineering required for MRI as well — almost all of it involving the generation of magnetic fields by passing current through electric coils — including so-called shim coils to achieve the 0.5 ppm static-field uniformity, transmission coils to transmit electromagnetic radiation into the nuclei, receiver coils to detect radiation,vii and gradient coils to vary the static field both spatially and temporally. Localization by Means of Gradients Once a static field is established and shimmed to make it spatially uniform, it is deliberately and precisely made spatially nonuniform. A controlled spatial nonuniformity is imposed on the static field by passing current through sets of coils arranged around the periphery of the bore to produce a uniform gradient vi
in the static field strength across the imaging field. There are three sets of the “gradient coils”, one set for each of the three components of the gradient — in the x, y, and z directions. When these gradient coils are switched on, the field at any point (x,y,z) in the scanner can be calculated as follows (Figure 2–13): M(x,y,z) = M0 + Gxx + Gyy + Gzz, where M0 is the magnetic field when there are no gradients and Gx, Gy, and Gz are the gradients in the x, y, and z directions with x = y = z = 0 at the center of the scanner’s bore and with the z axis pointing in the same direction as the static magnetic field, which is directed along the axis of the horizontal cylinder of the bore (for an open-bore scanner, the field direction is vertical), and the x axis pointing horizontally, as shown in Figure 2–13. As mentioned above, in our examples, we will use M0 = 1.5 T. The gradients are small, typically only a few hundredths of a tesla per meter, but without them, it would be impossible to determine the location of any of the nuclei because the meterslong radiation that they emit bathes the receiver coils uniformly. The determination of the position of a hydrogen nucleus in the scanner is based on the fact that its precession frequency is proportional to the strength of the field to which it is exposed. In mathematical terms, this proportionality can be stated as f = γ M, where f is the frequency in oscillations per second, γ equals 42,576,000 Hz/T (Hz = 1/s)viii = 42.576 MHz/T (MHz = million Hz),
As a measure of the improvement in MR scanner technology in two decades, the homogeneity level for MRI scanners suggested by the American Association of Physicists in Medicine in 1992 was 10 ppm — 50 times less stringent than the criterion of 0.5 ppm in 2010. vii In some cases combined transmitter-receiver coils are used. viii γ is equal to an empirically determined property of the proton — its gyromagnetic ratio — divided by 2π
Chapter 2
ct and mri
n
Figure 2–13. Schematic showing the position and orientation of a coordinate system within the bore of an MRI scanner. The three axes, x (red ), y (green), and z (blue) intersect at the origin of the coordinate system in the center of the bore. Each point, x, y, z, within the scanner is calculated relative to these three axes.
and M is the magnetic field strength. When a gradient is turned on, nuclei located at points in the scanner where the field strength is higher will precess at higher frequencies, whereas those at points where the field strength is lower will precess at lower frequencies. At the center of the scanner, where M = M0 = 1.5 T, the nuclei precess at f0 = 63,864,000 revolutions per second = 63.864 MHz, but at other points, the precession frequency will depend on the gradient. For example, if Gx = 6.00 mT/m and Gy = Gz = 0, then all the nuclei on the plane perpendicular to the x axis and located at x = 120.0 mm (Figure 2–14) will experience a magnetic field M equal to 1.5 T + ∆M, where ∆M = xGx = 0.1200 m × 6.00 mT/m = 0.720 mT, and they will precess at f = f0 + ∆f, where ∆f = γ × 0.720 mT = 30,650 Hz = 30.650 kHz (kHz = 1000 Hz). So at x = 120.0 mm, they will precess at = 63.864 MHz + 30.650 kHz =
63.89465 MHz. Note that precession frequency effectively “encodes” position only in the direction of the magnetic gradient, which is this example is along the x axis. It tells us nothing about position along the other axes, which are y and z in this example. Slice Selection, Frequency Encoding, and Analysis. Before scanning, a deci-
sion is made regarding the orientation of the slices. In practical terms, a slice is typically the plane in which a clinician would like to have the highest in-plane resolution at the possible sacrifice of through-plane resolution because slice thickness is the largest dimension of an MRI voxel. For illustrative purposes, we will choose z as the slice axis (ie, the one perpendicular to the plane of the slice) generating xy image planes, but any choice is possible including directions between the x, y, and z axes.
43
44
Image-Guided Surgery
A
B
Figure 2–14. A. Schematic showing the plane (pink with dashed outline) that is perpendicular to the x axis and located at x = 120.0 mm. This plane intersects the x axis at the point x,y,z = 120,0,0 (mm), which is pointed at by the arrow. B. View from the feet of the patient along the z axis.
The most common direction for z relative to the scanner is in the direction of the magnetic field. If this direction is inferior-to-superior, as shown in Figure 2–14, then axial slices are produced. The most common direction for x is right to left, and y is usually anterior to posterior, both of which are shown in Figure 2–14. However, the only requirement of the x, y, and z axes is that they are mutually perpendicular. With the z axis chosen as the desired slice orientation, the scanner performs “slice selection”, in which the proton nuclei within a slice are excited via NMR. In this process, the z gradient coil and the transmission coils are turned on simultaneously. The transmission coils produce a pulse of electromagnetic radiation that contains only a very narrow range, ∆f, of frequencies. Thanks to NMR, only those nuclei precessing at frequencies within the range ∆f will absorb energy, so, with the z gradient causing the field, and hence the
precession frequency, to vary along z, the positions of the set of nuclei that are precessing within ∆f will be confined to a narrow range of z values. In other words, the responsive nuclei will all lie within a thin slice perpendicular to the z axis, parallel to the xy plane. A simple relationship holds for slice width, which is also known as the field of view in the z direction (FOVz), the frequency range, and the gradient strength: ∆f = γ Gz × ∆z. For example, if Gz = 7.00 mT/m and the desired slice width is ∆z = 2.00 mm, then the width of the frequency range of the slice-selection pulse must be ∆f = 42.576 MHz/T × 7.00 mT/m × 2.00 mm = 596 Hz. As they absorb energy, the nuclei in this z slice will begin immediately to emit energy, each at its own precession frequency. After the excitation coil is switched off, the intensity of the signal received by the receiver coils will be proportional to the number of nuclei within the z slice.
Chapter 2
Next, with the desired slice activated and the slice-selection gradient switched off, a second gradient, perpendicular (aka orthogonal) to the first, is turned on. In our illustration, the second gradient is along the x axis. While it is on, the signal arising from the selected slice is collected by the scanner’s receiver coil. As the signal is collected, it is sampled at evenly spaced time intervals — often 256 of themix — and analyzed into 256 frequency components by an on-board digital computer to ascertain the relative intensity of the signal received at each frequency. Each such intensity is proportional to the number of nuclei precessing at the corresponding frequency.x Given an x gradient (Gx) of 6.00 mT/m, the intensity received at a frequency of f0 + 30.650 kHz (calculated above) reveals the proportion of nuclei located at x = 120 mm, and the proportion at x = 120 mm + 1 mm is given by the intensity at f0 + 30.650 kHz + 255 Hz. Thus, each 255-Hz change in frequency will be interpreted by computer analysis to represent a displacement of 1 mm in x, and all 256 of intensities for the x values separated by 1 mm within a 256-mm span can be determined during the analysis by calculating the intensity for each of their frequencies over a range ∆f = γ Gx × 256 mm = 65.397 kHz. This range of frequencies is known as the “bandwidth” of the measurement, and the span of x values is known as the “field of view” in the x direction (FOVx) with ∆f = γ Gx × FOVx, which, it will be noted, has the same form as the equation relating the slice-selection frequency range to the slice thickness. When frequency analysis is complete, ix
ct and mri
n
the number of hydrogen nuclei in each of 256 “giant voxels” will be known. Figure 2–15 shows one of these giant voxels after slice selection in the z direction and frequency encoding and analysis in the x direction. The dimensions of each voxel in the x and z directions are 1 mm and 2 mm. In the y direction, the voxel extends over the entire region of detection by the receiver coils, but the only excited nuclei lie within the patient’s body. One application of frequency analysis to the signal produced by one selection and one x gradient yields the number of nuclei in each of the 256 such giant voxels as shown in Figure 2–16A. One method for MRI involves a repetition of the process of selection of the (same) slice followed by frequency analysis 255 times, but each time, instead of an x gradient, a gradient in the xy plane at a varying angle to the x axis is used such that the angles cover the range from 0 to 180 degrees uniformly (analogous to CT scanning). Each single repetition will yield a set of 256 giant voxels all oriented at the same angle. Figure 2–16A shows the voxels acquired for an angle of zero, and Figure 2–16B shows the voxels acquired for an angle of 25 degrees. When the process is carried out for all 256 angles, a total of 256 × 256 = 65 536 voxels are produced. The set of voxels acquired for each angle is equivalent to one CT projection, and these 256 projections, each comprising the signals from 256 giant voxels, can be input to the same reconstruction algorithm used in CT to form a two-dimensional image of the selected slice. By carrying out this process
256 = 28 is chosen, instead of, say, 250, because frequency analysis is faster for a power of 2. Actually, over a small range of frequencies centered on the corresponding frequency.
x
45
46
Image-Guided Surgery
Figure 2–15. A giant voxel (blue) resulting from slice selection in the z direction (perpendicular to the page), followed by frequency encoding and analysis in the x direction. The widths of the voxel in x and z are 1 mm and 2 mm. The width in the y direction is the length of the active region of the scanner, but the only excited proton nuclei lie within the body.
for multiple contiguous slices, we can construct a 3D image, just as for CT. Phase Encoding. The method that we have presented involving CT reconstruction was the original method used by Paul Lauterbur (see Chapter 1, under “Discovery of MRI”). It led to Lauterbur’s Nobel Prize, and it works to some extent, but clinical MRI images are rarely produced in this way today because the giant “voxels” are distorted geometrically because of spatial variations in the static field (see subsection “Geometrical Distortion” below). Their distortion causes the images produced by the CT reconstruction algorithm to
be blurred. As a result, since the 1980s, the MRI image formation process has almost always incorporated an additional process for gleaning spatial information from the signal. This process is called “phase encoding”, which, as we saw in Chapter 1, was introduced by Ernst. Phase, which can be measured in degrees, is the proportion of the trip around the axis completed by a given nucleus at a given time during its precession — much like longitude measures a trip around the Earth’s equator, as we mentioned earlier in Chapter 2. Phase encoding is accomplished during the time period between (a) the excitation of the nuclei by the transmitter coils
A
B Figure 2–16. Two MRI acquisitions in a process analogous to CT. A. A set of giant voxels (outlined in blue) acquired by slice selection with a z gradient (perpendicular to the image) followed by an x gradient and frequency analysis. B. A second set of giant voxels resulting from selection of the same slice, followed by a gradient in the xy plane at an angle (approximately 25 degrees in this example) to the x axis and frequency analysis.
47
48
Image-Guided Surgery
while the slice-selection gradient is on and (b) the readout of the signal from the nuclei while the frequency-encoding gradient is on. During this time period, a gradient in the y direction, Gy , is turned on for a specified period of time. Each of these events — excitation, readout, and gradient application — is accomplished by means of a brief pulse of current through a coil — a transmitter coil, a receiver coil, or a gradient coil. As a result, the time course of these events is called a “pulse sequence”. Figure 2–17 depicts one such pulse sequence that includes the application of a phase-encoding gradient. While this phase-encoding gradient is on, the nuclei at one value of y get out of phase with nuclei at other values of y because Gy causes them to precess at different frequencies. During readout, the difference between the phase of each nucleus and what it would be without the phase-encoding gradient affects its
contribution to the overall signal, effectively “encoding” its y position. Slice + Phase + Frequency + Fourier Transformation = Better Image. Just
as frequency encoding of position can be decoded by numerical processing, phase encoding can be decoded as well, but unlike frequency encoding, it is not possible to perform phase decoding from one excitation because, unlike the frequency-encoding gradient, the phase-encoding gradient is turned off before samples of the signal are acquired. As a result, instead of contributing a different level of dephasing for each of a series of samples within one pulse sequence, it contributes only one. Therefore, many different phase encodings must be performed during a series of excitations and signal acquisitions. When the resulting set of signals is processed numerically using Fourier analysis (a linear-algebra technique in
Figure 2–17. Pulse sequence for MRI with phase encoding. Each line shows the time schedule for excitation (first line), gradients (middle lines), or readout (last line). Phase encoding is accomplished during the time between excitation and frequency encoding. During this time, a gradient in the y direction is turned on for a brief period.
Chapter 2
which calculations are done in the frequency domain as opposed to the time and/or space domainxi), a 2D image of the selected slice is produced. The process of excitation, phase encoding, and readout is repeated many times with the strength of gy being varied on each repetition to include negative values, positive values, and zero. Figure 2–18 depicts a pulse sequence that includes multiple strengths for the phase-encoding gradient. The change in the gradient changes the distribution of phases along the y axis. If, for example, 192 phase encodings are performed and each resulting signal is sampled 256 times (as described under “frequency analysis/encoding” above),
ct and mri
n
then the resulting 256 × 192 = 49 152 measurements can be subjected to Fourier transformation to produce a twodimensional image with an array of 256 × 192 voxels (x by y) that is considerably sharper than that which would be attained using the CT reconstruction technique. The strength and duration of the phase-encoding gradients are typically adjusted so that the voxel size in x and y will be the same; so in the example we are describing, we have voxels of size 1 mm × 1 mm × 2 mm. Figure 2–19 illustrates the result. For this example, the patient’s head was turned 90 degrees so that x is anteriorto-posterior and y is left to right. The reason for the patient reorientation is to
Figure 2–18. Pulse sequence for MRI that is identical to that in Figure 2–17, except that it shows multiple phase encodings, each of which is effected by means of a phase-encoding gradient of a different strength, including zero and negative values. The negative values represent gradients in the opposite direction from the positive-value gradients. xi
The human cochlea performs a Fourier transformer on audible sounds by taking time varying sound and extracting the individual frequency components to specific regions of the tonotopically arranged basilar membrane. The MRI receivers do essentially the same thing but over a different, inaudible frequency range.
49
50
Image-Guided Surgery
tions for each of the 50 slices and will readout 256 samples from each. Then each set of 49 152 samples is processed to produce one 2D image, and these 50 images are “stacked” in the computer memory to produce one 256 × 192 × 50 array of voxels that spans a 256 × 192 × 100-mm volume. Volume Imaging
Figure 2–19. Depiction of the voxel array that results from the pulse sequence of Figure 2–17. Phase encoding is in the vertical direction (yellow hatching ), and frequency encoding is in the horizontal direction (blue hatching ). The patient’s head is turned to the left, so the x axis (horizontal arrow) points from anterior to posterior and the y axis (vertical arrow) points from the patient’s left to right.
ensure that 192 mm in the phase encoding direction encompasses the entire head (the ear-to-ear distance being less than the front-to-back distance). Alternatively, the patient can remain supine with the frequency and phase-encoding directions interchanged. At this point, we have one image of one 2-mm slice of the patient. If we want a volume image, then we need to repeat the process for additional slices. The same slice-selection gradient is used, but the center of the frequency range of the slice-selection pulse is shifted. It retains the same width, so that the width of each z slice will remain at 2 mm, but since the center of the range is changed, the selected slice is changed. If, for example, 50 slices are desired, then we will carry out 192 slice selec-
The imaging protocol that we have just described involves the acquisition of 256 × 192 × 50 = 2 457 600 signals comprising 50 subsets of 49 152 samples, and each of these subsets contains all the phase and frequency encodings that pertain to one slice and one slice only. This protocol is sometimes called “2D acquisition” because the two-millionplus signals obtained during the scan can be partitioned into subsets such that each subset is used to produce one 2D array of intensities to depict the contents of one slice. This partitioning is possible because at any instant while the image is being acquired, the signal being sampled arises from just one excited slice. An alternative approach is to use socalled volume imaging or 3D imaging protocols. In volume imaging, the sliceselection process is replaced by “slab” selection. Instead of a slice, which might be 1 mm to 4 mm thick, a slab that is 100 to 200-mm thick is excited. To get 3D information from within this slab, phase encodings are performed in two directions instead of one. Frequency encodings are performed in the remaining third direction. So, if we wish to use volume imaging to replicate the 2D arrays produced for each slice in our slice-selection example, frequency encoding will be done with an x gra-
Chapter 2
dient and 256 samples acquired, phase encodings will be done in the y direction with a set of 192 phase encoding gradients with the gradients and timings chosen to produce an in-plane resolution of 1 × 1 mm, and a second set of phase encodings will be done in the z direction. Specifically, for every one of these 192 phase encodings in y, a separate set of 50 phase encodings is performed in the z direction (with a z gradient). Figure 2–20 depicts the resulting pulse sequence. To replicate the fifty 2-mm slices of our slice-selection example, we will assume that the selected slab is 100 mm wide and that we want to divide this width into fifty 2-mm voxels. In this case, volume imaging requires a total of 192 × 50 = 9600 phase encodings. Before each of these phase encodings, the slab-selection gradient is turned on and the slab is excited with the same slab-selection pulse, and, immediately after each phase encoding,
ct and mri
n
256 samples are acquired from the signal. The total number of signal samples is 256 × 192 × 50 = 2 457 600, just as in the slice-selection approach, but for volume imaging, these samples cannot be partitioned into slice-specific subsets. No such partitioning is possible because in volume imaging, the entire volume of the slab is excited during the acquisition of each signal. Instead, the entire set of signals is subjected to a three-dimensional Fourier transform to produce a 256 × 192 × 50 array of voxels that spans a 256 × 92 × 100-mm volume. After the processing is complete, we have arrived at the same image with the same number and size of voxels. Thus, we have two approaches to choose from. Although each method has its advantages, volume imaging is more common because of a higher speed of acquisition, as we will see in the section entitled “Image Acquisition Time” below.
Figure 2–20. Pulse sequence for MRI volume imaging. The first application of the z gradient selects a slab; the second encodes phase.
51
52
Image-Guided Surgery
Echoes Each of the above approaches — slice selection and volume imaging — allows us, in principal, to spatially localize tissue based on its response to magnetic fields, which is all we need to produce an image. However, with these methods, as we have described them, the signal received would be so weak that it would be overwhelmed by background noise and be useless for creating images. As will be explained in this section, the weak signal is due to “dephasing” in which the nuclei become misaligned in their precession. But — the good news is — dephasing can be overcome using “echoes”. When an electromagnetic pulse at the resonant frequency of a set of nuclei impinges on them, it not only excites them but it also puts them all in phase so that they precess in unison around the magnetic field that they are exposed to. In other words, at any given point in time, they all have the same phase as they precess together like dancers twirling in unison. While the pulse is on, it keeps the nuclei in this uniform dance, and it remains on long enough to allow the nuclei to absorb the maximum possible energy — a few milliseconds. When the transmission coils are then turned off, the fact that all the nuclei are in phase means that their individual signals add together to produce a strong signal. This situation does not last, however, because without the radiation from the transmission coils forcing them to stay together, they begin to lose their alignment. They become misaligned because of slight variations in the magnetic field that each of them is exposed to, which cause corresponding
variations in their precession frequencies. In our twirling dancers analogy, this might resemble some dancers being slowed down by variation in local conditions (eg, the roughness of the floor). As a result, over a time interval that may be as short as a few milliseconds to as long as several seconds (depending on the magnetic environment), their phases will become randomly distributed over the entire 360-degree range. Even though each is still precessing and giving off energy, their individual magnetic fields cancel each other because for every nucleus pointed in one direction, there is another pointed in the opposite direction. At this point, their cumulative signal will have faded to zero. This fading of the signal is called, for historical reasons, “free-induction decay”, and it is universally observed in any MRI scanner or NMR system that employs pulsed excitation (Figure 2–21). Spin Echoes. The differences in the
magnetic fields experienced by these nuclei are of two types — time invariant and time variant. The time-invariant differences arise from the fact that each nucleus is located at a different position in the body. Spatial variation in the magnetic susceptibility of the body, which is a function of its molecular makeup from point to point, will cause the static field to deviate from the almost perfectly uniform strength (eg, 0.2 T, 1.5 T, 3.0 T) produced by the exquisitely shimmed main field coils when there is no object in it. This spatial deviation from perfection is called “static-field inhomogeneity”, and because of it, the static field experienced by nuclei at two different points differs by a constant amount. The time-variant differences arise from
Chapter 2
ct and mri
n
Figure 2–21. Free-induction decay (FID). Depiction of a pulse of electromagnetic energy that excites the nuclei at their resonant frequency (Excitation) and causes them to precess in phase and the subsequent signal that they give off. Because of dephasing among the nuclei due to spatial variations in the static magnetic field, known as “inhomogeneity”, the signal begins decaying immediately (FID) and drops to zero.
the fact that the precessing nuclei interact with each other (much like spinning tops bumping into each other), causing slight variations in precession speeds over time leading to phase spread. These are known as “spin-spin” interactions, and they are random processes. The time-invariant differences can be corrected by applying a second electromagnetic pulse similar to the excitation pulse but twice as strong, as shown in Figure 2–22. The effect of applying a second pulse after the signal has disappeared was first observed in 1949 by Erwin Hahn,6 and it is as surprising now as it was then: after a short wait the signal reappears! The effect is analogous to what happens when one shouts in the direction of a high rock wall that is a few hundred meters away. The sound of the shout dies out, then
there is a period with no sound at all, and then there is an echo. The reappearance of the signal in nuclear magnetic resonance is likewise referred to as an “echo”. It is sometimes called a “Hahn echo” but is more commonly called a “spin echo”. Like a sound echo, its strength is smaller than the original signal, but unlike a sound echo, its shape is roughly that of the original juxtaposed next to its mirror image, and so it lasts twice as long as the original signal. The time between the beginning of the peak original signal and the peak of the echo is called the “echo time”, and it is usually symbolized with two uppercase letters: “TE” for “time of echo”. The two pulses from the transmission coil are distinguished by the names “90-degree” pulse for the excitation pulse and “180-degree” pulse for the
53
54
Image-Guided Surgery
Figure 2–22. Spin echo. At a time TE/2 after the 90-degree excitation pulse, a 180-degree pulse is transmitted. The initial signal (FID) decays to zero because of dephasing caused by static-field inhomogeneity. The 180-degree pulse causes rephasing, and as a result, after an “echo time”, TE, the signal reappears, approximating a slightly diminished form of the original signal plus its mirror image, and so it is twice as long.
echogenic pulse. These names are based on 90-degree and 180-degree rotations of the precessing magnetic vectors, which take place while the pulses are being transmitted but not about the precession axis. These rotations are instead about an axis perpendicular to the precession axis that is itself rotating at the precession rate. The spin echo occurs because the 180-degree rotation reorients the precessing nuclei so that the ones that were in the lead are moved to the rear of the pack and vice versa, but their precession rates remain the same. So now the slowest are in the lead, the fastest are at the back, and in exactly the same amount of time that it took them to disperse, they rephase, and the signal builds to a peak. The rephasing is not perfect because of the time-variant portion of the field differences, so the peak of the echo is never quite as high as the
original peak, and the height of the peak decreases exponentially with TE. Gradient Echoes. In addition to static field inhomogeneity and spin-spin interactions, there is a third source of variation in the magnetic field that causes dephasing — the impressed gradients. Because a gradient is by definition a spatial variation in the static field, it is also by definition a form of static-field inhomogeneity, and so it also causes dephasing. Thus, gradients, which are necessary for slice selection, frequency encoding, and phase encoding, all cause dephasing that reduces the signal. The dephasing caused by phase encoding, and its resulting signal reduction is a necessary part of image production, but the dephasing and signal reduction caused by the other two gradients are deleterious
Chapter 2
side effects. There is a trick for reducing those side effects that involves a reversed gradient. The forward and reversed gradients are referred to as positive and negative gradient “lobes”, and, in the case of frequency encoding, the negative lobe causes an echo, which is called a “gradient echo”. Figure 2–23 depicts a pulse sequence that includes such negative lobes. The sequence is a slice-selection pulse sequence, and it is identical to the slice-selection sequence of Figure 2–18 except that there is a negative lobe included with the sliceselection gradient that follows the positive lobe, and there is a negative lobe included with the frequency-encoding gradient that precedes the positive lobe. The vertical dashed line labeled TE shows the time center point of the gradient echo. Figure 2–24 also shows a sequence that produces a gradient echo.
ct and mri
n
It is a volume-imaging sequence, and it is identical to the volume-imaging sequence of Figure 2–20, except that the slab-selection and frequency-encoding gradients include negative lobes. Combined Echoes. Virtually every pulse sequence includes gradient echoes wherever possible, so when a 180-degree pulse is included to produce a spin echo, there is inevitably a gradient echo as well. Since gradient echoes are always included, any sequence that includes a combination of spin echo and gradient echo is called simply a “spin-echo” sequence with the idea that gradient echoes are also included being understood. Figure 2–25 depicts a spin-echo sequence, complete with 180-degree pulse and ubiquitous negative gradient lobes for the slice-selection and frequency-encoding gradients. The
Figure 2–23. Pulse sequence for gradient-echo slice-selection MRI. The sequence is identical to that depicted in Figure 2–17 except that the slice-selection gradient and the frequency-encoding gradient each includes a negative lobe to cancel the dephasing caused by the positive lobe. The result is an echo (not shown). The echo time (TE) is at the center of the positive lobe of the frequency-encoding gradient (vertical dashed line).
55
Figure 2–24. Pulse sequence for gradient-echo volume-imaging MRI. The sequence is identical to that depicted in Figure 2–18 except that the slab-selection gradient and the frequency-encoding gradient each includes a negative lobe to cancel the dephasing caused by the positive lobe. The gradient echo occurs at time TE, located at the center of the positive lobe of the frequency-encoding gradient.
Figure 2–25. Pulse sequence for spin-echo MRI. The sequence is identical to that depicted in Figure 2–23 except that a 180-degree pulse has been added. A signal echo occurs at the center of the positive lobe of the frequency-encoding gradient. The echo strength results from rephasing by the negative lobes of the slice-selection and frequency-encoding gradients and the 180-degree pulse. The pulses are synchronized so that both gradient echo and spin echo combine at TE.
56
Chapter 2
pulses are timed so that the spin-echo and the gradient echoes occur simultaneously, as shown in the figure. Relaxation Times MRI provides much more information about tissue than CT. CT images are maps of tissue density while images produced by MRI are maps that are modulated by both density and tissue relaxation times, and of the two, relaxation times convey considerably more information. The density in a CT image is roughly proportional to mass density, with bone being the densest tissue in the body and the brightest in a CT image. However, the density that modulates MRI is the number of hydrogen nuclei per unit volume, which is roughly proportional to the “wetness” of the tissue because these nuclei are found predominantly in water. As a result, bone typically produces very little signal in MRI and typically appears dark. T2 and TE. The dephasing caused by spin-spin interactions was modeled by Felix Blochxii in 1946.6 He found that
ct and mri
n
spin-spin dephasing reduces the signal exponentially with time. He introduced time constant T2, which is equal to the time at which the signal has been reduced to 36.8% of its starting value. T2 depends on the material of which the proton nuclei are a part. T1 and TR. Bloch also modeled a sec-
ond exponential time dependence with a second time constant, T1, but this second model reflects exponential growth (toward an upper limit) instead of exponential decay. After the signal has died away because of T2 decay, another 90-degree pulse can be applied immediately to re-excite the nuclei. Unfortunately, the resulting signal will be much smaller than the signal produced by the first excitation. Thus, repetition of excitations is delayed to allow the signal to grow. The delay is called the “repetition time” or “TR”, and T1 is the value of TR for which the signal grows to 63.2% of the original signal. Note that the signal grows faster if T1 is smaller and that T1, like T2, varies with tissue type (see Table 2–2). Table 2–2 gives ranges for
Table 2–2. Tissue Relaxation Times and Densities17 Tissue
T1, s
T2, s
Proton Densitya
CSF
0.8–2.0
0.11–2.0
1.9 ± 0.9
Gray
1.09–2.15
0.06–0.11
1.3 ± 0.2
White
0.76–1.08
0.06–0.10
1.0 ± 0.1
Adipose
0.2–0.75
0.053–0.094
0.9 ± 0.3
Muscle
0.95–1.82
0.002–0.067
0.8 ± 0.3
0.5–2.2
0.05–0.16
0.3 ± 0.2
Meninges
Note. aRelative to the mean value of white tissue (± gives approximate range).
xii
For this and other work regarding NMR, Bloch received the Nobel Prize in Physics in 1952.
57
58
Image-Guided Surgery
T1, T2, and density for hydrogen nuclei in selected tissues. Weighting. As mentioned above, the relaxations times, T1 and T2, play the major role in the determination of MR signal strength, which translates into image intensity. By changing TR and TE, images can be produced such that the contrast between the intensity of one tissue compared to another is weighted toward highlighting the differences in T1, in T2, or in proton density. From Bloch’s equations, the following relationship among these time constants can be derived:
(2–1)
where dn, T1n, and T2n are the density, T1, and T2 of tissue type n, and c is a proportionality constant that determines the brightness in the image of a unit of signal Sn. TR values range from 0.5 to 4 seconds, and TE values range from 0.02 to 0.1 seconds. By selecting values for TR and TE from these ranges and values for T1 and T2 from Table 2–2, inserting them into Equation 2–1, and comparing the signal strengths produced for tissues that differ strongly in T1 or T2, it can be seen that if TR and TE are both small relative to T1 and T2 respectively, the image will be
T1 weighted, and if TR and TE are both large, the image will be T2 weighted. If TR is large and TE is small, neither T1 nor T2 will have much effect, and the image will be density weighted. Since density in MRI refers to the number of hydrogen nuclei per unit volume, and since a hydrogen nucleus is a proton, a density weighted MRI image is said to be a “proton density” image. Rough guidelines for the meanings of “small” and “large” TR and TE to produce these weightings for intracranial imaging are given in Table 2–3. Image Acquisition Time TR and TE are each a couple seconds or less, and yet the typical image acquisition takes much longer. The reason is that in order to produce an image, it is necessary to perform multiple phase encodings, and for every phase encoding in a standard imaging protocol,xiii the entire sequence from the beginning of the excitation pulse to the end of the sampling time must be carried out. After one such sequence is concluded, the next begins, not necessarily at the end of the previous sampling time but possibly much later than that. In order to produce the desired weighting, as described above, it begins at time TR after the previous excitation pulse. Each
Table 2–3. Weighting Guidelines for TR and TE
xiii
Weighting
TR Range
TE Range
T1
TR < 0.5 s = 500 ms
TE < 0.01 s = 10 ms
T2
TR > 1.5 s = 1500 ms
TE > 0.05 s = 50 ms
Proton density
TR > 1.5 s = 1500 ms
TE < 0.01 s = 10 ms
“Standard” excludes multi-echo and echo-planar acquisitions.
Chapter 2
repetition is identical to the one before except that there is a change in gradient strength for each phase encoding. In 2D imaging, there is a large number, for example, 192 or 256, of phaseencoding gradients applied, so there is a large number, for example, 192 or 256, of repetitions. The reason for the large number of repetitions is that the number of phase encodings, N, is equal to the number of voxels in the phase encoding direction. Because each time interval between repetitions is TR, the total imaging time = (N − 1) × TR + the required time for one phase-encoded acquisition (ie, TE plus time to acquire one signal). For 192 phase encodings with TR = 3.8 seconds and time for one phase encoding of 0.1 second, this means that the acquisition of one slice requires 12.2 minutes. If 50 slices are desired, then 192 phase encodings are required for each slice, which equals 9600 phase encodings, which, with this protocol, would require about 10 hours! Multiplanar Imaging. Ten hours is clearly
not feasible for human imaging. However, it is possible to perform phase encodings for multiple slices during one TR because the time for one phase encoding is, as in our example above, usually much shorter than TR. This technique is called “multiplanar” imaging. When multiplanar imaging is used in our ongoing example, 38 phase encodings, which take 0.1 second each, will fit within each 3.8-second TR. So, one phase encoding can be acquired from each of 38 slices in 3.8 seconds. As a result, instead of acquiring the 192 signals for just one slice in xiv
ct and mri
n
12.2 minutes, the 7296 signals required for 38 slices can be acquired in 12.2 minutes. Since 50 slices are desired, the scan can be completed in 24.4 minutes (12.2 minutes for the first 38 slices and 12.2 minutes for the remaining 12 slices) — considerably better than 10 hours. Our multiplanar acquisition protocol produces a volume image comprising 50 slices, each of which utilizes 192 phase encodings and collects 256 samples during frequency encoding. We used a gradient strength and bandwidth for slice selection that produces a 2-mm slice, which means that the voxel size in the z direction is 2 mm. We used a gradient strength and bandwidth for frequency encoding that produced a1-mm spacing in the x direction. We did not give the specifications for phase encoding, but they are always chosen so that the spacing is the same as for the frequency encoding to keep the in-slice voxel shape square. Thus, the volume image constructed from these 50 slices has a voxel dimension of 1 × 1 × 2 mm (x × y × z), and its voxels are assembled into a 256 × 192 × 50 three-dimensional array. Averaging. In some cases, the entire
process that we have described will be repeated and the resulting two arrays of voxel intensities will be averaged in order to get a higher signal-to-noise ratio (S/N).xiv “Noise” is simply random fluctuations in the signal unrelated to the radiation from the nuclei, and unfortunately, such fluctuations are inevitable. They are inevitable because the patient, the bed, the walls, floor and ceiling, the lights, and even the scanner itself all emit random elec-
The averaging is typically applied to the signal samples before image construction, but the effect is the same.
59
60
Image-Guided Surgery
tromagnetic energy by virtue of the fact that their atoms are moving. The radiation, which is called “black-body radiation”, is emitted across a broad spectrum determined by the temperature of the body, and the portion of that blackbody radiation that strikes that scanner and happens to fall within the range of the proton precession frequencies will be picked up by the receiver coils. This random contribution to the signals in that range causes what is sometimes called a “salt-and-pepper” effect in the image because the intensity change is positive in some voxels and negative in others and appears to be sprinkled among them randomly. However, because the added values are random and the true signal is consistent, averaging two acquisitions has no effect on the signal (unless the patient moves), but the cancelling of positive and negative fluctuations from the noise will tend to reduce the added random component by a factor of about √ 2, which is approximately 1.4. Thus, doubling the number of signals acquired during the scanning of a patient has the advantage of increasing the image clarity by about 40%. The number of identical acquisitions for noise reduction is typically labeled NEX (Number of EXcitations). For example, with two acquisitions, NEX = 2. The disadvantage of the additional acquisition (or acquisitions) is that the patient must lie still in the scanner longer, which, in addition to causing discomfort, increases the likelihood of image artifacts being caused by patient motion. However, the acquisition of two images does not necessarily take twice as long as a single acquisition. In the example we have been considering, the slice acquisitions had to be
broken into two blocks — each of which numbered no more than 38 in order to fit the phase encodings into 1 TR. We chose blocks of 38 and 12, but we could have divided it into, say, 25 and 25. If we now double the acquisitions to 100 phase encodings, instead of requiring four blocks, they will fit into three blocks — 38, 38, and 24 or, for example, 33, 33, and 34 — so to increase S/N by 40%, the patient’s time in the scanner is increased by 50% to 36.6 minutes. However, if a diagnosis or surgical plan could be made with thirty-eight 2-mm slices, or perhaps using a 2.6-mm slices in order to maintain FOVz at 100 mm with 38 slices, then the imaging time with NEX = 2 is back down to 24.4 minutes. And, if 128 voxels, instead of 192, are acceptable in the y direction, then only 128 phase encodings are needed, which reduces scan time from 24.4 to 16 minutes. Small-Angle Excitation. Now, we turn
our attention to the time required for a volume acquisition. If we return to the volume-acquisition example that we considered when we introduced this technique, in which we employed 192 × 50 = 9600 phase encodings in the y and z directions, the imaging time = 9599 × TR + time for one phase encoding. If TR were 3.8 seconds, as in the slice-selection example above, the imaging time would be again be 10 hours. The multiplanar trick used to speed up slice selection, in which one slice is excited while others are relaxing, does not work for volume imaging because an entire slab is excited at once, and clearly volume imaging would not be feasible if it were not possible to speed up the process. In the mid-1980s, a dramatic speedup was achieved by greatly
Chapter 2
reducing TR.7 The approach has different names when implemented on scanners from different manufacturers, but it was originally called FLASH (acronym for “Fast Low-Angle Shot”) and retains this name on Siemens machines, with GE using the acronym SPGR (for “SPoiled GRadient”) and Philips using the acronym T1-FFE (for T1-“Fast Field Echo”). The “low-angle shot” refers to the excitation pulse, which is reduced in strength so that, instead of producing the maximum excitation, which is achieved with a 90-degree rotation of the nuclei, it produces only a partial excitation by rotating them by only 5 to 20 degrees. A slight disadvantage is that a 180-degree pulse interferes with it, and so it is not possible to produce a spin echo. The great advantage of the small-angle excitation is that nuclei can recover more quickly from one excitation to be ready for the next one. As a result, TR can be a mere fraction of T1 and still produce a bright image, and that short TR is what put the “Fast” in Fast Low-Angle Shot. With a TR of 8 ms for TR, our example 256 × 192 × 50 array with voxel dimensions of 1 × 1 × 2 mm (x × y × z) can be acquired in only 9600 × 0.008 seconds = 1¼ minute. In fact, because the low-angle acquisition method affords such fast acquisitions, clinicians typically specify a higher resolution in z for volume imaging compared to slice selection. For example, a 256 × 192 × 256 array of isotropic 1 × 1 × 1-mm voxels can be acquired using FLASH or any of its related variants in under 7 minutes! The inherent disadvantage of fast volume imaging is that, with such a short TR, TE, which must be shorter than TR, cannot be long enough to produce a T2-weighted image, but this problem is usually con-
ct and mri
n
sidered a small sacrifice given the high resolution possible with such a short acquisition time. The Fast Lane In our calculations of imaging times, we have focused on the two workhorses of MRI — (a) multiplanar slice selection, which is limited in the number of slices per volume but can produce any of the weightings — proton density, T1, or T2, and (b) volume imaging, which can produce a large number of slices but is limited to T1 weighting. We have shown how each sequence is related to the acquisition time and how they can be sped up, but there are variations on these two themes that allow for even faster imaging. Multiple Phase Encodings per Excitation. Two of these variations accom-
plish their time savings by acquiring multiple phase encodings per excitation within the same slice. The first is “fast spin echo”, which on some scanners is called “turbo spin echo”, and the second is echo-planar imaging (EPI), which, as we saw in Chapter 1, was introduced by Mansfield. Each causes a decrease in S/N, and EPI suffers from increased geometric distortion, but they also result in considerable speedups. In either case, the idea is that after a signal is acquired and before the next excitation pulse is transmitted, the phase encoding gradient can be turned on again to effect a different phase encoding, the readout gradient can be reversed again to produce an additional gradient echo, and, if it is a spin-echo image, a 180-degree pulse can be generated again to produce a spin echo (synchronized as usual with the ever-present gradient echo). During
61
62
Image-Guided Surgery
the additional echo, a set of samples can be acquired. If this process is carried out N times per excitation, it will provide N signal acquisitions per excitation, where N is the maximum number before either the signal strength has decayed too much or before all of TR has been used up. Then it is time for another excitation. Repeating this process for each excitation reduces that image acquisition time by a factor of N. In some EPI sequence, N = the number of phase encodings, which is the maximum possible speedup factor, but, as we point out below under “Distortion”, it also causes the maximum geometrical distortion. Parallel Imaging. A third method for speeding things up, which has been available since the late 1990s, is called “parallel imaging” and is also variously known as SENSE (“SENSitivity Encoding”), SMASH (“SiMultaneous Acquisition of Spatial Harmonics”), or GRAPPA (“GeneRalized Autocalibrating Partially Parallel Acquisition”), each of which is a variation on the same theme.8,9 That theme requires a significant hardware upgrade in which multiple receiver coils are installed, each of whose sensitivity to the received signal is high in only a local region within the bore and each of whose signals can be sampled independently and simultaneously (ie, in parallel). Taking advantage of their localized sensitivities allows an image volume of a given number of voxels to be acquired with fewer phase encodings. Fewer phase encodings means shorter acquisition times — with speedup factors ranging from 2 to 3.9 There is a small sacrifice in S/N, but there is no increase in geometric distortion, and the acquisition time is reduced in direct proportion to the reduction in
phase encodings. With the speed advantage and with almost no disadvantage, parallel imaging is now used routinely and represents the major step forward in MRI technology in the last 15 years. Speed Limits. We have seen that judi-
cious choices in parameter settings can have a big impact on the contrast and the resolution of an MR image and also on the length of time the patient must lie still while the image is acquired. Each method for reducing that time adds to those choices, and each has its own set of issues, but all of them can be sped up further by increasing the static field, which increases the rate of precession, which increases the rate at which energy is received from the nuclei, shortening the time required to achieve an acceptable S/N. However, higher fields cause increased forces and torques on ferromagnetic clips and implants, and they cause increased vertigo resulting from magnetic forces within the labyrinth.10 Furthermore, faster imaging typically requires faster gradient switching. High-speed gradient switching represents an engineering challenge, but it also represents a potential physiological hazard because changing magnetic fields induce currents in conductors, and the human body is a conductor, as Peter Mansfield well knew while he was inside that scanner in Nottingham in 1978, as described in Chapter 1. The levels of the currents induced in Mansfield’s body were negligible then and they remain negligible today for any patient in a routine scan today except for those with pacemakers, cochlear implants, and the like, which have metal conductors, but it does represent a potential limiting factor to future MRI speedups.
Chapter 2
Examples
ct and mri
n
Figure 2–26 shows examples of weighting in MR images. The figure includes four images acquired from the same
patient. The upper left image is a CT, included for comparison; the other three images are MRI slices. All four are axial orientations, each at approximately the same anatomical position
A
B
C
D
Figure 2–26. Four axial images, each taken from a set of slices obtained using slice selection of the same patient and each at approximately the same anatomical position. A. CT 200 kVp, dimensions = 512 × 512 × 29 with 0.65 × 0.65 × 4-mm voxels. B–D. MRI images, dimensions = 256 × 256 × 26 with 1.25 × 1.25 × 4-mm voxels. NEX = 2. B. Proton density weighted, TR, TE = 3.0, 0.02 seconds. C. T1 weighted, TR, TE = 0.8, 0.015 seconds. D. T2 weighted, TR, TE = 3.0, 0.09 seconds.
