You are in: eMedicine Specialties > Radiology > BREAST Mammography - Computer-Aided DetectionArticle Last Updated: Jan 26, 2005AUTHOR AND EDITOR INFORMATIONSection 1 of 9
  Author: Michael J Ulissey, MD, Assistant Professor of Radiology, Director of Breast Imaging, Department of Radiology, The University of Texas Southwestern Medical Center at Dallas Michael J Ulissey is a member of the following medical societies: American College of Radiology, Radiological Society of North America, and Texas Medical Association Coauthor(s): Jimmy Roehrig, PhD, Chief Science Officer, R2 Labs, R2 Technology Editors: John M Lewin, MD, Associate Clinical Professor, Department of Preventative Medicine and Biometrics, Director of Teleradiology, Co-director of Breast Imaging Section, Director of Breast Imaging Research, Department of Radiology, University of Colorado Health Sciences Center; Consulting Radiologist, Diversified Radiology of Colorado; Bernard D Coombs, MB, ChB, PhD, Consulting Staff, Department of Specialist Rehabilitation Services, Hutt Valley District Health Board, New Zealand; Edward Azavedo, MD, PhD, Director of Clinical Breast Imaging Services, Associate Professor, Department of Radiology, Karolinska University Hospital, Sweden; Robert M Krasny, MD, Consulting Staff, Department of Radiology, The Angeles Clinic and Research Institute; Eugene C Lin, MD, Consulting Staff, Department of Radiology, Virginia Mason Medical Center Author and Editor Disclosure Synonyms and related keywords: breast imaging, CAD, computer-aided diagnosis, mammographic analysis, image processing, IP, image segmentation, artificial intelligence, AI, neural network INTRODUCTIONSection 2 of 9
  Computer-aided detection (CAD) for mammography is a new and evolving topic in the realm of breast radiology. Despite our best efforts, radiologists have long known that some breast cancers remain undetected on screening mammograms. As the public grew more aware of this problem, medical/legal aspects of mammographically missed breast cancers became a pressing issue for both radiologists and lawyers. A variety of reasons may explain why breast cancer is missed on mammograms. Some breast cancers simply are not seen on mammograms and may remain hidden by dense tissue until a lump is felt. Other cancers are difficult to see because they blend into the background of fibroglandular tissue and are overlooked at screening. On retrospective evaluation, these cancers are occasionally detected; however, they might be missed a second time. Still other cancers are located in areas difficult to visualize (eg, subtle calcifications on the burned-out edge of the image). Occasionally, cancers are missed for no other reason than momentary distraction or inattention of the screening radiologist. Reasons for false-negative findings on mammography are difficult to determine because months to a year may pass before the cancer is detected, by which time it may have progressed to a more advanced state. Unfortunately, negative mammograms do not guarantee the absence of cancer, despite the wishes of both doctors and patients. As a result, years ago researchers began to develop better methods to make lesions more conspicuous on mammograms. Film and screen technology improved, quality standards were enacted, and breast radiology progressively became a more specialized field, allowing some radiologists to focus primarily in that area. Nevertheless, false-negative rates in mammography remain too high (Burhenne et al, 2000). Therefore, computers and CAD were applied to help detect breast cancer at an earlier stage. For CAD to be effective, it must be used widely. CAD can achieve its greatest contribution to medicine in the field of breast cancer, where it can help radiologists to detect cancers at the earliest stages possible. It is in this stage when most mammograms are read and most cancers missed. CAD is no longer a new modality. The US Food and Drug Administration (FDA) approved the first CAD system as an aid to the radiologist in screening mammography in June of 1998. As of 2001, only 130 CAD units were in clinical operation in the United States, many in academic centers. Now, more than 1600 units are being used in active clinical practice, and many new vendors have entered the CAD arena. Still, the concept of computers reading mammograms has been received with mixed feelings in the radiology community. The primary authors' personal experience in giving lectures is that many radiologists are afraid of the technology. For radiologists to accept CAD, they may need a better understanding of what it is, how it works, and how to use it in the everyday environment of breast radiology. Regarding nomenclature, CAD stands for computer-aided detection. In its present form, CAD should be used only for detection and never for diagnosis or reassurance. Regrettably, the erroneous term computer-aided diagnosis has been written into the lay and radiologic literature. HISTORY AND TECHNICAL ASPECTSSection 3 of 9
