1 Introduction Medical Imaging Informatics: From Theory to Application There are two arms to medical imaging informatics: the development of core informatics theories and techniques that advance the field of informatics itself; and the translation of these techniques into an application that improves health. To demonstrate, we first consider the reasons for the improper use of imaging today, and then how imaging informatics can impact these issues Improving the Use of Imaging The process of providing an accurate, expedient medical diagnosis via imaging can fail for several reasons(Fig. 1. 2): Sub-optimal study selection. The first potential point of failure arises when an imaging study is requested. Given the fairly rapid changes across all elements of imaging technology, it is unrealistic to believe that a physician can always make up-to-date if not optimal decisions about an imaging exam [9]. Thus, the wrong study may be requested for a given patient. To reduce this problem, practice guidelines have been introduced, but are often generic and do not take into account the specific condition of the patient. Poor acquisition. The next potential point of failure occurs during study acquisi tion. Problems arise due to poor instrumentation (e.g, sensitivity), equipment calibration, poor data acquisition methods, or poor technique. For example, due to the very technical nature of imaging procedures, the average clinician is unable to determine the most specific diagnostic protocol; this process is often left to a technologist or radiologist, who without fully knowing the context of the patient, may not use ideal acquisition parameters Poor interpretation. Study interpretation presents an additional point for potential failure. Poor study interpretation can be due to inadequate historical medical information, poor information filtering/presentation, or poor/mismatched skills by the study reader. Studies have shown that historical clinical information can improve the perception of certain radiographic findings [3]. Poor information presentation often leads to important data being buried within the medical record. nally, study reading itself can be improved by providing users with the facility to retrieve relevant data from online medical literature, or by choosing the best matched readers (i.e, generalist vS. specialist) for a particular exam. However, currently available search techniques do not support specific and directed retrievals nd no electronic framework exists for efficiently matching a given exam with the most appropriate reader for that exam. insufficient poor reading background wrong study poor study poor study acquIsition documentation/ reque interpretation communication alibration Figure 1.2: Identification of potential problems in the diagnostic process. In emergency cases, the process may also fail due to excessively long times to completion
1 Introduction 5 Medical Imaging Informatics: From Theory to Application There are two arms to medical imaging informatics: the development of core informatics theories and techniques that advance the field of informatics itself; and the translation of these techniques into an application that improves health. To demonstrate, we first consider the reasons for the improper use of imaging today, and then how imaging informatics can impact these issues. Improving the Use of Imaging The process of providing an accurate, expedient medical diagnosis via imaging can fail for several reasons (Fig. 1.2): Sub-optimal study selection. The first potential point of failure arises when an imaging study is requested. Given the fairly rapid changes across all elements of imaging technology, it is unrealistic to believe that a physician can always make up-to-date if not optimal decisions about an imaging exam [9]. Thus, the wrong study may be requested for a given patient. To reduce this problem, practice guidelines have been introduced, but are often generic and do not take into account the specific condition of the patient. Poor acquisition. The next potential point of failure occurs during study acquisition. Problems arise due to poor instrumentation (e.g., sensitivity), equipment calibration, poor data acquisition methods, or poor technique. For example, due to the very technical nature of imaging procedures, the average clinician is unable to determine the most specific diagnostic protocol; this process is often left to a technologist or radiologist, who without fully knowing the context of the patient, may not use ideal acquisition parameters. Poor interpretation. Study interpretation presents an additional point for potential failure. Poor study interpretation can be due to inadequate historical medical information, poor information filtering/presentation, or poor/mismatched skills by the study reader. Studies have shown that historical clinical information can improve the perception of certain radiographic findings [3]. Poor information presentation often leads to important data being buried within the medical record. Finally, study reading itself can be improved by providing users with the facility to retrieve relevant data from online medical literature, or by choosing the bestmatched readers (i.e., generalist vs. specialist) for a particular exam. However, currently available search techniques do not support specific and directed retrievals and no electronic framework exists for efficiently matching a given exam with the most appropriate reader for that exam. Figure 1.2: Identification of potential problems in the diagnostic process. In emergency cases, the process may also fail due to excessively long times to completion
