Machine Learning for Tomographic Imaging. Professor Ge Wang

patients, determine the best scan parameters, perform image reconstruction, enhance images, analyze the quality of images, perform a diagnosis, compute a treatment plan, etc.

      Dr Ge Wang was one of the earliest innovators who has been creatively exploring various ways of using deep learning in tomographic imaging since 2016, which is a long time ago given how fast the field has evolved. His team’s main impacts range across the topics of CT, MRI, and optical imaging, but they have also made fundamental contributions to the science of neural networks, such as through the investigation of quadratic deep neural networks.

      Dr Wang teamed up with Dr Zhang, Dr Ye and Dr Mou to write this book on machine learning for tomographic imaging. The authors are uniquely qualified to undertake this major project, given their combined expertise in mathematics, computer science, image processing, and tomographic imaging, as well as their pioneering research in machine learning for medical imaging. At a high level, this book teaches medical imaging from the perspective of machine learning, and hence covers two important fields: machine learning and tomographic imaging.

      The book’s first part gives a colorful and inspiring introduction to machine learning and tomographic imaging. Using the Human Vision System as reference, the authors take us on a journey from sparse representations, through dictionary learning, to neural networks and deep learning. Many ‘textbook’ deep learning architectures are covered at an introductory level. Parts two and three provide an in-depth tutorial of CT and MR image reconstruction, followed by a wide range of machine learning techniques that were developed in recent years. The fourth part further enriches the content of this book by elaborating on other imaging modalities, image quality evaluation, and quantum computing.

      Overall, this book provides an amazingly comprehensive overview of neural networks and tomographic reconstruction methods. It is written in an engaging and accessible style, without lengthy mathematical derivations and proofs. This makes it ideal for introducing machine learning and tomographic imaging in the more applied disciplines (physics and engineering), and also for bringing application contexts into the more theoretical disciplines (mathematics and computer sciences).

      Every medical imaging scientist who graduated before machine learning was taught in college should probably learn about this area in order to remain competitive. To my knowledge, this book is the first and only publication capturing all important aspects of machine learning and tomographic imaging in one place.

      I highly recommend this book for any medical imaging students/professionals with a STEM background. Start with chapters 1–3. Then, depending on whether you are a CT, PET, MRI, ultrasound, or optical imaging aficionado, you may select one or more of the other chapters for further study. Before you know it, you will ‘deeply learn’ this exciting new science, be able to talk intelligently about it, and perform state-of-the-art research in a world that can no longer be imagined without neural networks.

      Bruno De Man, October 2019

Preface

      This book arose from discussions among four colleagues with a long-standing collaboration and interest in advanced medical image reconstruction methods and applications. Beginning in 2018, our group realized the gap in the literature and in particular among technical books on the emerging technologies that develop and apply artificial intelligence/machine learning (AI/ML) techniques to tomographic reconstruction or tomographic imaging.

      As early as 2012, we recognized the opportunity presented by machine learning in formulating plans for doctoral dissertation research where dictionary learning can be used to recover images from projection measurements. By connecting several contemporary image recovery and signal processing methods, in particular compressed sensing, neural network, and deep learning techniques, our discussions and projects converge to develop and apply ML methods to advance the frontier of image reconstruction, with an emphasis on medical imaging.

      The interested reader entering this field may have a background in artificial intelligence or mathematical knowledge of tomographic reconstruction, but few will have all of the knowledge needed to understand the field of ML for tomographic image reconstruction. Hence, we have now created this book to cover what we believe to be a comprehensive collection of key topics in a logical and consistent manner.

      The prerequisites for reading this book include calculus, matrix algebra, Fourier analysis, medical physics, and basic programming skills. We believe that PhD candidates in the imaging field are generally well prepared to understand all of the content through serious effort, while advanced undergraduate students can also learn essential ideas and capture selected materials (you can skip the chapters/sections/subsections marked with an ‘*’). To facilitate teaching and learning, most relevant numerical methods are described in appendix A, and hands-on projects are suggested in appendix B, with sample codes and working datasets.

      The logical dependence between the key components of this book is illustrated in the diagram in the introduction below. We strongly recommend that you read this introduction first to obtain an overall perspective. It is also recommended to read the first three chapters sequentially so that you are well prepared with both the imaging context and network basics. However, chapter 3 alone is a good introduction to general knowledge on artificial neural networks. Then, we can proceed in parallel to CT, MRI, or other tomographic modalities, which are covered in parts II, III, and IV of the book, respectively. It would be the best to read part IV after reading parts II and III as deep reconstruction networks are clearly explained for CT and MRI in these two parts. Appendix A can be read as needed, but appendix B is strongly recommended, and should be at least consulted to run the basic networks explained in chapter 3. The network examples for CT and MRI can be adapted for independent class projects.

      We hope that this book will be useful for a review course at the graduate level, but it has not been tested yet. As teaching experience is accumulated using this book, homework problems and solutions will become available, along with example class project reports and codes. A book-related website is maintained on the Fully3D community website: http://www.fully3d.org/rpi/.

      The materials contained in this book are presented in their first version. As a result, a number of topics are not treated in detail or in depth. Nevertheless, after reading this book you should have state-of-the-art knowledge of a broad spectrum of methods. We welcome your critiques and suggestions so we can make future versions better, with key references cited in a more balanced way.

      July 2019

Acknowledgments

      The four parts of this book were initially drafted by Professors Mou, Zhang, Ye, and Wang respectively, based on a collectively developed overall layout. The appendices were drafted by Professor Ye. Hands-on examples were developed by Professors Zhang, Ye, and Mou collectively, and integrated by Professor Zhang. All parts were internally reviewed and revised by the four co-authors, and editorially refined by the staff of IOP Publishing.

      We would like to express our sincere gratitude to all individuals, publishers, and companies for permission to reproduce some of the images and figures in this book, IOP Publishing staff for guidance during the development of the book, and importantly our students, other lab members, and collaborators for their significant contributions, including but not limited to Hongming Shan, Qing Lyu, Christopher Wiedeman, Harshank Shrotriya, Huidong Xie, Fenglei Fan, Mengzhou Li, and Varun Ravichandran. Drs Michael Vannier and Hengyong Yu offered insightful advice on the strengths and weaknesses of this book for improvements. Last but not least, the following leading companies have graciously given permission to reproduce some of the best figures/images in this book: Cannon, General Electric, Siemens, and Phillips (in alphabetical order). Without these, this book would not have been created in its current form. We are happy that the first version of this book is now complete, and look forward to producing future versions and more excitement in the years to come.

Author biographies

      Ge Wang

      Ge Wang,