History of CAD The history of CAD involves different but intertwining and parallel developments in the research community, in the clinical community, and finally in the practical world of commercial development. In the research community, Winsberg et al may have been the first investigators to report on CAD. They considered the difficulty of viewing large volumes of screening mammograms before such screening was accepted in the clinical community. Other researchers, such as Ackerman and Gose, slowly expanded the concept of CAD. These investigators used a computer to extract 4 properties of lesions including (1) calcification, (2) spiculation, (3) roughness, and (4) shape. Eventually, these properties and many others came to be termed features in the CAD community, and the number of features gradually but substantially expanded. Early attempts of CAD Early attempts to apply computer vision were crude by modern standards because of the general state of computer and related technologies. For example, computer vision depends on digital images, and the scanners able to convert analog film data to digital numbers were not adequate to cover the resolution and dynamic range found on mammograms. Such devices did exist (primarily for use in astronomy as densitometers that digitized astronomic photos), but they were too expensive and cumbersome to achieve wide usage. In addition, the computers needed to analyze the large volume of digital data were expensive and usually too slow to be applied in a clinical setting. Almost concurrently with this early research, intriguing and often controversial studies were performed by using clinical data, and the results hinted at a difficulty in assessing radiologic findings. In 1981, Forrest and Friedman noted that the interpreting radiologist often missed as many as 60% of all nodules that were clearly present on chest radiographs. In 1993, Harvey et al performed a study of mammograms and found similar results. Possibly because it was controversial, Forrest found it difficult to publish the study. Perhaps for similar reasons, the article by Harvey et al was slanted more toward pointing out the difference between retrospective and blind interpretations than toward the clinical process of missing lesions on radiographs. Studies performed in early years indicated the need for double readings with human observers rather than computers to conduct the second reading. Most studies showed that sensitivity improved somewhat as a result of double reading; this was another indication that radiologists can fail to notice important findings on images. In the world of research, CAD slowly was becoming a respectable field of study. Fuji produced a drum-film scanner that a group at the University of Chicago used under the leadership of Kunio Doi, whose findings were published by Ishida et al in 1982. Studying the specifications of devices such as the Fuji drum scanner is of historic interest, given the comparative abundance of technology available today. The Fuji scanner digitized films with a sampling rate of 100 µm, with a pixel depth of 10 bits (1024 gray levels). This scanner was sensitive to optical densities in the range of 0.2–2.75. The host computer used in an early University of Chicago study was a VAX 11/750 computer interfaced with a high-resolution (1280 X 1024) Ramtek 9460 dual-user image processor. Later, this group ported their code to an IBM RS-6000 computer that could be purchased for a modest $100,000 (Fuji scanner not included). The computing power was such that a single mammographic image was analyzed in 10 minutes or longer, which made this approach impractical for use in the average breast imaging facility. Despite the inadequacy of these early attempts, the pioneering groups should be acknowledged for their foresight and tenacity. Although a few early papers suggested a false-negative rate, the clinical community did not agree that lesions were missed on radiographs. The present authors have received several private communications from radiologists indicating that they either never missed lesions or, if they did, that the effect of misses was negligible. Early researchers explored the use of various features to describe and detect lesions. Image processing (IP) was a natural tool for this purpose because it provides a set of techniques that enhances features of interest and de-enhances others with the application of many types of filters. IP also has techniques for quantifying visual features and for providing metrics to measure geometric, topologic, or other characteristics by which images are described. These various techniques became the foundations of CAD technology. Technical aspects Segmentation, which means separating an image into regions of similar attributes, is a typical early-stage operation under active study in IP. A common example of segmentation is the automated detection of the skin line of the breast and the separation of the image into the breast area, the directly exposed (background) area, and perhaps the unexposed (film edge) area. This particular segmentation method limits the area of the mammogram that the computer analyzes to save computation time. It also may avoid the mistake of producing marks on objects that are not in the breast. In the early detection of breast cancer, segmentation is used to determine the boundary of masses, separating the image into regions inside and outside of the mass. This is a particularly challenging type of segmentation because of the proximity of masses to normal parenchymal tissue. Once a lesion is segmented, the computer algorithm uses IP techniques to measure the pertinent features and to describe features such as the border of the lesion. Feature extraction, or the quantification of relevant features, is the portion of the computer code that is most important for good