A.A.T. Bui et al Poor reporting. The last potential point of failure concerns reporting of study results, which is a key concern in the coordination of care as related to the diagno- sis and intervention for a given case. This lack of coordination is due to: 1)poor documentation of study results; and 2)difficulties communicating the results of tests to referring healthcare providers. These inefficiencies can lead to problems such as initiating treatment before a definitive diagnosis is established, and From this perspective, medical imaging informatics aims to improve the use of imaging throughout the process of care. For example, what is the best imaging method to assess an individual's given condition? Are there image processing methods that can be employed to improve images post-acquisition (e.g, histogram correction, denoising, etc)? These and other questions motivate medical imaging informatics research Indeed, imaging plays a significant role in the evaluation of patients with complex diseases. As these patients also account for the majority of expenses related to health care, by improving the utility of imaging, cost savings can potentially be realized Choosing a Protocol: The Role of Medical Imaging Informatics To further highlight the role of medical imaging informatics, we consider the task of choosing an imaging protocol when a patient first presents in a doctor's office, ad dressing issues related to sub-optimal study design. When a primary care physician (PCP) decides to obtain an imaging study to diagnosis or otherwise assess a problem, the question arises as to which imaging modality and type of study should be ordered. Furthermore, the ability to make the best decisions regarding a patient is variable across individual physicians and over time. Individual physician biases often creep into decision making tasks and can impact the quality and consistency of healthcare provided [1, 6 To ground this discussion, we use an example of a 51 year-old female patient who visits her PCP complaining of knee pain. The selection of an appropriate imaging protocol to diagnosis the underlying problem can be thought of in three steps: 1)standard- izing the patient,s chief complaint, providing a structured and codified format to understand the individual's symptoms: 2) integrating the patient,'s symptoms with past evidence(e.g, past imaging, medical history, etc. )to assess and to formulate a differ ential diagnosis; and 3) selecting and tailoring the imaging study to confirm (or deny) the differential diagnosis, taking into account local capabilities to perform and evaluate an imaging study(there is no point in ordering a given exam if the scanner is unavailable or unable to perform certain sequences). We elaborate on each of the steps below, illustrating current informatics research and its application Capturing the chief complaint. As mentioned earlier, a patient's description of his or her symptoms is very subjective; for physicians-and computers more so-translating their complaints into a"normalized" response(such as from a controlled vocabulary) is tricky. For instance, with our example patient, when asked her reason for seeing her doctor, she may respond, "My knee hurts a lot, frequently in the morning. "Consider the following two related problems: 1)mapping a patient-described symptom or condition to specific medical terminology/disease (e.g, knee hurts= knee pain>ICD-9 719.46 Pain in joint involving lower leg); and 2)standardizing descriptive terms(adjec tives, adverbs) to the some scale(e.., Does"a lot"mean a mild discomfort or a cr pling pain? Does"frequently"mean every day or just a once a week
6 A.A.T. Bui et al. Poor reporting. The last potential point of failure concerns reporting of study results, which is a key concern in the coordination of care as related to the diagnosis and intervention for a given case. This lack of coordination is due to: 1) poor documentation of study results; and 2) difficulties communicating the results of tests to referring healthcare providers. These inefficiencies can lead to problems such as initiating treatment before a definitive diagnosis is established, and duplicating diagnostic studies. From this perspective, medical imaging informatics aims to improve the use of imaging throughout the process of care. For example, what is the best imaging method to assess an individual’s given condition? Are there image processing methods that can be employed to improve images post-acquisition (e.g., histogram correction, denoising, etc.)