reader of scientific publications and to understand how systems are evaluated during development and after deployment Structured in this way, this book forms a unique and valuable resource both for the trainee who intends to become an expert in medical imaging informatics and a refer ence for the established practitioner Steven C. Horii. MD. FACR. FSIIM Professor of Radiolog Clinical Director, Medical Informatics Group, and Department of Radiolog University of Pennsylvania Medical Center
Foreword ix development and after deployment. Structured in this way, this book forms a unique and valuable resource both for the trainee who intends to become an expert in medical imaging informatics and a reference for the established practitioner. Steven C. Horii, MD, FACR, FSIIM Professor of Radiology, Clinical Director, Medical Informatics Group, and Modality Chief for Ultrasound Department of Radiology University of Pennsylvania Medical Center reader of scientific publications and to understand how systems are evaluated during
Preface This book roughly follows the process of care, illustrating the techniques involved in medical imaging informatics. Our intention in this text is to provide a roadmap for the different topics that are involved in this field: in many cases, the topics covered in the ensuing chapters are themselves worthy of lengthy descriptions, if not an entire book As a result, when possible the authors have attempted to provide both seminal and current references for the reader to pursue additional details For the imaging novice and less experienced informaticians, in Part I of this book Performing the Imaging Exam, we cover the current state of medical imaging and set the foundation for understanding the role of imaging and informatics in routine clinical Chapter 1(Introduction) provides an introduction to the field of medical imaging informatics and its role in transforming healthcare research and delivery. The interwoven nature of imaging with preventative, diagnostic, and therapeutic elements of patient care are touched upon relative to the process of care. a brief historic perspective is provided to illustrate both past and current challenges of the discipline Chapter 2(An Introduction to Imaging Anatomy Physiology) starts with a review of clinical imaging modalities(i.e, projectional x-ray, computed tomography (CT), magnetic resonance(MR), ultrasound) and a primer on imaging anatomy and physiology. The modality review encompasses core physics principles and image formation techniques, along with brief descriptions of present and future directions for each imaging modality. To familiarize non-radiologists with medical imaging and the human body, the second part of this chapter presents an overview of anatomy and physiology from the perspective of projectional and cross- sectional imaging. A few systems(neurological, respiratory, breast) are covered in detail, with additional examples from other major systems(gastrointestinal urinary,cardiac, musculoskeletal More experienced readers will likely benefit from starting with Part Il of this book Integrating Imaging into the Patient Record, which examines topics related to communicating and presenting imaging data alongside the growing wealth of clinical Once imaging and other clinical data are acquired, Chapter 3(Information Systems Architectures) tackles the question of how we store and access imaging and other patient information as part of an increasingly distributed and heterogeneous EMr. A description of major information systems(.g, PACS; hospital informa- tion systems, HIS; etc. )as well as the different data standards employed today to represent and communicate data(e. g, HL7, DICOM) are provided. A discussion of newer distributed architectures as they apply to clinical databases(peer-to-peer grid computing) and information processing is given, examining issues of scal- ability and searching. Different informatics-driven applications are used to high- light ongoing efforts with respect to the development of information architectures, including telemedicine, IHE, and collaborative clinical research involving imaging After the data is accessed, the challenge is to integrate and to present patient information in such a way to support the physicians cognitive tasks. The longitud- inal EMR, in conjunction with the new types of information available to clinicians, has created an almost overwhelming flow of data that must be fully understood to
xi Preface This book roughly follows the process of care, illustrating the techniques involved in medical imaging informatics. Our intention in this text is to provide a roadmap for the different topics that are involved in this field: in many cases, the topics covered in the ensuing chapters are themselves worthy of lengthy descriptions, if not an entire book. As a result, when possible the authors have attempted to provide both seminal and current references for the reader to pursue additional details. For the imaging novice and less experienced informaticians, in Part I of this book, Performing the Imaging Exam, we cover the current state of medical imaging and set the foundation for understanding the role of imaging and informatics in routine clinical practice: Chapter 1 (Introduction) provides an introduction to the field of medical imaging informatics and its role in transforming healthcare research and delivery. The interwoven nature of imaging with preventative, diagnostic, and therapeutic elements of patient care are touched upon relative to the process of care. A brief historic perspective is provided to illustrate both past and current challenges of the discipline. Chapter 2 (An Introduction to Imaging Anatomy & Physiology) starts with a review of clinical imaging modalities (i.e., projectional x-ray, computed tomography (CT), magnetic resonance (MR), ultrasound) and a primer on imaging anatomy and physiology. The modality review encompasses core physics principles and image formation techniques, along with brief