performance. When as many features or descriptors as possible have been computed, a decision has to be made whether the object, with all its features, is suggestive enough to be brought to the radiologist's attention. Artificial intelligence (AI) techniques are used to make these decisions. AI techniques include rule-based codes or expert systems, decision trees, linear or higher-order classifiers, and neural networks. AI and IP techniques have been in development for decades and are used in diverse fields such as astronomy, image recognition for satellite and reconnaissance photos, tracking of objects on battlefields, and (more recently) fingerprint and face recognition. A neural network is the term for a computer code that helps make a decision based on the value of certain features. For example, imagine a code for separating apples from beans on the basis of color, where apples are defined as red and beans are green. The required neural network collects input (eg, the color of the object), and provides output (eg, "apple" if the color is red or "bean" if the color is green). Next, imagine that the color criterion is not foolproof, but rather another feature (eg, volume) is available to distinguish apples from beans, where apples are defined as big and beans are small. In this situation, the neural network collects both color and size as input. If the color is red and the size is big, the neural network selects "apple" as the output; otherwise, it selects "bean." Similarly, in mammography, a particular point on an image must be determined to be either a lesion or normal. Therefore, a large number of features are required for the distinction, possibly more features than a human observer can track. The result is a complicated computer code with an ability to make a decision based on the values of many features. However, the concept of multiple features, though simple, rarely makes the difference between good and poor performances of neural networks. Instead, the quality of decisions primarily depends on the features the programmer chooses. The first step is to concisely describe a lesion or suspected feature in English, and this description is then translated into computer code. For this reason, a radiologist's careful description of the appearance mammographic findings is important for improving algorithms. Although remarkable advances have been made in medical imaging to deliver more sophisticated and complex images from modalities such as CT and MRI, few commercially available methods make it easier for the radiologist to handle the developing information overload. To address this need, a new class of products delivering CAD capabilities has become available. Fortunately, computer and associated technologies have advanced to the point that CAD is no longer a curiosity of the research laboratory. Over the last 10 years, CAD has become clinically feasible. CAD technology In 1993, when the founders of R2 Technology, Inc, were searching for investors, the mammography code developed at the University of Chicago was running on the IBM RS-6000 computer, which cost approximately $100,000. Within 1 year, an equally powerful code had been run on a dual Sparc 20 machine with the aid of a hardware accelerator. At the time, its configuration cost approximately $50,000. Within another year, a more powerful code was running on a much less expensive dual-Pentium personal computer with no hardware accelerator and with faster processing times than that of the original IBM machine. During these few years, processing power increased (as predicted by the Moore law), and memory access times decreased approximately 7-fold. Prices of computers simultaneously decreased to the point to which a personal computer with 512-Mb memory that cost approximately $1500 was not uncommon. During this period of rapid development in computer technology, the number of groups actively pursuing research at universities grew dramatically. Currently, approximately 100 groups are studying CAD. The advance in computer technology, as well as new research in AI and IP, will enable radiologists to use the vastly increasing amount of information most effectively. With these new CAD technologies, the radiologist is no longer limited by image quality or the amount of information, but the inherent boundaries (eg, psychophysical, ergonomic boundaries) in human perceptual processes remain. Computer processing, either preprocessing or concurrent processing, may reduce the vast amount of information to manageable levels, and in this regard CAD may be useful. Mammography provides 1 example of when CAD may be useful because it is a screening modality involving a large volume of images. Projection chest radiography may be another because interpreting such radiographs can also be a high-volume procedure. Yet another example is low-dose multisection CT, which produce several hundred, if not thousand, images. With any modality in which the radiologist confronts a large volume of imaging information under conditions that may challenge normal perceptual abilities, CAD may be a useful adjunct to traditional image reading. Some considerations regarding digital mammography may provide a partial understanding of the benefits gained from digital imaging compared with CAD. Digital detectors are slowly being used in mammography, despite the fact that no clinical evidence demonstrates improved cancer detection with digital