63
64
Image-Guided Surgery
and orientation. Each MR image was obtained using slice selection. Panel (A) is a CT, acquired with a tube voltage of 200 kVp, dimensions = 512 × 512 × 29 with 0.65 × 0.65 × 4 mm voxels. Panels (B) through (D) are MR images. Each MRI was acquired with the same dimensions: 256 × 256 × 26 with 1.25 × 1.25 × 4-mm voxels, and each is the result of the averaging of two acquisitions (NEX = 2). Panel (B) is a proton density–weighted image with TR, TE = 3.0, 0.02 seconds. Panel (C) is a T1-weighted image with TR, TE = 0.8, 0.015 seconds. Panel (D) is a T2-weighted image with TR, TE = 3.0, 0.09 seconds. As can be seen in this figure, and in an overly simplified sense, CT scans are preferred for the evaluation of bone, T1 weighting shows tissue with high fat content as bright and tissue with high water content as dark, T2 weighting show tissue with high water content as bright, and proton-density weighting best differentiates tissue based on proton density (eg, differentiating muscle from bone). Distortion Distortion in MRI can be divided into two categories — geometric distortion and intensity distortion. Either or both are present to some extent in every image of every patient. The distortion is small enough to be ignored in most clinical applications, but there are concerns, particularly arising from geometric distortion, when MRI is used for surgical guidance. Geometric distortion is the spatial shift of intensity to an erroneous point in the image, while intensity distortion is an erroneous decrease or increase in intensity at a given point. For example, if the facial nerve appears xv
Subject to relatively small amounts of scatter.
shifted laterally from its true position by 1 mm, then the problem is geometrical distortion. If the nerve appears at its true location but is only half as bright as it should be, then the problem is intensity distortion. Because geometric distortion has the greater potential for harm in conjunction with IGS, we will consider it first. Geometric Distortion. CT is typically
preferred over MRI for highly accurate surgical guidance because of its geometrical fidelity. Geometrical distortion is far more serious for MR than for CT, and to understand why, it might help to review what we have learned about tomographic imaging via CT and MR. Tomographic imaging is a 3D mapping from physical space to virtual space, and, as we have seen, CT bases its mapping on the detection of beams of x-rays, which comprise electromagnetic radiation with wavelengths that are far smaller than any anatomical feature. These beams pass through the tissue at multiple directions and positions. Thanks to that small wavelength, the fact that x-ray beams travel in almost perfectly straight lines,xv and the geometric accuracy of modern CT scanners, these directions and positions can be analyzed geometrically to produce a mapping that, for the purposes of surgical guidance, is perfect. MRI, on the other hand, bases its mapping on the frequencies of electromagnetic radiation whose wavelengths are far too large to be subjected to such direct geometrical constructions. Each radiation frequency is equal to the precession frequency of some set of hydrogen nuclei, and both are proportional to the magnetic field felt by the nuclei.
Chapter 2
In MRI, that field is made spatially uniform with magnetic shimming when the scanner is calibrated, and then, during image acquisition, it is varied spatially by the MR scanner’s gradient coils so that the field strength can be systematically calculated at every position in the scanner at any instant of time. Knowledge of the field strength conveys knowledge of the precession frequency, which conveys knowledge of the position of the precessing nuclei, and by that indirect process, an MR image emerges from the scanner. In order for the mapping to image space to be geometrically true, the gradient in the magnetic field must be true, but there is in fact always some degree of distortion in that gradient, and unlike the situation in CT, distortion in MR can be large enough to compromise the accuracy of any guidance system that relies on it. Geometric distortion can arise from either or both of two sources: (1) a gradient whose strength differs from its nominal value and (2) static-field inhomogeneity. In either case, nuclei precess at frequencies that are offset by an unknown amount from their calculated frequencies because the field strengths at their true positions are offset by an unknown amount from their expected strengths. Distortion from (2) can be a shift in the slice-selection direction or in the readout direction. In the phase-encoding direction(s), distortion may arise from Source (1)— a gradient whose strength xvi
ct and mri
n
differs from its nominal value — but not from the Source (2) — static-field inhomogeneity — except for EPI, in which both sources can cause distortions in all directions.xvi If the error causes an increase in the field, then the nuclei will erroneously appear in the image at a position that is shifted in the direction in which the field increases —“up the gradient”. If the error causes a decrease in the field, they will be shifted down the gradient. The amount of the shift ∆s depends on two things — the error in the field ∆M and the size of the gradient G: ∆s = ∆M/G. Suppose, for example, that at some point in the image, the error is six-millionths of 1 tesla, which is equal to 4 ppm for a 1.5 T scanner and 2 ppm for a 3.0 T scanner; then, if the readout gradient is 6 mT/m, which, as we saw earlier, corresponds to a bandwidth of about 65 kHz for a 256-mm FOV, the received readout signal error will be 1 mm. The distortion calculation is the same for the gradient in the readout direction and in the sliceselection gradient and is zero for the phase-encoding gradient. Increasing the gradient slightly to 7 mT/m would reduce the error slightly to 0.86 mm. For EPI gradients, the effective gradient strengths are much smaller and cause much larger distortions.xvii Clearly, the geometrical error is reduced by increasing the gradient, which corresponds to an increase in the bandwidth, but when the bandwidth is increased, the image is noisier. So, increasing the
For EPI, the distortion in the phase-encoding direction arising from static-field inhomogeneity is in fact worse than that in the readout direction. This occurs because phase encoding relies not on an accurate precession rate for a given phase-encoding gradient strength but on accurate changes in the precession rate with changes in the strength, and, except in EPI, field inhomogeneity neither adds to nor subtracts from that change. xvii The effective phase-encoding gradient in the case of blipped EPI is less than the applied gradient. It equals the applied gradient times the blip duration divided by the time between blips and so is lower and results in higher distortion.
65
66
Image-Guided Surgery
gradient strength is not a simple panacea for distortion. It might be thought that geometric distortion can be corrected by simply acquiring an image of a known shape, determining the nature of the distortion by comparing the known shape with the observed image, and then applying a correction to patient images. That approach works to some extent for the geometrical distortion arising from Source (1) which is error in the applied gradients, and some scanners provide a postprocessing option to correct that type of distortion. However, the distortion arising from Source (2), which is static-field inhomogeneity, is caused by the patient and so varies with the patient and with the position and orientation of the patient in the scanner. A recent test (2014) of multiple scanners and standard spin-echo and gradient-echo, slice-selection, and volume acquisition protocols found that standard spin echo and gradient echo, slice selection, and volume imaging produce typical geometrical distortions on the order of 0.5 mm to 2.5 mm within 100 mm of the center of the scanner and 2 mm to 4 mm at 200 mm from the center.11 These are not EPI images, which are known to be geometrically untrustworthy and are avoided with IGS. They are images acquired using protocols commonly used by surgical-guidance systems. Distortion was found to be decreased in most cases by utilizing the postprocessing gradient correction options. As expected, it was observed that decreasing the bandwidth (and therefore the gradient) was associated with higher average distortions and vice versa. The lowest bandwidth of 7.7 kHz showed average distortions of 2 mm and the highest of 77 kHz showing average distortion of 0.6 mm.
These values of geometric distortion in MRI might seem surprisingly large to many clinicians. For diagnostic imaging, they do not usually represent a problem because radiologists “read” the entire scan as opposed to individual voxels when identifying pathology. For image guidance, however, these inaccuracies do come into play and are the primary reason that CT is typically the preferred imaging modality for IGS. As pointed out in Chapter 1 and revisited in Chapter 4, entitled “Registration”, there is method for combing CT and MRI. That combination provides a relatively straightforward way to utilize both the superior geometrical fidelity and spatial resolution of CT and the superior soft-tissue identification of MRI. The method is to “fuse” the MRI image to the CT image. This technique involves the exploitation of so-called mutual information in the two scans and using that information to superimpose one on the other. Although the typical mutual-information registration to CT does not correct the geometric distortion of MRI, it always allows for more accurate navigation than the use of MRI alone. Intensity Distortion. In the section
above, entitled “Spin Echoes”, we described a method to correct for dephasing caused by static-field inhomogeneity. The dephasing causes a reduction in the signal and the correction method is to transmit a 180-degree electromagnetic pulse, which refocuses the phase, and then wait until the signal regains strength. The peak of the signal is called the “spin echo”. This reduction in signal is more severe in those areas where the static-field inhomogeneity varies more rapidly from one point to another, and when the signal is pro-
Chapter 2
cessed to produce an image, the result is that the intensity at some places is abnormally low. This spatially varying diminution of the intensity is a form of intensity distortion, and it is present to some level in any gradient-echo image because gradient-echo sequences are by definition devoid of 180-degree refocusing pulses. Another source of intensity distortion, which can happen with both gradient-echo and spin-echo imaging, is an indirect consequence of geometric distortion. As explained above, when static-field inhomogeneity adds to the applied gradient, the apparent position of nuclei is shifted. If, for example, the static field in the x direction at a given position x,y,z = 100,100,100 mm is larger than it should be, then nuclei that should appear at x,y,z = 100,100,100 mm might appear in the image at, say, x,y,z = 100,100,102 mm. This is a simple geometric distortion of 2 mm, and there may be no intensity distortion associated with it. However, if at the same time, the x gradient at the position x,y,z = 100,100,105 mm is smaller than it should be, then the nuclei that should appear at x,y,z = 100,100,105 mm might also appear in the image at x,y,z = 100,100,102 mm — a shift in the opposite direction (of 3 mm). The result is that the intensity at x,y,z = 100,100,102 mm is a superposition of intensities from nuclei at two places, and so the resultant intensity at x,y,z = 100,100,102 mm is abnormally bright. This happens to some extent wherever the geometric distortion acts in a way that squeezes a larger physical volume into a smaller image volume. Concomitantly, whenever the distortion stretches a physical volume into a larger space, the intensity is abnormally dim. This distortion effect based on squeezing and stretching is sometimes
ct and mri
n
called the “Jacobian” effect based on the name of the mathematical quantity that calculates the ratio of physical to image volume when geometrical distortion occurs. Fat Shift So far we have implicitly assumed that the hydrogen nuclei involved in MRI are part of water molecules. This assumption is usually good because it is usually true, but another substantial source of the MRI signal comes from oil. Fat tissue is full of oil, and as a result, it shows up in MRI just like water. In fact, as can be seen from Table 2–2 (Adipose), its T1 value is quite short, which makes it bright in both proton density–weighted and T1-weighed images, as can be seen in Figure 2–26 panels B and C. Unfortunately, the fat in an MR image is not where it appears to be! The molecular environment of fat molecules affects the magnetic field that is experienced by the hydrogen atoms inside them such that it is always a bit less than that experienced by the hydrogen atoms in water at the same place, and thus their precession frequency is a bit less as well. The field and the frequency are each reduced by about 3.5 ppm. Thus, in our example above (Slice Selection plus Frequency Encoding and Analysis), with a gradient of 6 mT/m, at 120 mm in the readout direction, fat protons would precess more slowly than water protons by γ × 1.5 T × 0.0000035 = 224 Hz. As we saw in that example, a change of 255 Hz is interpreted by the computer analysis of the signal as a 1-mm displacement in x. Thus, the fat will appear to be 224/255 = 0.88 mm in the negative x direction from where it really is, and the shift will be larger with smaller gradients. Since
67
68
Image-Guided Surgery
fat is usually not the surgical target, this fat shift may seem not to be a problem, but in the image, the bright fat blob can obliterate the water-based tissue, which is actually there next to it and may be a surgical target. For this reason, special sequences, called “fat-suppression” sequences, are sometimes employed to reduce the signal of the fat. It’s still in the wrong place, but its dimmer. There is another problem caused by the fat shift. The problem is that oilfilled fiducial markers have been used in some IGS applications in the past and even today show up occasionally. Their images will be shifted, and when they are used to register MR images to other MR images, to CT images, or directly to the patient anatomy in the operating room, their shift translates to an opposite shift of every part of the (water) image relative to the fiducials. For this reason, oil should never be used in MRI fiducials. Functional Imaging So far we have focused on standard MRI protocols, which are formulated to depict static anatomical shape and tissue type. However, there are additional protocols that highlight underlying temporal changes that are taking place within the anatomy while the patient is being imaged. Because the changes are part of the physiologic function of the anatomy, the result is called a “functional image”, and the process is called “functional imaging” or “functional MRI”, which is abbreviated fMRI. We describe two of the most popular protocols for fMRI — BOLD imaging and diffusion-weighted imaging. BOLD Imaging. When the brain pro-
cesses stimuli or generates nerve impulses from within, local metabolic
processes must produce the energy to drive the brain’s activity. These processes affect the level of oxygenation of the blood, and the level of oxygenation affects the susceptibility of the blood to magnetization by the applied static field. This local magnetization causes a temporary additional static-field inhomogeneity in nearby tissue. As we pointed out above in the section called “Intensity Distortion”, static-field inhomogeneity causes a reduction in the intensity of signal and varies spatially with the severity of the inhomogeneity. In that section, we described this effect as a distortion, but, when the change in intensity is correlated with brain activity, what was a flaw is transformed into an asset. This effect, which is called “blood oxygen level-dependent” (BOLD) contrast, was first observed by Ogawa and colleagues in 1990 in rats with 7 T imaging,12 but it is commonly used today at 3 T on humans. BOLD imaging is performed by acquiring images without spin-echoes (which would cancel the intensity reduction), namely, gradient-echo volume images of the brain, both during brain activity and during a quiescent state. The images are then registered to correct for any motion of the head between acquisitions, and then one image is subtracted from the other. Regions in which the BOLD effect has reduced the intensity are thus highlighted in the resulting difference image. Diffusion-Weighted Imaging (DWI).
BOLD imaging takes advantage of the dephasing that occurs in gradient-echo images, and only gradient-echo images are used for BOLD imaging because spin-echo images do not exhibit BOLD contrast. The BOLD effect is absent in spin-echo imaging because, as we
Chapter 2
pointed out above in the section entitled “Intensity Distortion”, dephasing is removed by means of the 180-degree refocusing pulse. However, as we also pointed out earlier in the section entitled “Spin Echoes”, this refocusing works only for those nuclei that are stationary or move only a negligible distance during the pulse sequence. Diffusion-weighted imaging (DWI) employs a spin-echo sequence in which a 180-degree pulse is employed to eliminate the dephasing effect for stationary nuclei (including those that are participating in BOLD effects) while allowing dephasing for moving nuclei — in particular for those nuclei that are moving because of diffusion. Diffusion occurs naturally to some degree for all water molecules at all points in the body at all times, and the hydrogen nuclei in them ride along inside. Diffusion is simply the random migration of molecules from one spatial point to another that results from their constant thermal motion. It takes place continuously in all liquids (and gases), but ordinarily the effect is manifested only when molecules of one type that are initially spatially segregated migrate from their initial positions and, as a result, mix with molecules of different types, like ink squirted into a bottle of water that diffuses into previously clear regions, which then become clouded. In MRI, diffusion is manifested when nuclei with one phase that are initially spatially segregated migrate from their initial positions and, as a result, mix with nuclei of other phases. The resulting dephasing results, as it always does, in a reduction of signal. In 1985, Dennis Le Bihan developed a method to enhance this effect in MRI and produced the first DWI images.13 The effect of diffusion is enhanced in
ct and mri
n
this method by imposing a very strong gradient for a finite time both before and after a 180-degree refocusing pulse. The strong gradient is imposed so that nuclei do not have to diffuse very far to find themselves in a region where other nuclei have significantly different phases. Those nuclei that are inside approximately stationary molecules are rephased, while the phases of those nuclei that have diffused in a direction along the gradient are not restored to their original phase by the 180-degree pulse. Thus, the dephasing due to diffusion remains, and so does the reduced intensity that it causes. The effect is proportional to the level of diffusion in the direction of the gradient. By repeating the process with gradients in other directions, the full threedimensional diffusion pattern can be determined. Post-processing can produce color-coded images to show the spatial variation in the level of diffusion or the direction of greatest diffusion. Diffusion-weighted images are often employed in the evaluation of acute ischemic stroke, after which affected regions, in which diffusion pathways have been disrupted, have higher intensity than surrounding regions. It is also used to evaluate white-matter disease and for differential diagnosis between brain abscess and cystic tumor.14 Trade-Offs As we have seen, there is a myriad of variations in acquisition protocols and parameter settings in MRI, and every one of them represents a compromise involving time in the scanner, resolution, field of view, tissue contrast, geometric distortion, and signal-to-noise ratios. These compromises represent trade-offs that can affect diagnostic
69
70
Image-Guided Surgery
accuracy, surgical success, and patient comfort. An ideal image will always be elusive, but until a better modality supplants MRI, these trade-offs will require a negotiation between the physician ordering the scan and the technician acquiring the scan, subject to the limits imposed by T1 and T2 for the tissues being imaged and the immutable laws of physics. .
Why Is Understanding Imaging Important in Using IGS? Because IGS is dependent upon linking an image to a patient’s intraoperative anatomy, IGS can only be as accurate as the image. As a result, there are two main points that we would like to leave you with from this chapter: (1) images have inaccuracies, and (2) images are typically presented as 2D representations of a 3D structure and have to be interpreted with this in mind. Although (1) may be obvious, (2) may not be and deserves further attention.
Inaccuracy in Images The images that are produced by CT and MRI are inherently digital. They are nothing more than sets of calculated numbers, representing voxel values stored in computer memory and displayed on demand as 2D arrays of intensity, as shown, for example, by Figure 2–4. Each number gives us the only information available in the image about one voxel. The critical implication of this limited information for IGS is that one can never know the exact location of a structure. Take, for example, the image of a spherical marker shown schematically in Figure 2–27. This fig-
Figure 2–27. Shown is a 2D depiction of a spherical marker as would be obtained in CT scanning. Pixelation of the image makes it impossible to precisely localize the exact center of the circle, but FLE (arrow) is smaller for larger markers. Voxel a is brighter than voxel b because it overlaps the marker more.
ure is a 2D depiction of a 3D marker, but the problem is the same. We want to know the position of the center of the sphere so that it can serve as a reliable point of reference, which we will call a “fiducial” point, where “fiducial” means “trustworthy”. (Fiducial points are treated in detail in Chapter 4.) But, given that the image is pixelated (voxelated in 3D), where is that center? We can estimate the set of voxels that overlap the sphere and then calculate the mean of the centers of those voxels. We can weight each voxel according to its intensity, we can fit a model of the sphere to the intensity pattern,15 and we can reduce error by making the marker larger so it encompasses more roxels, but regardless of the approach, the determination of any fiducial point will be fraught with error. This voxelation error is compounded by noise that erroneously adds to, or subtracts
Chapter 2
from, the image intensity and, for MRI, by geometrical distortion. It should be clear from our earlier sections on CT and MRI that error can never be eliminated during imaging with either of these modalities. The impact of this error on IGS will be examined in much more detail in Chapter 5, where we show that IGS systems rely on the identification of fiducial points to register the image in CT or MRI to the patient in the operating room. Thus, regardless of the magnitude of other errors that may be present in the system, error in locating these points will inevitably lead to error in IGS intervention. This error in identifying fiducial points in the image is technically known as “fiducial localization error” in image space and is designated FLEimage. The identification of a fiducial point in an image, no matter how accurate it is, does no good unless a corresponding identification is performed in the operating room. Fiducial localization error in the operating room is known as FLEphysical. As pointed out in Chapter 1, the key to IGS intervention is to align these points by means of registration (Chapter 4), and the impact of FLEimage and FLEphysical on registration and on the total error of the IGS system is covered in Chapter 5.
2D Presentation of 3D Images Perhaps the biggest problem with images is that, because the human visual system is inherently 2D, the IGS system must display 2D representations of the 3D images. Through practice, surgeons become very good at assimilating 2D slices into 3D volumes within their minds by paging through the slices in the set. Although this skill is a testament to the remarkable abilities of
ct and mri
n
the human brain, most surgeons feel uncomfortable with this task when the 2D projections differ from the traditional axial, coronal, and sagittal sections that they are accustomed to viewing, further attesting to the difficulty of the task of inferring a 3D pattern from a set of 2D pictures. From an IGS standpoint, issues arise when an object is not entirely “in-slice” (ie, not completely encompassed in the one 2D projection) because a surgeon’s appreciation of the geometric relations for two objects out of plane is, at best, limited. Clinically, this problem can be appreciated by comparing the difficulty of diagnosing superior semicircular canal dehiscence (SSCD) on traditional axial, coronal, and sagittal cuts in which that the superior semicircular canal is not “in-slice”, to the ease of the same diagnosis when the images are reformatted into the plane of the semicircular canal (Figure 2–28). Further attesting to how difficult this task is the fact that SSCD was not even identified as a clinical entity until 199816 — 25 years after the first clinical CT scanners became available! Regarding IGS, the difficulty with inferring 3D relationships from 2D images means that subjectively verifying anatomy is much easier than precisely navigating to anatomy. Suppose, for example, that a surgeon knows the coordinates (x1, y1, z1) of the current location of a surgical pointer or tool and wishes to move to a desired anatomical point (x2, y2, z2). The surgeon could first move in the xy plane by the diagonal displacement, x2 − x1, y2 − y1, and then move along the z axis by z2 − z1, but the orientation of the individual axes relative to patient anatomy in the operating room is not clear to the surgeon (which way is z?).
71
72
Image-Guided Surgery
A
B Figure 2–28. 2D representation of 3D objects can lead to difficulties interpreting geometric relationships when an object is not completely “in-plane” as demonstrated by this false-positive diagnosis of a superior semicircular canal dehiscence (yellow arrows) in the coronal (A) and sagittal (B) views. In the parasagittal view (C), in which the entire semicircular canal can be seen, intact bone over the top is confirmed, thus ruling out the diagnosis.
C
Similarly, the surgeon could move in one motion to the desired point if given that the distance is √ (x2 − x1)2 + (y2 − y1)2 + (z2 − z1)2 , but, since the direction of the axes relative to patient anatomy in the operating room is unclear, the direction of the desired motion in the operating room is unclear. What is lacking is an anatomical orientation analogous to a true north direction on a map as will be discussed in further detail in Chapter 8. Certain IGS systems try to overcome this display mismatch using “line of sight” views, which orient the anatomy along the line from point 1 to point 2, but this direct routing (“as the crow flies”) assumes that one can move in the desired direction unimpeded by
intervening tissue. Other systems utilize 3D renderings of the anatomical field as will be presented in Chapter 8 (Figures 8–6 and 8–7). Renderings are helpful because surface depictions are easily processed by our brains, which evolved without x-ray vision and therefore survived by developing internal algorithms for inferring the 3D shape of a shaded 2D surface. However, surface rendering for IGS is challenging because it depends upon image-processing algorithms to predict continuity of structures (eg, the facial nerve) and hierarchies of display (eg, Is the surgeon interested in the surface of the skull or the surface of the brain? Which of these surfaces should be rendered and how transparent should they be made?).
Chapter 2
As amazingly useful as anatomical imaging is, a disconnect remains between the rich 3D information present in a tomographic image and the conveyance of that 3D information to the user. Computer systems are not constrained by this disconnect because they can “see” in any number of dimensions. Because of this ability to process large volumes of information, the maximization of the power of IGS will occur only when the computers that were trusted to generate these 3D images are also trusted to direct precise anatomical navigation via the images. Such delegation of control is contrary to the surgical mind-set, but it likely represents the next quantum leap forward in surgical interventions, as discussed in more detail in Chapter 8.
ct and mri
n
Guided Procedures, Robotic Interventions, and Modeling. In press. 5. Jackson EF, Bronskill MJ, Drost DJ, et al. Acceptance Testing and Quality Assurance Procedures for Magnetic Resonance Imaging Facilities. College Park, MD: American Association of Physicists in Medicine; 2010. 6. Mattson J, Simon M. The Pioneers of NMR and Magnetic Resonance in Medicine: The Story of MRI. Jericho, NY: BarIlan University Press and Dean Books Co; 1996. 7. Haase A, Frahm J, Matthaei D, Hanicke W, Merboldt KD. FLASH imaging: rapid NMR imaging using low flipangle pulses. J Magn Reson. 1986;67(2): 258–266. 8. Pruessmann KP, Weiger M, Scheidegger MB, Boesiger P. SENSE: sensitivity encoding for fast MRI. Magn Reson Med. 1999;42:952–962. 9. Larkman DJ, Nunes RG. Parallel magnetic resonance imaging. Phys Med Biol. 2007;52:R15–R55. References 10. Roberts DC, Marcelli V, Gillen JS, Carey JP, Della Santina CC, Zee DS. MRI mag 1. Labadie RF, Balachandran, R, Noble JH, netic field stimulates rotational senet al. Minimally-invasive, image-guided sors of the brain. Curr Biol. 2011;21(19): cochlear implantation surgery: first 1635–1640. report of clinical implementation. Laryn- 11. Walker A, Liney G, Metcalfe P, Hollogoscope. 2014 Aug;124(8):1915–1922. way L. MRI distortion: considerations 2. Siewerdsen JH. Cone-beam CT with a for MRI based radiotherapy treatment flat-panel detector: from image science planning. Australas Phys Eng Sci Med. to image-guided surgery. Nucl Instrum 2014;37(1):103–113. Methods Phys Res A. 2011;648(suppl 1): 12. Ogawa S, Lee TM, Kay AR, Tank DW. S241–S205. Brain magnetic resonance imaging with 3. Balachandran R, Schurzig D, Fitzpatrick contrast dependent on blood oxygenJM, Labadie RF. Evaluation of portable ation. Proc Natl Acad Sci USA. 1990;87(24): CT scanners for otologic image-guided 9868–9872. surgery. Int J Comput Assist Radiol Surg. 13. Bihan DL, Breton E. Imagerie de diffu2012;7(2):315–321. sion in-vivo par résonance. CR Acad Sci 4. Dillon NP, Siebold, MA, Mitchell JE, Paris. 1985;301(15):1109–1112. Fitzpatrick JM, Webster RJ. Increasing 14. Mascalchi M, Filippi M, Floris R, Fonda safety of a bone-attached robotic system C, Gasparotti R, Villari N. Diffusionfor inner ear surgery using probabilisweighted MR of the brain: methodoltic error modeling near vital anatomy. ogy and clinical application. Radiol Med. Proc. SPIE Medical Imaging 2016: Image2005;109(3):155–197.
73
74
Image-Guided Surgery
15. Fitzpatrick JM, Hill DLG, Maurer CR Jr. Image Registration. In: Sonka M, Fitzpatrick JM, eds. Handbook of Medical Imaging: Medical Image Processing and Analysis. Vol 2. Bellingham, WA: SPIE; 2000:337–513. 16. Minor LB, Solomon D, Zinreich JS, Zee DS. Sound- and/or pressure-induced
vertigo due to bone dehiscence of the superior semicircular canal. Arch Otolaryngol Head Neck Surg. 1998;124(3): 249–258. 17. Hornak JP. The Basics of MRI. http:// www.cis.rit.edu/htbooks/mri/inside .htm.
3 Tracking Systems is the most expensive component, often costing many thousands of dollars in comparison to the screen display and computer, which can be purchased for hundreds of dollars each. As noted in the prior chapter, early solutions came in the form of articulated arms with joints precisely encoded to allow knowledge of the tip of a pointer relative to its base. Although these arms achieve remarkable accuracy, on the order of hundredths of millimeters,i the inconvenience of the physical linkage combined with the availability of other solutions has led to the point that no clinical systems exist that use mechanical arms. At present, the two most widely used tracking systems in ENT IGS are optical tracking, which utilizes either visible light or infrared radiation, and electromagnetic (EM) tracking, which utilizes a magnetic field produced by electromagnets. Although optical tracking is more accurate, it suffers from the need for line-of-sight (LOS) in which the tracking system “sees” the object
Overview To achieve IGS, one must know where objects are in space. These objects include fiducial markers (Chapter 4) that are used to identify the location of the patient in the operating room and, subsequently, after registration (Chapter 4), are used to navigate using a tracked surgical tool. Tracking, although highly accurate, can never be done perfectly, and there is always associated error whose magnitude depends on many factors including the type of tracking used, the relative location of the tracked object to the tracker, and even the temperature within the operating room. As a result, it is critical that surgeons understand how to minimize tracking error for given applications and, more importantly, to understand when tracking error will be clinically significant. Given its critical importance in registration as well as active navigation, it should come as no surprise that the tracking component of an IGS system i
The FARO Gauge, for example, reports an accuracy of 0.018 mm (http://www.faro.com/en-us/ products/metrology/measuring-arm-faro-gage/overview. Accessed February 5, 2015).
75
76
Image-Guided Surgery
being tracked. Anything that blocks this LOS temporarily disables IGS. The lower accuracy of EM tracking can be somewhat overcome by placing the electromagnets closer to the patient, but a downside is that metallic objects within the area can disrupt the field, thus making tracking erratic. IGS manufacturers now offer both optical tracking and EM tracking in the same unit to allow the flexibility afforded by EM tracking when needed and the accuracy of optical tracking when LOS is available. This flexibility, of course, raises the cost of the system. A novel, inexpensive tracking approach that utilizes digital cameras tracking geometric shapes may hold out the prospect of very inexpensive IGS in the not too distant future as will be presented below when discussing the ClaroNav system. Regardless of the tracking system used, the protocol for actual use is remarkably similar. The tracking system is turned on and allowed to warm up because certain critical components are temperature dependent. The tracking system reports the position of a trackable tool, or “probe”, which is typically moved by the surgeon on or inside the anatomy. A trackable tool is either a pointer or a surgical device (eg, drill, endoscope, or surgical debrider) that is outfitted with a rigidly attached trackable component. The trackable component comprises a rigid configuration of fiducial markers or sensors whose positions in space can be determined continually (typically 30–60 times per second) by the tracking system. For electromagnetic tracking, the trackable component is ideally located near or at the “tool tip”, which is the distal end of the tool, which makes contact with anatomy and may even enter it, but for optical track-
ing, in order to remain visible to the tracking system, it is necessary that the trackable component be located on the tool distally to the surgeon’s hand holding the tool, which puts that component typically 8 to 10 inches away from the tool tip (Figure 3–1). The optical tracking system can “see” the trackable component and can determine its pose (location and orientation), but it is blind to the whereabouts of the tool tip. Instead of tracking the tip directly, it must infer each instantaneous position of the tip by combining the position and orientation of the trackable component with knowledge of the geometric relationship between the trackable component and the tool tip. The determination of this relationship is called “tool calibration”. Some tools arrive precalibrated by the manufacturer, but all other tools must be calibrated by the user. These latter tools include every tool that is assembled by clamping a trackable component to a device that lacks a built-in one. Tool calibration is in some special cases accomplished by clamping the tip into a precalibrated and tracked holder, so that the tracking system can simultaneously track the tool and the holder for a few seconds to determine the relationship between the trackable component of the tool and its tip. However, in the vast majority of cases, calibration is accomplished by “pivoting”. Pivoting is accomplished by placing the tool tip into a divot on a stationary object and tilting the tool in multiple directions while maintaining the tip stationary in the divot. The tracking system compares the detected pose of the trackable component at two or more instants during pivoting and, by performing geometrical calculations,
Chapter 3
Tracking Systems
n
? trackable component co
? ? ?
? ? ? ?
??
?
? tool p
? ?
A
B
? ?
C
trackable component
? ? ? D
tool p
E
Figure 3–1. Probe during calibration by pivoting. Red = first pose. Green = second pose. Yellow = third pose. A–C: Pivoting for Fiducial Configuration 1. D–E: Pivoting for Configuration 2. A and D each show the first pose (trackable component shown as red). B and E each show the 1st and 2nd pose (red and green). C shows the 1st and 3rd pose for Configuration 1 (red and yellow). Two poses are sufficient for Configuration 2 because it is assumed that the centroids of the markers lie on a line passing through the probe tip. Three poses are required for Configuration 1 because the position of the tip relative to the markers is assumed to be unknown (bent probe). In B, the green line shows the locus of possible tip positions for the pivot from 1st to 2nd pose. Dashed lines show possible (but incorrect) probe positions. In C, the yellow line shows the locus of possible tip positions for the pivot from 1st to 3rd pose. The true tip position is located at the intersection of the green and yellow lines.
determines the stationary point on the tool. That point is the tool tip. There are some minor constraints on the motion during pivoting that are determined by the configuration of fiducial markers on the trackable component. There are two common configurations — Fiducial Configuration (1), which comprises three or more markers not lying on the same straight line, and Fiducial Configuration (2), which comprises two or more markers lying on a straight line whose extension passes through the tool tip.
Fiducial Configuration (1) is more common and generalizable, and Figure 3–1 A–C depicts it. Before the tool is tilted (red), even though the tracking system can determine the pose of the trackable component, it is impossible for it to infer any information at all from that pose about the tip location. It can be anywhere in space, as indicated by the question marks sprinkled through space in Figure 3–1A. However, after detection of a second pose (green) produced by pivoting, as shown in Figure 3–1B, followed by a mathematical
77
78
Image-Guided Surgery
comparison of the two poses, the set of possible tip locations is reduced to a straight line, as indicated by the questions marks along the (green) line. That line passes through the pivot point, but the position of the pivot point on that line is still unknown. If the user’s second pivot moves the tool to a pose (yellow) away from the plane of the first pivot, as shown in Figure 3–1C, then the comparison of this third pose to the first pose will produce a second line (yellow), which intersects the first. Similarly, the line produced by comparing poses 2 and 3 (not shown in the figure) will intersect the line produced by comparing poses 1 and 2 at a single point. In either case, intersection of the lines is the location of the pivot point, which is also the position of the tool tip. Fiducial Configuration (2) requires only two poses because the tip is known to be co-linear with the marker centroids. Pose 1 produces a line of possible tip locations (question marks) as does Pose 2, and the intersection of these lines determines the tip location. The reader will realize that during a single pivot many intermediate poses can be detected by the tracking system. This extra information would seem to be redundant, given the analysis above, but because of small errors in the detection of the positions of the fiducial markers on the trackable component, a mathematical combination of the minimum number of poses is a suboptimal approach. Instead, the combination of many “redundant” detected poses during pivoting will reduce the error of the determination of the position of the tip. Thus, the redundant poses are in fact not redundant at all, and a well-engineered tracking system will use hundreds of poses obtained while
the user is pivoting the tool in order to reduce that error as much as possible during tool calibration. At this point, the tool is ready to be tracked within the specified field of view (FOV) of the system. However, tracking is not without error, and this error is dependent on the type of tracking used. Although we may be getting ahead of ourselves, the error of the tracking system plays a central role in the total error of the IGS system. We will learn in Chapter 5 that tracking error causes misidentification of the position of fiducials, which are used to align the patient in the operating room with his or her preoperative images. As pointed out in Chapter 2, this misidentification is known as fiducial localization error (FLE), and it and a similar error in image space cause misalignment of the images of the anatomy displayed in the operating room with the anatomy itself. Then, when the surgeon directs a probe to an anatomical point of interest, the tracking error compounds the image alignment error to produce the total IGS error, which causes a misidentification of the positions of surgical targets in the operating room. And, the size of the total error relative to a targeted region determines whether the IGS system is a useful adjunct to the surgery or a misleading detriment.
Optical Tracking Optical tracking utilizes the principle of triangulation to localize objects in space. Triangulation dates back millen-
Chapter 3
nia when it was used to estimate nautical distances. The underlying principle has not changed and is shown in Figure 3–2 in a 2D example. The distance or depth, z, from the tracking system to the unknown object, the star, is sought, and the distance between the cameras (d) and focal length (f) is known. By measuring where the image of the
Tracking Systems
n
unknown object is sensed on the camera lenses (CL and CR) and using the geometric principle of similar triangles, the unknown distance can be calculated as described in the caption of Figure 3–2. For the general 3D case, the other two dimensions, x, horizontal position, and y, height, can be found by knowing z,
Figure 3–2. Triangulation in 2 dimensions. The distance, d, between two cameras (lenses shown as gray half-donuts) with known focal lengths, f, is known, and what is desired is z, the perpendicular distance from the line joining the cameras to the point of interest, in this example the centroid of a star. d, being bisected by z, can be broken down into dR and d L. The star projects onto the left camera at cL and on the right camera at −cR, both of which are measured. (Note: moving to the left is considered the positive direction because the orientation is taken as if the user is standing at the star looking at the cameras.) Using similar triangles, we have z/dL = f/cL and z/dR = −f/cR. Rearranging these and substituting into d = dR + dL, we get d = (cL − cR)z/f or z = fd/(cL − cR).
79
80
Image-Guided Surgery
depth, and the projection in the x and y directions on either camera. When we use the term “optical” in referring to tracking, we mean it to include both visible and infrared light.ii Although visible light can be used, the task is made simpler by using infrared
Although IGS has been compared to the global positioning system (GPS), the underlying principles regarding localization are quite different. GPS utilizes a series of at least four satellites to determine the location of a receiver, which receives signals from each satellite. Those signals include the times the signals were sent and the positions of the satellites when they transmitted them. If signals are received from at least four satellites, it is possible to determine the position of the receiver, information that is unavailable to the IGS system. By dividing the time difference between reception of each signal and its transmission time by the speed of travel of the signal (the speed of light), the distance between the receiver and each satellite can be quickly calculated. Yet, because the wavelength of the signal (about 20 cm) is much greater than the typical receiving antenna (1 cm), it is not possible to detect the direction from which a given signal arrives. This situation is identical to that of the MR scanner, which is unable to determine the direction from which NMR radiation comes because its wavelength is too ii
(IR) radiation because (a) filters can be used to remove other, unwanted wavelengths, and (b) signals can be used that are much brighter than any other IR source in the OR (most everything else in the OR, including the patient and surgical team, emits IR as well!) and
large (Chapters 1 and 2). As with the MR scanner, the GPS detector must resort to an indirect calculation in which many signals are compared to determine a position. With IGS, on the other hand, although the distance to the signal source is not known, the direction of each received signal is known. This directional information makes it possible to determine the receiver’s position from just two or three transmitters (usually infrared or visible light sources), while a GPS receiver requires a minimum of four and typically uses five or six satellites’ signals to determine its position. Because there is no directional information from a given satellite, the receiver can determine only that it sits on a sphere, the center of which is the satellite and the radius of which is the calculated distance. Intersecting two of these spheres creates a circle on which the receiver deduces that it must be located. Adding a third sphere limits its possible positions to two points, and adding the fourth sphere specifies which of the remaining two is the receiver’s location. Utilizing additional satellite signals increases the accuracy of the calculated position.
A more precise term might be “refractable-beam” tracking, meaning that the radiation can be bent and focused via refraction through passive lenses, but the literature uses the term “optical” for such radiation.
Chapter 3
Tracking Systems
n
Figure 3–3. Infrared (IR) tracking systems utilize two cameras situated a known distance d apart (same as in Figure 3–1). Concentrically arranged around the IR cameras are IR emitters shown as the gray donuts. The IR light (gray arrows) reflects off a reflective sphere (dark gray sphere), also known as a passive marker, and those reflections (black arrows) are detected by each IR camera.
yet neither harmful nor even visible to the human eye. These two features enable the use of simple thresholding to detect only the IR signal of interest (Figure 3–3). Optical tracking can utilize either “active” tools, which generate an IR signal, or “passive” tools, which merely reflect IR that is emitted from another source — usually a source that is rigidly fixed close to the sensor. Passive IR tracking with disposable reflective spheres is used in most IGS systems because of convenience and low cost, and it is a common misconception that passive IR is inherently less accurate than active IR. Although in practice a lower accuracy may be observed with a passive system, it is likely due simply to poor handling of the reflective IR markers, which results in their being physically marred, deformed, or dirtied (Figure 3–4). In controlled experiments, no statistically significant difference can be demonstrated
Figure 3–4. Image of a passive marker as detected by an infrared (IR) camera. Although no statistically significant difference has been shown between passive and active IR tracking, in practice, smudging of passive IR markers, as shown in this figure (upper right ), may make them less accurate. This inaccuracy occurs because the calculated centroid of the image of the smudged marker (circled x ) is not the true centroid (gray diamond ). Image provided courtesy of ClaroNav, Inc.
81
82
Image-Guided Surgery
between active and passive IR when the reflective markers are properly handled.1 Optical tracking systems typically update at approximately 20 to 60 times per second, more technically known as an update rate of 20 to 60 Hz. A key component of triangulation (see Figure 3–2) is a fixed and known relationship between some set of points that are part of the measuring system. In optical tracking, the cameras are located at those points, and the distance and angulation among them must be very accurately known. Because temperature changes cause expansion and contraction of the metal used to hold the cameras in a fixed relationship to each other, optical tracking exhibits temperature-dependent accuracy, so most systems are designed to work optimally only after “warming up” to a stable operating temperature. Another important consideration is that tracking error varies with location within the FOV of the tracker. Variation in tracking error occurs because triangulation becomes less accurate as the angles between beams arriving from an unknown loca-
tion to the camera become smaller (ie, as the angle between the lines extending from the star in Figure 3–2 approach 0°). Manufacturers of optical tracking systems provide the manufacturers of IGS systems with detailed specification of the FOV as well as statistics on the error of tracking within that FOV. It is important for the IGS user to know the bounds of the FOV because error typically increases, sometimes dramatically, as the probe moves beyond that boundary. Unfortunately, while some manufacturers provide methods for centering the FOV on the patient, the dimensions of the FOV itself — if available — are explicitly specified only in the technical manuals. The leading manufacturer of optical tracking systems is NDI (Northern Digital, Inc, Waterloo, Ontario, Canada), whose systems are often incorporated into commercial IGS machines. NDI offers two medical models, the Polaris Spectra and the Polaris Vicra (Figure 3–5). The larger Spectra has more widely spread cameras (approximately 60 cm) and a larger FOV (Figure 3–6)
Figure 3–5. The NDI Polaris Spectra (in back) and Vicra (in front ) are the industry standards for infrared (IR) tracking and are used in both Brainlab’s and Medtronic’s image-guidance systems. The larger Spectra has an intercamera distance of 60 cm compared to the Vicra’s 25 cm. This difference allows the Spectra to have a larger field of view (FOV), or working volume, as shown in Figure 3–6. Image provided courtesy of Northern Digital, Inc.
A Figure 3–6. Working volume, also known as field-of-view (FOV) or pyramid, of NDI’s larger Spectra (A) and smaller Vicra (B). It is important for surgeons to ensure that the pertinent anatomy is within the FOV to maximize tracking accuracy. This includes knowing both the distance from the camera system to the beginning of the FOV (950 mm for the Spectra and 557 mm for the Vicra) and the end of the FOV (2.4 m for the Spectra and 1.3 m for the Vicra). Images provided courtesy of Northern Digital, Inc. continues
83
B Figure 3–6. continued
84
Chapter 3
while the Vicra has a smaller spread between cameras (approximately 25 cm) and a smaller FOV. The error commonly measured in the FOV is the distance of the measured position from the true position: d = √ ∆x2 + ∆y2 + ∆z2, where ∆x, ∆y, ∆z are the displacements from the true postion along the x, y, and z axes fixed in the camera. Distance, which is simply the magnitude of the error, is the usual measure, instead of the individual components of error in the x, y, z directions, because the direction of the error relative to the camera is typically not as important as the size of the error. To assess the accuracy within the camera’s FOV, many such error measurements must be made at many positions. It can be expected that the error components, Δx, Δy, and Δz, which can have positive or negative values will be normally distributed, and their means can be expected to be zero. (This mean of zero is artificially enforced by the manufacturers during calibration by adding constant offsets in x, y, and z.) The mean of the distances, however, will always be non-zero (unless the camera is perfect!) because distances are by definition non-negative. Therefore, the distribution of distance error, d, cannot be normal because a normal distribution always includes both positive and negative numbers. Instead, the distance error, d, exhibits a Maxwell distribution,iii an example of which is shown in Figure 3–7, which unlike a normal distribution, has no negative values. While it is necessary to provide both a mean and a standard deviation to specify a normal distribution, surprisingly only one parameter is required to iii
Tracking Systems
n
specify the Maxwell distribution, and the most common parameter used is the root mean square (RMS). To calculate RMS, the measurements are squared and then averaged, and then the square root of the mean of the squares is taken, as shown in the following equation:
The RMS of error is typically specified instead of the more familiar mean because it exhibits simpler relationships to other quantities, as we will see in Chapter 5, but, if the mean of a Maxwell distribution is needed, it can be easily calculated from the RMS as follows: mean = √ 8/(3p) × RMS ≈ 0.921 × RMS. Furthermore, the standard deviation = √ 1 − 8/(3p) × RMS ≈ 0.389 × RMS or 0.422 × mean. So, what can we make of the RMS error specified for a tracking camera in the OR? First, it is important to note what it does not mean. It does not mean that the errors will be limited to that specified value. Instead, it means that they follows a statistical distribution like that of Figure 3–7. If RMS = 1.0 mm, then the distribution will be the same as that of the figure. The mean ± standard deviation will be 0.921 ± 0.389 mm, and there is a 5% chance that the error will be greater than 1.61. If RMS is larger than 1.0 mm, then the distribution along with its 5% line and every other percentile line will be uniformly stretched to the right away from zero, and the mean and standard deviation will be proportionally larger. If RMS is smaller than 1.0 mm, then everything
The Maxwell distribution is also known as the “Maxwell-Boltzmann” distribution because, while Maxwell discovered it, Boltzmann showed that it had much broader application.
