Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
5 Lessons from Medical Imaging for Explosive Detection Systems Imaging technologies are used extensively in medicine for the early detection of disease (screening), for the diagnosis and characterization of disease, for the monitoring of therapy, and for post- therapy surveillance for disease recurrence. These imaging technologies include not only computed tomography (CT) but also more traditional x-ray modalities, magnetic resonance imaging, and various nuclear scanning modalities. The medical uses of imaging have fueled major advances in technology (hardware and software) and in methods of quality control and performance monitoring. An extensive research apparatus, in place for the study of all aspects of medical imaging, has given rise to a large body of scientific literature. This chapter provides a comparative review of key features of medical imaging systems and explosive detection systems (EDSs). Although the discussion focuses primarily on medical CT because of its technological proximity to CT for EDSs, many of the conclusions apply more broadly to other imaging modalities. COMPUTED TOMOGRAPHY IN MEDICINE AND IN EXPLOSIVE DETECTION SYSTEMS Diagnostic radiology was revolutionized with the introduction of computed tomography in the early 1970s 1 because this technology provided two important characteristics: (1) the ability to display relatively high resolution cross sections of human anatomy while (2) assigning quantitative values to the pixels. CT images are scaled in Hounsfield units, which approximately represent the mass density of the object being scanned. No prior diagnostic radiology instrument was capable of quantifying localized tissue characteristics such as density, and thus many studies were initiated to determine the value of this new kind of information. The cross-sectional nature of the images was immediately recognized as a remarkable breakthrough by the medical imaging community, and within only several years, CT was adopted by radiology departments both large and small. This development fostered intensive research and development by vendors of medical imaging instruments, with relatively few contributions being made by the academic community. Much of the engineering development in medical CT was therefore considered proprietary and has remained unpublished. In the same way, CT-based EDS vendors have sequestered details of their instruments from public knowledge, though with EDSs this is done for security purposes in addition to proprietary concerns. 1 Overviews of the history of medical CT can be found in the following: X. Pan, J. Siewerdsen, P. La Riviere, and W. Kalendar, Anniversary Paper, Development of x-ray computed tomography: The role of Medical Physics and AAPM from the 1970s to present, Medical Physics 35(8):3728-3739, 2008; E. Krupinski and Y. Jiang, Anniversary Paper, Evaluation of medical imaging systems, Medical Physics 35(2):645-659, 2008; M. Giger, G.P. Chan, and J. Boone, Anniversary Paper, History and status of CAD and quantitative image analysis, Medical Physics 35(12):5799-5820, 2008; and S. Armato III and B. van Ginneken, Anniversary Paper, Image processing and manipulation through the pages of Medical Physics, Medical Physics 35(10):4488-4500, 2008. 45
THE TECHNOLOGY OF MEDICAL COMPUTED TOMOGRAPHY SCANNERS Modern radiological CT scanners are similar to CT scanners for EDSs. They use rotating cone- beam geometry, with one or two x-ray tubes operating at 80 to 140 kilovolt peak voltage, and the same number of arrays of multi-channel x-ray detectors are arranged on a gantry that rotates at speeds of up to several hundred revolutions per minute. A bowtie filter may be used to reduce the dynamic range and homogenize the beam hardening of the projections. Projection data are typically acquired from many slices simultaneously, so that the entire anatomic volume of interest can be interrogated in under a second. Images are reconstructed by variants on filtered back-projection methods with corrections for scatter, soft- tissue beam hardening, off-focal-spot blurring, detector channel gains, gravity effects on the gantry, tube and/or modulator intensity drifts, and other, more subtle imperfections. The reconstructions typically utilize purpose-built hardware to perform the calculations rapidly. Whole-body scans have reconstructed voxels (volume pixels) of less than 1 cubic millimeter and have selectable algorithms that are optimized for either soft tissue or bone conspicuity. Each vendor of medical CT scanners initially had a proprietary image file format, but the Digital Imaging and Communications in Medicine (DICOM) format was adopted in 1993 as the industry standard through promulgation by a committee of the National Electrical Manufacturers Association (NEMA). This move was critical, because the standardization allowed third-party vendors of image- processing software and hardware to burgeon, and it provided flexibility to hospitals by equipping radiology departments with a mixture of vendorsâ devices optimized for their needs. QUANTIFICATION WITH COMPUTED TOMOGRAPHY Initially, there was great enthusiasm for exploring the quantitative nature of CT, given that Hounsfield units apparently provide high precision in depicting tissue density. It was hoped that tissue characterization based on these numbers would allow physicians to make informed decisions in detecting and staging pathology as well as for monitoring therapy. For example, studies examined whether a threshold Hounsfield unit boundary value could be set and used to diagnose lung cancer nodules. 