mammography. Digital detectors have many physical characteristics superior to those of film-screen methods. These characteristics include high detective quantum efficiency (DQE) and wider latitude; however, these advantages have not translated into better performance on the part of radiologists. Therefore, poor performance might be attributed to the radiologist and not the detecting technology. CAD aims to resolve this concern. In the future, the true benefit of digital mammography may be increased productivity and flexibility. If so, adoption of the technology will probably parallel the development of other enabling infrastructure, such as picture archiving and communication systems (PACS) and Digital Imaging and Communications in Medicine (DICOM) protocols for communication among devices. Such developments will further encourage the use of CAD by making it faster, more convenient, and more flexible. CAD is closely related to another component of the process, ie, the display of information. The limitations of current display technology (eg, resolution, dynamic range, ergonomics) may be the Achilles heel of digital technology. An inferior display can negate the benefit of superior detection because the radiologist must ultimately be able to see the object to make a diagnosis. Brightness, contrast, resolution, and reliability are concerns the display vendors are addressing. Other notable improvement in displays may result from the development of CAD-enabled displays that take the CAD information and properly displays any interesting or suggestive features. CAD is just one tool among many practical clinical applications in screening mammography. CAD is used in conjunction with the radiologist, functioning as a second reader. In this way, CAD indicates findings that have features suggestive of malignancy. These findings may be of concern, but the experienced radiologist makes the final call. Either way, CAD does not guarantee perfect screening performance. Some cancers will remain undetected. CAD is simply another tool that aids the radiologist in his or her interpretation of mammograms. CAD IMPLEMENTATION IN THE COMMUNITY BREAST CENTERSection 4 of 9
CAD unit A CAD unit consists of 3 main parts (see Multimedia 1-3): (1) the scanner, (2) the software, and (3) the viewer (see Multimedia 4-7). The scanner is used to scan and digitize the mammogram, similar to a desktop scanner used to digitally save photographs. Some mammograms already are captured digitally, in which case this step does not apply. The software includes sophisticated computer programs analyze the film or image and prompt the radiologist to review areas that may suggest a lesion. The software function displays the mammograms on viewers, or small, low-resolution monitors. Lesion identification Each vendor has its own way of displaying the CAD-generated information. For example, 1 unit places a tiny asterisk on a possible mass and a tiny solid triangle on a potentially malignant calcium deposit or mass. Although most marks do not represent cancer, they do represent areas that are somewhat important. For instance, they may indicate converging lines and shadows that can be a spiculated mass, densities with irregular margins that can be the start of a subtle cancer, or clusters of bright spots that can be malignant calcium masses. The experienced radiologist can readily determine whether a mark is truly important. Only occasionally does the computer indicate a mass that the radiologist did not previously detect. A review of this area may reveal that the mass is normal or stable or that repeat imaging of the patient might be warranted. Sensitivity and specificity CAD is not 100% sensitive. Occasionally, CAD fails to note an area the radiologist thinks may be a cancer (see Multimedia 7-8). For this reason, CAD is not a diagnostic tool. The erroneous term computer-aided diagnosis has been applied to CAD. In its current form, CAD should never be used for diagnosis but only for detection. However, this does not mean that CAD cannot be used on a diagnostic mammogram. A mammogram is considered diagnostic when it depicts a particular finding (eg, a recall finding from a screening image, a palpable mass). The radiologist does not need the help of CAD to evaluate the area of concern, but the remainder of the bilateral mammogram is effectively a screening study. It is easy to become entirely focused on 1 area of a mammogram and forget the rest of the study or simply give it a cursory review. In summary, if CAD can help with a screening mammogram, it can also help on the screening part of a diagnostic mammogram. Once the questionable spot is detected, the radiologist analyzes, diagnoses, and makes the final determination as to whether an area merits recall for further evaluation. If a density, mass, area of architectural distortion, or calcium is worrisome, the patient should be recalled regardless of whether CAD marked the finding. A finding that merits recall should never be ignored because CAD did not mark it (see Multimedia 7-9). Please do not use CAD for reassurance. CAD marks many areas that the radiologist may dismiss, and some of these findings may later be confirmed to be cancer (see Multimedia 8-11). The computer algorithm may be able to detect findings that the human eye still cannot perceive on the screening mammogram, which is unfortunate; however, this limitation must be accepted for now. Radiologists simply must do their best with the tools available to them. If the human eye cannot see an actionable finding, it does not matter whether CAD can detect it. Technology will improve over time, and CAD algorithms will improve in their ability to detect masses and to prompt radiologists regarding a finding. Years of data collection and experience are necessary to learn about CAD and how it can be developed to best fit the everyday needs of breast radiologists. CLINICAL USE OF CAD: A RETROSPECTIVE STUDYSection 5 of 9