85
86
Image-Guided Surgery
Figure 3–7. Probability function for the magnitude of 3D tracking error. The 3D error is the length of an error vector whose x, y, and z components are each distributed normally and independently with mean = 0 and variance = 1/3. The magnitude of the resulting 3D error vector has an RMS value = 1 and a mean of 0.92 — neither of which are located at the peak (0.82). The 3D error magnitude is plotted horizontally, and its probability per unit error, which has the form of a Maxwell distribution, is plotted vertically. The plot is divided into two parts by a vertical line at 1.61. The dark part labeled “95%” and the white part labeled “5%” indicate that 95% of the time the 3D error will have a value in the range from 0 to 1.61, and 5% of the time, it will have a value greater than 1.61.
will be uniformly contracted to the left toward zero and the mean and standard deviation will be proportionally smaller. As an example, NDI specifies that both the Spectra and the Vicra will provide an accuracy for which the RMS distance error is no higher than 0.35 mm within their respective FOVs. With an RMS of 0.35 mm or less, the Maxwell distribution of errors for these devices will be a contracted version of Figure 3–7 with the 5% line shifted to 0.56 mm or less, and their mean ± stan-
dard deviation will be 0.32 ± 0.15 mm or less. A closer examination of the Maxwell distribution reveals further that for an RMS of 0.35 mm, 0.1% of measured values will exceed 0.82 mm, and 0.01% will exceed 0.93 mm. Thus, the knowledge that the distribution is Maxwellian with a given RMS gives us a lot of information. However, the Maxwell distribution does not give direct information about spatial variation of errors. They may tend to be greater in one region than
Chapter 3
another. Indeed the errors measured at the edges of the FOV tend to be greater than those measured near the middle. The hallmark of IGS error is that it can be characterized only statistically. For any region and for any system, it is important to remember that IGS error refuses to be confined to one value. It is always spread over a distribution. These statistics have been confirmed empirically for the Polaris by Wiles et al,1 who manually measured 1500 points within the FOV and reported an RMS of 0.255 mm with 5% > 0.451 mm and 0.1% > 0.7 mm. These two levels are slightly worse than that those predicted by the Maxwell distribution (RMS = 0.255 corresponds to 5% > 0.411 and 0.1% > 0.594). Note that each measured error involves the location of only a single marker, typically a sphere. To determine the orientation of a surgical tool, at least two markers must be tracked, and that is good news because using multiple markers reduces the tool-tracking error. In Chapter 5, we outline the method used to calculate the expected error at the tip of a tool to which multiple markers have been attached. In general, for a given RMS error of individual marker localization, the error distribution for tracking the tool tip can be contracted closer to zero (smaller RMS, mean, and standard deviation) by using more markers, by spreading them more widely, and by shortening the tool tip. Although NDI used to be the only game in town and continues to dominate the field, other companies (eg, Advanced Realtime Tracking GmbH, Weilheim, Germany; Atracsys, Pui doux, Switzerland) now offer optical tracking with similar, or better, accuracies.2 As noted above, the data can be
Tracking Systems
n
presented in various ways that make one tracking system appear more accurate than another based on the “application FOV”, which is the region over which the RMS is measured for a given application as opposed to the manufacturer’s recommend FOV (smaller application FOV typically translates to smaller RMS error for a given system with a given manufacturer’s FOV). In our opinion, the bottom line is that for IR optical trackers within reasonable application FOVs for IGS applications in ENT, RMS values should be on the order of 0.3 mm or less over an FOV that encompasses the targeted anatomy. Visible-Light Optical Tracking Visible-light optical tracking involves using the visible spectrum to visualize and locate markers using two digital cameras, typically video cameras, placed at known distances and orientations relative to each other. Unlike IR systems, which can incorporate extremely bright signals, making it relatively easy to isolate markers from background and accurately localize them, visible-light systems typically rely on ambient light, making it necessary to employ image-processing algorithms to isolate and localize markers. Computationally, this processing can be intensive because there is such a large amount of information coming from the visual field. An approach that makes this feasible is the use of geometric patterns imprinted on markers that are distinct enough to facilitate their isolation from the background and shaped so as to allow the determination of their location with both these operations being accomplished in a time short enough to allow accurate identification of fiducials
87
88
Image-Guided Surgery
attached to an instrument that is being moved through the field by a surgeon. Such visible-light systems are not new in industry (eg, the well-recognized geometric pattern used to track crashtest dummies) or medical applications,3 and they are available today. As of the time of this writing, ClaroNav (Toronto, Ontario, Canada) offers a visible-light optical tracking system for dental use, Navident, which is approved for use in Canada, and an ENT system, Navi-
ent, which is not yet clinically available. Both of these products utilize ClaroNav’s MicronTracker (Figure 3–8), which has been available for research use for over a decadeiv and consists of two or three digital cameras employing charged-coupled devices (CCDs) that track markers consisting of grids of alternating black and white rectangles that intersect in “Xpoints” (Figure 3–9). The appeal of such a tracking system lies in the low cost and the ubiquity of
Figure 3–8. The ClaroNav MicronTracker uses two or three digital cameras placed a fixed distance apart and ambient visible light to detect geometric patterns (Figure 3–9), which consist of intersecting black and white geometric patterns. The model shown has two cameras (the large circular openings). The small circular opening is an LED power indicator. The appeal of such a system is its low cost, small size, and field of view (FOV) (see Figure 3–10), which may be more appropriate for neurotology procedures because it may be mounted on an operating microscope. Image provided courtesy of ClaroNav, Inc. iv
The original producer of MicronTrakers, Claron Technologies (Toronto, Ontario, Canada), was acquired by Lexmark International, Inc (Lexington, Kentucky) in 2015. Lexmark incorporated Claron’s image-processing software into their product line, and a separate entity, ClaroNav, was created that continues to offer the MicronTracker, the Navident, and the Navient.
A
B
C
Figure 3–9. The ClaroNav MicronTracker tracks geometric patterns (A) consisting of intersecting white and black shapes, which create a so-called Xpoint that is detected by digital cameras, which, after image processing, identify the “Xpoint” (B), which is more resistant to smudging (gray area at upper right) than passive IR markers (see Figure 3–4). The markers are easily affixed to surgical tools (C), which are then calibrated by rotating the tip about a fixed divot (see Figure 3–1), allowing the tracking system to know where the end of the tool is located in relationship to the markers. Images provided courtesy of ClaroNav, Inc.
89
90
Image-Guided Surgery
the markers, which can be made using a home or office laser printer and waterproof paper (for steam sterilization). The Xpoint approach and the high contrast of these markers are designed to be more resistant to the problem of smudging (Figure 3–9), which for the passive markers used with infrared systems can cause a decrease in accuracy (Figure 3–4). Additionally, the FOV (Figure 3–10) and accuracy are quite promising for ENT applications, particularly neurotology, where a MicronTracker could be attached to an operating microscope, whose focal length would fall well within the tracker’s FOV. Accuracy with the MicronTracker is comparable to IR optical tracking systems4 with reported RMS values in the range from 0.20 to 0.35 mm depending on the application FOV and ClaroNav models involved, which vary based on number of CCDs, lens size, and distance between cameras.v
Electromagnetic (EM) Tracking The biggest downside to optical tracking is the need for an unobscured LOS between fiducials and sensors, which can be interrupted because of ergonomic issues within the OR (eg, the surgeon having to lean over to better access anatomy, thus blocking the optical LOS) or by the need to place instruments deep within surgical fields, as often happens in minimally invasive or endoscopic approaches. In these situations, optical tracking systems depend on markers placed on portions of the instrument that remain outside the patient’s body and on the rigidity of the instrument so that the geometric relationship between the tip hidden v
inside the patient’s body and the base, which the optical system can “see”, is not altered by the intervention. This requirement of rigidity is often not explicitly considered during minimally invasive or endoscopic approaches, but the reality is that if the instrument is bent — either permanently or temporarily — the tip will no longer be accurately tracked unless it is re-calibrated via a pivoting procedure (see Figure 3–1). Because of these issues, especially the desire to track intentionally deformable/flexible instruments, much emphasis has been placed on the development of EM tracking systems, for which there is no LOS requirement. The basic concept behind EM tracking is that (a) a magnetic field that varies in time and space bathes a portion of the surgical region, (b) a set of magnetic field sensors attached to surgical tools introduced into that region continuously measures the field at the point where the sensors are located, and (c) those same sensors continuously report their measurements either by wire or wirelessly to a system that infers the sensor’s position. Intuitively, this concept is easy to appreciate because magnetic fields have known flux patterns according to the laws of physics (Figure 3–11). But, turning this simple concept into useful technology is more involved. Use of EM tracking in interventional procedures is a very broad topic5 that includes various means of tracking, including placing passive EM transponders into tumor beds to guide radiation therapy (Calypso System; Varian Medical Systems, Inc, Palo Alto, California), placement of EM sensors on the tips of feeding tubes to guide placement (CORTRAK; Cardinal Health, Dublin, Ohio;
http://www.claronav.com/microntracker-specifications.php (Accessed February 11, 2015).
91
Figure 3–10. Four ClaroNav MicronTracker models are available that have different fields of view (FOVs), as shown in this figure. The “S” model was the first available and has a lower camera resolution of 640 × 480 compared to the H series, which has a camera resolution of at least 1024 × 768. The larger FOVs are achieved by placing the cameras farther apart or, in the case of the H3, adding a third camera. The FOV, especially the capability of this to be closer to the patient (15–40 cm) may be well suited for otologic cases because it may be mounted on an operating microscope, which is typically used at a focal length of 250 mm. Image provided courtesy of ClaroNav, Inc.
92
Image-Guided Surgery
varying fields, such as the field of a bar magnet shown in Figure 3–11 sets of electro-magnets designed to produce optimal magnetic fields for IGS over a given FOV. An example is a tetrahedron arrangement, Figure 3–12, which is designed to create magnetic gradients optimal for detection over a cubic FOV of approximately 0.5 m on each side.7 Magnetic fields are broadly classified as either static (eg, ferromagnetic objects such as the magnets that we stick on our refrigerators) or dynamic, which are created electromagnetically (ie, via varying flow of electricity Figure 3–11. A simple bar magnetic creates through a wire or coil of wire, creata magnetic flux field as shown. A probe (not ing a varying magnetic field). Because shown), containing a sensing electromagnet, these dynamic fields can be controlled, measures this magnetic flux, which is correlat- especially turned off when so desired ed to position. Not conveyed in this simplified (both not only when not in use but also illustration is that any ferromagnetic material that comes in close proximity to the magnetic for synchronizing detection), dynamic distorts the magnetic field, leading to error magnets are used in EM tracking. Furin the correlation of measured magnetic flux thermore, the EM tracker has differwith position. ent properties if the flow of electricity or current is alternating current (AC) http://www.corpakmedsystems.com/ or direct current. Direct current (DC), cortrak/ [retrieved May 30, 2015]), and which ironically is delivered in pulses coupling to flexible endoscopes to guide to generate EM fields for IGS systems diagnostic and therapeutic interven- (alternating between on and off), is very tions.6 Because we are focusing on IGS — susceptible to extraneous static magthat is, tracking and navigation of a sur- netic fields, including the Earth’s and gical tool within a surgical field — we will magnetic, predominantly iron-containing, materials in close proximity to the limit our discussion to this application. For EM tracking, there has to be (1) a EM field generator. This susceptibility magnetic field generator at a source loca- to other fields can be at least partially tion and (2) a magnetic field sensor, as overcome by delivering a pulse and well as (3) a computer to control gener- recording the baseline response, which ation of the magnetic field, process the is then subtracted from the tracking data to allow calculation of current loca- signal. Systems based on AC, although tion based on the signals from the sen- not affected by external static magnetic sor, and convey this information to the fields, induce eddy currents in nearby surgeon, typically through a visual display. conductive materials, which then genThe magnetic field generator can erate their own magnetic fields, which be as simple as one or more perma- interfere with the generated EM field. nent magnets, which produce spatially The pros and cons of AC and DC are
Chapter 3
Tracking Systems
n
A
Figure 3–12. Northern Digital, Inc’s (NDI’s) Aurora EMF generation unit consists of electromagnets arranged in a tetrahedron configuration to create a dynamic magnetic flux field over an approximately 0.5 × 0.5 × 0.5-m cubic FOV for IGS tracking. A is from US patent 6,625,563 issued Sep. 23, 2003 and B is the unit as commercially available but with cover removed. Image A from the US Patent Office. Image B provided courtesy of Northern Design, Inc.
B
debatable, but in practice it comes down to testing in realistic environments to determine which is best for a given application.
As noted above, there are various magnetic field sensors, but the smallest of these, and thus the most applicable for IGS, consist of three small coils of
93
94
Image-Guided Surgery
wire oriented so that their axes are all mutually perpendicular (aka orthogonal) and located within the tip of the surgical instrument (Figure 3–13). When these coils are in the presence of a temporally varying magnetic field, electrical currents are induced in them according to Faraday’s law of induction. These currents can be measured and correlated with position within the magnetic field, and the measurement process can be facilitated by using the control computer to cycle the field on and off and/or change its strength and/or orientation at specified times to both optimize the condition for the sensor and allow multiple sensors to be tracked. Update times for EM tracking systems are similar to that for optical systems. For medical EM tracking, the two big players in the field are Polhemus, Inc
(Colchester, Vermont) and NDI, which in 2013 acquired Ascension Technology Corp (Shelburne, Vermont). As in optical tracking, accuracy in EM tracking depends upon the application FOV, and it also depends on whether competing magnetic fields are present (eg, ferromagnetic material and/or current loops such as power cords or cell phones close to the FOV). NDI provides detailed information about its Aurora planar and tabletop models (Figure 3–14), including FOV (Figure 3–15). The planar system is designed to be mounted on a positioning arm so that it can be placed as close to the tissue of interest as possible, while the tabletop model is designed for placement near an operating-room table. Accuracy is better with the planar system than with the tabletop model but also depends
Figure 3–13. EM sensor consisting of three orthogonally arranged coils of wire. When situated within a changing magnetic field, electrical current is induced in the coils, which is proportional to the rate of change of the field strength. If the magnetic field parameters are known, the measured electrical current measured can be used to calculate geometric location of the sensor within the field. Image from the US patent 6,690,963. February 10, 2004.
Chapter 3
Tracking Systems
n
Figure 3–14. Northern Digital, Inc’s current commercial offering of EM tracking systems — the larger Aurora planar model (shown in the back ) and the smaller Aurora tabletop model (shown in the front ). The planar model is designed to be mounted on an adjustable arm, allowing it to be placed as close as possible to the pertinent anatomy. The tabletop model is designed to be placed in close proximity to an operating room table. Image provided courtesy of Northern Digital, Inc.
upon the number of combinations of independent EM field-sensor interactions referred to as degrees of freedom (DOF), with accuracy being better with increased DOF, but more DOF requiring more processing time (ie, the update rate gets worse). Accuracy of EM tracking is, in general, worse than that of optical tracking. Table 3–1 gives error statistics for the Aurora tabletop system when there are no magnetic objects in its vicinity. However, when used in various clinical settings, the RMS error of EM systems can increase dramatically. For example, when they are used near a CT scanner, RMS error can grow to 5 mm.8 Unfortunately, at present, there is no way to know in real time whether EM tracking is being distorted. Clinicians are encouraged to verify tracking with known anatomical points at regular intervals during the surgical intervention. Additionally, it is recommended that all potential interfering
magnetic sources (eg, electrical equipment) be moved as far away as possible from both the magnetic field generator and the tracked tools. Although such accuracy issues represent a serious shortcoming of EM tracking, one very exciting possibility afforded by EM tracking is that it can be used to track flexible objects that are outside the surgeon’s view (eg, flexible endoscopes).6 In our opinion, the bottom line is that for EM tracking for ENT procedures, real-world operating condition RMS tracking error is on the order of 1.5 mm. As will then be obvious, it is our recommendation that when accuracy is of utmost importance, optical tracking should be used instead of EM tracking given its superior RMS tracking error of approximately 0.3 mm. Interestingly, most IGS manufacturers are now incorporating both OPT and EM tracking into the same platform so that both are available during surgical interventions.
95
A
B Figure 3–15. Working volume, also known as field of view (FOV), of NDI’s tabletop Aurora (A.) and planar Aurora (B.) (see Figure 3–14). Units mm. As with optical tracking systems (Figure 3–6), it is important for surgeons to ensure that the pertinent anatomy is within the FOV to maximize tracking accuracy. This includes knowing both the distance from the camera system to the beginning of the FOV (50 mm for the tabletop Aurora and 120 mm for the planar Aurora) and the extent of the FOV (0.5–0.66 m for the tabletop Aurora and 0.6 m for the planar Aurora). All numbers in figure represent distance in mm. Images provided courtesy of Northern Digital, Inc.
96
Chapter 3
Tracking Systems
n
Table 3–1. Optimal Error Statistics for Aurora Tabletop EM Trackers RMS
95% CI
Sensor
FOV*
0.70 mm
0 to 1.40 mm
5 DOF
Cube
1.10 mm
0 to 2.00 mm
5 DOF
Dome
0.48 mm
0 to 0.88 mm
6 DOF
Cube
0.70 mm
0 to 1.40 mm
6 DOF
Dome
Note. *See Figure 3–15. See also Figure 3–7 for an explanation of 95% CI (Confidence Interval), which is the range in which 95% of the errors are observed. Source: Compiled from data at http://www.ndigital.com/medical/products/ polaris-family/, accessed February 5, 2015.
Summary
References
Although often not considered in the discussion of error in IGS, tracking systems have errors associated with them. Optical systems tend to have better accuracy (0.3 mm or better) for most ENT IGS applications, while EM tracking, which is not restricted to LOS, tends to perform more poorly (≈1.5 mm in optimal conditions). If no other error existed in IGS systems, the error of the IGS system would be the error of the tracking system. As we will find in Chapter 5, the situation is much more complex because there are others sources of error that must be accounted for, including the error in determining the position of a fiducial marker in either CT or MRI, as was noted above and in Chapter 2. In the next chapter (Chapter 4), we will begin to appreciate the error imparted by the configuration of fiducial markers used to identify the patient in the operating room, and that will lead us to error analysis (Chapter 5).
1. Wiles AD, Thompson DG, Frantz DD. Accuracy assessment and interpretation for optical tracking systems. Proc SPIE 5367, Medical Imaging. 2004;Visualization, Image-Guided Procedures, and Display:421. 2. Elfring R, de la Fuente M, Radermacher K. Assessment of optical localizer accuracy for computer aided surgery systems. Comput Aided Surg. 2010;15(1–3): 1–12. 3. Colchester ACF, Zhao J, Holton-Tainter KS, et al. Development and preliminary evaluation of VISLAN, a surgical planning and guidance system using intraoperative video imaging. Med Image Anal. 1996;1(1):73–90. 4. Balachandran R, Fitzpatrick JM, Labadie RF. Accuracy of image-guided surgical systems at the lateral skull base as clinically assessed using bone-anchored hearing aid posts as surgical targets. Otol Neurotol. 2008;29(8):1050–1055. 5. Franz AM, Haidegger T, Birkfellner W, Cleary K, Peters TM, Maier-Hein L. Electromagnetic tracking in medicine
97
98
Image-Guided Surgery
— a review of technology, validation, and applications. IEEE Trans Med Imaging. 2014;33(8):1702–1725. 6. O’Donoghue K, Eustace D, Griffiths J, et al. Catheter position tracking system using planar magnetics and closed loop current control. IEEE Trans Magn. 2014;50(7):1–9.
7. Kirsch SR, Schild HR, Schilling CJ. Gain Factor and Position Determination System. Alexandria, VA: US patent 6625563. September 23, 2003. 8. Yaniv Z, Wilson E, Lindisch D, Cleary K. Electromagnetic tracking in the clinical environment. Med Phys. 2009;36(3): 876–892.
4 Registration From Chapter 2, we have a “map”— either CT or MRI — and we know that the map is imperfect. That is, it has distortion, so we can never truly know exactly where we are on the map, but we get pretty close (≈0.25 mm for CT and ≈1 mm for MRI). From Chapter 3, we have a means of “tracking” — either optical or EM — and it too is imperfect, so we can never truly know exactly where we are pointing, but we get pretty close (≈0.3 mm for optical and ≈1.5 mm for EM). These inevitable imperfections cause navigational error, and we investigate that error in Chapter 5, but first we need to learn how the map — the CT or MRI — is linked to the terrain — the patient’s anatomy in the OR. The goal of this map-to-terrain linkage is to determine, for every point of interest in the anatomy, the unique point in the map that corresponds to it. When this point-to-point correspondence is determined, the map is said to be “aligned” with the terrain. i
As we have pointed out in the previous chapters, in image-guided surgery the point-to-point correspondence is accomplished via a process known as registration. In registration a few clearly discernable features of the map, more technically called fiducials,i are identified and aligned with those same fiducials on the patient. Under most circumstances, if enough fiducials are aligned, then all features of interest are aligned as well. For consistency, we will call the CT or MRI “image space” and the patient in the operating room “physical space”. So, in image-guided surgery, registration involves using fiducials to align image space to physical space for the purpose of aligning all points of interest. The fiducials used for registration can be broadly classified as either points or surfaces. Although we make this artificial distinction, the reality is that a surface is represented by a collection of points, so under the hood of surface registration, there is always a point-
From the Latin fidere, meaning “to trust” (and root of the famous dog’s name, Fido, which in Latin means, “I trust”).
99
100
Image-Guided Surgery
registration engine. Another important distinction that we discuss below is that a “point”, or “fiducial point”, may be the centroid of a 3D “marker”, or “fiducial marker”.
Fiducial Markers and Points As a general rule, one needs at least three points to align, or register, two 3D spaces. These points can be either anatomical (eg, the manubrium of the malleus, {aka the umbo}, the most inferior aspect of the infraorbital rim) or manufactured (eg, skin-affixed markers, bone-implanted markers). Although anatomical points would seem ideal, a major problem is that they cannot be identified with sufficient consistency in image and physical space. In other words, a surgeon has a difficult time identifying exactly the same anatomical point on both the patient in the operating room and in the image of that patient acquired by CT or MRI. A solution is to use manufactured fiducial markers. Each such marker provides a fiducial point, which is a point on or within the marker, typically at or near its center that can be reliably and precisely identified in both spaces. The downside is that these fiducial markers must be attached to the patient before imaging is performed. This requirement occasionally necessitates repeat imaging and/or intraoperative imaging.
Skin-Affixed, Fiducial Markers and Points The easiest fiducial markers to use for both patients and surgeons are skinaffixed markers, which adhere to the
patient’s skin via adhesive pads and are placed so as to surround the surgical field of interest. Recommendations for where to place fiducial markers are presented in detail in Chapter 5, but the general rule is to place them in a region surrounding the anatomy of interest and as far apart from each other as possible. The adhesive pad of a skin-affixed marker may comprise the entire marker as, for example, the Multi-Modality Fiducial Marker made by IZI (IZI Medical Products, Owings Mills, Maryland), shown in Figure 4–1. Alternatively, also shown in Figure 4–1, the pad may support a base plate to which sequentially an imaging marker and then a tracking marker are attached. A good fiducial marker, whether or not it is skin affixed, is easy to identify in CT or MRI because it is made of material that shows up very bright in the image and therefore can be easily identified either visually or by automatic computer algorithms that employ a simple threshold at the high end of CT and/or MRI intensity scales. Additionally, the marker must be relatively large in comparison to the voxel size of image space so that the centroid, which defines the fiducial point, can be accurately determined (see Figure 2–27). In the operating room, the imaging marker is replaced by the tracking marker, which has a shape complementary to a tool tip that, when tracked, identifies the centroid of the marker in physical space. Although the skin-affixed setup works well in terms of repeatedly and precisely identifying the markers’
A
B
C
Figure 4–1. Skin-affixed markers. A is a multimodality universal skin-affixed marker with 2-mm inner center hole. Image provided courtesy of IZIMed, Inc. B shows the markers placed as would be used clinically for endoscopic sinus surgery with C showing the radiopaque nature. Republished with permission of Wolters Kluwer Health, Inc from Labadie RF, David BM, Fitzpatrick JM. Image-guided surgery: what is the accuracy? Curr Opin Otolaryngol Head Neck Surg. 2005;13:27–31. continues
101
102
Image-Guided Surgery
D Figure 4–1. continued D shows the Brainlab marker system, which consists of a skin-affixed base plate (left ), imaging marker with radiopaque blue sphere (center ), and tracking marker (right ). The centroid of the radiopaque blue sphere and the centroid of the divot in the tracking marker are in an identical location relative to the base plate.
positions in both image and physical space, the substrate to which they are attached, skin, is deformable and can shift its position between imaging and the operating room. These movements occur because of internal changes such as hydration level and external forces, including gravity (eg, sagging skin when going from a horizontal to a vertical position) and surgical positioning (eg, drapes placed over or adhering to the skin, unintentionally causing it to deform). Although such movements may seem inconsequential, in a game of millimeters, every little bit counts!
Bone-Affixed, Fiducial Markers and Points The gold standard for IGS fiducial markers is a set of bone-affixed markers, which are screwed directly into bone, because they are far less likely than skin-affixed markers to move between imaging and surgical intervention. The downside of the bone marker is that an invasive procedure is necessary
prior to imaging. This problem can be overcome by using general anesthesia during imaging and placing the boneimplanted fiducial markers at this time. The typical protocol for IGS treatment of Parkinson’s disease with deep brain stimulators involves having the patient undergo general anesthesia in the radiology suite, following which small skin incisions are made and bone-implanted markers are placed before both CT scanning and MR imaging. General anesthesia is necessary primarily because the tremors associated with Parkinson’s disease create unacceptable motion artifact in the scans, and placement of the markers during this general anesthetic is an opportunistic benefit. An alternative is the use of intraoperative imaging, as discussed in Chapter 2, in which bone-implanted markers can be placed after induction of general anesthesia, followed by imaging, fol-
Chapter 4
lowed by the IGS intervention (noting that sterility can be maintained during imaging by covering either the patient or the scanner with a sterile bag). Boneimplanted fiducial markers may consist of two parts with the base screwed into the bone supporting two different fidu-
Registration
n
cial attachments — one for imaging and the other for tracking in physical space (Figures 4–2 and 4–3). The most basic bone-implanted fiducial marker is a self-tapping facial plating screw whose crosshatch serves as the fiducial point and can often be clearly identified in a
Figure 4–2. Bone-affixed markers for imaging. Upper panel shows three photos of a post that screws into bone (shown screwed into wood): (a) driver and post, (b) cap containing imaging liquid being attached to post, (c) assembled marker. Lower panel shows preoperative image slices containing one imaging marker from (left to right ) CT, T1-weighted MR, and T2-weighted MR images. The lower panels are close-ups with a cross superposed on the marker centroids. A protective cap can be discerned in the CT close-up (bottom left ).
103
A
C
B
D
Figure 4–3. Bone-affixed markers for imaging and tracking. A. Imaging cap is removed after imaging is complete. B. Physical-space cap for tracking is attached to post in preparation for registration in the operating room. C. Tip of tracked probe mates with physical-space cap via ball-andsocket joint. D. The localized centroid of the imaging cap is at the identical location relative to the implanted post as the localized point on the physical cap in the operating room. This method of consistent image and physical localization was invented over 20 years ago (US patent 5,551,429).
104
Chapter 4
CT scan. A tracked probe can be placed in the crosshatch in the OR to define the fiducial point in physical space (Figure 4–4). Although bone-implanted fiducial markers are considered the gold standard, they are not perfect. The end screwed into bone may be inserted too shallowly —“underinserted”— allowing the marker to become dislodged between imaging and operating. The bone in which the screw end is placed may not be dense enough to effectively hold the marker. Or the screw may be “overinserted”, stripping the hole and
loosening the marker. To minimize over- and underinsertion, at least one group has developed an insertion tool that drives the screw into the bone until the undersurface of the head rests against the cortical surface of the bone and then releases.1,2 Even with such precautions, bone-implanted fiducial markers become dislodged between 1%3 and 3.4%1 of the time. Dislodgement of an anchor may reduce the number of markers below three, necessitating aborting the surgical intervention. For this reason, a fourth boneimplanted marker may be utilized.
A Figure 4–4. Use of facial plating screws as fiducial markers during lateral skull base surgery. A. An intraoperative picture of a left ear with four facial plating screws placed (arrows). B. After intraoperative CT scanning, views can be aligned to show each screw in profile (top right panel ) or en face (bottom right panel ). En face viewing allows identification of the crosshatch of the screw in the CT scan. This crosshatch can also be identified in the operating room by placing a tracked probe on the same crosshatch. These points are then aligned via registration.
Registration
n
B
105
106
Image-Guided Surgery
Stereotactic frames, including the N-bar covered in Chapter 1, fall into the category of bone-affixed markers because the frame is attached to the patient with pins that impale the skull. Unlike screws, these pins do not require threads to hold them in place. Instead, they are held firm by pressure against the skull, which is maintained by arranging the pins as opposed pairs, with each pin of a pair driven toward the other from opposite sides of the head. Two pairs (four pins) are required for rigid frame fixation. Many neurosurgeons naively assume that frames, because of their rigidity, are superior to individual boneimplanted markers. However, that rigidity can unfortunately hide insidious navigational error: if the frame is bumped or one pair of pins lose their grip so that the frame rotates slightly relative to the head during the hours between imaging and surgery, there is
Dental Casts In an effort to provide the benefits of bone-implanted markers in a noninvasive fashion, groups have explored attaching fiducials to exposed bone, namely the teeth! Because the upper dentition is an integral part of cranial anatomy, rigidly affixing markers to upper teeth should achieve accuracies comparable to other bone-affixed markers.4–6 (The mandible, because it is mobile relative to the rest of the skull, will not work.) However, there are a host of practical problems associated with the use of dental casts for anchoring fiducials, including (a) the need for
no telltale effect to alert the surgeon. The movement is undetectable because the frame is itself a rigid object and so its parts maintain the same mutual configuration regardless of how the frame moves relative to the anatomy. When a stereotactic frame suffers such a subtle shift, there is no failsafe indication that stereotactic guidance is no longer trustworthy. The surgeon will be in trouble during the intervention without knowing it. By contrast, if independently attached markers are bumped or lose their grip and move during the hours between imaging and surgery, the shape of the fiducial configuration will change. Then, when the fiducials are localized in the operating room, the change from the imaged configuration will become apparent — and measureable. The surgeon will become aware of the problem during the fiducial registration step — before the intervention begins!
adequate dentition not only in number of teeth but also in anatomical shape (ie, if the patients’ teeth are tapered, having wider bases than occlusive edges, a dental cast is not very secure), (b) the fact that teeth move and have interventions done to them (eg, crowns), (c) the oral cavity may be in the operative field and not suitable for referencing during surgery, and (d) the location of the teeth as a fiducial system for interventions located more centrally (eg, the skull base) is not optimal and leads to larger error during navigation.7 (This last problem is covered in more detail in Chapter 5 where general rules for placing fiducial markers to minimize error are covered.)
Chapter 4
Rigid Point Registration Regardless of how the fiducial points are selected (anatomical markers, skinaffixed markers, bone-affixed markers, N-bar frames, etc), registration involves identifying pairs of corresponding points in image space and in physical space and aligning all pairs as well as possible. Although this process may sound simple, we must remember that there are errors associated with identifying fiducial points in image space (see Figure 2–27 and the lower middle and right panels of Figure 4–2) and physical space (eg, see Figure 3–4 and 3–9B) that conspire to spoil the registration. The result is that registration will never be perfect (the effect of fiducial error on registration accuracy is treated in detail in Chapter 5). Imperfect registration is a central issue in IGS because registration provides the mathematical transformation of each point in image space to a point in physical space. This transformation takes the form of instructions to rigidly rotate (angular alignment) and translate (linear alignment) one space to the other. More specifically, this relationship requires six motions: three angular motions about the x, y, and z axes and three linear motions in the x, y, and z directions (Figure 4–5). To get the best rigid point registration, we find the unique rigid transformation that minimizes the mean of the squares of the distances between corresponding points.ii This mean is expressed ii
Registration
n
in Equation 4–1, and its square root, which will be covered in more detail in the next chapter, is known as the “fiducial registration error” (FRE).
(4–1)
In Equation 4–1, N is the number of fiducials, pi and qi are points in image and physical space, p′i is point pi after transformation from image space to physical space, and |p′i − qi| means “distance between p′i and qi”. Once we have minimized FRE, we have a transformation matrix that will be applied to every data point in image space, thereby transferring it to physical space, and thus allowing navigation. However, the navigation will be subject to error, meaning that when the surgeon places a probe on the patient, the point identified in the image will be displaced from the true point. Regardless of whether fiducial registration or some other registration method is used (eg, surface registration, discussed below), the distance between the true point and the one identified by an IGS system is called “target registration error” (TRE) (see Figure 4–5). This name is used because it represents an error in targeting caused by an imperfect registration. Both FRE and TRE are discussed in detail in Chapter 5. FRE clearly measures the misalignment of the fiducials used for registration, and, as a result, a misconception
The justification for minimizing the mean of the squares of these distances is statistical and is based on the reasonable assumption that errors in identifying corresponding points in the two spaces are statistically distributed independently and normally. With these assumptions, it can be shown that minimizing the mean of the squares of these distances produces the transformation that has maximum likelihood of being correct.
107
108
Image-Guided Surgery
Figure 4–5. Rigid point registration. (a) Three fiducial points in a patient image (dashed, top left ) are matched (arrows) to their corresponding points in physical space (solid, top right ). Then, the transformation of the image begins (downward arrow ). (b) The image is rotated by R, in this case a 20-degree rotation clockwise in the plane of the picture. The plus signs show the true position of a surgical target in each space. (c) The rotated image is then translated by t, in this case a horizontal translation to the right by 300 mm. The plus signs are slightly misaligned. The distance between them is the target registration error (TRE) for this target. The fiducials are likewise not aligned. The root mean square distance between them is the fiducial registration error (FRE). Here, as is typical for point registration, TRE is smaller than FRE.
often arises concerning the number of fiducials to use for IGS. This misconception, along with others, is covered in detail in Chapter 5, but it deserves mentioning more than once. It has to do with the fact that omitting fiducials will always make FRE smaller.8 Most IGS systems offer the option for the user to ignore selected fiducials during the registration step, and since a smaller FRE means a better fiducial alignment, it is tempting to take them up on the offer.
In fact, however, unless there is clearly some problem with a fiducial (eg, it is loose or damaged), minimization of FRE should never be accomplished by using fewer fiducials! Reducing the number of fiducials used in the registration certainly improves the fiducial alignment (aka FRE), but it rarely improves the overall accuracy of the IGS system. In fact, removing fiducials will almost always make the accuracy worse! Intuitively, this concept can be envisioned
Chapter 4
by considering cases where we have 1, 2, 3, . . . , up to n fiducial markers available for registration (Figure 4–6). With just one fiducial marker, FRE is zero, but the image spaces are tied together at only one point and can be oriented in an infinite number of ways about that point, so they are likely to be closely aligned only at that one point (Figure 4–6a). With two fiducial markers, FRE can quite small but the spaces are tied together only along the line joining the two points and can be oriented in an infinite number of ways about that line, so they are likely to be closely aligned only along that line (Figure 4–6c). With three fiducial markers, FRE will be somewhat larger, but the spaces are likely to be closely aligned everywhere, and we finally achieve a true registration
Registration
n
(Figure 4–6d). Adding more fiducials, although it makes FRE worse because having more fiducials makes it more difficult to align them, provides more information to the registration process and tends to improve the registration accuracy everywhere (Figure 4–6e).
Surface Registration Surface registration is essentially pointbased registration using a configuration of many, many fiducial points that lie on a surface. The surface to be registered must have unique geometry to avoid ambiguity as to where it matches or what its orientation is (eg, the occiput of the head is relatively nonunique, and as a result a surface registration
Figure 4–6. Rigid point registration with varying numbers of fiducials. The smiling gray star is to be registered to the white star. (a) Before registration. (b) Registration using one fiducial (red ). The red point is aligned, but erroneous rotations about three axes can misalign the rest of the objects. Two of those rotations are illustrated with curved arrows — one about an axis perpendicular to the page passing through the green point and a second about an axis passing through both the red and green points (the third rotation axis, which is not shown, is about an axis that is perpendicular to these two). (c) Registration using two fiducials (red and blue). The rightmost point of the gray star is hidden behind the white star. All points on the axis passing through the red and blue points are aligned, but erroneous rotation about that axis is possible. (d) Three fiducial points are necessary to ensure correct geometric alignment. (e) Real-world registration, illustrating that, even with three fiducials, there is a slight misalignment because of fiducial localization error.
109
110
Image-Guided Surgery
algorithm might confuse it for the anterior cranial curvature above the forehead, and, even if occiput is matched to occiput, the relative orientations of the two might be considerably different). An area that nicely satisfies this requirement is the face because it has unique anatomy with both concavities as well as convexities. Surface registration can be thought of as taking a “virtual” plaster-cast mask of the face from the CT or MRI scan and then fitting that to the patient in physical space — the operating room. In image space, the interface between air and skin can be identified using boundarydetection algorithms that detect dramatic differences in intensity values over short distances. For example, the Hounsfield CT-intensity value for air is −1000 (negative because Hounsfield defined the intensity of pure water to be zero) and the values for skin fall
within the range from −100 to 100, allowing a computer program to detect points at which the intensity changes from ±100 to −1000 over the span of one or two voxels as we move from a point on the skin into the air. These points can be designated as lying on the air-skin surface, allowing creation of a virtual mask composed of the configuration of these points. In physical space, points on the skin surface can be collected either by physically touching the skin with a probe tip (suboptimal because it deforms the skin during acquisition) or by using a laser scanner to do the same in a noncontacting fashion (Figure 4–7). As with independent fiducial points, surface points have error associated with them in both image space and physical space, so the fit of the mask to the patient will never be perfect. Mathematical techniques are used to optimize the fit via the same sort of minimiza-
Spectra CRF
laser la aser po ointe pointer
Figure 4–7. Acquisition of points for surface registration. A Brainlab Kick system is being utilized to collect surface points in the operating room. The system employs a Polaris Spectra to detect infrared rays reflected from the spheres on a coordinate reference frame (CRF) that is attached to, and allows tracking of, the patient’s head. The surgeon casts infrared rays over skin with a laser scanner. The Spectra detects the reflected infrared rays from many points on the skin (one such point is depicted in the zoomed inset). In an alternative approach (not shown), instead of infrared rays being cast, the tip of a tracked wand is touched to the skin at many points.
Chapter 4
tion of the mean of the squared distances between pairs of corresponding points as that used in the registration of discrete fiducial points (see Equation 4–1), but their registration requires more work because finding pairs of surface points that correspond in the two spaces is much more difficult than pairing fiducial markers. The most widely used algorithm to accomplish surface registration is the “iterative closest point” algorithm,9 commonly referred to simply as “ICP”. In this algorithm, three preliminary steps are performed once: (a) a configuration of surface points is identified in physical space, (b) the same surface is identified in image space, and (c) a rough initial registration is performed manually or by some other means to bring the physical points in proximity to the
Registration
n
image surface. Then the following three steps are performed repeatedly: (1) each surface point in physical space is paired with the closest point on the surface in image space, (2) the rigid transformation is found that minimizes the sum of squares of the distances between these paired points resulting in an FRE value, and (3) the transformation found in step 2 is applied to the set of points in physical space. These last three steps are typically repeated, or “iterated”, dozens or even hundreds of times until the decrease in FRE is negligible from one iteration to the next (Figure 4–8). Although surface registration is fast and easy in clinical application, the accuracy achieved gets precipitously worse as a surgeon navigates farther and farther away from the registered surface. This degradation in accuracy
points acquired the OR being shifted and reoriented to fit the air-skin surface in the patient’s image
patient in operating room
air-skin surface in image
Figure 4–8. Surface registration. An iterative algorithm — typically the ICP algorithm — is applied to the set of points acquired in the operating room to fit them to the air-skin interface in the patient’s pre-operative image. Only a few steps in the registration are shown here, but there are typically dozens or even hundreds of steps required.
111
112
Image-Guided Surgery
occurs because the distant site of interest is cantilevered from the registered surface. Because of this cantilevering, even small errors in the surface registration can lead to large navigation error, as shown in Figure 4–9. This problem has been demonstrated clinically. The accuracy of skin-surface registration using IR optical tracking was measured in a patient study in which surfaces on the face were acquired with Brainlab’s VectorVision IGS system using the z-touch laser scanner (BrainLAB, Heimstetten, Germany).10 Targeting error was measured for each patient after registration had been performed by placing the probe on three anatomical landmarks — nasion and right and left external auditory canals — and measuring TRE. This study found that mean TRE in the surgical field of frontally located lesions was 1.8 ± 0.8 mm (n =13), while mean
TRE at points on temporal, parietal, occipital, and infratentorial lesions was 2.8 ± 2.1 mm (n = 21). Furthermore, at the external auditory canal mean TRE was (right) 3.3 ± 1.9 mm (n = 23) and (left) 3.9 ± 2.3 mm (n = 22). The level of error near the ear with this same system was confirmed in another study, which found that mean TRE was 3.21 ± 1.01 mm when navigating in the region of the mastoid located approximately 15 cm distal to the surface of the face.7
Accuracies for Various Types of Fiducials As is true for tracking systems, accuracies of differing types of fiducial markers depend on the localization technique used (eg, skin surface regis-
Figure 4–9. Impact of surface registration on misalignment of objects that are shallow or deep to the surface. (a) A rigid object consisting of a surface (black ), a shallow object (green), and a deep object (red ). (b) The same object rotated 10 degrees. (c) Superimposition of the two objects after a surface registration with the 10-degree rotation error illustrated in (c). It can be seen that, while the misalignment of the shallow object is approximately the same as the surface misalignment, the misalignment of the deep object is much worse.