2 Unfortunately, this study and others demonstrated that biodiversity precluded this simple level-set approach from being fully successful (the approach was 98 percent sensitive but only 58 percent specific). 3 With this degree of specificity, it was realized that the use of Hounsfield unit numbers alone for cancer screening would lead to unacceptable false-positive rates. Thus, diagnostic CT suffers from the identical problem that plagues CT-based EDSs in that density values overlap between benign and malignant target types. COMPARISON OF COMPUTED TOMOGRAPHY (CT) FOR MEDICAL USE AND CT FOR EXPLOSIVE DETECTION SYSTEMS It is already evident that medical CT and CT for EDSs share a number of similarities but also have dissimilarities. It is instructive to examine this comparison in detail to determine if there are opportunities by which medical CT experience can inform EDS design and operation. First, there is a difference in the nature of the target of detection for the two systems. For diagnostic CT, the type of object being imaged (e.g., a tumor embedded in tissue) is fixed. That is, human pathology does not adapt except through relatively slow evolution or response by mutation to therapies 2 Stephen J. Swensen, Robert W. Viggiano, David E. Midthun, et al., Lung nodule enhancement at CT: Multicenter study, Radiology 214:73-80, 2000. 3 Sensitivity = PD [probability of detection]; specificity = 1 â PD. 46
that have been introduced. Thus, the CT characteristics of brain tumors are the same now as when CT was first introduced, although biodiversity guarantees a range of presentations. By comparison, the âtargetâ for CT scanners used for explosive detection systems is material within a device constructed by humans and purposefully designed to be deceptive in appearance; it can be composed of a wide variety of materials and components and designs that continue to evolve. The nature of these devices therefore can rapidly be altered as their designers adopt new strategies in response to changes in EDS equipment or geopolitical and other stimuli. Because the range of density for many explosives overlaps that of common household materials in checked baggage, the explosives designer has many choices and can readily alter the device composition and disposition as EDSs become more sophisticated. This means that the EDS detection software should be flexible so that changes in algorithms can be rapidly installed in response to the evolution of the explosives threats. That is not the case at the present time because EDSs are supplied and certified as a single package consisting of hardware, reconstruction software, and post-processing software. Second, in addition to the differences in the nature of the target for the two modalities, there are differences in the conditions of use. Medical CT has traditionally been an imaging modality used in symptomatic cohorts (that is, a test is indicated because of prior clinical findings) and is used to confirm a diagnosis, to determine the stage of a disease, to delineate the site or extent of a pathology, or to monitor the progression of a disease or therapy. The use of CT for imaging asymptomatic cohorts (screening) is more recent and is still in development. For example, CT colonography is under consideration as a modality of colon cancer screening, 4 CT angiography is beginning to be used in the screening and detection of coronary artery disease, 5 helical CT is currently evaluated as a modality to screen for lung cancer, 6 and CT is also under consideration as a modality to screen for breast cancer. 7 The screening uses of CT are still evolving, and as a result the image-processing software is not as much of an intrinsic component of the medical CT scanner as it is of the CT scanner used in EDSs. However, the growing potential of CT as a screening modality is giving rise to the development of imaging software such as computer-aided diagnosis for CT colonography and software for volumetric CT. 8 A third difference between medical CT and CT used in explosive detection systems is that medical CT relies on human operators to read the scans and render a diagnostic decision. As noted below, an array of software for visualization and classification has been developed for medical CT. However, these systems are not used as a replacement for human judgment. Time pressure to render a diagnosis is not a significant factor except in emergency medical cases. By contrast, EDS use of CT scanners must rely on the automated threat recognition (ATR) algorithm to make the first decision, by nature of the sheer volume of baggage that must be rapidly processed. In order to respond to the compressed time frame in which they must operate, the current EDSs also provide the post-processing algorithms as an inherent proprietary component. 