? These and other questions motivate medical imaging informatics research. Indeed, imaging plays a significant role in the evaluation of patients with complex diseases. As these patients also account for the majority of expenses related to healthcare, by improving the utility of imaging, cost savings can potentially be realized. Choosing a Protocol: The Role of Medical Imaging Informatics To further highlight the role of medical imaging informatics, we consider the task of choosing an imaging protocol when a patient first presents in a doctor’s office, addressing issues related to sub-optimal study design. When a primary care physician (PCP) decides to obtain an imaging study to diagnosis or otherwise assess a problem, the question arises as to which imaging modality and type of study should be ordered. Furthermore, the ability to make the best decisions regarding a patient is variable across individual physicians and over time. Individual physician biases often creep into decision making tasks and can impact the quality and consistency of healthcare provided [1, 6]. To ground this discussion, we use an example of a 51 year-old female patient who visits her PCP complaining of knee pain. The selection of an appropriate imaging protocol to diagnosis the underlying problem can be thought of in three steps: 1) standardizing the patient’s chief complaint, providing a structured and codified format to understand the individual’s symptoms; 2) integrating the patient’s symptoms with past evidence (e.g., past imaging, medical history, etc.) to assess and to formulate a differential diagnosis; and 3) selecting and tailoring the imaging study to confirm (or deny) the differential diagnosis, taking into account local capabilities to perform and evaluate an imaging study (there is no point in ordering a given exam if the scanner is unavailable or unable to perform certain sequences). We elaborate on each of the steps below, illustrating current informatics research and its application. Capturing the chief complaint. As mentioned earlier, a patient’s description of his or her symptoms is very subjective; for physicians – and computers more so – translating their complaints into a “normalized” response (such as from a controlled vocabulary) is tricky. For instance, with our example patient, when asked her reason for seeing her doctor, she may respond, “My knee hurts a lot, frequently in the morning.” Consider the following two related problems: 1) mapping a patient-described symptom or condition to specific medical terminology/disease (e.g., knee hurts = knee pain → ICD-9 719.46, Pain in joint involving lower leg); and 2) standardizing descriptive terms (adjectives, adverbs) to the some scale (e.g., Does “a lot” mean a mild discomfort or a crippling pain? Does “frequently” mean every day or just a once a week?)
1 Introduction Several informatics endeavors related to the automated structuring of data are perti nent here. Electronic collections of validated questionnaires are being created, formally defining pertinent positive/negative questions and responses(eg, see the National Insti- tutes of Health(NIH) PROMIS project [7] and related efforts by the National Cancer Institute, NCD). Such databases provide a foundation from which chief complaints and symptoms can be objectified and quantified with specificity: duration, severity, timing, and activities that either trigger or relieve the symptom can be asked. Likewise, existing diagnostic guidelines intended for non-physicians, such as the American Medical Asso- ciation Family Medical Guide 5], can be turned into online, interactive modules with ecision trees to guide a patient through the response process. Markedly, an inherent issue with such questionnaires is determining how best to elicit responses from patients: aspects of visualization and human-computer interaction(HCD) thus also come into play (see Chapter 4). Apart from structured formats, more complicated methods such as medical natural language processing(NLP) can be applied to structure the statement by vides an overview of NLP research and applications Assessing the patient. The chief complaint provides a basis for beginning to under- stand the problem, but a clinician will still require additional background to establish potential reasons for the knee pain. For example, does the patient have a history of a previous condition that may explain the current problem? Has this specific problem occurred before (ie, is it chronic) or did any specific past event cause this issue(e.g trauma to the knee)? The answers to these questions are all gleaned from questioning the patient further and an exploration of the medical record An array of medical and imaging informatics research is ongoing to enrich the elec tronic medical records(EMR) functionality and to bring new capabilities to the point of care. A longstanding pursuit of the EMR is to provide an automated set of relevant information and a readily searchable index to patient data: rather than manually in spect past reports and results, the system should locate germane documents, if not permit the physician to pose a simple query to find key points. Informatics work in distributed information systems concentrates on the problems of data representation and connectivity in an increasingly geographically dispersed, multidisciplinary