Chapter 4
tration with laser scanning vs tracing tool) and the location in which the accuracy is measured. It is clear that overall IGS accuracy is better when using boneaffixed fiducial makers, which have an accuracy (aka TREs) of 1.0 to 1.5 mm11– 13 compared to skin-affixed fiducial markers, which have accuracies (aka TREs) of 1.3 to 4 mm.12,14,15 Less clear is the difference between skin surface and skin-affixed point fiducials, with select studies supporting each. Even less clear are the accuracies of novel fiducial systems such as dental biteblocks and proprietary masks, which under ideal laboratory conditions can be shown to have excellent accuracy but fail to make the transition to clinical realization usually because of the added burden necessary to achieve the desired accuracy (eg, have the patient not move or not breathe during the surgical intervention). Long-term solutions may involve the use of multiple techniques simultaneously. For example, Maurer et al16 showed in 1998 that skin-surface registration error could be reduced from 3.1 ± 1.4 mm to 2.2 ± 1 mm by the addition of just one boneaffixed marker. And there may exist yet-to-be-implemented techniques such as combining scanned cortical bone surfaces after surgical exposure with facial skin surface scanning.
Fiducials During Navigation After registration has been completed, the fiducials can be largely ignored as long as (a) the patient’s head does not move (eg, it is immobilized) or (b) the head’s position is tracked (eg, using a CRF). Immobilization with a Mayfield head holder is typical in neurosurgi-
Registration
n
cal procedures but unusual in ENT procedures, in which tracked coordinate reference frames (CRFs) (Chapter 1) are more commonly encountered. CRFs can be either skin affixed or bone implanted with pros and cons similar to those of skin-affixed versus boneaffixed fiducials. Namely, skin-affixed CRFs are noninvasive but less stable, and hence less accurate, while boneimplanted CRFs are more invasive but more stable, and hence more accurate. However, because they are needed only in the OR, once general anesthesia has been induced, there is less reticence on the part of surgeons and patients to use bone-implanted CRFs.
Fusion Before we leave registration, we consider the special case of registration from one image space to another image space, as frequently is done when a preoperative CT scan is registered to a postoperative CT or a preoperative CT is registered to a preoperative MRI. This type of registration uses an entirely different technique known as intensity registration in which similarities between the two images in voxel intensity values or similarities in the frequency of occurrence of those intensities are exploited. As mentioned in Chapter 1, since 1997, when a largescale study validated its accuracy,17 the method of choice for CT-to-MR registration has been the maximization of mutual information (MI), a method that had been discovered independently by two groups two years earlier.18,19 This method is based on a statistical analysis of the co-occurrence of overlapping intensity values in the two images. Like
113
114
Image-Guided Surgery
the ICP algorithm for surface registration, MI requires that the two images be initially brought into approximate alignment, but unlike ICP, it is not necessary to identify surfaces in the images or otherwise preprocess them. Such registration, when it occurs between scans of the same patient, is commonly referred to as “fusion”, and the accuracy of the registration depends upon how accurate the images are and how similar they are. For fusion registration of a pre- to a postoperative CT scan, as long as the majority of the tissue has not been altered during the operative procedure, fusion can be achieved with submillimetric error. Because MRI scans have more distortion than CT scans, CT-MRI fusion suffers from a larger registration error — on the order of 1.5 mm—but if measures are taken to correct the distortion, submillimetric accuracy has been demonstrated.17
References 1. Balachandran R, Fritz MA, Dietrich MS, et al. Clinical testing of an alternate method of inserting bone-implanted fiducial markers. Int J Comput Assist Radiol Surg. 2014;9(5):913–920. 2. Fitzpatrick JM, Labadie RF, Mitchell JE, inventors; Anchor driver with assured seating. US patent 8,231,636. July 31, 2012. 3. Konrad PE, Neimat JS, Yu H, et al. Customized, miniature rapid-prototype stereotactic frames for use in deep brain stimulator surgery: initial clinical methodology and experience from 263 patients from 2002 to 2008. Stereotact Funct Neurosurg. 2011;89(1):34–41. 4. Bale R, Burtscher J, Eisner W, et al. Computer-assisted neurosurgery by using a noninvasive vacuum-affixed dental cast that acts as a reference base:
another step toward a unified approach in the treatment of brain tumors. J Neurosurg. 2000;93(2):208–213. 5. Fenlon MR, Jusczyzck AS, Edwards PJ, King AP. Locking acrylic resin dental stent for image-guided surgery. J Prosthet Dent. 2000;83(4):482–485. 6. Labadie RF, Shah RJ, Harris SS, et al. Submillimetric target-registration error using a novel, non-invasive fiducial system for image-guided otologic surgery. Comput Aided Surg. 2004;9(4):145–153. 7. Balachandran R, Fitzpatrick JM, Labadie RF. Accuracy of image-guided surgical systems at the lateral skull base as clinically assessed using bone-anchored hearing aid posts as surgical targets. Otol Neurotol. 2008;29(8):1050–1055. 8. Synderman C, Zimmer LA, Kassam A. Sources of registration error with image guidance systems during endoscopic anterior cranial base surgery. Otolaryngol Head Neck Surg. 2004;131(3):145–149. 9. Besl PJ, McKay ND. A method for registration of 3-D shapes. IEEE Trans Pattern Anal Machine Intell. 1992;14(2):239–256. 10. Raabe A, Krishnan R, Wolff R, Hermann E, Zimmermann M, Seifert V. Laser surface scanning for patient registration inintracranial image-guided surgery. Neurosurgery. 2002;50:797–802. 11. Maurer CR Jr, Fitzpatrick JM, Wang MY, Galloway RL Jr, Maciunas RJ, Allen GS. Registration of head volume images using implantable fiducial markers. IEEE Trans Med Imaging. 1997;16(4):447–462. 12. Mascott CR, Sol JC, Bousquet P, Lagarrigue J, Lazorthes Y, Lauwers-Cances V. Quantification of true in vivo (application) accuracy in cranial image-guided surgery: influence of mode of patient registration. Neurosurgery. 2006;59(1): ONS146–ONS156. 13. Metzger MC, Rafii A, Holhweg-Majert B, Pham AM, Strong B. Comparison of 4 registration strategies for computeraided maxillofacial surgery. Otolaryngol Head Neck Surg. 2007;137(1):93–99. 14. Labadie RF, Davis BM, Fitzpatrick JM. Image-guided surgery: what is the
Chapter 4
accuracy? Curr Opin Otolaryngol Head Neck Surg. 2005;13(1):27–31. 15. Woerdeman PA, Willems PWA, Noordmans HJ, Tulleken CAF, Berkelbach van der Sprenkel JW. Application accuracy in frameless image-guided neurosurgery: a comparison study of three patient-to-image registration methods. J Neurosurg. 2007;106(6):1012–1016. 16. Maurer CR Jr, Maciunas RJ, Fitzpatrick JM. Registration of head CT images to physical space using a weighted combination of points and surfaces. IEEE Trans Med Imaging. 1998;17(5):753–761.
Registration
n
17. West J, Fitzpatrick JM, Wang MY, et al. Comparison and evaluation of retrospective intermodality brain image registration techniques. J Comput Assist Tomogr. 1997;21(4):554–568. 18. Collignon A, Maes F, Delaere D, Vandermeulen D, Suetens P, Marchal G. Automated multi-modality image registration based on information theory. Inf Process Med Imaging. 1995;3(6):263–274. 19. Viola P, Wells WM III. Alignment by maximization of mutual information. International J Computer Vision. 1997; 24(2):137–154.
115
5 Error Analysis IGS manufacturers take accuracy very seriously. To that end, they employ talented engineers and programmers who work together to produce highly accurate systems, and they are rightfully proud of the results, as their websites attest. However, the other side of the accuracy coin is error, and, as we have seen in Chapters 3 and 4, every IGS system is subject to it. The presence of error means that an IGS system is to some degree misleading the surgeon as to where something is located. That something may be a target that is to be resected or a critical structure that is to be avoided, but in either case, the importance of reducing the severity of the error is clear. What is sometimes less clear is the importance of estimating the severity of the error that can be expected to remain despite all efforts to reduct it. A surgeon who is ignorant of the expected error level of an IGS system may be too conservative with ablation because of too little faith in the system’s navigational accuracy or too aggressive because of too much faith. Predicting error in IGS is the subject of this chapter, but such predictions are at best statistical in nature (means, standard deviations, root mean squares, confidence limits, and the like), because, if we could predict the error itself, we could correct it. 117
For example, if we could predict that a tracked pointer will, during this morning’s surgery, overshoot a target on the stapes in a certain direction by 0.7 mm, we could simply adjust the nominal target by −0.7 mm in the overshoot direction. With that correction made, we would achieve IGS perfection! Corrections of predictable components of error constitute a form of calibration, and, although calibrations may be an essential part of the system's installation procedure or of a setup protocol that is carried out before each surgery, they do nothing to eliminate the unpredictable components of error. As a more precise example, if the tracker consistently reports a position of x,y,z = 14, 17, 11 every time it is actually placed at position 14, 17, 10, then we can calibrate it by subtracting 1 from z when it reports 14, 17, 11. On the other hand, if it randomly reports z = 11 sometimes and reports z = 9 at other times when it is actually at 14, 17, 10, then calibration is no help. So after calibration is complete, we are back where we started — with an imperfect system and the need to estimate the magnitude of its error.
118
Image-Guided Surgery
In this chapter, we describe what is known about the statistics of errors in IGS systems that employ fiducial markers, and we describe ways to estimate them statistically. The reader is advised that our description does involve some explanation of mathematical relationships via equations. But, have faith! As we pledged in Chapter 2 when explaining MRI, we are presenting things in the simplest way possible (but no simpler!), and we are doing the same in this chapter with error analysis. If you follow the discussion — equations and all — you will be better off in your understanding, and your understanding will allow you to reduce error on a daily basis — we promise! We begin, as we often have, with a dichotomy.
Extrinsic Versus Intrinsic Error Navigational error can be divided into extrinsic error, which is error not caused by the navigation system, and intrinsic error, which is error caused by the navigation system. Extrinsic error may arise from two sources: (a) user errors such as incorrect image parameter readings (eg, voxel dimensions, CT gantry tilt, etc), transcription errors, and/or incorrect setup of the tracking system in the OR, and (b) undetected motion including rigid motion, such as slippage of the head in the head holder or slippage of a reference frame relative to the head, and nonrigid motion when tissue changes shape between imaging and surgery. Both of these sources of error can be essentially eliminated by careful use of the system and the use of IGS only on bone, which is reliably rigid to within small fractions of a millimeter. For applications in otolaryngology and
orthopedics, for spine surgery, and to a lesser extent for minimally invasive brain surgery, the assumption of rigid anatomy is reasonable. Thus, for the rest of this chapter, we can ignore extrinsic error, which through proper use of the system is avoidable, and will concentrate instead on intrinsic error, which, even with the most careful application of IGS, is not. The assumption of rigidity holds for minimally invasive procedures that reduce brain motion relative to the skull. These procedures take advantage of the relative rigidity of the unexposed brain, which, while it is sealed in the skull, is rigid to roughly ±0.5 mm even when the head is reoriented from supine to prone.1 However, during wide-field neurosurgical procedures, after the dura has been opened, there can be dramatic brain movement, resulting in inaccuracy on the order of a centimeter. Our error analysis is not applicable to such procedures. As we pointed out in Chapter 4, each guidance system relies on some registration method in which an algorithm processes information obtained from physical devices, such as CT scanners and optical tracking systems, to produce an optimal transformation to map points from image space (CT or MRI) to physical space (the OR). An optimal transformation is simply the best guess, given the available information, and it is amost never perfect. A perfect transformation maps every point in the image space onto its correct counterpart in physical space and vice versa with perfect accuracy. For example, with a perfect transforma-
Chapter 5
tion, when the plus sign icon on the display screen representing the tip of the probe touches an anatomical target on the screen, the actual probe will be touching the actual anatomical target on the patient (ie, the plus signs in Figure 4–5 of the previous chapter will align perfectly). It’s a wonderful vision of perfection, but in the real world, the actual probe will almost always be displaced some distance from the actual target as can be seen in Figure 4–5(c). The standard term in the literature for this displacement is target registration error (TRE),2 and it is widely accepted as the best measure of the accuracy of an IGS system both for the fiducialmarker systems treated in this chapter and all other systems — whether they employ stereotactic frames or surfaceregistration methods. For all IGS systems, high accuracy means low TRE and low accuracy means high TRE. We cannot eliminate TRE, but we can estimate it. When we do that, it is important to note that the only errors that need to be estimated are the ones that are random, because consistent errors can be eliminated via proper calibration (we give an example in the first box in this chapter). Fortunately, random errors in IGS systems typically follow statistical patterns, and this is the key to providing the surgeon with estimates of what the error may be. Over the past 20 years, much progress has been made toward understanding these patterns, and we can bring those results to bear on any analysis of IGS error. As explained in Chapter 4, the registration of point fiducials is the basis for all IGS systems, from stereotactic frames to discrete fiducial markers to surface-matching systems, so our analysis of the statistics of IGS error is based entirely on point fiducials.
Error Analysis
n
Fiducial Localization Error (FLE) The intrinsic errors made by all these systems result from one simple problem — the inevitable, imperfect localization of fiducial points, both in the image (see lower panels of Figures 2–4 and 4–2) and in the operating room (Figures 3–4 and 3–9B). This imperfection in the system is called "fiducial localization error" (FLE) and it the source of all intrinsic errors. In short, when it comes to the intrinsic error of IGS, FLE is the root of all error! And it is inevitable. In fact FLE is worse than inevitable: it is also invisible and unmeasurable. When a fiducial position is determined erroneously by an IGS system, there is by definition no way to know how bad the error is or even its direction — whether it be along x, y, or z or any combination of these. It is unknown “by definition” because, if we knew it, we would correct it. FLE is what’s left over after we have eliminated every source of localization error that we can. FLE is depicted schematically in Figure 5–1A. Each circle represents a true fiducial point in an image, and the erroneous localized position pi in the image for that fiducial is the dot that is closest to it. The displacement from circle to dot is the FLE for that fiducial point (FLE image). The FLE situation for physical space (FLE physical) is depicted in the panel labeled Figure 5–1B (note: 1B is on the far right). Here again the circles are the true fiducial points in physical space (the operating room), but in this panel, the localized points qi are represented by plus signs. When the IGS system performs its registration step, it determines a rigid transformation (rotation and linear translation)
119
120
Image-Guided Surgery
FLEphysical
FLEimage
pi
×
×
T
TRE 4
A
pi’ qi
qi B
C
Figure 5–1. Depiction of fiducial registration. A. Image space (far left ): the circles are true fiducial points (centroids of fiducials), the dots are the localizations p i and the x is a target. The difference between true position (circles) and localization (dots) is the fiducial localization error of image space (FLE image). B. Physical space (far right! ): circles are true fiducial points (centroids of same fiducials), “plus” signs are the localizations q i and the diamond is the same target. The differences between true position (circles) and localization (“plus” signs) are the fiducial localization errors of physical space (FLE physical). C. Superposition of localized points and targets both from (A) image space after a rigid transformation T and from (B) physical space. The respective localizations, p i and q i (dots and plus signs) are imperfectly aligned (true fiducial location, ie, circles, not shown). Equation 5–1 gives fiducial registration error (FRE) in terms p'i and qi for i = 1 to N with N = 4. p'i. The displacement between the target positons—the diamond and the x—represents target registration error (TRE).
that will come as close as possible to aligning each of the localized points in image space (dots) with its corresponding localized point (plus signs) in physical space. The result is depicted in Figure 5–1C (in the middle). Here the localized points in physical space (plus signs) have been copied from Figure 5–1C, but the localized points in image space have been transformed (arrow with “T” for transformation) by being rotated and translated as a rigid unit to the positions (dots) shown in Figure 5–1C.
Fiducial Registration Error (FRE) After the transformation is complete, we can see the effects of FLE in Figure 5–1C. The first effect is that the localized fiducial points from image space (dots) fail to land exactly on top of their
corresponding fiducial points in physical space (plus signs). This effect proves to the observer of the registration that at least for some of the fiducial points, FLE is nonzero, and if both the physical fiducial points and the transformed image fiducial points are displayed by the IGS system, then this effect will be immediately apparent to that observer as soon as the registration step is complete (ie, after the last fiducial has been localized in the operating room and the system has calculated the transformation). There is a standard name for this effect in the literature — fiducial registration error (FRE). We introduced this term in Chapter 4 and gave an expression for it, which we repeat here:
(5–1)
Although this equation may appear intimidating, bear with us! N is the
Chapter 5
number of fiducials, pi is a fiducial point in image space (Figure 5–1), qi is the corresponding point in physical space and p′i is point pi after the transformation from image to physical space, and |p′i – qi| denotes the absolute distance between p′i and qi, recalling that absolute values are always positive. The Greek capital sigma — Σ — stands for “sum”, and, with its subscript “i = 1” and its superscript N, it means that we sum up all the values of |p′i – qi|2 from the first one (i = 1) to the last one (i = N). The sum is then divided by N, yielding the mean of the squared distances. The square root of the mean produces the root mean square (RMS, as defined in the Optical Tracking section of Chapter 3). Thus, FRE is the RMS of the distances between corresponding fiducials after registration. This expression for FRE provides an overall measure sometimes called the “goodness of fit” (“fit” meaning alignment), and some IGS systems actually report FRE to the user in millimeters. Note that each term in the sum is either positive or zero, so if, and only if, at least one distance is nonzero, meaning that there is at least one misfit in the points after registration, FRE will be nonzero. Thus, a nonzero FRE, like the visible misalignment of fiducial point pairs, proves that one or more FLEs are nonzero.
Target Registration Error (TRE) The second effect of FLE is TRE. Returning to Figure 5–1A, note the “x” near the center. This “x” represents the true position in the image of a targeted point. After the registration step has been completed, the calculated transformation can be applied to any such
Error Analysis
n
target point to transform it into physical space. The result of that transformation is the “x” in Figure 5–1C. Now note the diamond shape (◊) in Figure 5–1B. This diamond represents the true position of this same target in physical space (somewhere within the patient in the operating room). This diamond, like the plus signs, has been copied from Figure 5–1B to Figure 5–1C. If the transformation were perfect, the centers of the “x” and the diamond would be coincident in Figure 5–1C. They are not because the transformation is imperfect, and it is imperfect because FLE is not zero. Thus, FLE is the cause of both FRE and TRE. As we said before, FLE is the root of all error in IGS! Stating this principle another way. FRE and TRE are effects of FLE. Both of these effects are clearly visible in Figure 5–1C, and FRE is also visible in an IGS system. However, TRE, like FLE, is typically invisible because the target is not usually visible in physical space — either because it is hidden beneath occluding bone or soft tissue or because the target is an imaging phenomenon (eg, center of a bright spot in a T2-weighted MR image) that is indistinguishable visibly in the operating room. If the target were visible in the operating room, then there would be no reason for a surgeon who is seeking the target to use IGS to locate it! So, if IGS is useful, TRE is invisible, and when TRE is invisible, we can only estimate it statistically. The statistical nature of these quantities is an important distinction that can be missed by the end user. We discussed this point in Chapter 3, but it is worth repeating here: Although FLE, FRE, and TRE are typically specified by quoting a single number of millimeters (1 mm, 0.7 mm, 0.5 mm, etc), a system’s
121
122
Image-Guided Surgery
accuracy is characterized in terms of a statistical distribution of errors. Quoting a single number when describing a system is analogous to stating that the average grade in a certain medical school class is 79. Although 79 may indeed be the average for that class, it is understood that 79 is merely the mean of a distribution of grades and is not the grade of every student, or any particular student. Similarly, 0.5 mm may be the mean TRE for a given IGS system at a given target point with a given FLE and given set of fiducials, but, even with all those specifications given, that system will exhibit a distribution of TREs if the procedure is repeated many times. The shape of the distribution may vary somewhat from one situation to the next, but the distribution of grades is typically very close to being normal, while the TRE distribution is shaped more like that of the Maxwell distribution. In particular, unlike the normal distribution which has both positive and negative values, because TRE is never negative, its distribution has no negative values (see Figure 3–7).
Error Relationships To review the errors in IGS, they can be divided at the highest level into two types: n Extrinsic errors, which can and
should be eliminated through proper use of IGS n Intrinsic errors, which can be estimated only statistically and are of three types: n FLE, which is inevitable and invisible
n FRE,
which is caused by FLE, is visible, and reveals the presence of FLE n TRE, which is caused by FLE but is invisible whenever IGS is useful
These three types of intrinsic error are related statistically as described below. And, considerable progress has been made in the mathematical derivation of expressions for these relationships since the mid-1990s. These derivations require only the very basic assumptions that the FLEs are statistically uncorrelated (ie, one marker’s localization error is unrelated to another’s).
Relating TRE to FLE No formula relating FLE, FRE, or TRE was available to the IGS community until the publication in 1998 of the following formula, which relates the root mean square values of TRE and FLE.3 RMS TRE(p) =
(5–2) Again, bear with us as we explain this very powerful equation, which can be understood by anyone with a basic grasp of algebra. Understanding this equation will allow you to make your IGS system work optimally. Here, RMS, as usual, means root mean square, p is the location of the target, and N is the number of fiducials. The square-root factor on the right is explained below (see “TRE Dependencies”), but first we remind the reader that, as promised, this is strictly a statistical relation-
Chapter 5
ship. This formula says that RMS TRE is proportional to RMS FLE, and it provides us with the proportionality factor, which is 1 over the square root of N times the square root of the expression involving the ds and fs. Before we discuss this factor it is important to note what this formula does not say. It does not say that TRE is proportional to FLE. In other words, it does not say that, if FLE is twice as large for one registration as it is for another one, then TRE will be twice as large for the first registration as it is for the second one. Instead, it means that if the process of fiducial localization followed by registration is repeated many times, then the root mean square of the many TRE values that would result at point p will be given by Equation 5–2. In short: RMS errors are proprotional; individual errors are not. Assumptions Concerning FLE As discussed earlier, FLE arises from imperfections in both image localization and physical localization. These two errors can be expected to be uncorrelated with each other, and like all uncorrelated errors, their errors add “in quadrature”, which means that the terms are squared and the square root of the sum taken. Thus, the resulting RMS FLE in Equation 5–2 can be expressed in terms of the image and physical RMS FLE values as follows:
Error Analysis
n
each space (image being CT or MRI and physical being the OR). Thus, in addition to the basic requirement that FLEs are uncorrelated, this formula for RMS TRE is based on two simplifying assumptions regarding FLE: (1) the error probabilities are the same for all fiducials and (2) the error probabilities are the same for all directions. In the registration literature, (1) is called localization “homogeneity” and (2) is called localization “isotropy”. Since 1998, advances in the mathematical derivations have allowed both of these requirements to be removed.4 Furthermore, registration algorithms have been developed that allow for weighting some of the fiducials and some directions more highly than others for optimal accuracy, and a general formula has been derived to replace Equation 5–2 for all possible inhomogeneous, nonisotropic FLEs and all weightings. But, the resulting formula is many times more complex than Equation 5–2, and the difference between the values given by the more complex expression and Equation 5–2 are so small for typical IGS applications that it is (mercifully!) sufficient to limit our discussion to this simpler expression. TRE Dependencies We now turn to the proportionality factor from Equation 5–2, which can be written as
RMS FLE = √ (RMS FLEimage)2 + (RMS FLEphysical)2 (5–3) Note also that, although there are N fiducials, there is only one RMS FLE in
The first part of the factor, which is 1 divided by the square root of N, should not be surprising. It says that error decreases as N increases, in inverse proportion to √ N.
123
124
Image-Guided Surgery
This deserves repeating: increasing the number of fiducials reduces the overall error of the IGS system. So the first thing you can do to increase the accuracy of the system is to use more fiducials (and, as pointed out in the Rigid Point Registration section of the previous chapter, the last thing you should do is use fewer fiducials)! This approach is explained in much more detail in “Reducing TRE” below. This part of the relationship should not be surprising because inverse dependence on N also characterizes the behavior of the standard deviation of the mean of N numbers. In both cases — the calculation of (a) the mean of a set of N measured numbers and (b) the position of a target from a set of N measured fiducial positions — we would expect each result to deviate less from its true value when N is larger. Furthermore, if the measurements are statistically uncorrelated, as we are assuming that they are, the deviation from the truth will decrease in inverse proportion to √ N. The second part of the proportionality factor — the square root of the expression involving the ds and fs, is a “form factor” that depends on the configuration of the fiducial arrangement. Each d in this expression equals the distance of the target point p from one of the three “principal axes” of the fiducial configuration. These principal axes intersect at the centroid of the fiducial configuration (the mean of the N fiducials’ positions) and, like the typical set of coordinate axes, they are perpendicular to each other. Envision an XYZ coordinate system centered at the geometric middle of the configuration of N fiducials. The orientation of this
coordinate system (ie, the directions in which the three axes point) is related to the f1, f2, and f3 in the denominators. Each fi equals the RMS distance of the N fiducials from one of the principal axes, and the orientation of the principal axes is chosen to minimize the sum of the squares of f1, f2, and f3. What does all this mean?!? It means that if you spread your fiducials out widely but keep the centroid of the configuration of fiducials approximately at the target region of interest, you will optimize your accuracy (reduce your error). This is explained in much more detail in the next sections. The formula given by Equation 5–2 quantifies the effect on registration accuracy of the target’s position relative to the fiducial configuration. It shows, for example, that RMS TRE is minimal at the centroid of the configuration, because at this position in space, and only at this position, the ds are all equal to zero. At that centroid, RMS TRE(p) = × RMS FLE, so its value is unaffected by the shape of the configuration, but when the target is not at the centroid of the configuration, then, as the target position moves farther from the centroid, the ds get bigger so RMS TRE tends toward larger values. The sizes of these larger values is determined not only by the ds in Equation 5–2 but also by the fs. The fs are determined not by the target position but by the shape and size of the fiducial configuration. Those fs tend to be larger for more widely spread configurations, and since each f is in the denominator, spreading out the configuration to produce those larger fs reduces RMS TRE.
Chapter 5
Reducing TRE Equation 5–2 gives a quantitative value for RMS TRE at target point p, when the parameters describing the fiducial configuration and FLE are entered into the expression on the right. But, given our discussion above regarding the impact of each variable within the equation, we can draw some useful qualitative conclusions as to how an IGS surgeon can reduce RMS TRE: 1. Use a better localization method in image space or physical space or both. 2. Choose a fiducial configuration that puts the targets closer to the centroid. 3. Choose a configuration whose fiducials are more widely spaced. 4. Use more fiducials. These four approaches work independently to reduce RMS TRE because they each affect a different parameter in Equation 5–2: Number 1 makes RMS FLE smaller, Number 2 makes the ds smaller, Number 3 makes the fs larger, and Number 4 makes N larger. Regarding the first item in this list, it is worth pointing out that the primary way to effect a better localization method is to use bone-implanted fiducials if they are not already being used. Their advantage is that they are more likely to be in the same position relative to the anatomy in image space and physical space. If there is movement of fiducials relative to the anatomy between the two spaces, the effect is to add error to the localization. All four items in the list above should always be considered by any practitioner before placing, or overseeing the placement of, fiducials for IGS. The implementation of these approaches is usually obvious, but not
Error Analysis
n
always. For example, for procedures on the right ear, the second item tells us that most or all of the fiducials should be on the right side of the head. However, for more medial positions, the target may be closer to the centroid of the fiducial configuration if one fiducial is placed on the left side of the head, because that placement will cause a medial shift in the centroid from the right side, Figure 5–2 may help the reader develop an intuition about the effect of fiducial placement. The fourth item tells us that there is always a benefit to adding another fiducial, but placing additional fiducials without increasing their spread or moving the centroid closer to the target leads to diminishing returns because the percentage improvement goes down as N goes up. For example, if fiducials are added without changing their spread, the improvement in going from three fiducials to four is 13.4%, from four to five is 10.6%, and from five to six is 8.7%. Furthermore, patient discomfort and time for placement and localizations (one in each space) of each fiducial must be weighed in the balance. Coordinate Reference Frames So far, our description of the generic guidance system requires that once the fiducials have been localized in the operating room, the anatomy must remain stationary relative to the tracker (eg, via a Mayfield head holder). This onerous restriction can be removed, however, if the relative movement of the anatomy and the tracker is continuously monitored. This continuous monitoring is made possible by the addition of a coordinate reference frame (CRF), which is a set of fiducials rigidly attached to the head and tracked separately but simultaneously from the
125
126
Image-Guided Surgery
Figure 5–2. Placing markers surrounding the surgical target improves TRE. This improvement is demonstrated in these panels adopted from West et al (2001), which shows in the left panel fiducial markers as yellow spheres and in the right panel the calculated RMS TRE spread for RMS FLE = 1 mm. Using two fiducials on either side of the head results in estimated RMS TRE (recall that TRE is a distribution as shown in Figure 3–7) of 1 mm over much of the cranium (top row ). When one of these fiducials is removed, the RMS TRE spread increases to 2 mm (middle row ). When four fiducials are placed on only one side of the head, RMS TRE is near 1 mm at the surface but increases to 4 mm on the contralateral side of the head (bottom row). Republished with permission of the American Association of Neurological Surgeons, from West JB, Fitzpatrick JM, Toms SA, Maurer CR Jr, Maciunas RJ. Fiducial point placement and the accuracy of point-based rigid body registration. Neurosurgery. 2001;48:810–817.
surgical probe (see Figure 4–7). The use of a CRF allows the patient and/or bed to be moved and reoriented and the cart that bears the tracking system to be rolled around the room at will without compromising the registration. It adds, however, the constraint that two objects — probe and CRF — must now be tracked, which for opitical systems restricts somewhat the positioning of people and equipment so as to avoid obstructing LOS of either, thus compounding the LOS problem described in the Electromagnetic Tracking section
of Chapter 3. A second effect is that a new source of error appears — the registration of the CRF fiducials in their new positions to the positions that they had during the physical localization of the fiducials. The RMS values of these two errors combine to produce one RMS TRE in the same way that the FLEs in image and physical space combine in Equation 5–3,5 but the combination does not affect any of the four improvement methods given above. It simply adds an additional one: 5. Place the CRF as close to the target as feasible.
Chapter 5
Relationships Involving FRE So far, we have considered only the relationship between TRE statistics and FLE statistics. There are two remaining relationships (the last two equations in this chapter!): RMS FRE = √ 1 − (2/N) × RMS FLE (5–4) where RMS FLE is given by Equation 5–3, and RMS TRE(p) =
(5–5) The first of these, Equation 5–4, was derived in 1979 in a context unrelated to physical tracking or registration,6 and its significance in IGS was not recognized until 1998.3 The second, Equation 5–5, is merely the result of using Equation 5–4 in Equation 5–2. The most interesting feature of Equation 5–4 is its simplicity. In particular, the shape of the fiducial configuration does not appear in it! (There are no ds or fs!) This expression tells us that RMS FRE is proportional to RMS FLE and that the proportionality depends only on the number of fiducials. Note that if the number of fiducials is decreased (with no change in RMS FLE), then RMS FRE decreases, and if the number of fiducials is increased, then RMS FRE increases. At first this relationship may seem counterintuitive because FRE is a measure of the goodness of fit of the registration, and more fiducials are likely to give a better registration, while the larger FRE seems to indicate a worse fit and thus a worse registration. All that is true, but as fiducials are added,
Error Analysis
n
it becomes more difficult for the registration algorithm to find a transformation that will bring each pair in the two spaces together, and therefore the RMS of the distances between them after registration goes up (Figure 5–3) even as TRE goes down. Thus, one needs to note that comparison between two FREs is meaningful only for two registrations with the same number of fiducials. And, more importantly, using fewer fiducials will lower FRE but at the same time will increase TRE! The astute reader will note that since reducing N in Equation 5–4 reduces RMS FRE and reducing RMS FRE in Equation 5–5 reduces RMS TRE, it appears that reducing N will reduce RMS TRE. However, the influence of √ N − 2 in the denominator of Equation 5–5 is greater than the influence of √ 1 − (2/N) in Equation 5–4, and as a result reducing N actually increases RMS TRE. This relationship is not entirely obvious but is observable in Equation 5–2, where the square root N appears in the denominator and RMS FRE is replaced by RMS FLE — the root of all error. The relationship between RMS TRE and N will be discussed further in the “myths” section below. Another benefit of Equation 5–4 is that it enables researchers to estimate RMS FLE. We do that by performing many registrations in the laboratory with some number N of fiducials, calculating FRE for each registration, using all those FREs to calculate RMS FRE, and then inverting Equation 5–4 to calculate RMS FLE from N and RMS FRE. Once we have done that, we can use RMS FLE in Equation 5–2 to predict RMS TRE at point p. Alternatively, we can get exactly the same result by
127
128
Image-Guided Surgery
Figure 5–3. Illustration that FRE is a misleading indicator of registration accuracy. (a) The “star” on the far left is copied from Figure 4–6e. FRE is the square root of the mean of the sum of squares of the distances between the locations of corresponding fiducials in image and physical space represented by the spheres at the tips of the stars. (b) When we get above the minimum three fiducials required for correct 3D orientation, each additional fiducial adds another term to the sum of squares (the middle panel adds the difference from the black spheres and the right panel adds the black and yellow spheres), and FRE typically increases. But, each added fiducial (c) gives more information, and the likely effect is to decrease TRE (aka improve accuracy). Unfortunately, at least one eroneous study10 has been published stating that reducing the number of fiducials will result in lower FRE — which is correct — but then concludes that this reduces TRE — which is incorrect as mathematically shown by Equation 5–1 and empirically shown in Figure 5–7.
using RMS FRE directly in Equation 5–5 without the intermediate step of calculating RMS FLE. Regardless of the method we choose, however, we will obtain an estimate of the magnitude of the error that the practitioner can expect for one particular pair of methods of fiducial localization (image and physical), one particular fiducial configuration, and one particular target. Repeated experiments and computer simulations will then verify the benefit of the four (and bonus fifth!) recommendations to lowering TRE that we presented in bold above.
Probe Localization Error (PLE) There is one other important source of intrinsic error in point-fiducial systems. After registration is complete and the probe is used to identify targeted
positions, its physical position must be monitored continuously by the tracking system as the surgeon places it at one or more anatomical target positions. The measurement of its position in the operating room during tracking suffers from some level of error, and this error adds an additional error to the TRE resulting from the registration from image space to physical space. This error is a distinct contribution to the IGS overall error, so we give it a distinct name — probe localization error (PLE). Thus, if the probe is placed at physical position r in the operating room but the system reports that it is at position r′, then the system is adding a PLE of r′ − r. This error will normally be uncorrelated with TRE and therefore will add to it in the same form as that given in Equation 5–3. For tracked probes, the RMS PLE varies with position relative to the tracking sensors and can be obtained from the IGS manufacturer or the manufacturer from which the tracking subsystem was purchased.
Chapter 5
The reader might wonder how the errors of systems that do not employ discrete fiducial markers comport with the mathematical development in this chapter — in particular, systems that use stereotactic frames or surface registration. Well, the stereotactic frame fits it well. The six intersection points of the N-bars with a given CT or MR image slice (see Figure 1–8) are equivalent to six fiducial points lying on the plane passing through the center of that slice in image space that are registered to the corresponding points in physical space (based on a model of the frame). If there is no slippage of the frame relative to the head (an insidious problem with the frame discussed in Chapter 4) and the RMS FRE of these fiducials is determined from repeated registrations, just as for discrete markers, then the RMS FLE of the frame can be determined using Equation 5–4, just as for discrete fiducial markers. Then Equation 5–2 can be used with the frame’s RMS FLE to determine its RMS TRE for a given target, just as for discrete markers. In fact, every equation and every guideline and all the myths below (note especially Myth 3 below regarding planar fiducials!) can be applied to stereotactic frames, just as they are to discrete markers. On the other hand none of the formulas in this chapter applies to sur-
Stubborn Myths Because true fiducial positions and true target positions are never known in the clinical situation, it is difficult to pin down the relationships between their errors without rigorous mathemati-
Error Analysis
n
face registration, even though that registration method registers points in one space to points in another. It would be nice if they did, because for surface registration N is usually on the order of hundreds or thousands, so the RMS registration errors promised by Equations 5–2 and 5–4 would be tiny for these systems. The reason that the mathematical development in this chapter does not apply to surface registration is that there is no point correspondence enforced in surface registration. Thus, while the pairs of points that are aligned in fiducial registration and frame registration correspond to the same point relative to the anatomy in image space and physical space (except for localization error), the points that are aligned in surface registration are merely the closest points after the registration is performed. While it is hoped that the points brought close together correspond to the same point, the fact remains that they may not. Indeed the major difficulty with surface registration is that its closest points at all which are treated like corresponding points during the registration, may not in fact be corresponding points, and, when they are not, the effect is the same as if they were corresponding points suffering from huge FLEs.
cal analysis and carefully controlled studies. In the absence of knowledge of the results of such analysis and such studies, intuition is the only remaining guide. Unfortunately, it is often a poor one. We consider here three myths that have pervaded image guidance since its beginnings. They are intuitive, and
129
130
Image-Guided Surgery
they are beguiling. Unfortunately, they are also false.
Myth 1: FRE Is an Indicator of TRE In the operating room, when a frameless navigation system is used, FRE, or some measure derived from it (“zones of accuracy”, etc), may be shown to the surgeon after registration is complete, as for example via the simple statement, “FRE = 2.3 mm.” The surgeon must decide how to interpret this information, and at this point a common, even dangerous, misconception typically arises. It is intuitive to expect that, if FRE is smaller than usual, then TRE will be smaller than usual. This expectation is wrong! Although, as pointed
out above in the section "Relating TRE and FLE", there is a direct relationship between their RMS values, as shown in Equation 5–5, the registration process removes all statistical correlation between FRE and TRE.7 The absence of correlation between FRE and TRE means that fluctuations of one about its mean are completely unrelated to fluctuations of the other about its mean. The myth that they are correlated has been part of neurosurgical lore since the beginning of the frameless revolution, and only recently have a few (unfortunately, very few) researchers begun to notice that it disagrees with observations.8,9 It’s easy to construct examples in which FRE and TRE are clearly uncorrelated. Figure 5–4 illustrates two of them. In
Figure 5–4. Schematic depicting extremes of the lack of corrlation between FRE and TRE. A. Preoperative image. Blue circles are fiducials. Blue cross is a target. B. Patient in OR. Same features are shown in red. C. Preoperative image registered to a patient in the OR. Top row: low FRE but high TRE. FRE is almost zero because the fiducials are almost perfectly aligned, but TRE is large shown by the large distance between the blue and red plus signs. Bottom row: high FRE but low TRE. FRE is large because the fiducials are badly misaligned, but TRE is small because the blue and red crosses are almost perfectly aligned.
Chapter 5
each row of the figure, (A) is the preoperative image, (B) is the patient in the OR, and (C) shows the image registered to the patient. Differences between the upper and lower row exist because of differences in the FLEs. Their magnitudes are about the same in the two figures, but their directions are different. The result is that, when the fiducials are registered, FRE is small and TRE is large in the top row, but FRE is large and TRE
Error Analysis
n
is small in the bottom row. These examples show that it is possible for FRE and TRE to be uncorrelated statistically, but they do not tell us whether or not they are indeed uncorrelated. To prove that there is no statistical correlation, theoretical derivations are required, but it is also possible to gather convincing evidence via computer simulations. Figures 5–5 and 5–6 show evidence gathered from one such
Figure 5–5. Results of 200 simulated registrations are shown for RMS FLE = 1 with four fiducials placed at typical places on the human head (e.g. top left panel Figure 5–2). For each registration the TRE value at a point in the vicinity of the deep brain is plotted as a square at the top and the corresponding FRE value for that same registration is plotted at the bottom. Only horizontal position is significant with both TRE and FRE increasing toward the right. All units are millimeters. Vertical dashed lines mark the mean values. The presence or absence of correlation in fluctuations about the means can be observed by identifying pairs of TRE-FRE belonging to the same registrations. Thus, four randomly selected registrations are highlighted using larger squares with a distinct color for each registration. Note that red has a very large FRE but a very small TRE, blue has below average values for both FRE and TRE, green has a roughly average TRE but above-average FRE, and orange has a large TRE but a small FRE. For a given localization technique, there is, in fact, no correlation between TRE and FRE fluctuations. Thus, the size of FRE gives no information about the size of TRE (and vice versa).
131
132
Image-Guided Surgery
Figure 5–6. Result of simulation showing that FRE and TRE are uncorrelated. Fiducials with random FLE added were used in 100 000 registrations, and FRE and TRE were measured for each registration. A plot of TRE versus FRE is shown with mean values plotted as dashed lines. The correlation coefficient (CC) for fluctuations of each quantity about its mean is 0.004.
simulation.7 For the simulation, four true fiducial points were chosen in the upper half of a simulated patient head in roughly the vicinity in which fiducials are placed for neurosurgerical applications. A target position was chosen near the skull base in the vicinity of the targets of deep-brain stimulation surgery. Then, using a computer program (MATLAB, R2014b; MathWorks, Natick, Massachusetts), random perturbations were repeatedly applied to the true fiducial positions in such a way that RMS FLE equaled 1 mm. For each set of perturbations, a registration was performed. For each registration, FRE and TRE were measured. A total of 100 000 registrations were performed.