descriptions of present and future directions for each imaging modality. To familiarize non-radiologists with medical imaging and the human body, the second part of this chapter presents an overview of anatomy and physiology from the perspective of projectional and crosssectional imaging. A few systems (neurological, respiratory, breast) are covered in detail, with additional examples from other major systems (gastrointestinal, urinary, cardiac, musculoskeletal). More experienced readers will likely benefit from starting with Part II of this book, Integrating Imaging into the Patient Record, which examines topics related to communicating and presenting imaging data alongside the growing wealth of clinical information: Once imaging and other clinical data are acquired, Chapter 3 (Information Systems & Architectures) tackles the question of how we store and access imaging and other patient information as part of an increasingly distributed and heterogeneous EMR. A description of major information systems (e.g., PACS; hospital information systems, HIS; etc.) as well as the different data standards employed today to represent and communicate data (e.g., HL7, DICOM) are provided. A discussion of newer distributed architectures as they apply to clinical databases (peer-to-peer, grid computing) and information processing is given, examining issues of scalability and searching. Different informatics-driven applications are used to highlight ongoing efforts with respect to the development of information architectures, including telemedicine, IHE, and collaborative clinical research involving imaging. After the data is accessed, the challenge is to integrate and to present patient information in such a way to support the physician’s cognitive tasks. The longitudinal EMR, in conjunction with the new types of information available to clinicians, has created an almost overwhelming flow of data that must be fully understood to
Preface properly inform decision making. Chapter 4(Medical Data visualization Toward Integrated Clinical Workstations) presents works related to the visualiz- ation of medical data. A survey of graphical metaphors (lists and tables, plots and charts, graphs and trees, and pictograms)is given, relating their use to convey clinical concepts. A discussion of portraying temporal, spatial, multidimensional nd causal relationships is provided, using the navigation of images as an example application. Methods to combine these visual components are illustrated, based on a definition of(task) context and user modeling, resulting in a means of creating an adaptive graphical user interface to accommodate the range of different user Part Ill, Documenting Imaging Findings, discusses techniques for automatically nt fro In Chapter 5( Characterizing Imaging Data), an introduction to medical image understanding is presented. Unlike standard image processing, techniques within medical imaging informatics focus on how imaging studies, alongside other clinical data, can be standardized and their content(automatically) extracted to guide medical decision making processes. Notably, unless medical images are standard- ized, quantitative comparisons across studies is subject to various sources of bias/ artifacts that negatively influence assessment. From the perspective of creating scientific-quality imaging databases, this chapter starts with the groundwork for understanding what exactly an image captures, and commences to outline the dif- ferent aspects encompassing the standardization process: intensity normalization denoising, and both linear and nonlinear image registration methods are covered Subsequently, a discussion of commonly extracted imaging features is given, divided amongst appearance-and shape-based descriptors. with the wide array of image features that can be computed, an overview of image feature selection and dimensionality reduction methods is provided. Lastly, this chapter concludes with a description of increasingly popular imaging-based anatomical atlases, detailing their construction and usage as a means for understanding population-based norms and differences arising due to a disease process Absent rigorous methods to automatically analyze and quantify image findings, radiology reports are the sole source of expert image interpretation. In point of fact, a large amount of information about a patient remains locked within clinical documents, and as with images, the concepts therein are not readily computer un- derstandable. Chapter 6(Natural Language Processing of Medical Reports) deals with the structuring and standardization of free-text medical reports via natural language processing(NLP). Issues related to medical NLP representation. computation, and evaluation are presented. An overview of the NLP task is first described to frame the problem, providing an analysis of past efforts and applica- tions of NLP. A sequence of subtasks is then related: structural analysis (e.g,section and sentence boundary detection), lexical analysis(e. g, logical word sequences, disambiguation, concept coding), phrasal chunking, and parsing are covered. For each subtask, a description of the challenges and the range of approaches are given to familiarize the reader with the field Core to informatics endeavors is a systematic method to organize both data and knowledge, representing original(clinical) observations, derived data, and conclu sions in a logical manner. Chapter 7(Organizing Observations: Data Models) describes the different types of relationships between healthcare entities, particularly focusing on those relations commonly encountered in medical imaging. Often in