health care environment. Patients are commonly seen by several physicians, who are often at different physical locations and institutions. As such, a patient's medical history may be segmented across several disparate databases: a core challenge of informatics is to find effective ways to integrate such information in a secure and timely fashion (see Chapter 3). For imaging, past exams should be made available; but instead of the whole study, only(annotated) sentinel image slices that detail a problem could be re called. Although manual image capture and markup is presently used, automated tech- niques are being investigated to identify anatomical regions and uncover potential abnormalities on an image(e.g, CAD); and to segment and quantify disease based on domain knowledge(see Chapter 5). For textual data, such as generated from notes and consults(e.g, a radiology report), NLP techniques are being developed to facilitate content indexing(see Chapter 6). To aggregate the information into a useful tool data model that matches the expectations of the clinician must be used to organize the extracted patient data(see Chapter 7), and it must then be presented in a way con- ducive to thinking about the problem(see Chapter 4) Specifying the study. Based on the patient 's responses and review of her record, the PCP wishes to differentiate between degenerative joint disease and a meniscal tear. If
1 Introduction 7 Several informatics endeavors related to the automated structuring of data are pertinent here. Electronic collections of validated questionnaires are being created, formally defining pertinent positive/negative questions and responses (e.g., see the National Institutes of Health (NIH) PROMIS project [7] and related efforts by the National Cancer Institute, NCI). Such databases provide a foundation from which chief complaints and symptoms can be objectified and quantified with specificity: duration, severity, timing, and activities that either trigger or relieve the symptom can be asked. Likewise, existing diagnostic guidelines intended for non-physicians, such as the American Medical Association Family Medical Guide [5], can be turned into online, interactive modules with decision trees to guide a patient through the response process. Markedly, an inherent issue with such questionnaires is determining how best to elicit responses from patients; aspects of visualization and human-computer interaction (HCI) thus also come into play (see Chapter 4). Apart from structured formats, more complicated methods such as medical natural language processing (NLP) can be applied to structure the statement by the patient, identifying and codifying the chief complaint automatically. Chapter 6 provides an overview of NLP research and applications. Assessing the patient. The chief complaint provides a basis for beginning to understand the problem, but a clinician will still require additional background to establish potential reasons for the knee pain. For example, does the patient have a history of a previous condition that may explain the current problem? Has this specific problem occurred before (i.e., is it chronic) or did any specific past event cause this issue (e.g., trauma to the knee)? The answers to these questions are all gleaned from questioning the patient further and an exploration of the medical record. An array of medical and imaging informatics research is ongoing to enrich the electronic medical record’s (EMR) functionality and to bring new capabilities to the point of care. A longstanding pursuit of the EMR is to provide an automated set of relevant information and a readily searchable index to patient data: rather than manually inspect past reports and results, the system should locate germane documents, if not permit the physician to pose a simple query to find key points. Informatics work in distributed information systems concentrates on the problems of data representation and connectivity in an increasingly geographically dispersed, multidisciplinary healthcare environment. Patients are commonly seen by several physicians, who are often at different physical locations and institutions. As such, a patient’s medical history may be segmented across several disparate databases: a core challenge of informatics is to find effective ways to integrate such information in a secure and timely fashion (see Chapter 3). For imaging, past exams should be made available; but instead of the whole study, only (annotated) sentinel image slices that detail a problem could be recalled. Although manual image capture and markup is presently used, automated techdomain knowledge (see Chapter 5). For textual data, such as generated from notes and consults (e.g., a radiology report), NLP techniques are being developed to facilitate content indexing (see Chapter 6). To aggregate the information into a useful tool, a data model that matches the expectations of the clinician must be used to organize the extracted patient data (see Chapter 7), and it must then be presented in a way conducive to thinking about the problem (see Chapter 4). Specifying the study. Based on the patient’s responses and review of her record, the PCP wishes to differentiate between degenerative joint disease and a meniscal tear. If abnormalities on an image (e.g., CAD); and to segment and quantify disease based on niques are being investigated to identify anatomical regions and uncover potential