Figure 5–5 examines 200 randomly selected TRE and FRE values to show that fluctuations about their means are unrelated to each other. In the upper and lower plots, pairs of TRE and FRE values are examined and found to vary independently. Figure 5–6 shows a plot of TRE versus FRE, from which the correlation coefficient (CC) was calculated. Possible CC values range from −1 to 1, where −1 and 1 each indicate perfect corelation, meaning that FREcan be used to predict TRE exactly. A coefficient of zero indicates that there is no correlation, meaning that FRE cannot be used to predict TRE at all. For this computer simulation, CC equaled 0.004, meaning that cor-
Chapter 5
rela-tion is negligible, and therefore predictability of TRE on the basis of FRE is negligible. For visual appreciation of these data, Figure 5–6 provides a plot of FRE versus TRE. (Only 1% of the points are plotted because 100 000 points cannot be distinguished in a plot small enough to fit on a page of this book, but the 1% were randomly chosen so as not to bias the plot.) It can be seen from this plot that for any value of FRE, there are roughly as many large TRE values as small values and that prediction of TRE from FRE is not possible. In fact, a paper published in 20114 proved that no formula exists which relates TRE to FRE. Not only is FRE not correlated with TRE, but no measure based on any combination of fiducial alignments or their positions is correlated with TRE. All such correlation is removed by the registration process itself as a side effect of finding the most accurate transformation. The lack of correlation means that, for a properly working IGS system, there
The independence of a single value from a statistical distribution in reference to another single value from another statistical distribution may be a hard concept to grasp in this context because the statistical distributions of TRE and FRE are related. If one considers unrelated distributions, for example the distribution of cars’ exhaust emissions and their drivers’ hat sizes, one would intuitively sense that individual values selected from these distributions would not be related, and that certainly makes sense. But what about distributions
Error Analysis
n
is no information given about TRE by FRE or by any other measure based on any of the individual FREs. Unfortunately, manufacturers have not made this fact clear, and, as a result, surgeons are being misled every day. There is, however, some value in knowing FRE. An abnormally large value, say, 8 mm for a system with which the usually observed FRE is 2 mm, for example, suggests that the system has been compromised in some way: perhaps a marker became displaced while the patient was being transported from the imaging suite to the operating room, fiducials in the image were wrongly paired with fiducials in the operating room, one or both of the two localization systems failed, or undetected patient motion has taken place during fiducial localization. In any of these situations, the system will be untrustworthy, and the abnormally large FRE is the evidence. Thus, FRE may be the bearer of very bad news, but otherwise it gives no news at all.
that, like the distributions of FRE and TRE whose RMS values are related? How can it make sense that their individual values are unrelated? Returning to our hypothetical medical school used above in the TRE section, let’s consider the relationship of the grades of two students, Brian Labb and Matt Ronic, in two clinical electives, say Facial Reconstruction Elective and Tumor Radiology Elective. Suppose that each course has a weekly quiz. Each student is analogous to one IGS system (eg, a system in OR-14 and a system in OR-20). Each week
133
134
Image-Guided Surgery
is analogous to one IGS surgery, the quiz scores in Facial Reconstruction Elective are analogous to FRE measurements, the quiz scores in Tumor Radiology Elective are analogous to TRE values, and a student’s innate intelligence is analogous to an IGS system’s FLE. Furthermore, just as we assume that the performance of each IGS system has been optimized by proper calibration to eliminate extrinsic error, we assume that the performance of each student has been optimized by proper study habits. As a result, just as FLE determines the average accuracy of an IGS system, intelligence determines the average performance of a student. However, in medical school, as in the OR, unknown random influences will cause performances to fluctuate independently about their average values. In this scenario, if Brian Labb has a lower intelligence than Matt Ronic, then Brian’s average grades can be expected to be lower than Matt’s for both Facial Reconstruction and Tumor Radiology. Similarly, if OR-14 has a lower FLE than OR-20 (eg, bone-
Myth 2: Dropping a Fiducial Will Increase Accuracy Most image guidance systems rightly encourage the surgeon to check the accuracy of the registration by touching some visible, easily identifiable points of the anatomy with the probe, preferably close to the region of interest, while watching the probe as depicted on the computer screen. If the system were perfect, the probe on the screen would be on the same anatomical point as the
implanted markers are used in OR-14 and skin-affixed markers in OR-20), then OR-14’s average TRE and FRE values can be expected to be lower than OR-20’s. On the other hand, for a given week, if Brian’s grade in Facial Reconstruction is lower than usual, Brian Labb’s grade in Tumor Radiology would not be expected to be lower than usual. Similarly, if for a given surgery, OR-14’s FRE is lower than usual, then OR-14’s TRE would not be expected to be lower than usual. So although average grades in the two courses are correlated because of their dependence on intelligence, fluctuations of the grades in the two courses are unrelated, and although changes in average FRE and TRE in the two OR’s are correlated because of their dependence on FLE, fluctuations of FRE and TRE in the two ORs are unrelated. So, in answer to our original question posed in the first paragraph of this boxed text, individual values selected from two mathematically related statistical distributions (see Figures 5–4 through 5–6) are not necessarily related.
probe on the patient. If not, then the distance on the screen between the depicted probe and the depicted anatomical point is a visual estimate of TRE at that point. TRE tends to vary slowly as the probe is moved from point to point, so a check of two or three points in the vicinity of the region of interest is sufficient validation. If TRE appears abnormally large, then the system may be compromised; if not, then the system is likely to be operating correctly. This “sanity check” is an excellent idea.
Chapter 5
However, users may then be tempted to veer away from this sensible process by touching the fiducials themselves to make a visual estimate of the individual FRE for a specific fiducial and then concluding that the registration can be improved by relocalizing the fiducial whose individual FRE is largest. Unless FRE (ie, given by Equation 5–1) is far greater than that which is normally observed, as in the example of 8 mm, given above, indicating that something has gone wrong with the system, such relocalization is likely a waste of precious surgical time. Worse still, the surgeon may decide to omit the fiducial with the largest error! The notion that dropping that fiducial is likely to increase guidance accuracy of a properly working IGS system is reinforced when the surgeon observes a concomitant reduction in FRE, and such a reduction is quite likely, because RMS FRE decreases as N decreases, as we pointed out in Equation 5–4. However, it is quite natural for the size of the individual FRE to vary from one fiducial to the next, depending on their relative positions within the fiducial configuration, even when there is no problem at all with the localization procedure, and so the individual FRE for a specific fiducial gives no information at all about its FLE. Thus, dropping a fiducial based on its FRE makes no sense at all! In fact, it makes much more sense to keep it, because, while Equation 5–4 shows that dropping a fiducial is likely to decrease FRE, which seems to indicate that the registration just got better, Equation 5–2 shows that dropping a fiducial is likely to make TRE larger. This latter effect shows that dropping a fiducial is far more likely to reduce accuracy than to increase it.
Error Analysis
n
When the worst fiducial is dropped, FRE will get smaller, making the registration look better, but TRE — and thus the likely outcome of the surgery — is likely to get worse, as shown in Figure 5–7 via another simulation. As in the previous simulation, four fiducials were repeatedly registered with a random perturbation whose RMS equals 1 mm being added to the true position each time. For each such four-fiducial registration, the fiducial with the worst individual FRE was omitted, and a second registration was performed using just the remaining three fiducials (without relocalizing any of them, as would be the case in the clinical situation). TRE is plotted versus FRE twice in the figure — once for registration performed before the “worst” fiducial was omitted and once for the registration performed afterward. The green and red lines show the means of FRE and TRE. The green arrow shows that the mean FRE was decreased, thus seemingly making the registration better; the red arrow shows, however, that mean TRE was increased, thus actually making the registration worse. Thus, this simulation clearly shows that dropping the worst fiducial is a bad idea. The finding above — that it is better to keep even the worst fitting fiducial — is certainly hard to swallow, but get ready for something even harder! The fiducial that appears to be the worst (ie, the one with the largest individual FRE) is actually likely to be the best! It is the best fiducial because it provides the most helpful information to the registration algorithm. The reason is that the fiducial that is hardest to align is typically farthest from the centroid of the fiducials, and fiducials that are spread out from the centroid give the most information.3 In a simple extension of
135
136
Image-Guided Surgery
Figure 5–7. Result of simulation showing that dropping the fiducial with the worst individual FRE tends to decrease FRE while increasing TRE. The simulation method is similar to that described for Figures 5–5 and 5–6. A plot of TRE versus FRE is shown before (a) and after (b) dropping the fiducial with the worst individual FRE and performing a new registration — with four fiducials in (a) and with the remaining three fiducials in (b). The green and red lines show the FRE and TRE means. The green arrow shows that mean FRE was decreased, but the red arrow shows that mean TRE was increased.
the experiment presented above (not shown), dropping the fiducial whose alignment is the worst is compared The concept of dropping fiducials to improve FRE was proposed by Snyderman et al in 2004.10 Although they correctly stated that dropping fiducials would improve FRE (see Equation 5–4), they erroneously concluded that this would then improve accuracy, that is, that it would lower TRE. This is not the case, as shown mathematically by Equation 5–2 and empirically by Figure 5–7. It is difficult to stamp out an erroneous clinical recommendation published in a peer-reviewed journal because the original paper gets cited in later papers, perpetuating the fallacy. This paper by Snyderman et al has i
with dropping the one whose alignment is the best. We'll spare you the plot for this one, because it looks very simibeen cited 21 times, and four of the citing papers explicitly perpetuate the fallacy by stating that dropping a fiducial will improve accuracy. These four studies have themselves been cited 68 times and so on and so on.i We sincerely hope that the intensive (and somewhat earnest) analysis in this book might make some progress toward effectively stamping out the pernicious myth that accuracy is improved by dropping fiducials, and we hope that IGS users will use all the fiducials they can for their patients, secure in the knowledge that even the worst fiducial is helping them achieve the best TRE.
And, as we show in Chapter 7, this fallacy has been incorporated into commercial software.
Chapter 5
lar to Figure 5–7, and by now you can probably guess what to expect: Dropping the worst-aligned fiducial causes more harm than dropping the best-aligned fiducial. So, regarding fiducial registration, when entering the OR, it is probably best to check your intuition at the door. Then, if something compels you to drop a fiducial even after you have read this book, then, unless you have some convincing evidence (eg, loose marker or ambiguous anatomical point) other than mere misalignment, please, resist the urge! And, if you just cannot resist, then, please, drop the one that fits best. Your patient will thank you.
Myth 3: Planar Fiducial Configurations Are Bad When choosing the placement of the fiducials, another misconception continually arises. It is the assumption that a planar configuration is somehow inferior to a nonplanar one. This idea may be based on the intuitive notion that threedimensional guidance cannot be based on a two-dimensional configuration (not true!), or it may arise from the fact that some (inferior) registration algorithms will fail when all the fiducials in one space or the other are planar (true). A few minutes of thought will convince almost anyone that this myth is false, if one considers the fact that any configuration of three markers is planar. Furthermore, stereotactic frames would not work if this myth were true, because they rely on the registration of six points that lie in the same planar image slice. In any case, with the algorithms used in the IGS systems of leading manu-
Error Analysis
n
facturers, planar arrangements are no worse than nonplanar ones. Indeed, in some cases, planar arrangements are slightly better than non-planar ones.5 Figure 5–8 presents the results of a simulation that shows the equivalence of planar and nonplanar image configurations. In this case, two sets of four fiducials were used in two separate sets of registrations. The first set was arranged as an equilateral tetrahedron; the second set was arranged as a square. The spread between the fiducials was the same in each set. As for Figures 5–6 and 5–7, the fiducials were repeatedly registered (separately) with random 1-mm RMS FLE being added each time, and then TRE was plotted versus FRE. The distribution of FRE-TRE points can be seen to be the same in the two plots, and the FRE and TRE means, shown by the green and red lines, are the same. The similarity between these two plots shows that there is no disadvantage to planar fiducial configurations.
Summary In this chapter, we have examined error in IGS. We have introduced the error triad of IGS — FLE, FRE, and TRE — and identified FLE as the cause of the other two errors. We gave expressions for FRE and TRE in terms of FLE but cautioned that these expressions relate only RMS (aka, statistical) values. We derived a set of rules for minimizing RMS TRE via a careful examination of its formula — namely, 1. Reduce FLE (eg, use boneimplanted markers), 2. Spread markers widely, 3. Arrange markers such that their centroid is close to the surgical target, 4. Use more markers, and 5. Put the CRF close to the surgical target. We exploded some myths surrounding the
137
138
Image-Guided Surgery
Figure 5–8. Results of a simulation showing that there is no disadvantage to planar fiducial configurations. The simulation method is similar to that described for Figure 5–5 through 5–7. A plot of TRE versus FRE is shown for (a) a tetrahedral (nonplanar) arrangement of 4 fiducials and (b) a square arrangement. The green and red lines show the FRE and TRE means. The distributions and the means are the same, showing that there is no disadvantage to planar fiducial arrangements.
errors of IGS, and we explained how FRE can be used to determine whether the IGS system is likely to be working properly. The purpose of this error discussion is to help the surgeon improve outcomes both by improving the likelihood that the tracked pointer is where the computer says it is and by knowing when that likelihood is so low that the best approach is to set the IGS probe aside and use anatomical knowledge.
References 1. Hill DLG, Maurer CR Jr, Maciunas RJ, Barwise JA, Fitzpatrick MJ, Wang MY. Measurement of intraoperative brain surface deformation under a craniotomy. Neurosurgery. 1998;43(3):514–526. 2. Fitzpatrick JM, Hill DLG, Maurer CR Jr. Image registration. In: Sonka M, Fitzpat-
rick JM, eds. Handbook of Medical Imaging: Medical Image Processing and Analysis. Bellingham, WA: SPIE; 2000;2:337–513. 3. Fitzpatrick JM, West JB, Maurer CR Jr. Predicting error in rigid-body pointbased registration. IEEE Trans Med Imaging. 1998;17(5):694–702. 4. Danilchenko A, Fitzpatrick JM. General approach to first-order error production in rigid point registration. IEEE Trans Med Imaging. 2011;30(3):679–693. 5. West JB, Maurer CR Jr. Designing optically tracked instruments for imageguided surgery. IEEE Trans Med Imaging. 2004;23(5):533–545. 6. Sibson R. Studies in the robustness of multidimensional scaling: perturbational analysis of classical scaling. J R Stat Soc B. 1979;41(2):217–229. 7. Fitzpatrick JM. Fiducial registration error and target registration error are uncorrelated. Proc SPIE 7261, Medical Imaging. 2009:Visualization, Image-Guided Procedures, and Modeling, 726102.
Chapter 5
8. Steinmeier R, Rachinger J, Kaus M, Ganslandt O, Huk W, Fahlbusch R. Factors influencing the application accuracy of neuronavigation systems. Stereotact Funct Neurosurg. 2000;75(4): 188–202. 9. Woerdeman PA, Willems PWA, Noordmans HJ, Tulleken CAF, Berkelbach van der Sprenkel JW. Application accu-
Error Analysis
n
racy in frameless image-guided neurosurgery: a comparison study of three patient-to-image registration methods. J Neurosurg. 2007;106(6):1012–1016. 10. Snyderman C, Zimmer LA, Kassam A. Sources of registration error with image guidance systems during endoscopic anterior cranial base surgery. Otolaryngol Head Neck Surg. 2004;131(3):145–149.
139
6 Best Practices for Use of IGS So now that we understand how IGS works, and before we begin to review clinical applications, perhaps we should ask an obvious question: “Why use IGS at all?” We propose this question in order to avoid falling into the trap of using new technology just because it’s there. As we mentioned in Chapter 1, where we presented Dr Bull’s use of IGS in 1896 to direct surgical extraction of buckshot from a patient’s hand, and elsewhere throughout this book, IGS seems to offer an obvious advantage of localization of anatomy. But, underlying this obvious assumption is the existence of a perfect IGS system without error, which, as we know from the prior chapters on error in IGS systems, cannot exist in the real world. So the more pertinent question is, “Given the error associated with IGS systems, should we use them?” This question takes us from the ideal world to the real world and leads us to the following very practical questions: n Is IGS safer for patients? n Does IGS make better surgeons? n Is there a medicolegal reason to
use IGS?
Within this chapter, we review the current literature on each of these questions and try to reach evidence-based answers.
Who Is Using IGS? Knowing who uses IGS in otolaryngology is pertinent to our goal because that knowledge will help us answer the questions posed above. By far the largest group within otolaryngology using IGS is rhinology. Utilizing the membership of the American Rhinologic Society (ARS), two groups have conducted surveys via mailings — one in 20051 and a second in 2010.2 In 2005, 86.4% of respondents reported having IGS systems available to them, and 18.1% reported never using IGS. Five years later, in 2010, 94.6% reported availability of IGS, with approximately 7% never using IGS. In both surveys, exposure to IGS technology in residency did not correlate with use or nonuse of IGS. Comparing the two data sets, IGS is most frequently used for revision endoscopic sinus surgery (ESS) — excluding revision maxillary antrostomies — and
141
142
Image-Guided Surgery
complex cases including skull base tumor resections, skull base defects with cerebrospinal fluid (CSF) leaks, Lothrop procedures, and orbital and optic nerve decompression. What about the use of IGS in other subspecialties of otolaryngology? Perhaps the next most relevant subspecialty would be neurotology because the vast majority of these procedures are performed in a rigid (ie, boneencased) operative field, where IGS has been shown to be most accurate. The 2005 ARS survey cited above was estimated to have occurred 10 years after early adoption of IGS at tertiary care centers, and over the next five years there was a notable but not enormous increase in terms of the availability of IGS (86.4% had access to IGS in 2005 and 94.6% in 2010). Given that rhinologists typically operate in the same operating centers as their neurotology colleagues, IGS has likely been available for their use over this time frame. To date, no surveys regarding the use of IGS in the field of neurotology have been undertaken. In fact, there are only a few reports regarding its use, and these come from tertiary care centers describing use of IGS in only a few clinical scenarios including middle cranial fossa approaches and/or cases involving complex surgical anatomy.3,4 This lack of routine use of IGS in neurotology appears to be confined to the “otology” realm because the “neuro” realm has likely been using IGS during combined cases with neurosurgery. And, as mentioned in Chapter 1, use of IGS is relatively ubiquitous in neurosurgery. Thus, the vast majority of the data we have available to answer the i
questions posed above come from the rhinology literature, with some input from the neurosurgical literature.
Is IGS Safer for Patients? In the ideal world, technological advances improve patient outcomes and overall health. However, just as IGS must face error associated with its realworld existence, technological advances must face the real world where marketdriven economics impose an often paradoxical mission to medical device companies: improve health but also make a profit). We can safely assume that both the desire to improve health and the desire to make a profit drove IGS technological development and clinical acceptance. Additionally, when medical devices are ready for clinical testing and commercialization, they must meet regulations governed by the Food and Drug Administration (FDA) via either 510(k) clearance, which requires a demonstration of substantial equivalence to a device that was commercially available prior to May 28, 1976, or via the more demanding Premarket Approval Route, which involves explicit demonstration of safety and effectiveness of the device.i Most, if not all, IGS systems that are currently commercially available were cleared via the 510(k) route and entered the market without demonstration of superiority over existing surgical paradigms. Early clinical reports of IGS in ENT were focused on describing the use of the technology in select, often unique, cases rather than randomized controlled trials (RCTs), which are neces-
http://www.fda.gov/downloads/MedicalDevices/.../UCM284443.pdf (Accessed February 26, 2015).
Chapter 6
sary to demonstrate superiority over existing techniques. Moreover, early reports predicted that almost unachievable enrollment figures5,6 would be necessary to perform a definitive study showing superiority. The consensus was that the benefits of IGS were obvious and a definitive trial to demonstrate superiority would be unethical.6,7 Nonetheless, data are beginning to emerge suggesting that interventions with IGS, at least in rhinology, are better for patients than ESS without IGS. Fewer Complications? Surgeons, first and foremost concern is with patients’ well-being. Thus, an appropriate first question to ask is, “Are there fewer complications with IGS?” Given that the incidence of complications from ESS is quite small (< 1.5% incidence of major complications),8 the size of an RCT required to detect a statistically significant difference would be quite large. Smith et al6 predicted that such an RCT would require 3017 patients in each arm of a trial of IGS versus non-IGS in order to detect a difference in complication rate of 1% versus 2% with a power of 90% and P set to .05 and would require 2254 in each arm for a power of 80%. Thus, we will have to settle for the next best thing, which is meta-analysis. Dalgorf et al9 performed one in 2013 and concluded that the use of IGS is associated with fewer majorii complications and a smaller total number of complications compared to nonIGS ESS. This meta-analysis included 13 articles, pared down from 192, with one being a randomized controlled ii
Best Practices for Use of IGS
n
trial. Included in these 13 studies were 1119 patients who underwent IGS ESS and 1282 who had ESS without IGS. Fourteen (1.25%) major and 44 (3.9%) total complications were noted in the IGS group and 42 (3.3%) major and 81 (6.3%) total in the non-IGS group. The meta-analysis showed that use of IGS was associated with a statistically significant lower relative risk of major complications (risk ratio [RR], 0.48; 95% confidence interval [CI], 0.28–0.82; P = .007) as well as total complications (RR, 0.66; 95% CI, 0.47–0.94; P = .02). The incidence of major complications for ESS is quite small — 1.5% — but the use of IGS would be expected to reduce it to 0.75%. Recently, a sizable retrospective review concluded that CT-guided IGS was associated with a decreased rate of CSF leak after transsphenoidal pituitary resection. Chung et al10 analyzed 48 848 transsphenoidal pituitary resections be-tween 2007 and 2011 using the Nationwide Inpatient Sample. The use of CT-guided IGS (n=1768 cases) was associated with a lower incidence of CSF leaks (1.1%) compared to no IGS (1.9%). Chung et al10 also report that the use of MRI guided IGS (n =1228) was associated with a statistically significant higher incidence of CSF leak (2.5%, P < .001). They hypothesize that this is due to either more co-morbidities, more complex disease, or the lack of definition of bony anatomy in MRI. We would add that the lower overall fidelity of MRI as described in Chapter 2 (ie, more distortion and larger voxels in comparison to CT) may have contributed to this finding as well. In short, the literature seems to support that
Major complications were defined as “(1) inadvertent entry into an area beyond the nasal cavity and/or paranasal sinus, (2) postoperative bleeding requiring surgical or angiographic intervention, and (3) the necessity to abort the procedure for any surgical reason.”
143
144
Image-Guided Surgery
CT-based IGS is associated with lower complication rates but the jury remains out regarding MRI-based IGS (P < .001). Less Need for Subsequent or Revision Surgery? Another concern for patients is that the surgical intervention might not achieve its intended goal and as a result additional or revision surgery will be necessary. IGS theoretically would allow more complete dissection, obviating the need for revision surgery. Is this the case? As with complications, this question is difficult to answer without large RCTs, and although cadaveric studies suggest that IGS allows more complete resection of tissue, such studies have been criticized because cadaveric complications have no sequelae prompting surgeons to be more aggressive in the laboratory. To date, the best clinical study on this subject is that by Sunkaraneni et al11 who performed both a retrospective chart review of 355 cases over a 3.5-year period of patients who underwent ESS at their institution and also performed a meta-analysis on eight studies including their own data. Reviewing their cohort, which was disproportionate with 333 having IGS ESS and 22 not having IGS, they found the need for revision surgery within 1.5 years of the original surgery was statistically significantly higher in those treated without IGS (9 of 22, 40.9%) versus those treated with IGS (79 of 333, 23.7%), P = .001. Their meta-analysis found that the majority of the included studies showed a reduction in the need for revision surgery when IGS was used, but this finding was not statistically significant. iii
Does IGS Help Make Better Surgeons? As early as 2000,12 there appeared substantial evidence that IGS helps make better surgeons. In this study, residents with limited ESS experience were given instruction and then allowed to perform surgical dissections on cadavers with or without IGS. Use of IGS was associated with statistically significant improvement in identification of specified anatomy (97% correct with IGS versus 76.8% without IGS). Additionally, in the non-IGS group there were three major complications (all intracranial violations) while no major complications were noted in the IGS group. These findings were further substantiated in 201113 and most recently in 2015 in the neurosurgical literature where augmented reality systems (see Chapter 8) were shown to improve performance in less-experienced surgeons as demonstrated by decreased time of task completion and total tool-path distance.14
Professional Society Position Statements It is clear from the above that the evidence supporting use of IGS is, well, unclear! So, what are professional societies’ positions? The American Academy of Otolaryngology–Head and Neck Surgery (AAO-HNS) first opined on this subject in a position statement in 2002, revised it in 2005, reaffirmed it in 2012, and most recently revised it in 2014. Quoting from their websiteiii:
http://www.entnet.org/?q=node/929 (Accessed February 24, 2015).
Chapter 6
The American Academy of Otolaryngology–Head and Neck Surgery endorses the intraoperative use of computer-aided surgery in appropriately select cases to assist the surgeon in clarifying complex anatomy during sinus and skull base surgery. There is sufficient expert consensus opinion and literature evidence base to support this position. This technology is used at the discretion of the operating surgeon and is not experimental or investigational. Furthermore, the American Academy of Otolaryngology–Head and Neck Surgery is of the opinion that it is impossible to corroborate this with Level 1 evidence. These appropriate, specialty specific, and surgically indicated procedural services should be reimbursed when used by qualified physicians regardless of the specialty. Examples of indications in which use of computeraided surgery may be deemed appropriate include: 1. Revision sinus surgery. 2. Distorted sinus anatomy of development, postoperative, or traumatic origin. 3. Extensive sino-nasal polyposis. 4. Pathology involving the frontal, posterior ethmoid and sphenoid sinuses. 5. Disease abutting the skull base, orbit, optic nerve or carotid artery. 6. CSF rhinorrhea or conditions where there is a skull base defect. 7. Benign and malignant sino-nasal neoplasms.
The American Rhinologic Society further adds the followingiv: iv
Best Practices for Use of IGS
n
The American Academy of Otolaryngology–Head and Neck Surgery and American Rhinologic Society endorse the intraoperative use of computeraided surgery in appropriately select cases to assist the surgeon in clarifying complex anatomy during sinus and skull base surgery. This technology is used at the discretion of the operating surgeon and is not experimental or investigational. These appropriate, specialty specific, and surgically indicated procedural services should be reimbursed whether used by neurosurgeons or other qualified physicians regardless of the specialty.
The importance of professional society position statements is evident when comparing the above to insurance carriers’ coverage of benefits. Coverage Position CIGNAv HealthCare covers image-guided sinus surgery as medically necessary for any of the following indications: n revision sinus surgery n distorted sinus anatomy of
developmental, postoperative, disease, or traumatic origin n extensive sinonasal polyposis n pathology involving the frontal, posterior ethmoid, or sphenoid sinuses n disease abutting the skull base, orbit, optic nerve, or carotid artery n CSF rhinorrhea or conditions where a skull base defect exists
http://www.american-rhinologic.org/image_guided_surgery (Accessed February 24, 2015). https://my.cigna.com/teamsite/health/provider/medical/procedural/coverage_positions/ medical/mm_0257_coveragepositioncriteria_image_guided_sinus_surgery.pdf (Accessed February 24, 2015).
v
145
146
Image-Guided Surgery
n benign and malignant sino-
nasal neoplasms CIGNA HealthCare does not cover image-guided endoscopic sinus surgery for any other indication because it is considered not medically necessary. An almost verbatim posting exists under United HealthCare’s online reimbursement policy.vi
Is IGS Financially Sustainable? IGS systems typical cost over $100 000, including hardware and software, and they require annual maintenance and updates costing thousands of dollars per year. And, reimbursement via the appropriate Current Procedural Terminology (CPT) add-on code +61782 (replaced +61795 in 2011) for “Stereotactic computer-assisted (navigational) procedure, cranial, extradural” is sporadic. So, does it make financial sense to use IGS? Because IGS systems are not free, the question becomes: Do the benefits of IGS translate into cost savings that offset the expenditures? Early reports indicated that IGS ESS took longer and was more expensive than non-IGS ESS. In 1999, Metson et al15 reported that IGS took 17.4 minutes longer than non-IGS cases primarily due to setup of the system and cost $496 more per case. This report, however, failed to factor in the potential benefit of the decreased need for revision surgery secondary to use of IGS. vi
A more recent study that did include the decreased need for revision surgery found that IGS was financially more favorable than non-IGS interventions.16 This study equally distributed the cost (purchase price plus annual maintenance) of the IGS system among the 19 IGS enrollees and reported a net savings of £ 71 000 (≈$110 000). It is notable that the vast majority of this savings was realized by the insurer, who did not have to pay for additional procedures or additional time off work. A small savings was realized by the medical center because operative time with IGS was noted to be quicker than that without IGS. This finding of cost savings secondary to decreased operative time enabled by IGS has also been demonstrated in neurosurgery. In patients undergoing transsphenoidal resection of pituitary tumors, the additional cost of intraoperative CT scanning and use of an IGS system was more than offset by the cost savings from decreased operative time of approximately 10%. Overall cost savings was approximately 3% when intraoperative CT and IGS were used.17 In the recent retrospective study of 48 848 transsphenoidal pituitary resections mentioned above, CT-guided IGS resections had both significantly shorter length of hospitalization (2.9 days) compared to non-IGS (3.7 days, P < .001) and lower costs and charges with CT-guided IGS charges and costs $47 589 and $16 748, respectively, and non-IGS charges and costs $62 629 and $20 530, respectively.10 Additionally, the use of image guidance has been shown
https://www.unitedhealthcareonline.com/ccmcontent/ProviderII/UHC/en-US/Main%20 Menu/Tools%20&%20Resources/Policies%20and%20Protocols/Medicare%20Advantage%20 Reimbursement%20Policies/S/StereoCAVNav_10162012.pdf (Accessed February 24, 2015).
Chapter 6
to obviate the need for neuropathologic confirmation of tumor biopsy, resulting in substantial cost savings for institutions.18
Less Litigation? Perhaps the least important of all the questions posed in this chapter, but an important reality check as to what society “thinks” of IGS, is this one: Do surgeons get sued more if they do not use IGS? There is some precedence for defensive actions in otolaryngology, where use of facial-nerve monitoring, which is near ubiquitous in parotid and otologic surgery, is often employed for medicolegal reasons.19 Regarding IGS, Eloy et al20 surveyed the Westlaw legal database over the period from January 2004 to April 2013 and identified thirty jury verdicts and settlements involving claims of medical malpractice in sinus surgery. Only four of these mentioned IGS, and all four were resolved with a verdict in favor of the plaintiff — 100% of those who used IGS lost their lawsuits! Eloy et al conclude that if widespread use of IGS is motivated, at least in part, to limit legal liability, there is no evidence to support this stance and that “IG should be considered as a tool used by a skilled surgeon rather than a means to decrease litigation.” Like most others who report in the medicolegal realm, they conclude that adverse outcomes and inadequate informed consent are the primary causes of litigation. Another legal concern regarding IGS is the question of whether it already is, or should be, considered “standard of care”. Although the definition of “standard of care” has evolved and will likely continue to evolve, at present it
Best Practices for Use of IGS
n
means “that which a minimally competent physician in the same field would do under similar circumstances”.21 The concept of the “minimally competent physician” is often interpreted to mean one who practices in the same community as the physician under legal scrutiny. Given this definition, IGS may be considered “standard of care” in certain situations (eg, complex revision skull base surgery) and not in others (eg, maxillary antrostomies), and it is unlikely that IGS will ever become “standard of care” for all surgical interventions.
Overview To show definitively that IGS is superior to non-IGS, a large RCT would be necessary. How might non-IGS be better than IGS? Well, one might suppose that surgeons could become too dependent on IGS, become too aggressive, and have more major complications when inaccuracies arise during surgical dissection. To disprove this supposition, or disprove the contrary one that IGS is superior to IGS, would require large sample sizes with thousands of patients in each treatment arm. It appears unlikely that such a study will ever occur given the position statement from the AAO-HNS that definitive evidence, via an RCT, is “impossible” and suggestions by others6 that such a study would be unethical. Thus, we will have to settle for, at best, meta-analyses and well-controlled prospective studies that strongly suggest that IGS is beneficial. One concern expressed by virtually all involved in IGS and reiterated by the current authors is that IGS should
147
148
Image-Guided Surgery
never be used as a surrogate for intimate knowledge of the anatomy of the surgical field. And, if ever a discrepancy arises between IGS and anatomical knowledge, the surgeon should pause, reassess, and, if conflict between anatomical knowledge and IGS persists, trust anatomical knowledge!
References 1. Orlandi RR, Petersen E. Image guidance: a survey of attitudes and use. Am J Rhinol. 2006;20(4):406–411. 2. Justice JM, Orlandi RR. An update on attitudes and use of image-guided surgery. Int Forum Allergy Rhinol. 2012;2 (2): 155–159. 3. Wanna GB, Carlson ML, Blachon GS, et al. Implantation of the completely ossified cochlea: an image-guided approach. Otol Neurotol. 2013;34(3):522–525. 4. Balachandran R, Tsai BS, Ramachandra T, et al. Minimally invasive imageguided access for drainage of petrous apex lesions: a case report. Otol Neurotol. 2014;35(4):649–655. 5. Kingdom TT, Orlandi RR. Image-guided surgery of the sinuses: current technology and applications. Otolaryngol Clin North Am. 2004;37(2):381–400. 6. Smith TL, Stewart MG, Orlandi RR, Setzen M, Lanza DC. Indications for imageguided sinus surgery: the current evidence. Am J Rhinol. 2007;21(1):80–83. 7. Smith GC, Pell JP. Parachute use to prevent death and major trauma related to gravitational challenge: systematic review of randomised controlled trials. BMJ. 2003;327(7429):1459–1461. 8. McMains KC. Safety in endoscopic sinus surgery. Curr Opin Otolaryngol Head Neck Surg. 2008;16(3):247–251. 9. Dalgorf DM, Sacks R, Wormald PJ, et al. Image-guided surgery influences perioperative morbidity from endoscopic
sinus surgery: a systematic review and meta-analysis. Otolaryngol Head Neck Surg. 2013;149(1):17–29. 10. Chung TK, Riley KO, Woodworth BA. The use of image-guidance during transsphenoidal pituitary surgery in the United States. Am J Rhinol Allergy. 2015; 29(3):215–220. 11. Sunkaraneni VS, Yeh D, Qian H, Javer AR. Computer or not? Use of image guidance during endoscopic sinus surgery for chronic rhinosinusitis at St Paul’s Hospital, Vancouver, and metaanalysis. J Laryngol Otol. 2013;127(4): 368–377. 12. Casiano RR, Numa WA Jr. Efficacy of computed tomographic image-guided endoscopic sinus surgery in residency training programs. Laryngoscope. 2000; 110(8):1277–1282. 13. Stelter K, Ertl-Wagner B, Luz M, et al. Evaluation of an image-guided navigation system in the training of functional endoscopic sinus surgeons: a prospective, randomised clinical study. Rhinology. 2011;49(4):429–437. 14. Marcus HJ, Pratt P, Hughes-Hallett A, et al. Comparative effectiveness and safety of image guidance systems in neurosurgery: a preclinical randomized study. J Neurosurg. 2015;123(2):307–313. 15. Metson R, Cosenza M, Gliklich RE, Montgomery WW. The role of imageguidance systems for head and neck surgery. Arch Otolaryngol Head Neck Surg. 1999;125(10):1100–1104. 16. Masterson L, Agalato E, Pearson C. Image-guided surgery: practical and financial experiences from a UK centre 2001–2009. J Laryngol Otol. 2012;126(12): 1224–1230. 17. Eboli P, Shafa B, Mayberg M. Intraoperative computed tomography registration and electromagnetic neuronavigation for transsphenoidal pituitary surgery: accuracy and time effectiveness. J Neurosurg. 2011;114(2):329–335. 18. Shooman D, Belli A, Grundy PL. Imageguided frameless stereotactic biopsy
Chapter 6
Best Practices for Use of IGS
n
guidance in endoscopic sinus surgery: without intraoperative neuropathois it associated with decreased medicological examination. J Neurosurg. 2010; legal liability? Int Forum Allergy Rhinol. 113(2):170–178. 2013;3(12):980–985. 19. Lowry TR, Gal TJ, Brennan JA. Patterns of use of facial nerve monitoring during 21. Moffett P, Moore G. The standard of care: legal history and definitions: the parotid gland surgery. Otolaryngol Head bad and good news. West J Emerg Med. Neck Surg. 2005;133(3):313–318. 2011;12(1):109–112. 20. Eloy JA, Svider PF, D’Aguillo CM, Baredes S, Setzen M, Folbe AJ. Image-
149
7 Surgical Systems
The image-guided surgical market for otolaryngology is dominated by three companies: the surgical navigation division of Medtronic, Brainlab, and the surgical navigation division of Stryker. In addition to these big three, at least two smaller players are entering the market — Fiagon and ClaroNav. Market reports regarding IGS are readily available but expensive because their intended use is to guide large-
At one point in the last ten years, the surgical division of General Electric, which had purchased Visualization Technologies, Inc with its InstaTrak system (Figure 1–12 and Figure 7–1), was the largest supplier of IGS for ENT needs, but it is now noticeably absent from the market. Why this occurred is not entirely clear, but GE
scale investors, who are willing and able to pay for the considerable effort required to put these reports together. There are at least two very relevant, relatively recent market reports. The first is from Medtech, entitled US Markets for Image-Guided Surgery. They offer a 2012 analysis (forecasts through 2016) containing 270 pages with 56 exhibits for a single-site license fee of $4750 or for a corporate-wide fee of $8550.ii The
Healthcare did recall certain OEC InstaTrak 3500 Carts at least in part because of the potential for the cart to tip over when the arm of the imaging device was extended during use, and their manufacturing site was incapacitated for some time secondary to issues raised during an FDA inspection.i
i
http://www.fda.gov/NewsEvents/Newsroom/PressAnnouncements/2007/ucm108828.htm (Accessed January 4, 2016); http://www.accessdata.fda.gov/scripts/enforcement/enforce_ rpt-Product-Tabs.cfm?action=select&recall_number=Z-0524-2014&w=12262013&lang=eng (Accessed December 29, 2015).
151
152
Image-Guided Surgery
Figure 7–1. The GE Instatrack IGS dominated the market during the 1990s and early 2000s. It is no longer commercially available. Image provided courtesy of GE Healthcare, Inc.
second is available from Millennium Research group and is entitled Surgical Navigation Systems, US, 2015, Market Analysis. This 87-page report, complete with 47 tables, 17 figures, and three appendices, has a retail price of $7995.iii (Subliminal authors’ message: Our book is a bargain!) The data contained in such reports, although aimed at investors, can nevertheless convey some information that may be of interest to clinicians, including market size, market share, and future trends (eg, prognostications about future market activity). What is perhaps most striking about the various offerings is that they are more similar than they are different. Brainlab and Medtronic each use NDI subsystems for optical tracking. Accordingly, if their systems are each used correctly, their accuracies are strikingly ii
similar. Stryker uses an optical tracking system developed in-house. Brainlab uses an NDI subsystem for electromagnetic tracking, and Medtronic does development of electromagnetic tracking systems in-house. As is common in many successful businesses, Apple, Inc, being a prime nonmedical example, each of these companies exploits compatibility among its own devices. Thus, the engineers at each of these companies design their devices to interface especially easily with other products that it sells within the same product line, and it emphasizes this feature in its marketing campaigns. This strategy allows the companies to maximize sales in IGS areas and increase face-time with their consumers (IGS surgeons) so that additional products can be promoted, leading to potentially more sales. For example, Medtronic’s IGS system, the
https://lifescienceintelligence.com/market-reports-page.php?id=A540 (Accessed December 29, 2015). h ttp://www.mrg.net/Products-and-Services/Syndicated-Report.aspx?r=RPUS51IG14 (Accessed August 10, 2015).
iii
Chapter 7
StealthStation S7, interfaces nicely with the company’s own portable CT scanner, the O-arm; Brainlab offers both the Curve and the Kick, two IGS systems that interface nicely with the Airo portable CT scanner, which is sold by Mobius, a company that has a business relationship with Brainlab to offer the system; and Stryker, which offers both the iNAV3 and the ADAPT Systems, has recently associated itself with Neurologica and its line of portable CT scanners. Note that analysis of the geometric fidelity of the Neurologica scanners, discussed in Chapter 2, shows that without correction, the images are not suitable for IGS because the tank-tread drive system that advances the scanner across the body during acquisition of a set of slices can lead to slice misalignment, leading to geometric distortion.
Current FDA-Cleared IGS for ENT Medtronic Medtronic is one of the largest companies in the world — recently listed at #249 by Forbes magazine with a market cap of over $100 millioniv— but it has very humble beginnings being founded
iv
Surgical Systems
n
by an electrical engineer, Earl Bakken, and his brother-in-law, Palmer Hermundslie, in 1949 in a garage in Minneapolis. They first focused on repairing medical equipment and subsequently, at the request of a prominent cardiothoracic surgeon, Walton Lillehei, built a miniaturized pacemaker power supply in 1957 and then one with an implantable power supply in 1960, which led Medtronic to exponential growth and a global presence.v As often happens with large corporations for a myriad of reasons but predominately to diversify risk, radical innovation typically comes via acquisition of small, startup companies, and that is indeed how Medtronic got into the surgical navigation business. The startup was an offshoot of an engineering laboratory at Southern Illinois University run by Kurt Smith, PhD. His uncle was Chief of Surgery at St Louis University and put him in contact with Richard Bucholz, MD, and with several other inventors, he received US Patent 6491699 entitled Instrument guidance method and system for image guided surgery. According to a Medtronicproduced video on the history of the company, Stealth Technologies began in 1991 with engineer Kurt Smith tinkering with computers in his Missouri garage.vi Smith moved the company to Colorado in 1994. In 1995, SofamorDanek bought Stealth, and in 1998, Medtronic bought SofamorDanek.vii
http://www.forbes.com/global2000/list/#industry:Medical%20Equipment%20%26%20Supplies (Accessed December 30, 2015). v http://www.medtronic.com/about-us/company-profile/medtronic-history/index.htm (Accessed December 29, 2015). vi https://www.youtube.com/watch?v=8Nmdiq7LyRI (Accessed January 4, 2016). vii http://www.seattlepi.com/lifestyle/health/article/Stealth-navigation-device-promisesfewer-surgical-1161819.php (Accessed December 29, 2015).
153
154
Image-Guided Surgery
Medtronic StealthStation and Fusion The current neurosurgical IGS system offered by Medtronic is the portable SteathStation S7 mentioned above (Figure 7–2). A model with similar features that is permanently mounted in the OR is the i7, which — with appropriate installation — can be used in intraoperative MRI suites. In either system, underneath the neat external packaging is a relatively standard IGS system, as depicted in Figure 1–1. In fact, Section 5–2 of the StealthStation S7 Treatment Guidance System Manual details how to access the inner components, including a computer with connections to the tracking equipment and video monitor
A
displays. And, in a play to their target market — s urgeons — a high-fidelity stereo system that can play MP3 files is also included! (As is a warning about the possible deleterious effects of loud noise exposure [pp. 1–6, StealthStation S7 Treatment Guidance System Manual]!) The system offers IR and/or EM tracking, with the IR tracking consisting of an NDI Spectra camera (Figure 3–5; pp. 4–5, StealthStation S7 Treatment Guidance System Manual) and the EM tracking system produced in-house by Medtronic. Imaging data can be uploaded via a USB port or via a network connection directly to the hospital’s Picture Archiving and Communication System (PACS). For ENT procedures, either of two Medtronic IGS systems are
B
Figure 7–2. Medtronic’s IR and EM offerings — StealthStation s7 (A) and Fusion (B). Images provided courtesy of Medtronic, Inc.