4 C.D. Johnson, M.H. Chen, A. Toledano, et al., Accuracy of CT colonography for detection of large adenomas and cancers, New England Journal of Medicine 359(12):1207-1217, 2008. 5 G.L. Raff and J.A. Goldstein, Coronary angiography by computed tomography: Coronary imaging evolves, Journal of the American College of Cardiologists 49:1830-1833, 2007. 6 O. Brawley and B. Kramer, Cancer screening in theory and in practice, Journal of Clinical Oncology 23:293- 300, 2005. 7 K.K. Lindfors, J.M. Boone, T.R. Nelson, K. Yang, A.L. Kwan, and D.F. Miller, Dedicated breast CT: Initial clinical experience, Radiology 246(3):725-733, 2008. 8 Much of the early computer-assisted diagnostics in medicine can trace its roots to the militaryâs automated target-recognition programs. Such screening and detection programs may also have application to threat recognition in baggage screening. See, for example, John M. Irvine, Targeting breast cancer, IEEE Engineering in Medicine and Biology 21(6):36-40, 2002. Many of the issues discussed by the author (including appropriate ROC settings, the role of the screener, and the problems of missed detections and false alarms) are relevant to an airport setting. Additionally, the militaryâs experience in identifying targets in a cluttered environment and with forces determined to defeat it can inform the aviation security setting. 47
A fourth difference is that standardization of image output exists only for medical CT. Medical scanners typically stop at simply producing images, without post-processing as an inherent component. As a result, it has been possible to standardize the image file format since the mid-1990s. This has led to a vast array of third-party products for visualizing and classifying images, with input from the academic and business communities key to their development. In addition, scanner vendors have developed workstations with post-processing software for the visualization and analysis of cardiovascular, angiographic, dynamic contrast, and many other functional assessments. Thus commercial pressures have led to a wide range of innovation in the products that are available. By contrast, CT scanners for EDSs at present have built-in, proprietary classification algorithms, and thus only a small fraction of the national expertise in image analysis has been brought to bear on the explosives-detection problem. In recognition of the success of competition in the medical arena with the adoption of the DICOM file format as the industry standard, a similar plan, the Digital Imaging and Communication in Security (DICOS), has been proposed by the NEMA for standardizing EDS images and allowing competition that includes the academic community for the development of post-processing algorithms. The proposed DICOS standard relies heavily on the work that has already been done in DICOM and adapts those standards for security applications. 9 In the opinion of this committee, this is a welcome and necessary development. As a fifth difference between the uses of CT for medical and for explosives-detection purposes, false positives and false negatives have different implications for the two modalities. EDS scanners have very strict limits on the time allotted (seconds) for scanning and processing before a decision must be made. Failure to clear a bag in that time means either that the bag is rescanned or that it must undergo manual inspection, both of which can lead to flight delays and passenger inconvenience and have major human resource implications related to the costs in personnel time for resolving false alarms. The requirement of a high probability of detection and the limits in the overall CT examination time augur for caution among those creating the processes and lead to the high false-positive rates by nature of the receiver operating characteristic (ROC) curve. A sixth difference is that medical CT scanners have dose limitations, but these can be relaxed for CT-based EDSs. This difference enables the use of dual-energy approaches for providing a second degree of freedom in the information available to the ATR algorithm. It is by no means clear that the present design of CT scanners for EDSs is optimized to take advantage of this added freedom. As a seventh difference, the patient is positioned in the medical CT scanner by a technologist who is trained to do so uniformly and is provided with adequate time for ensuring highly repeatable, diagnostic-quality images. Thus repeated scans are rarely required because of the human precision exercised in the scan setup. With CT-based EDSs, the bags have many styles and shapes, and the bags and their contents are not oriented in a consistent manner when entering the gantry on the conveyor to the scanner. As a result, the same bag scanned repeatedly by a CT-based EDS can have a high degree of inter-scan inconsistency in the images due to image artifacts and finite image resolution; consequently the system might alarm on one scan but not the next. As described in Chapter 3, there is thus an opportunity to improve detection efficiency (to reduce the rates of false positives while maintaining or increasing the true-positive rates) by using repeated scanning of bags. A final difference is that the