xii Preface properly inform decision making. Chapter 4 (Medical Data Visualization: Toward Integrated Clinical Workstations) presents works related to the visualization of medical data. A survey of graphical metaphors (lists and tables; plots and charts; graphs and trees; and pictograms) is given, relating their use to convey clinical concepts. A discussion of portraying temporal, spatial, multidimensional, and causal relationships is provided, using the navigation of images as an example an adaptive graphical user interface to accommodate the range of different user goals involving patient data. Part III, Documenting Imaging Findings, discusses techniques for automatically extracting content from images and related data in order to objectify findings: In Chapter 5 (Characterizing Imaging Data), an introduction to medical image understanding is presented. Unlike standard image processing, techniques within medical imaging informatics focus on how imaging studies, alongside other clinical data, can be standardized and their content (automatically) extracted to guide medical decision making processes. Notably, unless medical images are standardized, quantitative comparisons across studies is subject to various sources of bias/ artifacts that negatively influence assessment. From the perspective of creating scientific-quality imaging databases, this chapter starts with the groundwork for understanding what exactly an image captures, and commences to outline the different aspects encompassing the standardization process: intensity normalization; denoising; and both linear and nonlinear image registration methods are covered. Subsequently, a discussion of commonly extracted imaging features is given, divided amongst appearance- and shape-based descriptors. With the wide array of image features that can be computed, an overview of image feature selection and dimensionality reduction methods is provided. Lastly, this chapter concludes with a description of increasingly popular imaging-based anatomical atlases, detailing their construction and usage as a means for understanding population-based norms and differences arising due to a disease process. Absent rigorous methods to automatically analyze and quantify image findings, radiology reports are the sole source of expert image interpretation. In point of fact, a large amount of information about a patient remains locked within clinical documents; and as with images, the concepts therein are not readily computer understandable. Chapter 6 (Natural Language Processing of Medical Reports) deals with the structuring and standardization of free-text medical reports via natural language processing (NLP). Issues related to medical NLP representation, computation, and evaluation are presented. An overview of the NLP task is first described to frame the problem, providing an analysis of past efforts and applications of NLP. A sequence of subtasks is then related: structural analysis (e.g., section and sentence boundary detection), lexical analysis (e.g., logical word sequences, disambiguation, concept coding), phrasal chunking, and parsing are covered. For each subtask, a description of the challenges and the range of approaches are given to familiarize the reader with the field. Core to informatics endeavors is a systematic method to organize both data and knowledge, representing original (clinical) observations, derived data, and conclusions in a logical manner. Chapter 7 (Organizing Observations: Data Models) describes the different types of relationships between healthcare entities, particularly focusing on those relations commonly encountered in medical imaging. Often in on a definition of (task) context and user modeling, resulting in a means of creating application. Methods to combine these visual components are illustrated, based
Preface XIIL clinical practice, a disease is studied from a specific perspective (e.g, genetic, pathologic, radiologic, clinical). But disease is a phenomenon of nature, and is thus typically multifaceted in its presentation. The goal is to aggregate the observations for a single patient to characterize the state and behavior of the patient's disease, both in terms of its natural course and as the result of (therapeutic)interventions The chapter divides the organization of such information along spatial (e.g physical and anatomical relations, such as between objects in space), temporal (e.g, sequences of clinical events, episodes of care), and clinically-oriented models (i. e, those models specific to representing a healthcare abstraction a discussion of the motivation behind what drives the design of a medical data model is given, leading to the description of a phenomenon-centric data model to support healthcare research Finally, in Part IV, Toward Medical Decision Making, we reflect on issues pertain- ng to reasoning with clinical observations derived from imaging and other data sources in order to reach a conclusion about patient care and the value of our decision A variety of formalisms are used to represent disease models, of these, probabilistic graphical models have become increasingly popular given their ability to reason in light of missing data, and their relatively intuitive representation. Chapter 8 Disease Models, Part 1: Graphical Models) commences with a review of key concepts in probability theory as the basis for understanding these graphical models and their different formulations. In particular, the first half of the chapter handles Bayesian belief networks(BBNS), appraising past and current efforts to apply these models to the medical environment. The latter half of this chapter addresses he burgeoning exploration of causal models, and the implications for analysis and positing questions to such networks. Throughout, a discussion of the practical considerations in the building of these models and the assumptions that must be Following the discussion of the creation of the models, in Chapter 9(Disease Models, Part 1: Querying Applications), we address the algorithms and tools that enable us to query BBNs. Two broad classes of queries are considered: belief updating, and abductive reasoning. The former entails the re-computation of pos- terior probabilities in a network given some specific evidence; the latter involves calculating the optimal configuration of the bBn in order to maximize