Chapter 7
appropriate — the S7 or a scaled-down version called the Fusion, which provides only EM tracking. Pearls From Medtronic System and Training Manuals User manuals are supposedly available online, but they can be frustratingly difficult to find.viii The title page of all the system manuals states, “Read this manual completely prior to using this device.” We hypothesize that few surgeons have in fact read any of these manuals. Nonetheless, they are full of useful information, and any serious IGS user should read them in their entirety. Medtronic also offers a training manual with questions after each chapter that we found useful. Pearls that stood out from our review of the StealthStation S7 Treatment Guidance System Manual include the following (similar information is provided in the StealthStation i7 System Manual): n The intended use is explicitly
stated: “Your Medtronic computer-assisted surgery system and its associated applications are intended as an aid for precisely locating anatomical structures in either open or percutaneous procedures. Their use is indicated for any medical condition in which the use of stereotactic surgery may be appropriate, and where reference to a rigid viii
Surgical Systems
n
anatomical structure, such as the skull, a long bone, or vertebra, can be identified relative to a CTor MR-based model, fluoroscopic images, or digitized landmarks of the anatomy” (pp. 1–4, StealthStation 57 Treatment Guidance System Manual). n A caveat that use of the system is “as an adjunct for surgical guidance” and is “not a replacement for the surgeon’s knowledge, expertise, or judgment.” Furthermore, “if system navigation seems inaccurate and recommended steps to restore accuracy are not successful, abort use of the system” (pp. 1–5). n The need for the patient reference frame to be rigidly affixed to the patient and the warning that any motion of the reference frame relative to the patient after registration results in “inaccurate navigation” (pp. 2–3). n The need for the system to warm up for at least 2 minutes to allow the IR trackers to thermally expand to a stable configuration (see Chapter 3) (pp. 2–4). n The need to keep “ferrous and other conductive metals” at least 6 inches away from the EM source (pp. 2–7). n The need for all instruments to be in correct geometric alignment (ie, “not bent or otherwise
The authors could not find accessible copies online despite being assured of their availability. Medtronic provided us with the following electronic versions upon request, which are referenced in this chapter: (1) Fusion Navigation System Manual; Part Number 9733478, revision 11 (2) StealthStation AXIEM™ System Manual; Part Number 9732573, revision 8 (3) StealthStation i7 System Manual; Part Number 9734120, revision 6 (4) StealthStation S7 Treatment Guidance System Manual; Part Number 9733782, revision 16 (5) StealthStation/Fusion ENT Navigation Solution Training Manual
155
156
Image-Guided Surgery
damaged”) to avoid the risk of loss of navigational accuracy (pp. 2–13). n To reduce the risk of electrical shock, no one should touch the patient and the machine at the same time (pp. 3–2, 3–6). n That the EM source, named the AXIEM, is compatible “with Medtronic implantable cardiac device families” (pp. 4–11). Additional information conveyed in the StealthStation/Fusion ENT Navigation Solution Training Manual, a 222-page “active learning” workbook, includes the following: n Medtronic’s EM source, which the
company refers to as the AxiEM, has a FOV of approximately 50 cm in diameter, starting approximately 50 mm from the face of the EM source (p. 8). This FOV is similar to that of NDI’s Aurora systems (Figure 3–15), although the shape of the EM volume appears subjectively different (Figure 1–3 in the StealthStation/ Fusion ENT Navigation Solution Training Manual). n Ferromagnetic objects must be kept more than 6 inches away from the EM source (p. 8). including the holder post (p. 59). n The system can track four instruments at once (p. 14). n Workflow for functional endoscopic sinus surgery (FESS) as well as anterior skull base (ASB) and lateral skull base (LBS) options is available, as are various registration options (p. 36; See section regarding Medtronic registration below).
n After registration, the system
will display predicted error graphically, although the method for its calculation is not specified. Predicted errors greater than 5 mm will cause one not to be able to navigate with the system and will require reregistration (p. 112). n For lateral skull base cases, the system is claimed to be most accurate when positioned pointing at the patient’s ear from the front of their face and 4 to 8 inches away from the ear (p. 61). n Interference with the EM system can be caused by metal objects nearby (eg, OR table, Mayo stand, towel clips, magnetic tool pads) and by cell phones, which many surgeons keep in the chest pocket of their scrubs (p. 63). n Both pressure-affixed and boneimplanted options are available for the reference frame (p. 82) (Figure 7–3). Registration With Medtronic Systems Registration is central to IGS. It’s so important that we spent an entire chapter, Chapter 4, on it, and Medtronic spends two of its learning modules on it (Self-assessments 11, pp. 97–105, and 12, pp. 107–155, StealthStation/ Fusion ENT Navigation Solution Training Manual). So how does one perform registration with the Medtronic system? Recalling from Chapter 4 that all registration occurs via point-based registration, Medtronic allows selection of points for registration as (a) a set of points distributed over an anatomical surface, (b) a set of discrete anatomi-
Chapter 7
A
Surgical Systems
n
B
Figure 7–3. Coordinate reference frames (CRFs) for Medtronic system — bone-affixed (A) and pressure affixed (B). Images provided courtesy of Medtronic, Inc.
cal landmarks, or (c) a set of fiducials. Medtronic refers to its surface registration technique as “Tracer” (p. 98). With Tracer, once the CRF is attached to the patient, a probe is employed to trace over the supraorbital rim and nasal dorsum, with the manual noting that, if one pushes too hard with the probe, the skin will be deformed and the registration will be poor. Another way to initiate “Tracer” is “3-Point Tracer” (p. 102), in which selection of three anatomical points (tip of nose and two points on the forehead) is followed by using a probe to trace over the skin of the brow. Finally, all Medtronic systems allow point-based registration with any combination of four to eight anatomical landmarks, skin-affixed fiducials, and/or bone-implanted fiducials via a process they refer to as “PointMerge” (p. 107). One of the features (or lack thereof) of all commercial IGS systems is that they provide only indirect evidence of the quality of a registration (eg, good, medium, or failed).
Users may find the subjective reporting of FRE as frustrating as it might be to find that instead of a speedometer, their new car was outfitted with only a light that showed red for traveling way too fast, yellow for traveling a little too fast, and green indicating traveling below the speed limit — frustrating because we want to know exactly how fast we are traveling, but perhaps understandable because, as we have pointed out in Chapter 5, error estimation is only statistical. The astute reader might ask, “Well in Chapter 5 you stated that a better registration, as measured by fiducial registration error (FRE), does not correlate to better targeting accuracy, aka target registration error (TRE), so what can any system tell us about registration accuracy?” The answer is that it can give us only FRE, but FRE can be useful to a well-prepared person, who will understand its limitations. Such a per-
157
158
Image-Guided Surgery
son might be a surgeon who has studied Chapter 5 of this book, including its section entitled “Myth 1: FRE Is an Indicator of TRE.” In that chapter, we learned that FRE can reveal the level of fiducial localization error (FLE) when its RMS is computed over a large number of registrations. That level is important because, as pointed out in that chapter, FLE is the root of all error in IGS. However, in the “Myth 1” section, we also learned that for a specific case FRE is not a reliable correlate to target registration error (TRE), which is the error that the surgeon most wants to know. So, if the IGS system reports FRE, what can FRE tell this well-prepared surgeon? It tells the surgeon one thing and one thing only: whether to use the system on a given case or to set it aside! An exceptionally large FRE tells the surgeon that something is wrong with the system. Perhaps some markers have moved, or an ambiguous anatomical fiducial was mislocalized, or the image or tracking system suffers from geometrical distortion, or the probe’s markers are flawed, or some other fault has occurred. But, whatever the problem, the system should not be relied on for any guidance at all. An FRE that is not exceptionally large tells the surgeon that the IGS system can be relied on to deliver its typical targeting accuracy. In other words, if FRE is in its typical range then TRE is likely to be in its typical range. However, an exceptionally small FRE does not indicate that TRE will be exceptionally small. In fact, even if it is zero, it gives no additional information at all. Thus, to repeat the summary of the situation from Chapter 5: FRE may be the bearer of very bad news, but otherwise it gives no news at all. This discussion of error and typical error values of course assumes that
we have done everything feasible to optimize registration and thus reduce TRE, as detailed in Chapter 5, but it is worth repeating our guidelines one last time: use as many fiducials as is feasible, spread the fiducials as far apart as feasible, arrange the fiducials so that their centroid is as close to the target of interest as feasible, place the CRF as close to the surgical target as feasible, and reduce FLE by using bone-affixed fiducials if feasible. Accuracy of Medtronic StealthStation Equation 5–2 of Chapter 5 provides an estimate of expected TRE when expected FLE is known, but the FLE for a clinical case is never known — for the StealthStation or any system. So, how are we to use Equation 5–2, which requires knowledge of RMS FLE, to determine what TRE to expect for a given system in a given setup? The answer is that we must depend on phantom studies to get an estimate of FLE using similar setups. Although one might imagine that documentation of accuracy would be routinely available from the regulatory clearance of every IGS system, no such information is revealed publicly by any of the companies for any of their IGS systems. So we are limited to information published in peer-reviewed journals, and such information is hard to find. Galloway and Maciunas1 noted as early as 1990 that given the crucial nature of the systems’ accuracy, there is a surprising paucity of independent confirmation of the devices’ accuracy values. Unfortunately, 25 years later, a paucity of information still remains. The relatively limited amount of data available on the accuracy of these systems is to an important extent due to the fact that carrying
Chapter 7
out such studies is extremely difficult. It is different for a number of reasons including (1) ground truth for absolute position is impossible to define and (2) the understanding of what is being measured and reported is often poor
Surgical Systems
n
(see boxed text). For all of the IGS systems presented, we have selected the most pertinent published studies, and in what follows, we present a critique of the method of each study along with a summary of its findings.
Common to many papers reporting accuracy of IGS systems is an erroneous attribution of a total distance to a single projection of a total distance. We will refer to this error as the “subtotal measurement error”. We begin with the proper way to determine the distance between two points, for example a point localized in the OR (xOR, yOR, zOR) and a point in the image (xIM, yIM, zIM). That distance can be calculated as the square root of the sum of the squares of the differences in each orthogonal direction: d = √ (xOR – xIM)2 + (yOR – yIM)2 + (zOR – zIM)2 If this calculation cannot be done, an equally correct approach would be to adjust the orientation and position the plane of the image such that both the centroid of the target (not merely some portion of the target) and the centroid of the tip of the probe (not merely some portion of the tip) are visible in the same plane and then to use a measurement tool in the viewer to determine the distance in millimeters between the two centroids. Unfortunately, what many papers report is either one or multiple distances measured from a monitor in arbitrary image orientations (eg, axial, coronal, and sagittal). For example, one study cited in the Brainlab accuracy section2 reported TRE as the greatest distance between a portion of a target and the tip probe in the axial, sagittal, or coronal view. This erroneous approach discounts the contributions from the other views. If we take the averaged data from that study (Ledderose et al2) with a TRE of 1.13 mm and assume that the measurements in the other two projections were of similar magnitude, then the correct TRE to report would be about 22.5% larger — 1.225 × 1.13 mm = 1.38 mm. To see why, we orient our coordinate system so that the axial view = the xy plane, sagittal = yz plane and coronal = xz plane. In this case the errors measured in the axial (A), sagittal (S), and coronal (C) views are: dA = √ (xOR – xIM)2 + (yOR – yIM)2 dS = √ (yOR – yIM)2 + (zOR – zIM)2 dC = √ (xOR – xIM)2 + (zOR – zIM)2
159
160
Image-Guided Surgery
By squaring these three equations, adding them, dividing by 2, and taking the square root, we find that d = √ (dA2 + dS2 + dC2 /2 = √ (xOR – xIM)2 + (yOR – yIM)2 + (zOR – zIM)2 and, if dA ≈ dS ≈ dC, then d ≈ √ (3dA2 /2 = √ 3/2dA ≈ 1.225dA. Such mistakes would likely be less common if research groups were composed of multidisciplinary groups — namely, engineers and surgeons.
Phantom Studies of Medtronic StealthStation. Perhaps the best phantom
study done to date was completed by Steinmeier et al in 2000.3 They built a phantom consisting of two parts: a clear acrylic cylindrical lid on which they affixed adhesive fiducial markers and a base containing 32 rods of varying length with notched tips that could repeatedly and accurately mate with an IR tracked probe. MRI and CT scans were acquired, and for the MRI the container was filled with gadolinium-water solution so that the solution would appear bright relative to the MRI-invisible plastic. Then, 2816 measurements were made with two configurations of markers: either widely spaced configurations surrounding the rods — they term this configuration “generalized”— or with markers placed on one side of the phantom away from the rods — they term this configuration “clustered”. In addition to giving TREs for the StealthStation, their data confirm that utilizing more fiducials results in better accuracy (sounds like something from Chapter 5!). Their results are shown in Table 7–1, which is adapted from their paper. Note that ix
the generalized configurations give better results for CT (something else from Chapter 5). This trend is reversed for MRI, but the authors point out that their MRI images were distorted on the order of 1 mm by gradient error and static-field inhomogeneity (See Chapter 2), which probably caused this paradoxical result (similar to the seemingly paradoxical result of more CSF leaks when using MRI-guided IGS reported in Chapter 6). Their best-case scenarios are mean TREs of 1 to 2 mm: 1.59 ± 0.68 mm, which yields an RMS of 1.73 mm for CT, and 1.65 ± 0.61 mm, which yields an RMS of 1.76 mm for MRI.ix� But, even in these ideal conditions, as expected, individual TRE values are observed, with the 95th percentile and maximum values for these best-case scenarios being 2.75 mm and 3.6 mm, respectively, for CT and 2.75 mm and 3.2 mm for MRI. Another phantom study shows similar results for CT with Eljamel4 reporting a mean TRE of 1.96 mm. Combining these findings with the evidence of achievable accuracy for IR trackers (see Chapter 3), we conclude that in best-case scenarios in a laboratory setting, the
For ease of comparison regarding the information presented in Chapter 5, we are calculating RMS of reported FRE and TRE values with: RMS = √ Mean2 + Standard Deviation2, recalling that this follows the statistical distribution shown in Figure 3–7.
Chapter 7
Surgical Systems
n
Table 7–1. TRE of the StealthStation Depending on Fiducial Number, Attachment Pattern, and Image Modality Fiducials n
Attachment Pattern
Imaging
TRE
SD
Median
95th Percentile
Maximum
4
Generalized
MR
2.45
0.94
2.3
3.9
6.15
4
Clustered
MR
1.99
0.86
1.98
3.6
4.1
6
Generalized
MR
2.40
0.99
2.3
4.25
5.15
6
Clustered
MR
2.18
0.92
2.03
3.95
5.4
8
Generalized
MR
1.93
0.74
1.78
3.1
3.95
8
Clustered
MR
1.65
0.61
1.58
2.75
3.2
10
Generalized
MR
1.98
0.68
1.9
3.1
3.85
4
Generalized
CT
1.65
0.82
1.45
3.2
4.15
4
Clustered
CT
3.75
0.84
3.75
5.25
5.9
8
Generalized
CT
1.59
0.68
1.55
2.75
3.6
8
Clustered
CT
3.95
0.98
4.0
5.55
6.7
Note. All measurements in millimeters. Source: Adapted from Steinmeier R, Rachinger J, Kaus M, et al. (2000). Factors influencing the application accuracy of neuronavigation systems. Stereotact Funct Neurosurg. 2000;75(4):188–202. Table 3.
Medtronic IGS system can consistently achieve RMS TREs between 1.5 and 2.0 mm. In the real world, however (and as we will see below), things are a bit more complicated. Clinical Studies of Medtronic StealthStation. Clinically, in 2004, Henderson5
reported on the accuracy of a Medtronic StealthStation using a CRW stereotactic frame as ground truth and a motorized micromanipulator/microdrive system to target intracranial locations. Participants were receiving deep brain stimulators as a means of treating neurological chronic neurological conditions with tremors. Error was defined as the difference between the target location determined by a StealthStation and
where the CRW frame indicated the target should be located. The study is somewhat flawed in that there is error associated with the CRW frame — so the study really shows the disagreement between two systems, a CRW frame and a StealthStation, with each system’s error contributing to the overall observed errors. Because these errors can be expected to be uncorrelated and add in quadrature, RMS TRE for the StealthStation (and for the CRW frame) would be expected to be less than that reported in the paper. The data in Table 7–2, adapted from the paper, show TREs (column 5) that range from 0.7 to 3 mm. It can be seen that FRE (column 4) and TRE reveal the characteristic that we discussed in Chapter 5 (see Figure 5–4) — namely,
161
162
Image-Guided Surgery Table 7–2. StealthStation FRE and TRE During Clinical Neurosurgical Procedures Patient No.
Diagnosis
Operation
FRE
TRE
1
Chronic pain
DBS
1.5
2
Parkinson’s disease
Pallidotomy
0.7
3
Posttraumatic tremor
DBS
1
4
Essential tremor
DBS
2.3
5
Essential tremor
DBS
0.4
1.6
6
Parkinson’s disease
Pallidotomy
0.3
1.8
7
Parkinson’s disease
Pallidotomy
0.4
1.7
8
Essential tremor
DBS
0.7
1.6
9
Parkinson’s disease
DBS
0.9
2.7
10
Essential tremor
DBS
0.4
3
11
Parkinson’s disease
DBS
0.1
1.6
Note. All measurements in millimeters. DBS, deep brain stimulator. Source: Adapted from Henderson JM. Frameless localization for functional neurosurgical procedures: a preliminary accuracy study. Stereotact Funct Neurosurg. 2004;82(4):135–141. Table 2.
that FRE is not predictive of TRE: one of the lower FREs of 0.4 mm is associated with the largest TRE of 3 mm. The largest FRE, 0.9 mm, is associated with one of the largest TREs, 2.7 mm, and the second largest FRE, 0.7 mm, is associated with one of the lower TREs, 1.6 mm. Within otolaryngology, a paper by Metson et al6 has been frequently cited (63 times as of the writing of this chapter in September 2015) regarding the accuracy of the Medtronic LandmarX system, which was the predecessor to the StealthStation. Although details of the methods are not clearly presented in the paper, PointMerge registration was used, and the “accuracy of anatomical localization at the start of the surgery was 1.69 ± 0.38 mm (RMS 1.73 mm)
(range, 1.53 ± 0.41 mm (RMS 1.58 mm) to 1.79 ± 0.53 mm (RMS 1.88 mm)),” but our interpretation is that the reported values are not TRE but rather FRE.
Brainlab Brainlab is a German company currently located just outside Munich and founded in 1989 as a startup company providing planning software for stereotactic biopsies. Their business soon expanded into navigation within the cranium through collaborations with the University of Vienna. In addition to neurosurgical and otolaryngologic IGS, additional service lines include stereotactic radiosurgery and orthopedic IGS. Although the company is privately held and its financial statements
Chapter 7
are closely guarded, some information is publicly available (eg, they have over 9000 systems installed in about 100 countries and employ 1200 people in 18 offices worldwide).x Regarding the otolaryngology market, most reports put Brainlab in second place regarding market share, trailing Medtronic by single percentage points. This ranking, despite the size mismatch between Brainlab and the far larger Medtronic, probably reflects Brainlab’s intense focus on image guidance. For other areas of expertise, Brainlab has adopted
A
Surgical Systems
n
a strategy of partnership such as those with Olympus,xi Boston Scientific,xii and Mobius.xiii Regarding IGS, it is this last partnership that is most intriguing given the unique design of the Mobius Airo intraoperative CT scanner, as noted in Chapter 2. Brainlab Curve and Kick For the otolaryngology market, Brainlab pared down its high-end neurosurgical unit, VectorVision, into the award-winning systems, Curve™, and a
B
Figure 7–4. Brainlab Curve (A) and Kick EM (B). Images provided courtesy of Brainlab, Inc. x
https://www.brainlab.com/en/press-releases/brainlab-mobius-receive-ce-mark-approval- airo-mobile-intraoperative-ct/ (Accessed December 30, 2015). xi http://www.olympusamerica.com/corporate/corp_presscenter_headline.asp?pressNo=1009 (Accessed December 30, 2015). xii http://www.bloomberg.com/article/2015-04-27/abTZ1qrtomE0.html (Accessed December 30, 2015). xiii A complete list of partnerships can be found at https://www.brainlab.com/en/about-brain lab/partners/ (Accessed December 30, 2015).
163
164
Image-Guided Surgery
basic system intended for ESS, Kickxiv (Figure 7–4). The design of the surgical interfaces for these systems is analogous to that of an iPad. For example, a single, centered button on the bottom of the display sends the user back to the home page, and planning software, only available on the VectorVision, is even referred to as iPlan. As with the Medtronic support literature, the Brainlab literature is supposedly available online but difficult to find.xv Also similar to the Medtronic literature (and likely stemming from regulatory requirements), Brainlab dedicates an inordinate amount of space in these guides to logistics; for example, six pages of the 130-page Curve 1.0 System User Guide (pp. 40–45) are spent on positioning the user cart arms, and an entire chapter (Chapter 6, pp. 97–118) is devoted to transporting the system, while much less space is allocated to issues central to IGS (eg, only one page is spent on optimizing the FOV of the IR cameras [p. 84]), but we do note that in operation the system has a rough check that the fiducials on both the CRF and instrument are within the FOV (pp. 54 and 66, Curve 1.0 System User Guide). xiv
Pearls From Brainlab Users Guides n The system is intended for use
only within an operating room (p. 11, Curve 1.0 System User Guide) with example procedures listed including skull-base surgery under “cranial procedures” and sinus surgery under “ENT procedures” (p. 17). n “Trained hospital personnel (e.g., nurses) are responsible for setting up the system before the surgical procedure and for removing the system after the surgical procedure” (p. 11). n “This system solely provides assistance to the surgeon and does not substitute or replace the surgeon’s experience and/ or responsibility during its use” (p. 14). n Brainlab instructs that “all users of system, instruments and software: Read the user guides carefully before handling the equipment and have access to the user guides at all times.” It adds that “disregarding information in the user guides, in particular the disregard of warning and
Curve won the award in 2012 (http://red-dot.de/pd/online-exhibition/work/?lang=en &code=2012-15-3764 [Accessed December 30, 2015]) and Kick in 2015 (http://red-dot.de/pd/ online-exhibition/work/?lang=en&code=26-00316-2015&y=2015&c=167&a=1001 [Accessed December 30, 2015]). xv The authors could not find accessible copies online despite being assured of their availability. Brainlab provided us with the following electronic versions upon request, which are referenced in this chapter: (1) Curve 1.0 Technical User Guide; Revision 1.1, Copyright 2012 (2) Curve 1.0 System User Guide; Revision 1.0, Copyright 2011 (3) Kick 1.1 Technical User Guide; Revision 1.1, Copyright 2014 (4) Kick 1.1 System User Guide; Revision 1.1, Copyright 2014 (5) Cranial ENT 2.1 Software User Guide; Revision 1.3, Copyright 2013 (6) Cranial ENT 3.0 Software User Guide; Revision 1.2, Copyright 2015
Chapter 7
cautions, is considered to be abnormal use,” and explicitly noting that “Quick Reference Guides do not replace the user guides” (p. 15). n There are multiple warnings concerning the system not physically harming the patient, including “Do not position the camera, monitors or any other part of the Curve system directly over the patient” (p. 21) and “Ensure that the system is set up so it is not possible for the patient to touch or come in contact with the equipment” (p. 21). n Regarding “Expected Service Life. Brainlab provides a minimum of five years of service for platforms. During this period of time, spare parts as well as field support are offered” (p. 29). n The display screen does not comply with DIN EN 6868, which is a German standard for testing of imaging systems using a specially designed phantom to measure image quality parameters (p. 38). In other words, radiologists cannot make clinical reads from the system. n As with the Medtronic device, the Brainlab offers a “Sound Panel” that will allow the linking of various music-playing devices to the “Sound Panel”. The Brainlab systems are also compatible with Apple iPod touch and iPhones. Five warnxvi
Surgical Systems
n
ings (safety-critical information) and two warnings (possible problems with the device) are given regarding use of an MP3 players with the Curve mostly concerning avoidance of loud audio signals, which may be disruptive or dangerous (eg, played so loud that alarm signals in the OR are not audible) (pp. 53–54). n Multiple cautions and warnings concern the IR camera, including, “The camera’s infrared light may interfere with other IR-based OR equipment, such as remote controls, pulse oximeters or IR-sensitive microscopes” (p. 22). “The infrared LED array is located around the inner ring of the lenses. Do not directly look into the infrared LED array at a distance less than 15 cm” (p. 66).xvi “An additional Curve system in close proximity may interfere with wireless camera communication” (p. 69), although “close proximity” is not defined (will a unit operating next door interfere?). “Reflections of IR light, e.g., from drapes or shiny surfaces, may interfere with the camera’s ability to correctly track instruments” (p. 76). n As noted in Chapter 3 and in the Medtronic manuals regarding optical trackers “Every time the camera is powered on, it generally requires a warm-up time of
Although unlikely, IR light can cause visual damage, especially with near IR (the NDI IR system is a near-IR system with wavelength of 635 nm but a low power of 1 mW), which is focused on your retina by your eye’s lens almost as sharply as visible light, but the retina has no pain receptors, so you don’t realize the IR light is causing irreparable damage!
165
166
Image-Guided Surgery
two minutes. If camera is stored at low temperatures, warm-up time may be longer” (p. 72). n Because the Brainlab system scans the surface of a patient’s face to achieve surface registration, there are warnings to avoid shining the class II positioning laser into either the user’s or the patient’s eyes (p. 79). n Page 84 has very detailed information about the FOV of the IR cameras, which are very similar to the NDI Polaris Spectra FOV data (see Figure 3–5). And, the user guide points out that “measurement precision in a plane perpendicular to viewing direction of camera is higher than precision along the direction of camera view.” n When the Error LED (rightmost light under laser aperture) is flashing amber or shows steady amber, even if navigation is possible, the user must contact Brainlab support (p. 86). Information conveyed in the Curve 1.0 System Technical User Guide is intended for clinical engineering support staff and covers topics including electrical safety, troubleshooting, and maintenance. Additional information conveyed in the 2013 Cranial ENT 2.1 Software User Guide (references to version 3.0 published in 2015 are included in parentheses), a 394 (314)–page manual, includes the following: n Color coding of trackable fidu-
cials is provided on page 51 (39) (eg, red for reference array and yellow for instrument array) with the admonition that “color blind users may not be able to clearly
distinguish color differences of the represented instruments.” Additionally, the status bar below the live video image of the trackable markers must be green to indicate normal tracking (yellow — camera has lost sight of fiducial, red — registration has not been performed, and black — camera communication error). n Page 53 (41) points out that “the software is not able to distinguish between instruments with identical tracking array geometries.” This ambiguity is explicitly shown in Figure 17 of the Cranial ENT 2.1 Software User Guide (Figure 10 of version 3.0) and is repeated herein as Figure 7–5. This is an excellent point that was presented in Chapter 3. Optical IR systems see only the markers and, based upon calibration (a process in which the instrument’s tip is placed in a “divot” of a calibration block and then pivoted around so the tracking system can determine the relationship between the tracked fiducials and the point of pivot of the instrument tip — see Figure 3–1), extrapolate the location of the tip. Should two different instruments have the same fiducial configuration, the tracking system has no way of knowing which tip is on which instrument. n In addition to “Probe’s Eye” view (display perpendicular to the probe’s trajectory) Brainlab offers “Split Probe’s Eye”, p. 150 (123), in which four perpendicular displays are shown at 0, 5, 10, and 15 mm, to allow one
Chapter 7
Surgical Systems
n
patient (Chapter 13, pp. 293–300 (Chapter 12, pp. 234–244)). n Operating microscopes can be tracked as if the focal point were the tip of a handheld probe (Chapter 15, pp. 333–358 (Chapter 14, pp. 267–272)). Registration With Brainlab Systems
Figure 7–5. Figure 17 from the Brainlab Cranial ENT 2.1 Software User Guide illustrates the fact that the tracking array needs to be calibrated to identify the different instruments since the optical tracker only sees the IRreflective spheres. Image provided courtesy of Brainlab, Inc.
to anticipate what is ahead (aka navigate), as well as “Autopilot”, pp. 161–162 (131–132), in which the user is prompted to move the probe tip and orientation to reach a preselected target. (Note: This approach is discussed in more detail in Chapter 8, where an example of “Autopilot” navigation is shown in Figure 8–2.) n A surface rendering of pathology (eg, tumor or polyp) termed “paint” is available to allow removal of a section of tissue from the rendering once it has been removed from the
Brainlab registration can occur via either “standard” to “surface”. In “standard”, a few markers worn by the patient are localized by the software in image space and by the practitioner touching a probe to them in the operating room and then registered. In “surface matching”, the system localizes hundreds or thousands of points on a surface in image space and the practitioner localizing hundreds of points on the patient’s face in the operating room by either physically touching the face or laser-scanning the face. Both registration techniques have entire chapters dedicated to them in the Cranial ENT 2.1 Software User Guide, Chapter 5 for “standard registration”, pp. 69–91 (57– 78), and Chapter 6 for “surface matching”, pp. 93–121 (79–100). In standard registration, some general recommendations are provided, including the following: “Do not place markers/landmarks very close to one another. Avoid placing markers/landmarks in a symmetric structure, for example, in a row or regular shape. Spread the marker/landmark positions around the head, avoiding areas with loose skin” (p. 74 (62)), all of which is in agreement with our recommendations from Chapter 5. “Standard” registration uses thresholding to automatically identify markers (p. 76 (64)). If thresholding fails, markers can be manually added (p. 77 (67)) and adjusted (p. 81
167
168
Image-Guided Surgery
(68)). Marker localizations can be performed in any order, an approach with the user guide calls “unsequential”, unless a regular configuration (eg, equilateral triangle, square) of points makes the correspondences between image and OR ambiguous (eg, an equilateral triangle can be aligned with a copy of itself in three ways), and unless more than seven points are localized for the registration (p. 83 (70)). The reason for the limitation to seven points is unclear, but it might be that the Brainlab algorithms are insufficiently clever to isolate the one point-to-point correspondence among the 5040 possibilities (the 1st fiducial can be fit to any of the 7 positions, the 2nd fiducial can be fit to any of the 6 remaining positions, the 3rd fiducial can be fit to any of the 5 remaining positions, and so on until the 6th fiducial which can be fit to any of the 2 remaining positions leaving the last fiducial only 1 option; thus 5040 possibilities = 7 × 6 × 5 × 4 × 3 × 2 × 1) on the basis of their geometric arrangement without spending an inordinate amount of computer time. For sequential registration, instead of the software automatically determining the one marker in the operating room that corresponds to each marker in the image, the location of each marker in image space is shown to the user, and the user is asked to identify the same marker in the operating room based on the user’s recognition of the corresponding anatomy that surrounds the marker in the two spaces. Although we know from Chapter 4, Figure 4–6, that only three fiducial points are necessary to ensure correct geometric alignment, Brainlab requires at least four points for standard registration. It is unclear why this number is chosen, but Brainlab may enforce the four-point minimum because, as is
pointed out in Chapter 5 (Equation 5–2), four fiducial points are likely to give a more accurate registration. This restriction is a defensible design decision, but the system seems to run off the rails at this point because the result of standard registration is then categorized as “good precision” (8 mm). The only information available to the Brainlab system for its categorizing of the level of registration accuracy is the set of fiducial positions and their alignment errors after registration. In Chapter 5 we named a fiducial’s alignment error its “individual FRE” (the displacement between corresponding image and physical fiducial points after registration). We refer the reader back to our discussion of FRE in Chapter 5, Figures 5–5 and 5–6, where we note that variation in TRE is not correlated with FRE or individual FRE and in fact TRE is not correlated with any formula based on either one.7 Thus, it can be seen that, according to registration theory, Brainlab’s categorization of registration accuracy as good or medium on the basis of fiducial alignment error is unfortunately meaningless and misleading. Theory does, however, support the claim of a threshold for failure (Brainlab chooses FRE > 8 mm for point based registration). A reality check is recommended (p. 88 (75)), in which the locations of at least three anatomical points are to be verified, following which the registration can be either accepted or rejected and tried again with a recommendation to try again if error of greater than 3 mm is indicated during the reality check. The “advanced registration” options (p. 90 (77)) indicate that some markers may be skipped by the software (either “considered” or “uncon-
Chapter 7
sidered”), unfortunately perpetuating one of the myths presented in Chapter 5 (Myth 2: dropping a fiducial will increase accuracy). This recommendation shows how ingrained these myths have become, even being incorporated into commercial software! Surface registration can be based on laser scanning of the surface of the patient (z-touch appears as red during scanning), physically tracing the surface of the patient (Softouch appears as blue during tracing), or a combination of both. Without providing supporting evidence, the guide states that “surface matching registration is more likely to fail if used with rotational angiography, C-arm, or DVT image sets” (p. 95 (80)). “If the software cannot proceed due to insufficient surface matching results, guided mode registration is activated; you are then prompted to acquire three anatomical landmarks (see page 107)” (pp. 102, 105 (86, 89)). These three anatomical points are from a list of four: lateral canthus right, lateral canthus left, nasion, and inion. Registration is noted as “good” (3 mm) (p. 111(95)). Brainlab explicitly acknowledges that these values specify only how well the surfaces were matched and do “not necessarily represent the overall error” (p. 111 (95)). It is interesting to note that these values allow roughly half the leeway of the values that are accepted for standard registration (8 mm). We interpret this difference as an attempt to factor in the cantilever effect that occurs with surface matching (see Figure 4–9). A technique to “improve” the registration is offered on page 112 (96), xvii
Surgical Systems
n
in which additional anatomical points may be added to the surface registration to improve accuracy. Remember what we learned in Chapter 5 — more points are better! It is interesting that Brainlab incorporates the correct concept of increasing the number of registered points to improve accuracy but then includes the fallacious concept of deleting fiducials to improve accuracy (see paragraph on “standard registration”, where points may be “considered” or “unconsidered”). Pages 116 to 120 (88–94) cover general considerations for optimal collection of points for surface registration in the operating room, especially the concept of collecting points that are over bone (eg, zygomatic arch) and avoiding areas of high deformability (eg, the ear other than the tragus), collecting points on uniquely shaped surfaces (eg, the nose), and the need to collect more points if working in nonuniquely shaped surfaces (eg, lateral or posterior skull base). Accuracy With Brainlab Systems Phantom Studies of Brainlab Systems.
In 2007, a group from Munich reported on the accuracy of laser surface scanning to register for ENT-based IGS surgeries, including ESS and lateral skull base surgery.3 Two cadaveric heads were prepared with 16 facial plating screws on the surface of facial bones, within the paranasal sinuses, and on the lateral skull base. After CT scanning (helical 1-mm slice thickness, overlap not specified), the image was registered to the cadaveric head using z-touch. The distance between where the Brainlab systemxvii identified the central depres-
The VectorVision Compact, which preceded the Curve, was used but consists of the same infrared optical tracking system.
169
170
Image-Guided Surgery
sions of the facial plating screws and where they were identified on the CT scan was determined as the maximal distance in the axial, coronal, or sagittal views. This study suffered from subtotal measurement error (see boxed text under Medtronic accuracy). However, it will give us some trends — a lbeit underestimated trends — about TRE. Their data are presented in Table 7–3 adapted from their paper. Averaging their TRE values, they calculated a mean TRE of 1.13 ± 0.53 mm yielding an RMS of 1.34 mm for the anatomical volume pertinent for ENT surgeons with accuracy better at the surface compared to deeper locations (see Figure 4–9) and better accuracy when laser scanning took place over both the right and left sides compared to just scanning over the side of interest (see Chapter 5, where we discuss using more points, spreading them out widely, and arranging them so that the target is near their centroid in order to achieve better accuracy). Clinical Studies of Brainlab Systems.
In 2002, a neurosurgical group from the University of Regensburg, Germany, reported on the TRE in clinical cases using registration with z-touch versus skin-affixed fiducials and subsequent point-based registration with the VectorVision Compact.8 Accuracy, in terms of TRE, was evaluated by measuring the distance to “three distant points on the scalp, the tragus and canthus lateralis on both sides, and the target fiducial”. As with the Ledderose et al3 study just referenced, the accuracy “was determined on the monitor display (600% zoom) for axial, sagittal, and coronal images. The highest error on these three images was determined for each of the
Table 7–3. TRE Using BrainLab VectorVision Compat Position of Facial Plating Screw (Target)
TRE (SD), mm
Frontal, left
1.01 (0.56)
Frontal, right
0.89 (0.42)
Temporal, left
1.12 (0.23)
Temporal, right
1.06 (0.40)
Preauricular, right
0.98 (0.38)
Preauricular, left
1.11 (0.39)
Retroauricular, left
1.74 (0.44)
Sinus maxillaris (left)
1.02 (0.34)
Sinus maxillaris (right)
1.02 (0.27)
Sinus frontalis (posterior left)
0.91 (0.30)
Sinus frontalis (posterior central)
0.78 (0.37)
Sinus frontalis (posterior right)
0.77 (0.30)
Sinus ethmoidalis (top)
1.27 (0.42)
Sinus sphenoidalis
1.25 (0.47)
Lateral skull base
1.40 (0.64)
Petrous bone
1.76 (0.45)
Source: Adapted from Ledderose GJ, Stelter K, Leunig A, Hagedorn H. (2007). Surface laser registration in ENT-surgery: accuracy in the paranasal sinuses — a cadaveric study. Rhinology. 2007;45(4): 281–285. Table 1 and Table 2.
above mentioned landmarks,” which incurs the problem of subtotal measurement error (see boxed text under Medtronic accuracy). They report the FRE of z-touch to be 1.36 ± 0.34 mm (RMS FRE 1.40 mm) and point-based registration with skin-affixed markers to have an FRE of 1.10 ± 0.53 mm (RMS
Chapter 7
FRE 1.22 mm). TRE was reported to be 2.77 ± 1.64 mm (RMS TRE 3.22 mm) for the target fiducials for z-touch. For point-based registration, the TRE was reported to be 1.31 ± 0.97 mm (RMS TRE 1.63 mm) for the target fiducials. Again, these measurements suffer from subtotal measurement error and thus underestimate the true RMS FRE and RMS TRE. In 2006, clinical testing on 368 patients undergoing ESS was undertaken using z-touch registration.9 Of the 368 patients, 202 underwent successful registration (FRE < 2.5 mm) with a single registration, 135 required repeat registration, and 31 required multiple re-registrations, with eleven cases failing to register to a clinically acceptable level and five cases with IGS aborted due to technical failures. They report values for a “clinical plausibility test”, which they define as follows: “Anatomical landmarks on the patient’s face were located by a specially marked pointer instrument (ie, lateral and medial orbital edge, tip of the nose, columella, nasion, head of the middle nasal concha, and anterior wall of the sphenoid sinus) and compared with the corresponding position provided by the navigation system.” They reported “an average deviation in the x and y axes of 1.3 mm and in the z axis of 1.4 mm”. Here again the approach suffers from subtotal measurement error, but we can add the error in the xy plane and the error along the z axis in quadrature to determine the TRE as √ 1.32 + 1.42 = 1.9 mm. Standard deviations for the measurements were not given, and thus RMS values cannot be calculated. Note that this value applies only anteriorly. Because registration consisted of facial surface points and only one deep
Surgical Systems
n
point — the anterior wall of the sphenoid sinus —TRE would be predicted to get worse as one progresses deeper in the surgical volume (see Figure 4–9). Perhaps the most rigorous testing was performed in 2008 by Balachandran et al10 in a study in which z-touch registration was performed on patients with bone-anchored hearing aids (BAHAs), each of which consists of a surgically implanted abutment screwed into the cranium behind the external ear. Before testing, patients were affixed with radioopaque markers on their BAHA abutment and then underwent CT scanning. After surface registration with the Brain lab VectorVision Compact, the centroid of the marker was localized in physical space using a system similar to that shown in Figure 4–3. TRE at the location of the BAHA abutment at the lateral skull base was determined to be 3.21 ± 1.02 mm (RMS 3.37 mm) with a range of 1.85 ± 0.04 mm (RMS 1.85 mm) to 4.42 ± 0.19 mm (RMS 4.42 mm). Given the cantilever effect (Figure 4–9), the TRE at locations closer to the plane of the face would be expected to be less than these values.
Clinical Accuracy of Electromagnetic (EM) Tracking Systems Although both Medtronic and Brainlab offer EM tracking within their systems (the Stryker systems do not), there is almost no information about accuracy of these clinical systems. To gain insight into the accuracy of EM systems in general, we are forced to look at an old but technically impressive study that was performed in 1997 with the InstaTrak system, which is no longer commercially
171
172
Image-Guided Surgery
available. This work involved placing a unique fiducial frame on the patient (see Figures 1–12 and 7–1) and adhesive markers located on the supraorbital rim of both the left and right eye. Fifty-five patients were enrolled, and average error between physical locations of the adhesive fiducials with EM tracking and those locations in the preoperative CT scan was 2.28 mm with a standard deviation of 0.91 mm (RMS TRE 2.45 mm) and a maximum value of 5.08 mm.11 Although not directly applicable to the accuracy of the Brainlab Kick EM or Medtronic Fusion system, the study does provide confidence that under certain conditions, EM systems can achieve clinical accuracy approaching that of optical tracking.
Stryker Stryker Corporation, now a Fortune 500 company most recently listed at #719,�xviii was founded by orthopedic surgeon Homer Stryker in 1946 and existed mainly as an orthopedic medical device company through international expansion in 1972 and initial public offering in 1979. During this growth, Stryker ventured into related product lines including endoscopy, medical and surgical hospital beds, pain management, and, in 2000, surgical navigation with the acquisition of Image Guided Technologies, a manufacturer xviii
of active optical tracking systems. (Recall from Chapter 3 that active tracking means that the fiducial marker actively emits a signal compared to a passive fiducial marker, which reflects a signal. And, while in practice, passive systems may be less accurate due to user error in marring of the surface of the reflector (see Figure 3–4), in theory, active and passive optical tracking systems should achieve similar levels of accuracy.) In 2014, Stryker announced a partnership with Neurologica, a manufacturer of portable CT scanners.xx As we noted in Chapter 2 and the beginning of this chapter, the Neurologica scanners — with their tankdrive systems advancing along the floor between image acquisitions — are dependent upon either a flat floor to acquire geometrically stackable images and/or algorithms to correct for imperfections in the floor. At present, we do not consider the Neurologica scanners suitable for IGS. xix
Stryker NAV3i and PROFESS Stryker has stuck with active optical tracking and sponsored a study showing that their active optical tracking has better accuracy than NDI,12 with the ac-curacy of the NDI Polaris Active IR tracking being 0.089 ± 0.061 mm (RMS = 0.11 mm)xxi and the Stryker navigation System II Camera being 0.058 ± 0.033 mm (RMS = 0.067 mm). Stryker offers neither passive optical tracking
http://www.forbes.com/global2000/list/16/#tab:overall (Accessed December 30, 2015). http://www.stryker.com/en-us/corporate/AboutUs/History/index.htm (Accessed December 30, 2015). xx http://www.neurologica.com/news-media/press-releases/stryker-and-neurologica-partner-provide-surgical-navigation-bodytom%C2%AE-porta (Accessed December 30, 2105). xxi Accuracy was measured over a smaller region of field of view than that of the measurement reported in Chapter 5 for the passive Polaris system. When active and passive Polaris systems are compared within the same field of view, there is no significant difference in accuracy (Chapter 3). xix
Chapter 7
nor EM tracking. Their active system, CranialMap, for use on the NAV3 and NAV3i computer platform (Figure 7–6), is primarily geared toward neurosurgery and skull-base surgery, although it has been used in otolaryngologic procedures. In 2015, they released a novel IGS system for otolaryngology, the PROFESS system, for use on the ADAPT, NAV3, or NAV3i computer platform, which provides tracking via video cameras embedded into disposable surgical tools and recognizes the type of tool based on the unique geometric fiducial configurations. This system appears to be a hybrid between the InstraTrak head-
Surgical Systems
n
set (see Figure 1–12) and ClaroNav video tracking (see Figures 3–8 and 3–9). One difference from the ClaroNav system is that the PROFESS video camera is integrated into the disposable surgical tools thus avoiding most line-of-sight issues. The PROFESS system and ADAPT computer platform is shown in Figure 7–7. The interface for both CranialMap and PROFESS is via a tablet and, both having actively tracked instruments, the user can interface with the system (eg, select options) by pushing buttons on the actively tracked tools. User manuals were provided to the authors upon request from Stryker and are referenced below.xxii
Figure 7–6. The Stryker NAV3i system. Image provided courtesy of Stryker, Inc. xxii
(1) CranialMap™ Express REF 6001-655-000 Version 2.1 (2013), (2) Instructions for Use PROFESS System (2014), (3) Stryker NAV3i Platform (REF 7700-800-000) (2013)
173
A
Figure 7–7. The Stryker PROFESS system (A) and ADAPT platform (B). PROFESS using skin-affixed geometric patterns for registration, a headpiece that projects from the nasal dorsum as a CRF, and video cameras in the individual surgical tools to achieve navigation. Images provided courtesy of Stryker, Inc.