development of medical CT scanners has occurred over a period of years, driven by a broad range of consumer needs and marketing preferences. The vendors of these scanners are incentivized by market pressures to deliver systems with the highest image quality and flexible features, and they have developed extensive research and development efforts to remain competitive. This competition is informed by well-publicized academic studies that examine and report in the open literature both the physical characteristics and the clinical performance of the CT machines under clinical (field) operating conditions. The open environment and standardized image format (DICOM) allow easy entrance into this arena to multiple vendors for supplying image-processing and computer-aided diagnosis processing software to the community, where efficacy can be tested and readily 9 National Electrical Manufacturers Association, NEMA Standards Publication [IC] v01. Rosslyn, Va., 2010. 48
reported. It also enables open access for research in image classification and analysis by any of the many experienced investigator groups around the world. EDS machines, by comparison, were developed over a much shorter period, in proprietary secrecy, by a small number of vendors, and although not in widespread use until after the attacks of September 11, 2001, these machines had a rapid research and development cycle and, subsequently, mass deployment. The vast general academic imaging community had no part in either the design of the characterization of these instruments, and it remains unengaged. There are many parallels between medical CT scanners and CT scanners for EDSs, including the equipment development history, the nature of the target or threat to be detected and classified, the time allotted for doing so, the need for automated detection in EDSs, and the difference in operating points on the ROC curve distinguish the two scanner applications. Freedom to increase the x-ray dose in the EDS application and to introduce dual-energy scanning leads to possibilities for additional information recovery for EDSs and may lead to further classification precision. Engagement of the substantial image reconstruction and/or processing community could lead to further beneficial evolution. Conclusion: The engagement of more members of the academic and industrial communities, as well as of those in the medical diagnostics and military communities having theoretical and applied expertise in image reconstruction and target recognition, could lead to increases in the effectiveness (and, in particular, decreases in false alarms) of CT-based explosives detection. EXPLOSIVE DETECTION SYSTEMS AND MEDICAL IMAGING The main points about the comparison of medical CT to CT for EDSs apply more generally to other medical imaging modalities. In particular: ⢠Screening for disease is the closest analogue to the use of CT-based EDS for baggage screening. However, the time frame for baggage screening is considerably more compressed than that for screening for disease, and the volume of items to be screened is considerably greater. As a result, baggage-screening modalities have higher reliance on automation and software that permit high throughput while maintaining desired accuracy. ⢠Exposure to radiation (e.g., x-ray, nuclear scans) and other kinds of harm from screening are a significant concern in screening for disease. However, such concerns are far less relevant in the baggage-screening context and permit the use of technology with higher radiation. ⢠Medical imaging technology undergoes a continuous process of evaluation through formal studies, often comparative, routinely published in the extensive literature on diagnostic imaging. These studies address a broad range of issues and span the developmental trajectory of imaging modalities, from early laboratory and engineering studies to advanced clinical trials evaluating the use of these modalities in a medical setting. The considerable methodological and practical expertise from medical imaging research can be put to good use in fostering the development of a rigorous evaluation of EDS for baggage screening. ⢠Various approaches have been developed and implemented to monitor the quality and effectiveness of medical imaging in daily medical practice. These approaches are often institution- specific, but they utilize standards and best practices developed and recommended by professional societies and other organizations. For the broadly used modality of mammography for breast cancer screening, a national system of monitoring quality has been in place since the early 1990s. The system was instituted by the Mammography Quality Standards Act (MQSA) of 1992 (P.L. 102-539) and is run by the U.S. Food and Drug Administration. The MQSA regulations include nation-wide quality standards for mammography, with annual inspections, accreditation and certification requirements, standards for reporting results, and requirements for the training, education, and experience of all personnel. Also for mammography, substantial data collection is conducted nation-wide through mammography registries. Data from these registries are used to monitor and evaluate the practice or mammography around the 49