A.A.T. Bui et al a patient complains of knee pain, then traditionally as a first step an x-ray is obtained But if the patient's symptoms are suggestive of pain when going up stairs, then a knee magnetic resonance(MR) imaging study is warranted over an x-ray(this symptom being suggestive of a meniscal tear). When asked whether going up stairs aggravates e knee pain, the patient indicated that she was unsure. Thus, her PCP must now make a decision as to what imaging test should be ordered. Furthermore, the selection of the imaging exam must be tempered by the availability of the imaging equipment, the needed expertise to interpret the imaging study, and other potential constraints (e.g, cost, speed of interpretation, etc. First, supporting the practice of evidence-based medicine(EBM) is a guiding principle of biomedical informatics, and hence medical imaging informatics. The development and deployment of practice guidelines in diagnosis and treatment has been an enduring effort of the discipline, suggesting and reminding physicians on courses of action to improve care. For instance, if the patient,'s clinician was unaware of the sign of a meniscal tear, the system should automatically inform him that an MR may be indicated if she has knee pain when climbing stairs; and supporting literature can be automatic- ally suggested for review. Second, formal methods for medical decision-making are central to informatics, as are the representation of medical knowledge needed to inform the algorithms [10]. Techniques from computer science, ranging from rudimentary rule- bases to statistical methods(e.g, decision trees); through to more complex probabilistic hidden Markov models(HMMs) and Bayesian belief networks(BBNs) are finding applications in medicine(see Chapter 8). For example, the evidence of the patients medical history, her response to the physician,s inquiries, the availability of imaging, and the relative urgency of the request can be used in an influence diagram to choose between the x-ray and MR (see Chapter 9). Such formalizations are providing new tools to model disease and to reason with partial evidence. Essential to the construction of many of these models is the compilation of large amounts of(observational)data from which data mining and other computational methods are applied to generate new knowledge. In this example, these disease models can be used: to identify further questions that can be asked to further elucidate the patients condition(improving the likelihood of choosing an optimal imaging exam); and to select the type of imaging udy, and even its acquisition parameters, to best rule in/out elements of the differential diagnosis Ultimately, an electronic imaging infrastructure that expedites accurate diagnosis can improve the quality of healthcare; and even within this simple example of choosing an imaging protocol, the role of informatics is apparent in enhancing the process of care When used appropriately, medical imaging is effective at objectifying the initial diagnostic hypothesis(differential diagnosis) and guiding the subsequent work-up Given a chief complaint and initial assessment data, one can envision that specialists or software algorithms would select an imaging protocol for an appropriate medical condition even before a visit to the PCP. The PCP can then access both objective imaging and clinical data prior to the patients visit. Medical imaging informatics research looks to improve the fundamental technical methods, with ensuing translation Cost Considerations Some have targeted the cost of imaging as a major problem in healthcare within the United States: one 2005 estimate by the American College of Radiology (ACR) was that $100 billion is spent annually on diagnostic imaging, including computed
8 A.A.T. Bui et al. a patient complains of knee pain, then traditionally as a first step an x-ray is obtained. But if the patient’s symptoms are suggestive of pain when going up stairs, then a knee magnetic resonance (MR) imaging study is warranted over an x-ray (this symptom being suggestive of a meniscal tear). When asked whether going up stairs aggravates the knee pain, the patient indicated that she was unsure. Thus, her PCP must now make a decision as to what imaging test should be ordered. Furthermore, the selection of the imaging exam must be tempered by the availability of the imaging equipment, the needed expertise to interpret the imaging study, and other potential constraints (e.g., cost, speed of interpretation, etc.). First, supporting the practice of evidence-based medicine (EBM) is a guiding principle of biomedical informatics, and hence medical imaging informatics. The development and deployment of practice guidelines in diagnosis and treatment has been an enduring effort of the discipline, suggesting and reminding physicians on courses of action to improve care. For instance, if the patient’s clinician was