B
174
Chapter 7
Pearls from NAV3i CranialMap Express User Manual n “The CranialMap Express
Navigation system supports, but is not limited to, the following surgical procedures: Endoscopic Sinus Surgery (ESS), Intranasal procedures, Ear implant procedures, Craniotomies, Removal of foreign objects, Skull base procedures, Transnasal neurosurgical procedures, Transsphenoidal pituitary surgery” (p. 4). This is the first time we have seen “ear implant procedures” listed in any of the IGS manuals we reviewed. n “Contraindications: Surgical situation where increasing surg-ical time may be detrimental to the patient” (p. 4). This is the first time we have seen explicit recognition that the use of IGS might result in longer surgical times. n “Segmentation: Segmentation allows you to accentuate anatomical structures of interest such as tumors, the skull, and brain tissue. Additionally, for each segment, you can define a safety margin and enable a collision warning. During navigation, the software gives a warning as soon as the navigated tool tip collides with the safety margin associated with the segment” (p. 22). No details are provided regarding how segmentation is achieved, but it is likely due to “thresholding”, a method in which an area of interest (eg, the skull) is selected and all voxels touching the area of interest either directly or indirectly whose intensity values are within a range of the Hounsfield unit
Surgical Systems
n
of the selected area are included. These techniques are relatively standard with the precision of segmentation highly dependent upon the image quality and proximity of tissue with different Hounsfield units (eg, change from skin to air is large while change from subcutaneous fat to muscle is relatively small). n Although information about the FOV of the camera is not explicitly given in the manual (it is provided in the aforementioned reference),13 during setup, the camera system provides feedback regarding patient positioning with a central green circle on the video display indicating that the patient and tool are within the working space of the camera, a central red circle indicating that the patient and/or tool is too close to or too far away from the camera, and a central white circle indicating that the patient and/or tool tracking is not visible (p. 30). n “To ensure that navigated instruments were not bent during reprocessing, they require validation or initialization” (p. 31). n A microscope can be linked to the IGS system for craniotomies (pp. 59–60). n Calibration of straight and curved instruments is facilitated with a tracked calibration tools (pp. 61–67).
Registration with CranialMap Registration (pp. 31–55) can be accomplished via either fiducials alone, where the fiducials are either anatomical landmarks (eg, the canthi, nasion,
175
176
Image-Guided Surgery
and nasolabial junction) or active IR fiducials on a surface mask with an attached communication unit (Figure 7–8), or via fiducials supplemented by surfaces. In practice, a “tracker”, which in our terminology is a CRF, is first affixed to the patient consisting of either (1) a separate piece with active IR emitters rigidly affixed to the skull typically via a Mayfield head holder or (2) the surface mask itself. When using the sur-
face mask, registration is achieved with surface matching automatically, and the mask can be used as a non-invasive patient tracker. All other registration begins with point-based registration using anatomical fiducials. The user is asked to perform a reality check of the point-based registration by touching a reasonably well-defined anatomical feature (eg, the tip of the nose, with the tip of the tracked probe, and
Figure 7–8. The disposable Stryker face mask consists of 30 active IR markers (green indicating that the camera sees the markers and red indicating that the markers are not visible), which serve both as the fiducials for registration as well as the CRF during tracking. The white box is a battery-powered communication box. In practice, the left and right pieces are affixed to the side of the face but are shown nonaffixed in this illustration so that they can be seen. Registration takes place via a three-step process during which the camera acquires fiducial locations on the right (or left) side, then the mid-face, then the left (or right) side, noting that overlap with each area (right, mid-face, and left) is necessary to create the composite surface. Image provided courtesy of Stryker, Inc.
Chapter 7
confirming the location of the probe tip on the display screen). If the location on the screen appears subjectively accurate, the user can “accept” and proceed. If it does not, the point-based registration is “rejected” and the registration procedure is repeated until a subjectively acceptable level of accuracy is achieved. As we have seen before, and showing how ingrained Myth 2 of Chapter 5 has become, the Stryker manual explicitly states, “After digitizing three points, the mean deviation is displayed. This information is used to decide which, if any, points should be re-digitized. The number shows relative deviation between the reference points defined on the images and those digitized on the patient” (p. 37). In other words, if one point has a higher deviation, it should be re-acquired or dropped and another point selected.” Although we learned in Chapter 5 that the point with the highest deviation is likely to be the one that is contributing the most accuracy to the registration and thus is the last point that should be dropped, and, although we learned that dropping that point is likely to make the registration worse, we also learned that dropping it can be expected to reduce FRE and, as a result, will give the practitioners that drop it (presumably not having read this book!) the false sense that they are helping their patients. So unfortunately, this practice is likely to continue producing worse registrations and worse outcomes.
Surgical Systems
n
Next, the user is prompted with the following query: “Do you want to perform surface registration to improve the point-to-point registration or do you want to continue with navigation?” (p. 39). Surface registration is accomplished by touching the contour of the face with an actively tracked probe, in which case the manual advises (a) collecting more than 50 surface points, (b) avoiding areas of high skin deformability, and (c) concentrating on areas of prominence such as the nose and orbital rims. An optional and unique aspect of the Stryker system is the patient registration and tracking mask (Figure 7–8), which affixes flexibly to the skin of the face and provides active-IR fiducial markers that serve as surface points for registration of the patient to the facial surface identified in the preoperative image. In practice, the Stryker mask is affixed to the patient in the operating room using adhesive tape on the undersurface of the mask, and then the navigation camera is used to capture the IR signals emitted from the fiducials embedded in the face mask. Usually the patient can be registered with a single click on “Register” automatically. If the navigation camera cannot detect a sufficient number of mask fiducials from one direction, the software offers capturing from the right, from the front, and from the left, which allows it to receive the IR signal from all of the fiducials in order to localize them (see legend of Figure 7–8). Crucial to this process is that, as the camera is moved during registration (and this is perhaps the ONLY application where camera motion during registration is acceptable), the individual views that it acquires of the fiducials (ie, the
177
178
Image-Guided Surgery
right-side and left-side acquisitions) overlap the frontal acquisition, so that the system can piece together a panoramic registration. As with all activeIR systems, each fiducial emits a burst of radiation while the other fiducials are dark. The timing of these bursts is governed by the box at the top of the mask, which also provides the power to generate the IR signal. The main tracking system communicates continuously with this box, synchronizing IR generation and IR reception to identify which fiducial is emitting each burst. Additionally, and separate from its role as fiducial markers, 10 fiducials of the mask that is adhering to the patient’s forehead also serve as the CRF by continuing to emit IR during the intervention so that the tracking system can detect, and compensate for, any motion of the head relative to the camera. This very clever arrangement essentially “kills two birds with one stone.” The patient registration and tracking mask is available for ENT and neurosurgical applications. In combination with intraoperative CT imaging the Stryker mask can also be used for automatic intraoperative mask (AIM) registration of the intra-operative image set. The patient is CT scanned with the mask applied to the head. Once uploaded to the system the CranialMap software automatically detects the scanned fiducials (IR LEDs on the mask flexible circuit) in the CT volume. The navigation camera detects the spatial position of the mask LEDs, and the software automatically matches the detected LEDs in the scan to the LEDs detected by the camera for registration. Unlike the automatic surface registration on CT or MRI data using the mask, the AIM registration does not rely on skin surface detection or surface match-
ing but rather performs discrete point based registration. Pearls From Instructions for Use PROFESS System n “Indications for use: Endoscopic
sinus surgery, intranasal procedures” (p. 4). n “Accuracy of ±2 mm and a latency of ≤ 4 seconds”. And “Within the nasal cavities, the system has a mean accuracy of ±2 mm. This value does not take into account user-dependent deviations of patient registration” (p. 4). No experiments or corroborating publications are cited. n The user’s manual and software recommend that CT scans be acquired within the past 60 days, and if longer than 60 days, the user is asked to “ensure that the 3D skin surface (registration surface) shown by the software matches the patient’s actual skin appearance” (p. 11). n Tracked tools consist of three suction devices (0, 70, and 90 degrees of angulation) and a frontal sinus seeker. These tools are factory calibrated, and the system requires a check to verify accuracy of tool-tip tracking before they can be used. If this accuracy check fails, the user is instructed to discard the tool. n An accuracy check is strongly recommended. “Always perform a landmark check after registration. Failure to comply may leave inaccurate registration undetected and compromise navigation accuracy” (p. 20), and “To ensure continued accuracy, perform a landmark check before
Chapter 7
each critical step of the surgery” (p. 21). n The user is explicitly reminded that, if the CRF moves relative to the patient during the intervention, then the system will display inaccurate positional information (p. 21) and that re-registration can be performed at any point during the intervention (p. 22). Registration With the Profess System Pages 15 to 18 of the Instructions for Use PROFESS System describe registration. Before surface registration can begin, the user is prompted to place stickers containing unique geometric patterns on the forehead, cheeks, and nose (see Figure 7–7). These stickers are used for non-contact 3D scanning of the skin surface, and the manual cautions against anything that impedes this surface representation (eg, overlapping stickers, stickers not flush with the skin, and/or placing the stickers in areas where they may not adhere [eg, hair bearing or greasy skin]). Next, the CRF, which Stryker calls the “patient tracker”, is placed. Their CRF is similar to the InstaTrak headpiece (see Figure 1–12) in having earbuds and nasal dorsum rests (see Figure 7–7). Next, registration occurs using one of the tracked tools, which is plugged into the system and has an integrated video camera that identifies the skin-affixed stickers to be used to create a surface for registration, and the CRF, which must be at least partially visible throughout the intervention in order to continuously track the patient’s location in physical xxiii
Surgical Systems
n
space. If surface registration fails and/ or if use of the skin-affixed stickers is not desired, anatomical fiducial registration can be performed using the tip of the nose, lateral canthus (right and left), and nasolabial angle (right and left). Phantom Studies of Stryker Systems.
Accuracy studies of the Stryker active IR system are reported in Elfring et al.13 A complete study of all components of the Stryker IGS system, as opposed to only the optical tracker, as in Elfring et al, was performed by a group from Heidelberg, Germany, and published in 2011 in the neurosurgical literature.13 This well-done study utilized a customized phantom skull complete with deformable silicon surface layer to mimic skin and subcutaneous tissue in comparing surface to point-based registration. Additionally, they used a computer numeric control (CNC) positioning table to very precisely localize targets in physical space. These positions were compared to the locations obtained by the IGS system. The group rejected any registration for which FRE was greater than 2.5 mm. They report “navigation accuracy” in terms of the navigational error that we define as TRE, and they measured means of 1.45 ± 0.63 mm (RMS 1.58 mm) for point-based registration and 2.51 ± 1.49 mm (RMS 2.91 mm) for surface registration (surface registration was performed using surface tracing and not use of the mask system, which was not available at the time of the study).xxiii Other interesting findings from this study, which are somewhat buried in the prose, are the following: (1) FRE revealed no relationship to “navigation accuracy” (TRE),
The study also reported findings with the Brainlab VectorVision TRE = 1.27 ± 0.53 (RMS 1.38 mm) for point-based registration and 2.61 ± 1.56 mm (RMS 3.04 mm) for surface registration.
179
180
Image-Guided Surgery
which confirms the section entitled, “Myth 1: FRE Is an indicator of TRE” in Chapter 5; (2) using more fiducial points during registration resulted in better navigation accuracy (TRE), which confirms the section entitled, “Myth 2: Dropping a Fiducial Will Increase Accuracy”. also in Chapter 5; and (3) using a higher percentage of points for registration (ie, not dropping fiducials to improve FRE) resulted in better navigational accuracy (TRE) (p. 224, second bulleted paragraph), which also confirms the Myth 2 section. Clinical Studies of Stryker Systems.
Despite extensive efforts, we are unable to find a quantitative clinical study of either the CranialMap (or its predecessor systems) or the PROFESS system.
Smaller IGS Companies Fiagon Aimed at office-based interventions, an EM system from Fiagon received FDA clearance in 2014.xxiv Fiagon began in 2007 as a spinoff from the University of Berlin founded by doctoral students Dirk Mucha and Timo Kruger. Utilized at multiple sites in Germany, the system is penetrating the US market with 56 units sold as of this writing in 2015 and a recent agreement with Entellus,
xxiv
a large ENT product company, to be the exclusive distributor of the Fiagon system to physician offices and ambulatory surgical centers.xxv This agreement appears to be in place to facilitate sales of Entellus’s balloon sinuplasty sets — sets used for minimally invasive sinus dilatation for treatment of chronic sinusitis. These procedures can be performed under local anesthesia but require IGS to target the sinus cavity openings within the nose.xxvi In Europe, Fiagon’s system is also being used for otologic, spinal, and oral surgery applications. The Fiagon system (Figure 7–9) utilizes an NDI Aurora tabletop EM field generator and tracker (see Figure 3–14). Their reusable tools (ten uses per tool) have three embedded EM sensors located at the tip, mid-region, and base. This arrangement allows the tools to be malleable (ie, they can be bent during surgery and tracked in the new configuration). As of this writing, tool cost is $2400 apiece or $240 per case. Registration is achieved using a CRF affixed to the forehead with an adhesive pad and a combination of point-based and surface registration. The user first touches anatomical points — the lateral canthi and the nasion — and then traces a multilobed elliptical path over the supraorbital and infraorbital rims. Although no accuracy data are provided by the company and limited case reports are just beginning to surface, we expect the
http://www.accessdata.fda.gov/cdrh_docs/pdf13/K133573.pdf (Accessed December 30, 2015). http://investors.entellusmedical.com/phoenix.zhtml?c=253886&p=irol-newsArticle&ID= 2078508 (Accessed January 4, 2016). xxvi Interestingly, Entellus recently announced a strategic marketing agreement with Xoran, manufacturer of cone-beam CT scanners, for “ . . . cross-promotion and co-marketing of each company’s products as well as collaboration at key industry events.” (http://www.nasdaq. com/press-release/entellus-medical-announces-strategic-marketing-agreement-with-xorantechnologies-20151119-01032. Accessed January 4, 2016.) xxv
Chapter 7
Surgical Systems
n
Figure 7–9. Fiagon’s EM system has received regulatory approval and is being marketed for in-office and ambulatory surgical sinus interventions, including sinoplasty. Images provided courtesy of Fiagon, Inc.
accuracy to be approximately equal to other EM systems described earlier in this chapter.
ClaroNav Although not yet FDA cleared, we mention the ClaroNav system presented in Chapter 3 (see Figures 3–8 and 3–9), because the system designed for ENT surgery, NaviENT (Figure 7–10) is being evaluated via investigational studies in the United States in the hopes of achieving regulatory approval.xxvii The NaviENT systems utilize an optical tracking system (see Figure 3–8) with checkered patterns intersecting in so-called x-points, as shown in Chapter 3 (see Figure 3–9). A large potential xxvii
appeal of the NaviENT system is their low cost secondary to the use of optical cameras and simple black and white patterns, which can be printed with a laser printer on waterproof paper with adhesive backing and then placed on the CRF, fiducial markers, and tracked tools.
Summary In summary, we see that although each company has a uniquely branded and marketed system, when used correctly, they all achieve similar results. Results obtained in laboratory settings where conditions are highly controllable are
http://www.claronav.com/navient/ (Accessed December 30, 2015).
181
182
Image-Guided Surgery
Figure 7–10. ClaroNav’s NaviENT system is currently under clinical investigation and, at the time of this writing (September 2015), has not received regulatory approval. Images provided courtesy of ClaroNav, Inc.
much better than those obtained in the operating room. Accuracy studies with IGS systems are difficult to perform and are often fraught with technical errors (eg, subtotal measurement error) that would likely be minimized
if research teams included both engineers and surgeons. Surprisingly, there is no universal paradigm for measuring system accuracies or any other system characteristics — for example, a standardized phantom experiment
Chapter 7
that all manufacturers would be required to perform as in other industries (eg, road testing of cars with efficiency in terms of fuel consumption per mile and performance in terms of time to go from zero to 60 miles per hour). Furthermore, despite the use of target registration error (TRE) as a standard for assessing navigational accuracy in the medical engineering community, device manufactures have yet to agree on TRE or any other standard when making their accuracy claims. In closing, we will quote from neurosurgeon Gene Barnett’s review in 2001 of an IGS paper, which is as true today as it was fifteen years ago14. There is something particularly seductive about the information (both graphic and numerical) presented on surgical navigational systems. The images and crosshairs are crisp. The registration error is displayed in millimeters and fractions (tenths, hundredths) thereof. It is little wonder that surgeons want to believe that what they see is what they get. We hope that this chapter has allowed you, the reader, to better appreciate the complexities that lie beneath these deceptively crisp and clear images and to better understand that what you get is not always what you see.
References 1. Galloway RL, Maciunas RJ. Stereotactic neurosurgery. Crit Rev Biomed Eng. 1990;18(3):181–205. 2. Ledderose GJ, Stelter K, Leunig A, Hagedorn HG. Surface laser registration in ENT-surgery: accuracy in the
Surgical Systems
n
paranasal sinuses—a cadaveric study. Rhinology. 2007;45(4):281–285. 3. Steinmeier R, Rachinger J, Kaus M, Ganslandt O, Huk W, Fahlbusch R. Factors influencing the application accuracy of neuronavigation systems. Stereotact Funct Neurosurg. 2000;75(4): 188–202. 4. Eljamel MS. Validation of the PathFinder neurosurgical robot using a phantom. Int J Med Robot. 2007;3(4):372–377. 5. Henderson JM. Frameless localization for functional neurosurgical procedures: a preliminary accuracy study. Stereotact Funct Neurosurg. 2004;82(4):135–141. 6. Metson RB, Cosenza MJ, Cunningham MJ, Randolph GW. Physician experience with an optical image guidance system for sinus surgery. Laryngoscope. 2000;110(6):972–976. 7. Danilchenko A, Fitzpatrick JM. General approach to first-order error production in rigid point registration. IEEE Trans Med Imaging. 2011;30(3):679–693. 8. Schlaier J, Warnat J, Brawanski A. Registration accuracy and practicability of laser-directed surface matching. Comput Aided Surg. 2002;7(5):284–290. 9. Stelter K, Andratschke M, Leunig A, Hagedorn H. Computer-assisted surgery of the paranasal sinuses: technical and clinical experience with 368 patients, using the Vector Vision Compact system. J Laryngol Otol. 2006;120(12):1026–1032. 10. Balachandran R, Fitzpatrick JM, Labadie RF. Accuracy of image-guided surgical systems at the lateral skull base as clinically assessed using bone-anchored hearing aid posts as surgical targets. Otol Neurotol. 2008;29(8):1050–1055. 11. Fried MP, Kleefield J, Gopal H, Reardon E, Ho BT, Kuhn FA. Image-guided endoscopic surgery: results of accuracy and performance in a multicenter clinical study using an electromagnetic tracking system. Laryngoscope. 1997; 107(5):594–601. 12. Elfring R, de la Fuente M, Radermacher K. Assessment of optical localizer accu-
183
184
Image-Guided Surgery
racy for computer aided surgery systems. Comput Aided Surg. 2010;15(1–3):1–12. 13. Paraskevopoulos D, Unterberg A, Metz ner R, Dreyhaupt J, Eggers G, Wirtz CR. Comparative study of application accuracy of two frameless neuronavigation systems: experimental error assessment quantifying registration methods and
clinically influencing factors. Neurosug Rev. 2010;34(2):217–228. 14. West JB, Fitzpatrick JM, Toms SA, Maurer CR Jr, Maciunas RJ. Fiducial point placement and the accuracy of pointbased rigid body registration [Comments following the article]. Neurosurgery. 2001;48:810–817.
8 What Does the Future Hold?
Predictions about the future, although titillating, are often so far off the mark as to be useless. For example, who could have predicted the widespread use of the Internet and/or that a free, online encyclopedia (eg, Wikipedia) would replace expensive, bound sets? Nonetheless, given our experience in the field, we are often asked to make predictions about the future of IGS, and so, with some trepidation, we will do that in this final chapter. There are a number of interesting advances that are likely to be realized clinically within the next decade. These areas, broadly, include (1) computerassisted navigation, (2) augmented reality, and (3) robots. Although each of these topics could be considered independently, we find that there is more overlap than is typically recognized. Nonetheless, these three topics will form the framework for our discussion.
Computer-Assisted Navigation At present, IGS systems confirm where a surgical probe is located but provide little help in guiding a surgeon toward a specific target. Going back to the GPS analogy of Chapter 3, the current state-of-the-art for IGS is akin to a one-way interactive GPS machine. A user can turn on the GPS and ask, “Where am I on the map?” but cannot request directions to another location. Through a series of small trial motions and knowledge of anatomy, surgeons usually become adept at determining how to get to where they want to go. But, for small motions toward objects that may be one or two voxels away, the process can be frustrating because there are 26 voxels neighboring the current voxel and usually no good way to know which one lies in the correct direction. What is lacking is “true
185
186
Image-Guided Surgery
north”. During registration, the coordinate system is transferred from the CT or MRI to the patient but not typically conveyed to the surgeon in a precise fashion. The problem is that, although one typically knows the approximate orientation of superior/inferior, right/ left, and anterior/posterior one rarely knows it precisely. Figure 8–1A depicts a central voxel (black) surrounded by its 26 neighboring voxels (blue outlines). Each neighbor shares one, two, or four corners with the central voxel. As shown in Figure 8–1B, if the central voxel represents the current position in the OR, one can relatively easily get from that voxel to the block of nine voxels superior (red) or inferior (green) to the current location, but not easily to a specific red or green voxel. This sort of specific navigation might be achieved by providing the coordinate reference frame (CRF) first presented in Figure
1–1 as an xyz origin to provide a local, referable coordinate system visible to the surgeon so that directionality of movements in the OR with respect to the imaging (eg, CT) is known. This would require orienting the xyz coordinate frame after registration. Some systems do provide visual navigation assistance (eg, Brainlab’s Curve software’s “probe’s eye view”; Figure 8–2), in which a desired target is specified, the current location is tracked, and a trajectory tunnel is provided with the surgeon moving the probe in various directions to determine which direction is along the tunnel. Visual feedback is then provided (eg, green represents staying on course and red means off-course). Similar systems are used in non-ENT IGS systems such as that offered by Mako Surgical (Ft. Lauderdale, Florida; http://www .makosurgical.com/makoplasty/knee
A
B
Figure 8–1. A. Central voxel (black ) surrounded by its 26 neighbors (blue outlines). Neighbors are voxels that share at least one common corner. B. The same set of 27 voxels of A is colored to show the block of nine superior neighbors (red ) of the central voxel and the block of nine inferior neighbors (green). Although it may be easy to navigate to the superior or inferior block, navigation to a specific voxel in either block can be much more challenging without knowing the precise orientation of the coordinate system.
Chapter 8
What Does the Future Hold?
n
Figure 8–2. Brainlab’s probe’s-eye view, termed AutoPilot in ENT 3.0, helps surgeons navigate along a selected trajectory by giving guidance regarding movement of the tracked probe. As seen in the top left panel, the system is suggesting that the probe be moved 1.6 mm in the specified direction and the pose, or angular alignment, be moved 0.4° in the opposite direction to align the probe to the trajectory. Image provided courtesy of Brainlab AG.
[Accessed December 31, 2015]) where red and green visual feedback (Figure 8–3) is used to guide a surgeon to drill a planar tibial plateau for knee replacement surgery with green indicating that additional drilling needs to be done and red dots indicating that drilling has reached the desired depth and should proceed no further.1 The Mako system further uses auditory feedback and also tactile feedback — restricting motion of the drill by a robotic arm — to make the surgical intervention more precise. In otolaryngology, a similar procedure has been experimentally performed in cadavers. This procedure, termed “navigated mastoidectomy”,2 involves tracking the surgical drill and disabling the drill by turning it off when it approaches preselected boundaries
(eg, sigmoid sinus, tegmen). The group recently reported on its use in clinical trials, albeit in a German article, the abstract of which is in English.3 This concept has also been applied to endoscopic sinus surgery (Figure 8–4) where a microdebrider is activated or deactivated according to its position relative to the anatomy.4,5
Augmented Reality Augmented reality is a broad topic and implies that additional information is provided to the surgeon that is not directly available from the immediate environment of the OR. We divide this topic into visual and nonvisual
187
A
B
Figure 8–3. Mako Surgical’s system uses visual, auditory, and tactile feedback to guide a surgeon in milling of the tibial plateau (A) and articulating femur (B) for artificial knee replacement. These screenshots show bone that should be removed (green), boundary (red outline), and areas where too much bone has been removed (red dots in A). The system also uses auditory signals as well as an articulated arm/robot to restrict the surgeon’s motion to prevent removal of too much bone. Republished with permission of the Journal of Bone and Joint Surgery, from Conditt MA, Roche MW. Minimally invasive robotic-arm-guided unicompartmental knee arthroplasty. J Bone Joint Surg Am. 2009;91(suppl 1):63–68.
Figure 8–4. Navigated endoscopic sinus surgery. The workspace is defined by the white region, and only in this area is the surgical debrider permitted to be active (ie, turned on). Republished with permission of Taylor and Francis, from Koulechov K, Strauss G, Dietz A, Strauss M, Hofer M, Lueth, FESS control: realization and evaluation of navigated control for functional endoscopic sinus surgery. Comput Aided Surg. 2006;11(3):147–159.
188
Chapter 8
versions and treat each in turn in the subsections below.
What Does the Future Hold?
n
The most commonly used augmentation is visual information in the form of surface rendering of selected anatomy overlaid onto intraoperative anatomy to provide subsurface vision (aka x-ray vision) so that the surgeon can visualize structures before encountering them (Figure 8–5). This technology has been around for decades with the visual augmentation being provided either on a computer monitor (eg, the “fourth” panel of an IGS system display) or into one eyepiece of a binocular operating microscope using a half-mirrored lenses to allow projection of the subsurface anatomy overlaying the visually apparent anatomy.6,7 Such systems require intensive computing power to process images and constantly re-register them to the surgical field in order to update the subsurface anatomy in
reference to the surface anatomy. This challenging problem is being solved today, and commercially available systems are becoming available. In neurosurgery, at least one startup company has received FDA clearance to use its augmented reality system in preoperative planning and intraoperative navigation and guidance (Surgical Theater, Mayfield Village, Ohio; http://www. surgicaltheater.net [Accessed December 31, 2015]). So where do we stand in otolaryngology? Regarding rhinology, perhaps the most useful application of such a system would be for transsphenoidal approaches to the skull base to allow identification of the carotid artery, cavernous sinus, and optic chiasm. At least one group has used a custom platform for such on twelve clinical cases.8 The reasons that this system did not reach wider clinical use are unclear but likely relate to the intensive computer processing required as continues to be evident in other clinical arenas, including endoscopic laparoscopy.9 In addition to
A
B
Visual Augmented Reality
Figure 8–5. Augmented visual reality. A. The right temporal skull surface with projection of the location of an arteriovenous malformation (AVM) shown in yellow. B. The same view with the skull and dura removed. Clearly visible is the outflow track as well as the tangle of blood vessels comprising the AVM. Such visual augmentation helps guide interactions by providing subsurface identification of anatomy. Republished with permission of Springer, from Cabrilo I, Bijlenga P, Schaller K. Augmented reality in the surgery of cerebral arteriovenous malformations: technique assessment and considerations. Acta Neurochir. 2014;156:1769–1774.
189
190
Image-Guided Surgery
transsphenoidal approaches to the skull base, such systems may find utility in lateral approaches to the skull base via the mastoid with visual augmentation delineating the cranial nerves, blood vessels, and tegmen. Although the benefit of visual augmented reality may seem obvious, a recently raised caveat warns of “inattentional blindness”, which consists of missing unusual or atypical findings when the user’s attention is directed toward other objects or tasks. As studied by Dixon et al, 10 surgeons may fail
Perhaps the best example of inattentional blindness (aka “selective inattention”) consists of a study published in 1999 in which participants were asked to count how many times a basketball was passed by players wearing white shirts. During the video (http://www .theinvisiblegorilla.com/videos .html), a person dressed in a gorilla outfit clearly walks into the midst of the players pounding his chest or a woman carrying a white umbrella walks across the scene. Despite this clear abnormality, in the more difficult tasks, only 54% of 192 participants noticed the unexpected activity — the gorilla or the woman — when focused on the task of counting the basketball passes! Unusual findings such as these support our push for rigorous study and demonstration of utility of technology before it is widely adopted — similar to our call in Chapter 6 for a randomized controlled trial of the utility of “obviously beneficial” IGS in ESS!
to notice unusual findings — a foreign body consisting of a screw or a cut optic nerve — when focused on a task highlighted by augmented reality, which is usurping the surgeon’s concentration.
Nonvisual Augmented Reality A potentially more beneficial area of augmentation is the use of nonvisual information to guide interventions. Nonvisual augmented reality has been pioneered in neurosurgery where visual distinctions between tissues may be impossible. One such application is deep brain stimulation (DBS) surgery in which electrodes are inserted into various nuclei in the mid-to-deep brain, typically to treat movement disorders. The subthalamic nucleus is a common target for movement disorders, as are the globus pallidus for Parkinson’s disease, the ventro-intermediate thalamic nucleus for essential tremor, and the ventrolateral thalamus for intentional tremor. Because these targets are typically indistinguishable from neighboring tissue, IGS is the ideal tool for this application. To target these nuclei, physiologic atlases are used to estimate their locations (Figure 8–6).11 These atlases consist of statistical predictions of the location of the target based on (a) anatomical measurements and (b) electrophysiological records from hundreds of surgical cases. Based on these atlases and CT and MR images of the patient, pre-aimed, patient-specific fixtures are designed to guide probes to specified locations via linear trajectories. This approach has also been used clinically to access the cochlea (Figure 8–7) and may be useful to target more proximal sites along the auditory pathway,
191
Figure 8–6. Coronal (top left), sagittal (top right), and axial (bottom left) MRI panels showing trajectories (green and purple lines) targeting the right and left subthalamic nuclei. This information is used to design a pre-aimed, patent specific fixture and superimpose a model (green) of it on a patent rendering (bottom right) which will be attached in the operating room to fiducial markers that have been placed prior to imaging.
A Figure 8–7. Use of image guidance to target the cochlea. A. Planning software depicts the facial nerve outlined in magenta, chorda typmani in green, ossicles in cyan, external auditory canal in yellow, and cochlea in red (scala tympani) and blue (scala vestibuli) in orthogonal CT views and surface rendering (bottom right). The planned trajectory is depicted in yellow and is achieved using a patient-customized microstereotactic frame (B) to guide a surgical drill along the trajectory. Republished with permission of John Wiley and Sons, from Labadie RF, Balachandran R, Noble JH, Blachon G, Mitchell J, Reda F, Dawant B, Fitzpatrick JM. Minimally-invasive, image-guided cochlear implantation surgery: first report of clinical implementation. Laryngoscope. 2014;124(8):1915–1922.
B 192
Chapter 8
including implantation in the modiolous,12 brainstem, or midbrain.13 Other potential applications may include use of image-guided atlases to identify anatomical areas that have different, yet to be determined, biological function but similar visual features.
Robots In the mind of the layperson, robots are the topic of science fiction and futuristic
What Does the Future Hold?
n
visions. However, robots have been in use in manufacturing since the 1960s (Figure 8–8) where they are typically used to replace monotonous tasks requiring a relatively high level of precision (eg, assembling engines). A major problem when discussing robots is the lack of a universal definition. People usually have an idea of what a robot is,i but to define it in words is quite difficult and often controversial. A good working definition that most agree upon is an automated machine that either replaces or augments human activity.
Figure 8–8. The first robot installed in American industry, this Unimate “pick-and-place” first removed hot metal parts from a die-casting machine at a GM plant in Trenton, New Jersey, in 1961. http://www. madehow.com/Volume-2/Industrial-Robot.html#ixzz3f8aP5VZ5. From the collections of The Henry Ford. i
One pioneer in robotics, Joseph Engelberger, is credited with saying “I can’t define a robot, but I know one when I see one.” (http://science.howstuffworks.com/robot.htm; Accessed December 31, 2015).
193
194
Image-Guided Surgery
Central to the use of surgical robots is IGS because IGS provides the geometric directions as to where the robot should move in determining where the tissue the robot should target is located. Robotic guidance with IGS may be achieved either by (1) optically tracking the robot and patient or (2) rigidly affixing the robot to the patient. With the first option, users must be aware of update rates for the optical tracking systems (Chapter 3) and the image processing required to transform the image to anatomical space and determine the current robot position. As computer processing speeds continue to increase, these updates are going faster and faster; however, a sudden sharp movement of the patient or robot may cause a lag in updating the information. Regarding the second option, rigid fixation to the patient eliminates the need for tracking once the initial registration is performed.
History The history of surgical robotics is as fascinating as the history of imaging detailed in Chapter 1 and has been the topic of review papersii and textbooks.iii A brief summary pertinent to our discussion follows. Using the convention suggested by Russ Taylor,14 a pioneer in surgical robotics, we discuss only autonomous robots, which are devices that perform complete tasks on their own following humangenerated instructions with minimal ii
human intervention during the task. (Instead of “autonomous”, Taylor uses the phrase “Surgical CAD/CAM” robots with the acronym CAD standing for “computer-aided drafting” and CAM for “computer-aided manufacturing”.) We limit our discussion to autonomous robots because only autonomous robots require IGS. Perhaps the most widely known autonomous robot is ROBODOC (Curexo Medical Technologies, Freemont, California), which is used to precisely mill a receiving hole in the proximal femur during prosthetic hip surgery and is discussed below. We will not delve deeply into surgical-assist robots, also known as “master-slave” devices or “tele-manipulators” (Taylor uses the phrase “Surgical Assistant”). These devices merely modulate a surgeon’s motions (eg, removing tremor and/or minimizing motions to allow access when anatomical boundaries are tight). Perhaps the most widely known surgical-assist robot is the DaVinci Surgical System (Intuitive Surgical, Inc, Sunnyvale, California). Although the DaVinci Surgical System has been used in otolaryngologic procedures (eg, transoral excision of tumors),15 at present — and for the foreseeable future — the use of surgical-assist robots does not require, or even incorporate, IGS.
History of Autonomous Robots Figure 8–9 depicts early autonomous robots. The first widespread investiga-
At the time of this writing (July 2015), a PubMed search revealed 1743 review articles on “surgical robotics” dating back to 1991. iii As with the review articles, there exist a plethora of textbooks on medical and/or surgical robotics such as that edited by Troccaz J. Medical Robotics. Hoboken, NJ: John Wiley; 2012.
Chapter 8
What Does the Future Hold?
n
A
B
C
Figure 8–9. Early autonomous robots. A. The Programmable Universal Manipulator Arm (PUMA). © Roger Ressmeyer/Corbis. B. The Whole Arm Manipulator (WAM). Image provided courtesy of Barrett, Inc and William Townsend, CEO. C. Early ROBODOC. Republished with permission of Think Surgical.
tion of a surgical robot was of an articulated arm modified from an industrial robot. This robot — the “Programmable Universal Manipulator Arm” or PUMA
was manufactured by Westinghouse — Corporation, headquartered at the time in Pittsburgh, Pennsylvania. It was first used clinically by Kwoh in 1985 to
195
196
Image-Guided Surgery
biopsy a brain tumor.16 Drake et al.17 reported on the use of the PUMA on six children who underwent neurosurgical resections. The robot was compared with the ISG viewing wand presented in Chapter 1 (see Figure 1–11) as used on 43 similar pediatric surgical interventions. The PUMA was found to be bulky and cumbersome and added little to what the ISG viewing wand could accomplish.18 Because of these limitations, this group and others migrated away from industrial-based, articulated arm, robotic systems and toward tracked probes. Lagging just behind the adaptation of industrial articulated arms to the operating room was the design of a novel robot with a goal of providing machine precision in performing hip replacement surgery. Starting in 1986, researchers at the University of California at Davis began work on development of a robot19 subsequently named ROBODOC. Licensed to Integrated Surgical Systems (Davis, California), it was first used clinically in the early 1990s20 to bore a receiving well for a hip prosthesis. Following its initial success, it was lauded with a Computerworld Honors Awardiv and used in more than 28 000 surgical cases. In practice, the device was either rigidly affixed to the femur or registered using pins placed the day before the procedure prior to CT scanning. Use of the system was shown to be superior to what a human iv
operator could perform without it in terms of the precision of the receiving well for the prosthesis. However, a class-action lawsuit filed in Germany in 2003v claimed malpractice because patients were left crippled due to the surgical approach in which muscle and nerves were transected. The resolution led to a ruling in favor of the plaintiff in 2005,vi and the company ceased operation in 2005, following which the technology was acquired by Curexo Medical Technologies (Freemont, California). Curexo, renamed THINK Surgical, Inc. in 2014, received FDA clearance for the ROBODOC in 2008vii with the system currently commercially available. According to its website, ROBODOC has found a recent resurgence in clinical applications for total-hip arthroplasty using surface registration which obviates the need for rigid fixation or pins.vii At approximately the same time that the PUMA was being clinically tested, the Whole Arm Manipulator, or WAM, was being developed at the Laboratory for Electromagnetic and Electronic systems at the Massachusetts Institute of Technology21 and the Active Constraint Robot (ACROBOT) at Imperial College in London.22 Differentiating the WAM and ACROBOT from industrial robots was their back-drivable cable system, which allowed passive movement of the robot to desired locations by the human users. The WAM was licensed to Mako Surgical Systems in 1998, where
http://www.cwhonors.org/search/caa_5.asp (Accessed December 31, 2015). http://www.scotsman.com/news/world/crippled-german-patients-sue-cyber-surgeon- 1-1294536 (Accessed December 31, 2015). vi http://www.bizjournals.com/sacramento/stories/2005/06/06/daily42.html (Accessed December 31, 2015). vii http://www.robodoc.com/patient_about_history.html (Accessed July 8, 2015). v
Chapter 8
What Does the Future Hold?
n
it was redesigned, resulting in the RIO Robotic Arm Interactive System, and, after receiving FDA clearance in 2005, has been used in knee replacement surgery to limit surgeons’ motions to those that create an optimal seat for a prosthetic joint (see Figure 8–3). The ACROBOT was acquired by Mako in the settlement of a patent infringement lawsuit.
tem affixes directly to the patient, and while initially used for placement of pedicle screws in spine fixation surgery, it has received FDA clearance to target intracranial sites and likely will be used clinically for tumor biopsy and/or DBS placement.
Current FDA-Cleared Autonomous Robots
In terms of otolaryngology, the robots listed above are available, use IGS, and could be used in otolaryngologic surgeries today (eg, the NeuroMate, Rosa, and Renaissance are FDA cleared for intracranial targeting, which would include the cochlea and thus could be programmed to target the cochlea as shown in Figure 8–7). But, in light of the discussion above, it may be better to ask, “What surgical scenarios would robots be better suited at performing than human operators?” This question is perhaps best answered by citing the reason that robots were initially used in industry — namely, to replace monotonous human tasks requiring a high level of precision. Although surgeons may claim that their jobs are never “monotonous”, most would agree that many elements of surgery are, at the least, highly “repetitive”. So, what surgical tasks are highly repetitive and require a high level of precision? These tasks are ones that a robot may be best suited for. In otolaryngology, one such task is mastoidectomy. Familiar to all otolaryngologists, a mastoidectomy is a surgical procedure in which a specified volume of the temporal bone, located medial and posterior to the external ear, is milled
As in the field of IGS, neurosurgical applications continue to dominate the field of surgical robots relevant to otolaryngology. For the same reasons that IGS expanded into neurosurgery — that tissues could not be visually differentiated — robots that use IGS have expanded into this realm. The largest clinical applications are (1) tumor biopsy (ie, identify the tumor in the MRI and direct the robot to target the tumor for biopsy), (2) placement of DBS (see Figure 8–6), and (3) ablative treatment such as those used to treat intractable seizures secondary to epilepsy. At least three neurosurgical robots are FDA cleared for such indications (Figure 8–10): the NeuroMate (Renishaw PLC, Wotton-Under-Edge, United Kingdom), Rosa (MedTech S.A.S., Castelnau Le Lez, France), and Renaissance (Mazor Robotics Ltd., Caesarea, Israel).23 The NeuroMate and Rosa are articulated arm systems that can rigidly mate to the surgical bed on which the patient has been fixed, typically with a Mayfield head holder, or can be used in a “frameless” configuration with tracking. The Renaissance sys-
What Autonomous Robots Does Otolaryngology Need?
197
198
Image-Guided Surgery
A
B
C Figure 8–10. Current autonomous robots. A. The NeuroMate. Republished with permission of Renishaw, Inc. B. Rosa. Republished with permission of MedTech, Inc. C. The Renaissance system. Republished with permission of Mazor Robotics, Inc. The NeuroMate and Roas are articulated arm robots. The Renaissance system is a parallel-platform robot that is mounted directly on the patient and is much smaller (a little larger than a can of soda).
away (Figure 8–11). Given that the boundaries are known and can be identified preoperatively, the use of a robot following an IGS plan would seem to be an ideal application of such technology. Indeed, in engineering applications, similar tasks are performed
by articulated-arm robots (eg, cutting stone countertops, Figure 8–12). In 2003, Federspil et al24 reported on the use of a force-controlled, articulatedarm robot to mill the receiving well for a cochlear implant (Figure 8–13), and in 2011, Danilchenko et al25 reported
Figure 8–11. In a mastoidectomy, the temporal bone is milled away in a roughly pyramidal volume with the superior boundary being the tegmen (floor of the middle cranial fossa), the anterior boundary being the posterior aspect of the external auditory canal, and the posterior boundary being the intracranial continuation of the internal jugular vein, the sigmoid sinus. This procedure is typically performed manually using a high-speed surgical drill rotating at 80 000 RPM and magnified vision using an operating microscope that can magnify up to 6×.
Figure 8–12. Industrial robot cutting stone for a countertop. Robo SawJet image provided courtesy of BACA Systems (Auburn Hills, MI).