some specified criteria. Brief descriptions of exact and approximate inference methods are provided. Special types of belief networks(naive Bayes classifiers, influence diagrams, probabilistic relational models)are covered, illustrating their potential usage in medicine. Importantly, issues related to the evaluation of belief networks are discussed in this chapter, looking to standard technical accuracy metrics, but also ideas in parametric sensitivity analysis. Lastly, the chapter concludes with some example applications of BBNs in medicine, including to support case-based retrieval and image processing tasks Chapter 10(Evaluation) concludes by considering how to assess informatics endeavors. A primer on biostatistics and study design starts this chapter, including a review of basic concepts(e.g, confidence intervals, significance and hypothesis testing)and the statistical tests that are used to evaluate hypotheses under differ- ent circumstances and assumptions. a discussion of error and performance assessment is then introduced, including sensitivity/specificity and receiver opera- tive characteristic analysis Study design encompasses a description of the differ ent types of experiments that can be formed to test a hypothesis, and goes over the
Preface xiii clinical practice, a disease is studied from a specific perspective (e.g., genetic, pathologic, radiologic, clinical). But disease is a phenomenon of nature, and is thus typically multifaceted in its presentation. The goal is to aggregate the observations for a single patient to characterize the state and behavior of the patient’s disease, both in terms of its natural course and as the result of (therapeutic) interventions. The chapter divides the organization of such information along spatial (e.g., physical and anatomical relations, such as between objects in space), temporal (e.g., sequences of clinical events, episodes of care), and clinically-oriented models (i.e., those models specific to representing a healthcare abstraction). A discussion of the motivation behind what drives the design of a medical data model is given, leading to the description of a phenomenon-centric data model to support healthcare research. Finally, in Part IV, Toward Medical Decision Making, we reflect on issues pertaining to reasoning with clinical observations derived from imaging and other data sources in order to reach a conclusion about patient care and the value of our decision: A variety of formalisms are used to represent disease models; of these, probabilistic graphical models have become increasingly popular given their ability to reason in light of missing data, and their relatively intuitive representation. Chapter 8 (Disease Models, Part I: Graphical Models) commences with a review of key concepts in probability theory as the basis for understanding these graphical models and their different formulations. In particular, the first half of the chapter handles Bayesian belief networks (BBNs), appraising past and current efforts to apply these models to the medical environment. The latter half of this chapter addresses the burgeoning exploration of causal models, and the implications for analysis and positing questions to such networks. Throughout, a discussion of the practical considerations in the building of these models and the assumptions that must be made, are given. Following the discussion of the creation of the models, in Chapter 9 (Disease Models, Part II: Querying & Applications), we address the algorithms and tools that enable us to query BBNs. Two broad classes of queries are considered: belief updating, and abductive reasoning. The former entails the re-computation of posterior probabilities in a network given some specific evidence; the latter involves calculating the optimal configuration of the BBN in order to maximize some specified criteria. Brief descriptions of exact and approximate inference methods are provided. Special types of belief networks (naïve Bayes classifiers, influence diagrams, probabilistic relational models) are covered, illustrating their potential usage in medicine. Importantly, issues related to the evaluation of belief networks are discussed in this chapter, looking to standard technical accuracy metrics, but also ideas in parametric sensitivity analysis. Lastly, the chapter concludes with some example applications of BBNs in medicine, including to support case-based retrieval and image processing tasks. Chapter 10 (Evaluation) concludes by considering how to assess informatics endeavors. A primer on biostatistics and study design starts this chapter, including a review of basic concepts (e.g., confidence intervals, significance and hypothesis testing) and the statistical tests that are used to evaluate hypotheses under different circumstances and assumptions. A discussion of error and performance assessment is then introduced, including sensitivity/specificity and receiver operative characteristic analysis. Study design encompasses a description of the different types of experiments that can be formed to test a hypothesis, and goes over the
Preface process of variable selection and sample size/power calculations. Sources of study bias/error are briefly described, as are statistical tools for decision making. The second part of this chapter uses the foundation set out by the primer to focus specifically on informatics-related evaluations. Two areas serve as focal points evaluating information retrieval (IR) systems, including content-based image retrieval; and assessing(system)usability
xiv Preface process of variable selection and sample size/power calculations. Sources of study bias/error are briefly described, as are statistical tools for decision making. The second part of this chapter uses the foundation set out by the primer to focus specifically on informatics-related evaluations. Two areas serve as focal points: evaluating information retrieval (IR) systems, including content-based image retrieval; and assessing (system) usability