country. For example, a host of studies conducted by the Breast Cancer Surveillance Consortium have been published in recent years on such topics as the rates of positive mammography findings, the diagnostic accuracy and predictive value of mammography, and factors associated with the variation in positivity rates and diagnostic performance across institutions and individual mammographers. 10 LESSONS LEARNED Image Standardization and Post-Processing Software Development Importantly, the introduction of a standardized image format opened the door to academic participation in post-processing innovations in three- and four-dimensional visualization and computer- aided diagnosis programs, 11 because details of the scanner process were divorced from those details related to the processing of the images for specific applications. This allowed the scanner vendors to retain control over propriety details of the acquisition of images, but it provided easy access to research on the application of these images from a large body of academic and industrial groups with experience in relevant fields of study. Based on the experience with DICOM, there is now a move toward standardizing the image format of CT used for EDSs for the same purpose: to foster participation by academic and other laboratories in the development of post-processing algorithms for explosives detection. The committee believes that proceeding with the plan to separate the acquisition of CT images from the post-processing programs will improve CT-based EDSs while at the same time inviting greater competition for the development of the post-processing programs. The existing medical image processing field is large and includes dozens of strong academic laboratories as well as well-supported industrial medical research and development programs that have been successful in providing excellent computer- aided diagnosis algorithms. Broader participation by these highly experienced groups with diverse backgrounds in image processing would make it likely that new methods would be developed that may improve the detection and classification efficiency of baggage scanners. However, one must exercise caution in this endeavor to sever the connection between acquisition and analysis operations. Post-processing success depends on the quality and completeness of the images themselves. It is by no means clear that the image quality is fully appropriate and optimized in current baggage scanners, and thus limiting the availability of the information to only the reconstructed images guarantees that any existing deficiencies in the acquisition and reconstruction processes will not be addressed by the larger community. To enable optimization, a form of raw data (such as sinogram 12) will need to be made available in a standardized format to a limited community of scientific experts so that they might assess current limitations of CT based EDS images acquisition and potentially derive novel solutions for improved image reconstruction as a part of the problem. Because of the importance of addressing the false-positive rate of CT-based EDS alarms, it will be critical to address the proprietary considerations in a way that allows such a larger community involvement. 10 R. Ballard-Barbash, S.H. Taplin, B.C. Yankaskas, et al., Breast Cancer Surveillance Consortium: A national mammography screening and outcomes database, American Journal of Roentgenology 169(4):1001-1008, 1997. 11 See, for example, âSecurView Diagnostic Workstations,â available at http://www.hologic.com/en/breast- imaging/diagnostic-workstations/, accessed September 12, 2010; and Fang-Fang Yin, Maryellen L. Giger, Kunio Doi, Charles E. Metz, Carl J. Vyborny, and Robert A. Schmidt, Computerized detection of masses in digital mammograms: Analysis of bilateral subtraction images, Medical Physics 18(5, September):955-963, 1991. 12 In this context, a sonogram is a three-dimensional visual representation of the x-ray signal as it is measured at a specified angle in the imaging plane at varying distances along the detector array. 50
Quality Control and Performance Monitoring The experience from the medical uses of imaging provides strong support for the feasibility of and the need for the establishment of nation-wide quality-control standards not only for equipment and processes but also for the training and continuous education of image-interpreting personnel. These standards will need to be based on the results of scientific research in both technology and human factors. The actual performance of the systems in practice will need to be monitored through systematic data collection and analysis, as discussed in Chapter 6 of this report. Finding: The introduction of an industry-standard medical image format (DICOM) in 1993 fostered the development of a diverse and innovative array of diagnostic and therapeutic image visualization, processing, and automated detection/diagnostic products, fueled by the panoply of academic and private-sector research laboratories with extensive experience in the field. Recommendation: The Department of Homeland Security should promote the rapid acceptance of a standardized format for EDS images for all TSA-certified machines. 51