199
200
Image-Guided Surgery
A
C
B
D
Figure 8–13. First use of a robot for otosurgery in a cadaver. Federspil et al modified a Staubli* robot (A) with a force and torque sensor and affixed a high-speed surgical drill (B) to mill out the receiving well for a cochlear implant (C and D). Republished with permission of John Wiley and Sons, from Federspil PA, Geisthoff UW, Henrich D, Plinkert PK. Development of the first force-controlled robot for otoneurosurgery. Laryngoscope. 2003;113(3):465–471. *Noting that Staubli purchased PUMA.
on the use of a modified industrial robot to perform a cortical mastoidectomy (Figure 8–14). Most recently, Dillon et al26 have reported on the design of a bone-affixed robot to perform a mastoidectomy and translabyrinthine approach to the internal auditory canal to access vestibular schwannomas (aka acoustic neuromas), which are benign
tumors located in the cerebellopontine angle with local growth causing compression of vital structures including the facial and auditory nerves (Figure 8–15). The hypothetical advantage of having a robot perform a mastoidectomy and translabyrinthine approach is time reduction for the access part of the surgery as well as preservation of the
Figure 8–14. First demonstration of robotic mastoidectomy using a modified Mitsubishi Industrial robot termed the Otobot. The temporal bone is mounted on the left and has a fiducial frame rigidly affixed to it, as does both the base of the robot (bottom right corner ) and the end effector, which holds the drill.
Figure 8–15. Bone-affixed robot designed to perform mastoidectomy and translabyrinthine approach to the internal auditory canal. The lower black component is a patient positioning frame (PPF), which is screwed into the skull cortex using facial plating screws. Spherical balls are used as fiducials in registering the patient’s anatomy to the robot, which is rigidly mounted to the PPF by the user (gloved hand ). The robot has four degrees of freedom — x, y, and z linear travel (double-headed arrows ) and rotation of the drill (curved arrow ).
201
202
Image-Guided Surgery
surgeons’ effort for the more delicate tumor removal. With demonstration of the ability to replicate standard surgical approaches, investigators have reached out to explore approaches that would not be possible without the use of IGS and robots. One such approach is minimally invasive access to targets within the temporal bone, including the cochlea. This approach was first clinically demonstrated using customized microstereotactic frames as discussed above27 (see Figure 8–7). At least three groups have experimented with minimally invasive robotic access to the cochlea: Vanderbilt University28; Medical Hospital of Hannover, Germany29,30; and the University of Bern, Switzerland,31,32 with the Bern group (Figure 8–16) having regulatory approval to move ahead with clinical testing (personal communication, 2015). Potential benefits of
such automated, minimally invasive interventions include shorter interventional time, standardization of surgical approach, and ability for surgeons with less training to perform more complex tasks that may have significant impact in developing countries where adequately trained surgeons are often in short supply.
Conclusions Although all of the above technology is achievable and probably inevitable, experienced surgeons have expressed concern that it will lead to dependence on the technology. This concern may be valid, but numerous studies have shown that using such technology can make it possible for novice surgeons to safely and effectively perform more difficult
Figure 8–16. Articulated arm designed for minimally invasive cochlear implantation by a group at the University of Bern, Switzerland and shown as set up in a mock OR with a phantom skull. Image provided courtesy of CASination (Bern, Switzerland).
Chapter 8
surgical interventions. Furthermore, for robotic interventions, a surgeon’s role may evolve into one of oversight of the intervention. Such conflicts, which often accompany technological advances that lead to paradigm shifts, will ultimately be resolved by studies of patient outcomes. However, most people — including the authors — agree that we have yet to even approach total automation in surgical interventions, and, as we pointed out early in this book (Chapter 1), knowledge of surgical anatomy coupled with an understanding of the limits of the accuracy of navigation through that anatomy is paramount for the use of these technologies now and for the foreseeable future. Thus, the field of Image-Guided Surgery has a way to go before it becomes ubiquitous. Still, it is our hope that by presenting the current state of the art of IGS, this book might inspire some who read it to make a reality of the advances envisioned here and others to pioneer methods that will give tomorrow’s surgeons abilities that can only be imagined today.
References 1. Conditt MA, Roche MW. Minimally invasive robotic-arm-guided unicompartmental knee arthroplasty. J Bone Joint Surg Am. 2009;91(suppl 1):63–68. 2. Strauss G, Koulechov K, Hofer M, et al. The navigation-controlled drill in temporal bone surgery: a feasibility study. Laryngoscope. 2007;117(3):434–441. 3. Strauss G, Schaller S, Naidu S, et al. Clinical experience with navigation functions for temporal bone surgery: interim results after 40 patients [in German]. HNO. 2012;60(12):1115–1121.
What Does the Future Hold?
n
4. Koulechov K, Strauss G, Dietz A, Strauss M, Hofer M, Lueth TC. FESS control: realization and evaluation of navigated control for functional endoscopic sinus surgery. Comput Aided Surg. 2006;11(3):147–159. 5. Strauss G, Hofer M, Fischer M, et al. First clinical application of a navigationcontrolled shaver in paranasal sinus surgery. Surg Technol Int. 2008;17:19–25. 6. Roberts DW, Strohbehn JW, Hatch JF, Murray W, Kettenberger H. A frameless stereotaxic integration of computerized tomographic imaging and the operating microscope. J Neurosurg. 1986;65(4): 545–549. 7. King AP, Edwards PJ, Maurer CR Jr, et al. Stereo augmented reality in the surgical microscope. Presence. 2000;9(4):360– 368. 8. Kawamata T, Iseki H, Shibasaki T, Hori T. Endoscopic augmented reality navigation system for endonasal transsphenoidal surgery to treat pituitary tumors: technical note. Neurosurgery. 2002;50(6):1393–1397. 9. Nakamoto M, Ukimura O, Faber K, Gill IS. Current progress on augmented reality visualization in endoscopic surgery. Curr Opin Urol. 2012;22(2):121–126. 10. Dixon BJ, Daly MJ, Chan HHL, Vescan A, Witterick IJ, Irish JC. Inattentional blindness increased with augmented reality surgical navigation. Am J Rhinol Allergy. 2014;28(5):433–437. 11. D’Haese PF, Pallavaram S, Li R, et al. CranialVault and its CRAVE tools: a clinical computer assistance system for deep brain stimulation (DBS) therapy. Med Image Anal. 2012;16(3):744–753. 12. Sobhy Afifi WF, Guigou C, Mazalaigue S, Camuset JP, Ricolfi F, Bozorg Grayeli A. Navigation-guided transmodiolar approach for auditory nerve implantation via the middle ear in humans. Audiol Neurotol. 2015;20(2):128–135. 13. Lim HH, Lenarz M, Lenarz T. Auditory midbrain implant: a review. Trends Amplif. 2009;13(3):149–180.
203
204
Image-Guided Surgery
14. Taylor RH. A perspective on medical robotics. Proceedings of the IEEE. 2006; 94(9):1652–1664. 15. O’Malley BW Jr, Weinstein GS, Snyder W, Hockstein NG. Transoral robotic surgery (TORS) for base of tongue neoplasms. Laryngoscope. 2006;116(8): 1465–1472. 16. Kwoh YS, Hou J, Jonckheere EA, Hayati S. A robot with improved absolute positioning accuracy for CT-guided stereotactic brain surgery. IEEE Trans Biomed Eng. 1988;35(2):153–160. 17. Drake JM, Joy M, Goldenberg A, Kreindler D. Computer- and robot-assisted resection of thalamic astrocytomas in children. Neurosurgery. 1991;29(1):27–33. 18. Drake JM, Prudencio J, Holowka S. A comparison of a PUMA robotic system and the ISG viewing wand for neurosurgery. In: Maciunas RJ, ed. Interactive Image-Guided Neurosurgery. New York, NY: Thieme Publishing Group; 1994: 121–133. 19. Paul HA, Bargar WL, Mittlestadt B, et al. Development of a surgical robot for cementless total hip arthroplasty. Clin Orthop Relat Res. 1992;285:57–66. 20. Bargar WL, Bauer A, Börner M. Primary and revision total hip replacement using the ROBODOC system. Clin Orthop Relat Res. 1998;354:82–91. 21. Townsend WT, Salisbury JK. Mechanical design for whole-arm manipulation. In: Dario P, Sandini G, Aebischer P, eds. Robots and Biological Systems: Towards a New Bionics? Berlin, Germany: Springer Berlin Heidelberg; 1993;102(NATO ASI Series):153–164. 22. Davies BL, Harris SJ, Lin WJ, Hibberd RD, Middleton R, Cobb JC. Active compliance in robotic surgery — the use of force control as a dynamic constraint. Proc Inst Mech Eng H. 1997;211(4):285–292. 23. Shoham M, Burman M, Zehavi E, Joskowicz L, Batkilin E, Kunicher Y. Bonemounted miniature robot for surgical
procedures: concept and clinical applications. IEEE Transactions on Robotics and Automation. 2003;19(5):893–901. 24. Federspil PA, Geisthoff UW, Henrich D, Plinkert PK. Development of the first force-controlled robot for otoneurosurgery. Laryngoscope. 2003;113(3):465–471. 25. Danilchenko A, Balachandran R, Toennies JL, et al. Robotic mastoidectomy. Otol Neurotol. 2011;32(1):11–16. 26. Dillon NP, Balachandran R, Fitzpatrick JM, et al. A compact, bone-attached robot for mastoidectomy. J Med Devices. 2015;9(3):0310031–310037. 27. Labadie RF, Balachandran R, Noble JH, et al. Minimally invasive imageguided cochlear implantation surgery: first report of clinical implementation. Laryngoscope. 2014;124(8):1915–1922. 28. Kratchman LB, Blachon GS, Withrow TJ, Balachandran R, Labadie RF, Webster RJ. Design of a bone-attached parallel robot for percutaneous cochlear implantation. IEEE Trans Biomed Eng. 2011;58(10):2904–2910. 29. Majdani O, Rau TS, Baron S, et al. A robot-guided minimally invasive approach for cochlear implant surgery: preliminary results of a temporal bone study. Int J Comput Assist Radiol Surg. 2009;4(5):475–486. 30. Kobler JP, Prielozny L, Lexow GJ, Rau TS, Majdani O, Ortmaier T. Mechanical characterization of bone anchors used with a bone-attached, parallel robot for skull surgery. Med Eng Phys. 2015;37(5): 460–468. 31. Bell B, Stieger C, Gerber N, et al. A selfdeveloped and constructed robot for minimally invasive cochlear implantation. Acta Otolaryngol. 2012;132(4): 355–360. 32. Bell B, Gerber N, Williamson T, et al. In vitro accuracy evaluation of imageguided robot system for direct cochlear access. Otol Neurotol. 2013; 3 4(7): 1284–1290.
Index Note: Page numbers in bold refer to figures and tables.
A Accuracy of optical (infrared) IGS systems Brainlab Curve and Kick, 169–171 Medtronic StealthStation, 158–162, 161, 162 Stryker systems, 179–180 of electromagnetic (EM) systems InstaTrack, 171–172. See also Errors, Target registration error (TRE) Acoustic neuroma, 200, 201 Active Constraint Robot (ACROBOT), 196, 197 Active IR tracking, 81–82, 172–173, 176 ADAPT system (Stryker), 153, 173, 174 Adipose tissue, magnetic resonance imaging (MRI), 57, 67–68 Advanced Realtime Tracking GmbH, 87 Airo Mobile CT scanner (Brainlab Corp.), 31, 33, 34, 39, 153, 163 American Rhinologic Society (ARS), 141, 142, 145–146 Armless tracking systems, 18 Articulated arm tracking systems, 19, 75, 188, 197 Artis Zee C–arms (Siemens), 31 Ascension Technology Corp., 94 Atracsys, 87 Augmented reality, 187, 189–190, 189–192, 192 Nonvisual augmented reality, 190, 191, 192, 193 Physiologic atlases, 190, 191 Surgical Theater, 189
Aurora EMF generation unit (NDI), 93, 94, 95–97, 180 Autonomous robots, 194–198, 195, 198–202, 200, 202 AutoPilot (Brainlab Corp.), 187 Averaging, MRI, 59–60
B Bakken, Earl, 153 Beam hardening. See Computed Tomography (CT) Best practices for IGS, 141–148 Black-body radiation. See Magnetic Resonance Imaging (MRI) Bloch, Felix, 57 Bocage, André-Edmund-Marie, 5, 6 BodyTom CT scanner (Neurologica Corp.), 31, 33, 34, 39, 39, 40 BOLD imaging, 68. See also Magnetic resonance imaging (MRI) Bone-implanted fiducials. See Fiducial markers and points; N-bar systems; Stereotactic frames Boston Scientific, 163 Brainlab Corp., 19, 31, 33, 34, 110, 112, 151, 152, 153, 159, 162–171, 186, 187 accuracy of Brainlab systems, 169–171 AutoPilot, 187 Curve system, 153, 163, 163, 164, 186 Kick system, 110, 153, 163, 164, 172 registration with Brainlab systems, 167–169
205
206
Image-Guided Surgery
Brainlab Corp (continued) user guides, 164–167, 164, 165 Cranial ENT 2.1 Software User Guide, 167, 167–168 Curve 1.0 System Technical User Guide, 166–167 Curve 1.0 System User Guide, 164–166 VectorVision Compact system (Brainlab Corp.), 163, 169, 170, 170, 171–172 Breakaway Imaging, LLC, 36, 37 Brown, Russell, 14, 14 Brown-Roberts-Wells frame (BRW frame), 14 Bull, Dr., 3, 4, 141
C C-arms (CT scanners), 31, 169 Calypso System (Varian Medical Systems), 90 Cantilever effect, 112, 112, 169, 171 Cerebrospinal fluid (CSF), 57, 142, 143, 145 CereTom CT scanner (Neurologica Corp.), 31, 33, 34 ClaroNav, 76, 77, 81, 88, 89, 91, 151, 173, 181, 182 Clayton, J.H., 3 Clinical studies of Brainlab systems, 170, 170–171 of InstaTrak system, 171–172 of Medtronic systems, 161, 162 unavailable for Stryker systems, 180 Closed-bore MRI scanners, 40, 41 Cochlea, image guidance to target the cochlea, 190, 192, 202, 202 Computed tomography (CT), 25–40, 63, 72, 103, 105, 187 attenuation coefficient, 27, 27–29, 35 beam hardening, 35 DICOM image, 30 discovery, 4–6 fiducial localization error, 70–71 field of view (FOV), 29 fusion with MRI, 16, 17, 114 geometric distortion, 13, 14, 33, 34, 39 helical scanning, 15, 29
history, 4–6, 5, 13–15 interoperative scanners stationary scanners, 30–31 Siemens Medical Systems, 31, 39, 61 VISIUS iCT (IMRIS), 31, 32 Zeego (Siemens), 31 portable scanners, 31, 33, 34–35, 35, 36, 37–39, 39–40, 153, 172 Airo Mobile CT scanner (Brainlab Corp.), 31, 33, 34, 39, 153, 163 BodyTom and CereTom CT scanners (Neurologica Corp.), 31, 33, 34, 39, 39, 40, 153, 172 Breakaway Imaging, LLC, 36, 37 C-arms (CT scanners), 31, 169 Mobius Imaging LLC (Brainlab Corp. Airo), 33, 34, 153, 163 O-arm CT scanner (Medtronic, Inc.), 31, 33, 36, 37 xCAT (Xoran, Technologies LLC), 31, 36, 38, 180 geometrical accuracy of commercial scanners, 33, 34, 39, 40 geometrical distortion of Neurologica scanners, 33, 34, 153, 172 method of operation, 25–30, 26–27, 64 motion artifact, 13, 29 multislice CT (MSCT) scanners, 15, 29, 33 N-bar systems, 13–14, 14, 16, 18, 106 obsolescence, 15, 17 noise, 35, 35 reconstruction, 6, 15, 36, 45–46, 49 slip ring, 15 scatter, 25, 26 spiral scanning, 15. See also helical scanning Computer-assisted navigation, 185–187, 186–189. See also Augmented reality Cone-beam scanners. See fpVCT scanners Coordinate reference frame (CRF), 2, 110, 113, 125–126, 137, 157, 158, 164, 174, 176, 176, 178–181, 186 Cormack, Alan MacLeod, 6 Cosman-Roberts-Wells frame (CRW frame), 14 Cox, John, 3
Index
Cranial ENT 2.1 Software User Guide, 166–167, 167 CranialMap Express Navigation system (Stryker), 173, 175 CRF. See Coordinate reference frame CRW frame. See Cosman-Roberts-Wells frame CSF. See Cerebrospinal fluid CT. See Computed tomography Cupping artifact, 35. See also Computed tomography (CT) Curexo Medical Technologies, 194, 196 Curve 1.0 System User Guide, 164, 166 Curve 1.0 System Technical User Guide, 166–167 Curve system (Brainlab Corp.), 153, 163, 163, 164, 186
D Damadian, Raymond V., 8, 9 DaVinci Surgical system (Intuitive Surgical), 194 Deep brain stimulation (DBS) surgery, 162, 190, 197 Degrees of freedom (DOF) of electromagnetic (EMF) systems, 95 Dental casts, 106 Dephasing, MRI, 52, 53, 54, 55, 56, 57, 66, 68–69, 79 DICOM image, 30, 30 Diffusion-weighted imaging (DWI), 68–69 Distortion in imaging magnetic resonance imaging (MRI), 64–68. See also Magnetic resonance imaging (MRI) computed tomography (CT), 13, 14, 33, 39. See also Computed tomography (CT) DOF. See Degrees of freedom DWI. See Diffusion-weighted imaging
E Echo-planar imaging (EPI), 10–12, 61, 65, 65. See also Magnetic resonance imaging (MRI)
Echo time, MRI, 53–54, 57–58, 53, 55–57, 58 Echoes, MRI, 52–57 Edelstein, William, 10 Electric and Musical Industries (EMI), Ltd., 4, 6 Electromagnetic (EM) tracking, 2, 2, 18, 19, 75–76, 90, 92, 94, 92–96, 152 clinical accuracy, 171–172 components, 92–94, 93–96 usefulness, 90 Electromagnetic field (EMF), 2, 2, 18, 19, 92, 93 EMI, 4–6, 5 EMIDEC 1100, 6 EM tracking. See Electromagnetic tracking. Endoscopic sinus surgery (ESS), 19, 21, 142–144, 146, 152, 156, 164, 169, 170, 171, 174, 175, 181, 190 Entellus, 180 EPI. See Echo-planar imaging Ernst, Richard, 10, 46 Error analysis, 117–139 Error extrinsic, 118, 122 intrinsic, 118, 122 Maxwell distribution, 85–86, 86, 122 statistical nature of prediction, 117–118 subtotal measurement error, 159–160, 170, 171 total IGS error, 78. See also Error analysis, Fiducial localization error (FLE); Fiducial registration error (FRE); Probe localization error (PLE); Target registration error (TRE) ESS. See Endoscopic sinus surgery
F Facial plating screws, 103, 105, 105 FARO Medical Technologies, 18 Fast spin echo, MRI, 61 – 62 FESS. See Endoscopic sinus surgery Fiagon IGS system, 151, 180–181, 181 Fiducial localization error (FLE), 119–120, 120
207
208
Image-Guided Surgery
Fiducial localization error (FLE) (continued) assumptions concerning, 123 in CT, 70 in general image, 71 of passive IR marker, 81 RMS relationship to FRE, 127 to TRE, 122–125, 126, 127, 137 root of all error, 119–122, 127, 158 of stereotactic frame, 129 Fiducial markers and points, 100–109 bone-implanted, 18, 30, 102–103, 103–105, 105–106, 113, 125, 134, 137, 156, 158 configuration, 76, 77, 77, 78, 87, 101, 105, 124, 125, 126, 138 placement related to TRE, 100, 124–126, 126, 137 dislodgement, displacement 105, 106, 133 during navigation, 2, 113 examples, 21, 89, 101, 102, 103–104, 105, 110, 167, 176 facial plating screws, 103, 105, 105 for fixture attachment, 191 for frameless stereotaxy, 17–18 IGS-manual instructions, 166–169, 177 IR reflective spheres, 81–82, 81, 110, 167 Mayfield head holder, 20, 113, 125, 176, 197 myths, 129–138 planar configurations, 137, 138 N-bar systems, 13–14, 14, 106 oil-based problem, 68 pixelated image, 70, 81, 89 IR reflective spheres, 81, 81, 167 improper handling, 81, 81, 82 patient positioning frame (PPF), 201 skin-affixed, 18, 100, 101–102, 102, 113, 157, 170, 174 IZI Medical Products, 100 using fewer fiducials, 108, 124, 127, 134–137, 136 using more fiducials, 109, 124–125, 127, 128, 137, 160
Xpoint, 88, 89, 181. See also Coordinate reference frame (CRF), Fiducial localization error (FLE); Fiducial registration error (FRE); Target registration error (TRE) Fiducial registration error (FRE), 108, 120–121 Brainlab categorization, 168 clinical reports, 162, 162, 170–171 information from, 157–158 minimization by registration, 107–108 myths, 130–138 fallacious TRE prediction, 128, 130–134, 130–132, 158, 168, 179–180 misguided fiducial omission, 134–137, 136, 138, 168–169, 177, 179–180 planar configurations, 137, 138 RMS relationship to FLE, 127 to number of fiducials, 109, 109, 127, 128, 135, 136 to TRE, 127 in surface registration, 111 Field of view (FOV), 29, 45, 82, 83, 85, 87, 90, 92, 94 FLASH, MRI, 61 FLE. See Fiducial localization error Fluoroscopy units, 31 fMRI. See Functional imaging Food and Drug Administration (FDA), 18, 142, 153, 180, 181, 189, 196, 197 FOV. See Field of view fpVCT scanners (flat-panel-volume CT scanners), 35, 35–36, 37, 38 “Frameless” stereotaxy, 17 “Frameless” tracking systems, 18, 197 Frank, Gabriel, 5 FRE. See Fiducial registration error Free-induction decay, 52, 53 Frequency encoding, MRI, 10, 42–45, 46, 47, 48 Functional endoscopic sinus surgery. See Endoscopic sinus surgery Functional imaging (fMRI), 68–69
Index
Fusion (Medtronic model name), 154, 155, 172 Fusion of CT an MRI images, 16, 17, 113–114
G Galloway, Robert L., 17, 158 General Electric (GE), 19, 41, 61, 151, 152 Geometric distortion computed tomography (CT), 13, 14, 33, 34, 39 magnetic resonance imaging (MRI), 10, 11, 14, 46, 61, 62, 64–66 Glasscock, Mike, 4, 5 Goodness of fit, 121 Gradient coils. See Magnetic resonance imaging (MRI) Gradient echo. See Magnetic resonance imaging (MRI) GRAPPA, GeneRalized Autocalibrating Partially Parallel Acquisition MRI, 62
H Hahn, Erwin, 53 Hahn echo, 53 Helical scan computed tomography, 15, 29 Hermundslie, Palmer, 153 Homogeneity level, 42 “Horseless carriage”, 17 Hounsfield, Godfrey Newbold, 4–6, 8, 25, 28. See also Computed tomography (CT) Hounsfield unit (HU), 28, 38, 110, 175
I ICP. See “Iterative closest point” algorithm IGS technology vendors, 151–183 Brainlab Corp., 19, 31, 33, 34, 110, 112, 151, 152, 153, 159, 162–171, 167, 186, 187 ClaroNav, 76, 77, 81, 88, 89, 91, 151, 173, 181, 182
cost, 146–147 Fiagon, 151, 180, 181 Mako Surgical Systems, 186–187, 188, 196–197 Medtronic, Inc., 19, 31, 36, 37, 151, 152–166, 154, 157 Neurologica Corp., 31, 33, 34, 153, 172 Northern Digital, Inc. (NDI), 82–84, 86, 87, 93, 93–96, 95, 96, 152, 172 Stryker Corp., 19, 151, 152, 153, 172–180, 173, 174, 176 Image acquisition time, 15, 58–64 Image-guided surgery (IGS) about, vii, 1–3, 2, 147–148 augmented reality, 187, 189–190, 189–192, 192 best practices, 141–148 complications, 143–144 computer-assisted navigation, 185–187, 186–188 errors, 70–71, 78, 82, 85, 87, 117–138 financial sustainability, 145–146 historical timeline, 16 history, 1, 12–15, 13, 14, 16, 17–19, 17, 19–21, 21 legal liability, 147 market reports, 151–152 need for revision surgery and, 144 professional society position statements, 144–145, 145 registration, 2, 17, 71, 75, 99–114 robots, 193–198, 193, 195, 198–202, 200, 202 safety, 142–144 standard of care, 147 surgeons’ performance and, 144 tracking systems, 75–97 2D presentation of 3D images, 71–73, 72 use of term, 17 users of, 141–142 viewing wand, 18, 20. See also Computed tomography; FDA approval; Fiducial markers and points; IGS technology vendors; Magnetic resonance imaging; X-rays Image Guided Technologies Inc., 172
209
210
Image-Guided Surgery
Imaging history, 3–12 tomography, 5 2D presentation of 3D images, 71–73, 72. See also Computed tomography; Image-guided surgery; Magnetic resonance imaging; X-rays IMRIS, Inc. (Innovative Magnetic Resonance Imaging Systems), 31, 32 Inattentional blindness, 190 iNAV3 System (Stryker), 153 Infrared (IR) tracking, 18, 19, 80–81, 81, 176, 176. See also Tracking InstaTrak system (GE/Visualization Technologies, Inc.), 19, 21, 151, 152, 171–172, 173 Integrated Surgical Systems, 196 Intensity distortion, magnetic resonance imaging (MRI), 66–67 Interactive Image-Guided Neurosurgery (Maciunas), 17 Intraopertive CT scanners 30–31, 32, 33, 34–35, 35, 36, 37–39, 39–40, 153, 172 geometric distortion, 33, 34, 39, 39, 153, 172 Intrinsic error, 118, 122 Intuitive Surgical, Inc., 194 ISG Technologies, 18, 20 “Iterative closest point” algorithm (ICP), 111, 111, 114 IZI Medical Products, 100, 101
J “Jacobian” effect, 67
K Kick system (Brainlab Corp.), 110, 153, 163, 164, 172 Kwoh, Y.S., 195
L LandmarX system (Medtronic, Inc.), 162 Lauterbur, Paul C., 8–9, 10, 46
Le Bihan, Dennis, 69 Leksell frame, 14 Lexmark International, Inc., 88 Line of sight (LOS), 72, 75
M Macintyre, John, 4 Maciunas, Robert J., 17, 158 Magnetic field, effect on body, 7, 13, 62 Magnetic resonance imaging (MRI), 40–70 Averaging, 59–60 BOLD imaging, 68 black-body-radiation noise, 60 coordinate axes, 43, 43, 44, 44 compromises, 69–70 dephasing, 52, 53, 54, 55, 56, 57, 66, 68–69, 79 diffusion-weighted imaging (DWI), 68–69 discovery, 6–8 distortion, 64–68 geometric, 64–66, 67–68 fat shift, 67–68 intensity, 66–67 echo, 52–57 echo-planar imaging (EPI), 10–12, 61, 65, 65 fast spin echo, 61 gradient echo, 54–55, 55, 56, 61, 62, 66, 67, 68 spin echo, 52–54, 55, 56, 61, 66, 67, 69 turbo spin echo, 61 excitation, 8, 40, 52 90-degree, 53, 54 number of (NEX), 60 pulse, 48, 48, 53 repetition, 49, 57 repetition time (TR), 58 slab selection, 50–51 slice selection, 8, 44, 46, 46 small-angle, 60–61 fat shift, 67–68 fiducial localization error, 70–71 field strength, 40, 42 FLASH (Siemens), 61 frequency encoding, 10, 42–45, 46, 47, 48
Index
functional imaging (fMRI), 68–69 fusion with CT, 16, 17, 114 giant voxels, 45, 46, 47 gradient coils, 42 history, 6–12, 9, 11, 12, 15, 16, 17 image acquisition time, 58–64 intraoperative, (VISIUS iMRI), 31 “Jacobian” effect, 67 localization by means of gradients, 42–51 magnet, 40, 42 multiplanar imaging, 59 noise, 52, 59–60, 69 parallel imaging, 62 phase encoding, 46–51, 48–50 multiple per excitation, 61 proton-density weighting, 58, 58 pulse sequence, 48–56, 48–51, 52–56 relaxation times, 57–58, 58 scanners, 40, 41 signal-to-noise (S/N), 59, 60 slice selection, 8, 10, 43–44, 44 SPGR (GE), 61 spin-spin interactions, 53, 57 static-field inhomogeneity, 52–54, 53, 54, 65–67, 68, 160 T1-FFE (Philips), 61 T1 weighting, 57–58, 57–58, 63 T2 weighting, 57–58, 57–58, 63 TE (time of echo), 53–54, 57–58, 53, 55–57, 58 TR (repetition time), 57–58, 58 volume imaging, 50–51, 51, 56, 60, 61, 66 Mako Surgical Systems, 186–187, 188, 196–197 Mansfield, Peter, 8, 9, 10, 11, 12, 12, 13, 62 Market reports, 151–152 Surgical Navigation Systems, US, Market Analysis (Millenium Research Group), 152 US Markets for Image-Guided Surgery (Medtech Insight), 151–152 Mastoidectomy, 187, 197–198, 199, 200, 201 Navigated mastoidectomy, 187 Maxwell distribution, 85–86, 86, 122 Mayfield head holder, 20, 113, 125, 176, 197
Mazor Robotics Ltd., 197 Medical Hospital of Hannover, Germany, 202 MedTech S.A.S., 197, 198 Medtronic, Inc., 19, 31, 36, 151, 152–162 accuracy of Medtronic StealthStation, 158–162, 161, 162 Fusion, 154, 155, 172 LandmarX system (Medtronic), 162 O-arm CT scanner (Medtronic, Inc.), 31, 33, 36, 37 registration with Medtronic systems, 156–158, 157 SofamorDanek, 153 Stealth Technologies, 153 StealthStation S7, 153, 154, 154, 158–162, 161–162 StealthStation S7 user manuals, 155–156, 164 StealthStation/Fusion ENT Navigation Solution Training Manual, 156, 157 StealthStation i7 System Manual, 155 StealthStation S7 Treatment Guidance System Manual, 154, 155–156 MicronTracker (ClaroNav), 88, 88, 89, 90, 91 Microstereotactic frames, 191, 192, 202 Millennium Research Group, 152 Mitsubishi Industrial Robot, 201 Mobius Imaging LLC, 33, 153, 163 Movement disorders, 190 MRI. See Magnetic resonance imaging Multi-modality fiducial markers, 100, 101, 103 Multiplanar imaging, MRI, 59 Multislice CT (MSCT) scanners, 15, 29, 33, 34 Mutual information (MI), 17, 66, 113–114 Myths. See Fiducials and point; Fiducial registration error (FRE); Target registration error (TRE)
N N-bar systems, 13–14, 14, 106 NAV3i system (Stryker, Corp.), 153, 173, 173, 175–179 Navident (ClaroNav, Inc.), 88, 88
211
212
Image-Guided Surgery
NaviENT (ClaroNav, Inc.), 88, 88, 181, 182 Navigation, 1–21 future, 185–187, 186–188 history, 3–15, 16, 17–19 “navigated endoscopic sinus surgery”, 188 “navigated mastoidectomy”, 187 NDI. See Northern Digital, Inc. Neodymium magnets in contrast to MRI magnet, 42 Neurologica Corp., 31, 33, 34, 153, 153, 172 NeuroMate (Renishaw), 197, 198 NMR. See Nuclear magnetic resonance Noise (in images) CT, 35, 35 inaccuracy cause, 70 MRI, 52, 59–60 signal-to-noise (S/N), 59, 60 Nonvisual augmented reality, 190, 191, 192, 193 Northern Digital, Inc. (NDI), 82, 82–83, 86, 87, 93, 94, 95–96, 152, 154, 156, 165, 166, 172, 180 Nuclear magnetic resonance (NMR), 7, 8, 10, 12, 40
O O-arm CT scanner (Medtronic, Inc.), 31, 33, 36, 37 Ogawa, S., 68 Olympus, 163 Open-bore MRI scanners, 40, 41 Optical tracking (visible and IR), 2, 2, 75, 76, 78–90 visible light tracking, 87–88, 88, 89, 90 Otobot, 201
P Parallel imaging, 62 Passive IR tracking, 81, 81 Patient positioning frame (PPF), 201 Phantom and cadaveric accuracy studies of Brainlab systems, 169–170, 170 of StealthStation S7 (Medtronic), 160–161, 161 of Stryker systems, 179–180
Phase introduced in MRI, 10 Phase encoding, MRI, 46–51, 48–50 Philips, 41, 61 Physiologic atlases, 190, 191 Pixel, 27, 28, 70 Pixelated, pixelation, 29–30, 30, 70 Planar fiducial configurations, 137, 138 PLE. See Probe localization error Polaris Active IR tracking (NDI), 172 Polaris Spectra (NDI), 82, 82–84, 86, 87, 110 Polaris Vicra (NDI), 82, 82–84, 85, 86, 87 Polhemus, Inc., 94 Portable CT scanners, 30, 31–40, 153, 172 accuracy, 33, 34, 39–40, 39, 40 fpVCT scanners, 35–36, 35, 37, 38 MSCT scanners, 33, 34 PPF. See Patient positioning frame Probe localization error (PLE), 126 PROFESS system (Stryker Corp.), 173, 174, 178–179 Programmable Universal Manipulator Arm (PUMA), 195–196, 195 Protons, magnetism of, 6 Pulse sequence, MRI, 48–56, 48–51, 52–56 Pupin, Michael, 3, 4
Q Quick reference guide, 165
R Registration, 2, 75, 99–115 with Brainlab systems, 167–169 fusion, 16, 17, 113–114 with Medtronic systems, 156–158, 157 rigid point registration, 107–109, 108, 109 with Stryker systems, 175–176, 176, 179–180 surface registration, 109–112, 110–112 z-touch registration, 110, 112, 169–171. See also Fiducial markers and points; Fiducial Localization Error (FLE); Fiducial Registration Error (FRE); Target registration error (TRE)
Index
Relaxation times, MRI, 8, 57–58, 57 Renaissance system (Mazor), 197 Renishaw PLC, 197 Rigid point registration, 107–109, 108, 109 RMS. See Root mean square RMS distance errors, 86, 86, 87, 95 RMS FLE, 122–125, 126, 127, 128, 129, 131, 132, 137 RMS FRE, 121,127–128, 129, 135, 170–171 RMS PLE, 128 RMS tracking error, 86, 86, 87, 90, 95, 97 RMS TRE, 122–126, 126, 127–128, 129, 137, 161, 161, 170, 170–172, 179 ROBODOC (Curexo), 194, 196 Robotic Arm Interactive System (Mako), 197 Robots, 193–198, 193, 195, 198–202, 200, 202 Active Constraint Robot (ACROBOT), 196, 197 autonomous robots, 194–198, 195, 198–202, 200, 202 Curexo Medical Technologies, 194, 196 DaVinci Surgical system (Intuitive Surgical), 194 Integrated Surgical Systems, Inc., 196 Intuitive Surgical, Inc., 194 Kwoh, Y.S., 195 Mako Surgical Systems, 186–187, 188, 196–197 “Master-slave” devices, 194 Mazor Robotics Ltd., 197 MedTech S.A.S., 197, 198 Mitsubishi Industrial, 201 NeuroMate (Renishaw), 197, 198 Otobot, 201 Programmable Universal Manipulator Arm (PUMA), 195–196, 195 Renaissance system (Mazor), 197 Renishaw PLC, 197 ROBODOC (Curexo), 194, 196 Robotic Arm Interactive System (Mako), 197 Rosa system (MedTech), 197, 198 Staubli robot, 200 Surgical-assist robots, 194 Tele-manipulators, 194 THINK Surgical, Inc., 196
Unimate “pick-and-place”, 193 Whole Arm Manipulator (WAM), 195, 196 Zeego (Siemens), 31 Röntgen, Wilhelm Conrad, 3, 3 Root mean square (RMS), 85 Rosa system (MedTech S.A.S.), 197, 198
S Salt-and-pepper effect, 60 Samsung Electronics Co., 33 SENSE, 62 Siemens Medical Systems, 31, 39, 61 Signal-to-noise ratio, 59 Skin-affixed fiducials, 18, 100, 101–102, 102, 113, 157, 170, 174 Slice selection, 8, 10, 43–44, 44 Slip-ring, CT, 15 Small-angle excitation, 60–61 SMASH, 62 SofamorDanek, 153 Sound localization, 18 Spectra (NDI), 82, 82–84, 86, 87, 110 SPGR, 61 Spin echo. See Magnetic resonance imaging (MRI), echo Spin-warp imaging, 10 SSCD. See Superior semicircular canal dehiscence Standard of care, 147 Static-field inhomogeneity, MRI, 52–54, 53, 54, 65–67, 68, 160 Stationary CT scanners, 30–31 Staubli robot, 200 Stealth Technologies, 153 StealthStation/Fusion ENT Navigation Solution Training Manual (Medtronic, Inc.), 156, 157 StealthStation i7 System Manual, (Medtronic, Inc.) 155 StealthStation S7 (Medtronic, Inc.), 153, 154, 154, 158–162, 161–162 StealthStation S7 Treatment Guidance System Manual (Medtronic, Inc.), 154, 155–156 Stereotactic frame, 12–14, 14 historical perspective, 16 insidious error, 106
213
214
Image-Guided Surgery
Stereotactic frame (continued) need for, 13–14 obsolescence, 15, 17 Stern, Otto, 6 Stryker, Homer, 172 Stryker Corp., 19, 151, 152, 153, 172–180 ADAPT system, 153, 173, 174 accuracy of Stryker systems, 179–180 CranialMap Express Navigation system, 173, 175 Image Guided Technologies, Inc., 172 NAV3i System, 153, 173, 173, 175–178, 177 PROFESS system, 173, 174, 178–179 registration with Stryker systems, 175–176, 179–180 System II Camera, 172 user manuals, 175–179 CranialMap Express Navigation System, 173, 175 Instructions for Use PROFESS System, 178–179 NAV3i CranialMap Express User Manual, 175–178 Subtotal measurement errors, 160, 170, 171, 182 Superior semicircular canal dehiscence (SSCD), 71 Surface registration, 109–112 ICP algorithm, 111, 111 improvement by adding fiducial, 113 laser scanning, 110 problem with deep objects, 112 Surgical approaches anterior skull base, 156, 171 CSF leak, 142, 143, 145 endoscopic sinus surgery (ESS), 19, 21, 142–144, 146, 152, 156, 164, 169, 170, 171, 174, 175, 181, 190 intracranial for DBS, 162, 190, 197 lateral skull base, 18, 19, 31, 33, 105, 106, 132, 142, 145, 147, 156, 164, 169, 170, 171, 173, 189, 190 acoustic neuroma, 200, 201 cochlea, 190, 192, 202, 202 mastoidectomy, 187, 197–198, 199, 200, 201 translabyrinthine, 200, 201 Surgical-assist robots, 194
Surgical Navigation Systems, US, Market Analysis (Millenium Research Group), 152 Surgical robots. See Robots Surgical Theater LLC, 189 System II Camera (Stryker Corp.), 172
T T1-FFE, 61 Target registration error (TRE), 108, 119–124, 120 bone markers vs skin markers, 113 dependencies, 123–124 inevitability, 107 myths, 130–137 counterproductive fiducial omission, 134–137, 136, 138, 168–169, 177, 179–180, 180 fallacious prediction from FRE, 128, 130–134, 130–132, 158, 168, 179–180 reducing, 100, 125–126, 126, 128 RMS relationships to FLE, 122–125, 126, 127, 137 to FRE, 127 studies of Brainlab Curve and Kick, 169–171 Medtronic StealthStation, 158–162, 161, 162 Stryker, 179–180 InstaTrack, 171–172 stereotactic frame, 129 Tele-manipulators, 194 THINK Surgical, Inc., 196 Time constant T1, MRI, 57–58, 57–58, 58 Time constant T2, MRI, 57–58, 57–58, 58 Time of echo TE, MRI, 53–54, 57–58, 53, 55–57, 58 Time of repetition TR, MRI, 57–58, 58 Tomography era before computed tomography (CT), 5–6. See also Computed tomography Toshiba Medical Systems, 15 Tracking, 18, 19, 75–97 “armless” tracking systems, 18 electromagnetic (EM) tracking, 2, 2, 18, 19, 75–76, 90, 92–96, 152
Index
Ascension Technology Corp., 94 Aurora, Northern Digital Inc. (NDI), 93, 94, 95–96, 97 Calypso System (Varian Medical Systems), 90 Polhemus, Inc., 94 eye damage, 165 “frameless” tracking systems, 18 infrared (IR) tracking, 18, 19, 80–81, 81, 165, 176, 176 active IR tracking, 81, 172 passive IR tracking, 81, 81 Advanced Realtime Tracking, GmbH, 87 Atracsys, GmbH, 87 Polaris Spectra (NDI), 82, 82–84, 86, 87, 110 Polaris Vicra (NDI), 82, 82–84, 85, 86, 87 mechanical linkage FARO Medical Technologies, 18 ISG Technologies, 18, 20 Viewing Wand (ISG Technologies), 18, 20 optical tracking (visible and IR), 2, 2, 75, 76, 78–90 pivoting, 76–78, 77 protocol for use, 76 tool calibration, 76 visible light tracking, 87–88, 88, 89, 90 ClaroNav, 76, 77, 81, 88, 89, 91, 151, 173, 181, 182 Translabyrinthine approach, 200, 201 TRE. See Target registration error Triangulation, 78–80, 79, 82 Turbo spin echo, 61 2D acquisition, 50
U Unimate “pick-and-place”, 193 University of Bern, 202 US Markets for Image-Guided Surgery (Medtech Insight), 151–152
V Vanderbilt University, ix–xi, 202
Varian Medical Systems, Inc. (Calypso), 90 VectorVision Compact system (Brainlab Corp.), 163, 169, 170, 170, 171–172 Vicra (NDI), 82, 82–84, 85, 86, 87 Viewing Wand (ISG Technologies), 18, 20 Visible light optical tracking, 87–88, 88, 89, 90 VISIUS iCT (IMRIS), 31, 32 VISIUS iMRI, 31 Visual augmented reality, 188–189, 188 Visualization Technologies, Inc., 19, 151 Volume imaging (in MRI), 50–51, 51, 56, 60, 61, 66 Voxelated, voxelation, 30, 30, 70, 70 Voxels, 27–28, 27, 28, 29, 30, 36, 45, 70, 100, 174, 185, 186, 186
W Whole Arm Manipulator (WAM), 195, 196
X Xpoint, 88, 89, 181 X-ray attenuation coefficient, 27 X-ray tomography, 5–6 X-rays action of passing through tissue, 25, 26 computed tomography, 25, 26, 27–28, 27 history, 3, 3 xCAT CT scanner (Xoran Technologies), 31, 33, 36, 38 Xoran Technologies, 31, 36, 38, 39, 39, 180
Y Your patient will thank you, 137
Z Z-touch registration, 110, 112, 169–171 Zeego (Siemens), 31
215