unaware of the sign of a meniscal tear, the system should automatically inform him that an MR may be indicated if she has knee pain when climbing stairs; and supporting literature can be automatically suggested for review. Second, formal methods for medical decision-making are central to informatics, as are the representation of medical knowledge needed to inform the algorithms [10]. Techniques from computer science, ranging from rudimentary rulebases to statistical methods (e.g., decision trees); through to more complex probabilistic hidden Markov models (HMMs) and Bayesian belief networks (BBNs) are finding applications in medicine (see Chapter 8). For example, the evidence of the patient’s medical history, her response to the physician’s inquiries, the availability of imaging, and the relative urgency of the request can be used in an influence diagram to choose between the x-ray and MR (see Chapter 9). Such formalizations are providing new tools to model disease and to reason with partial evidence. Essential to the construction of many of these models is the compilation of large amounts of (observational) data from which data mining and other computational methods are applied to generate new knowledge. In this example, these disease models can be used: to identify further questions that can be asked to further elucidate the patient’s condition (improving the likelihood of choosing an optimal imaging exam); and to select the type of imaging study, and even its acquisition parameters, to best rule in/out elements of the differential diagnosis. Ultimately, an electronic imaging infrastructure that expedites accurate diagnosis can improve the quality of healthcare; and even within this simple example of choosing an imaging protocol, the role of informatics is apparent in enhancing the process of care. When used appropriately, medical imaging is effective at objectifying the initial diagnostic hypothesis (differential diagnosis) and guiding the subsequent work-up. Given a chief complaint and initial assessment data, one can envision that specialists or software algorithms would select an imaging protocol for an appropriate medical condition even before a visit to the PCP. The PCP can then access both objective imaging and clinical data prior to the patient’s visit. Medical imaging informatics research looks to improve the fundamental technical methods, with ensuing translation to clinical applications. Cost Considerations Some have targeted the cost of imaging as a major problem in healthcare within the United States: one 2005 estimate by the American College of Radiology (ACR) was that $100 billion is spent annually on diagnostic imaging, including computed
1 Introduction 9 mography (CT), MR, and positron emission tomography(PET) scans [2. while acknowledging that many factors are contributing to these high costs it is, however, important to separate out two issues: the healthcare cost savings generated as a result of imaging, in light of earlier diagnoses and quality of life; and the true cost of performing an imaging study (i.e, versus what is charged An"appropriate"process of care that disregards issues related to utilization review ind approvals required for imaging studies can be very effective for care of the patient as well as cost-effective. In one study performed by us for a self-insured employer group, we removed all of the requirements for(pre-)approval of imaging studies and allowed primary care physicians to order imaging based on their diagnostic hypothesis and the need of the patient. The imaging costs were instead capitated for the employer group. The number of cross-sectional images, particularly CT and MR, more than doubled and the number of projectional images decreased. However, the net effect was not only significant cost savings to the employer group but also much higher quality and satisfaction by patients [12]. A follow-up study further showed improved he (lowered incidence of chronic disease, decreased number of hospitalizations and emergency room visits, etc. ) continued high levels of patient satisfaction, and lowered expenditures within the cost-capitated imaging environment relative to a control group [4]. All of this is to suggest that it is not necessarily the overuse of imaging that is inherently costly, and that there are in fact cost-savings introduced through the un- restricted use of imaging. Of course, a capitated cost agreement with unfettered usage of imaging is not the norm. Unfortunately, the cost of imaging studies is rarely the true ost of performing the study. As an example, presently charges for a brain MR imaging study with and without contrast are in excess of $7, 000 at some institutions largely because of professional fees and attempts to recoup costs(e.g, from non paying and uninsured individuals). Yet in one internal study we conducted in the 1990s to understand the real cost of CTs and Mrs, it was concluded that the price of an MR study is no more than S200 and the price of a CT less than $120. These costs included technologists time, materials used (e.g, contrast) and the depreciation of the scanning machines over five years. Even adjusting for inflation and a moderate profes- sional fee, one can argue that the charges seen today for imaging largely outpace ne true cost of the exam. Hence, a current practical challenge for medical imo s informatics is to develop new paradigms of delivery that will encourage the use imaging throughout the healthcare environment while still being cost-effective A Historic Perspective and Moving Forward Medical imaging informatics is not new: aspects of this discipline have origins span- ning back over two or more decades [14]. As such, it is useful to consider this field's interdisciplinary evolution to understand its current challenges and future. Below, we consider four different eras of technical research and development. PACS: Capturing Images Electronically Concurrent to the progress being made with respect to CT and MR imaging, initial efforts to create an electronic repository for(digital) imaging in the 1980s led to the creation of picture archive and communication systems(PACS).[8, 11] provide some perspective on the early development of PACS, which focused on linking acquisition devices (ie, scanners), storage, intra-site dissemination of studies, and display tech- nologies(soft and hard copy). With the introduction of PACS, some of the physical limitations of film were overcome: images were now available anywhere within an
1 Introduction 9 tomography (CT), MR, and positron emission tomography (PET) scans [2]. While acknowledging that many factors are contributing to these high costs it is, however, important to separate out two issues: the healthcare cost savings generated as a result of imaging, in light of earlier diagnoses and quality of life; and the true cost of performing an imaging study (i.e., versus what is charged). An “appropriate” process of care that disregards issues related to utilization review and approvals required for imaging studies can be very effective for care of the patient as well as cost-effective. In one study performed by us for a self-insured employer group, we removed all of the requirements for (pre-)approval of imaging studies and allowed primary care physicians to order imaging based on their diagnostic hypothesis and the need of the patient. The imaging costs were instead capitated for the employer group. The number of cross-sectional images, particularly CT and MR, more than doubled and the number of projectional images decreased. However, the net effect was not only significant cost savings to the employer group but also much higher quality and satisfaction by patients [12]. A follow-up study further showed improved health (lowered incidence of chronic disease, decreased number of hospitalizations and emergency room visits, etc.), continued high levels of patient satisfaction, and lowered expenditures within the cost-capitated imaging environment relative to a control group [4]. All of this is to suggest that it is not necessarily the overuse of imaging that is inherently costly, and that there are in fact cost-savings introduced through the unrestricted use of imaging. Of course, a capitated cost agreement with unfettered usage of imaging is not the norm. Unfortunately, the cost of imaging studies is rarely the true cost of performing the study. As an example, presently charges for a brain MR imaging study with and without contrast are in excess of $7,000 at some institutions – largely because of professional fees and attempts to recoup costs (e.g., from nonpaying and uninsured individuals). Yet in one internal study we conducted in the 1990s to understand the real cost of CTs and MRs, it was concluded that the price of an MR study is no more than $200 and the price of a CT less than $120. These costs included technologists time, materials used (e.g., contrast) and the depreciation of the scanning machines over five years. Even adjusting for inflation and a moderate professional fee, one can argue that the charges seen today for imaging largely outpace the true cost of the exam. Hence, a current practical challenge for medical imaging informatics is to develop new paradigms of delivery that will encourage the use of imaging throughout the healthcare environment while still being cost-effective. A Historic Perspective and Moving Forward Medical imaging informatics is not new: aspects of this discipline have origins spanning back over two or more decades [14]. As such, it is useful to consider this field’s interdisciplinary evolution to understand its current challenges and future. Below, we consider four different eras of technical research and development. PACS: Capturing Images Electronically Concurrent to the progress being made with respect to CT and MR imaging, initial efforts to create an electronic repository for (digital) imaging in the 1980s led to the creation of picture archive and communication systems (PACS). [8, 11] provide some perspective on the early development of PACS, which focused on linking acquisition devices (i.e., scanners), storage, intra-site dissemination of studies, and display technologies (soft and hard copy). With the introduction of PACS, some of the physical limitations of film were overcome: images were now available anywhere within an