Background of the retrospective study In 2002, Burhenne et al conducted an important retrospective study of mammographically missed cancers and the potential benefits of CAD. This was also one of the best studies to analyze radiologists' false-negative rate in reading mammograms. The investigation clearly demonstrated that radiologists—good radiologists from reputable centers in the United States and Canada—were missing cancers. Parameters of the retrospective study The researchers obtained more than 1000 mammograms that led to an initial diagnosis of breast cancer in a woman whose previous mammograms were normal. The researchers then obtained as many of the most recent normal mammograms they could for retrospective review. Findings of the retrospective study When an expert panel reviews these previous mammograms from several different angles, they found that 67% of the cancers were visible. Although many of these cancers were subtle, at least 27% were both visible and actionable. That is, they should have been noticed and acted upon instead of being called normal. When the researchers processed the mammograms showing the 27% visible and actionable findings through a CAD unit, it marked 77% of these cancers. Conclusion of the retrospective study The results of this study clearly indicated a potential for CAD to help the breast radiology community in detecting breast cancers. CLINICAL USE OF CAD: A PROSPECTIVE STUDYSection 6 of 9
Background of the prospective study According to reports by Beam and Hendrick (1999) and te Brake et al (1998), double reading of mammograms can increase breast cancer detection. Nevertheless, this practice has not gained wide acceptance in the United States, possibly because it is impractical in many breast imaging centers. By the time the FDA approved the first CAD machine for use in the United States in June 1998, retrospective evidence already indicated that CAD may be useful in the early detection of breast cancer. In 2000, Burhenne et al reported that CAD may significantly increase the early detection of breast cancer, possibly by 21%. This figure was higher than what most conventional double-reading studies had shown at that time. Radiologists at the Women's Diagnostic and Breast Health Center (WDC) in Plano, Texas, wanted to incorporate double reading but did not have sufficient radiologists on staff to do this by using traditional methods. Soon after CAD became commercially available, the WDC bought a unit (at approximately $200,000) and set out to verify the clinical usefulness of CAD and the benefit of ongoing costs such as those for upgrades and maintenance contracts. Parameters of the prospective study The authors wanted to know how CAD affected mammographic screening by asking several questions: (1) How does CAD affect the patient recall rate? (2) Does CAD affect decisions to perform biopsy, ie, does the positive predictive value (PPV) for biopsy decrease with the use of CAD? (3) Does CAD detect more occurrences of breast cancer than human reading alone? (4) If CAD helps in detecting more cancers, at which stages are the additional cancers found (as classified after final surgery)? (5) What is the cost per diagnosis without CAD versus the cost per diagnosis with CAD? The authors believed it important to answer these questions prospectively, not retrospectively. Axiomatic to the study was the fact that the recall rate would increase because CAD cannot eliminate patient recalls. Rather, it can only indicate that a recall is necessary by prompting an area of concern that the radiologist did not detect first. Images were read first without CAD. Mammograms were categorized as showing normal findings or as showing abnormal findings and therefore flagged for patient recall. Next, CAD was engaged (see Multimedia 5), and the radiologist reviewed the CAD-prompted images. If CAD marked an area, the radiologist reanalyzed the images to decide whether the patient recall was warranted. The authors kept track of the following data: (1) the nature of the finding (That is, did a mass or calcium deposit prompt patient recall?), (2) the mode of detection (That is, was the finding initially seen by the radiologist alone or by CAD alone, which then prompted the radiologist to recall the patient?), and (3) the final Breast Imaging Reporting and Data System (BIRADS) assessment at the recall evaluation, including any subsequent histopathology and tumor staging information. The CAD unit arrived in November 1998. After a ramp-up period (November 1998 to February 1999) to become familiar with the equipment, the WDC began collecting data on 12,680 consecutive screening mammograms. One of 2 experienced breast radiologists first analyzed the mammograms without the benefit of CAD. Then CAD was engaged, and the radiologists reanalyzed the mammograms after looking at the CAD-prompted images. The present authors believe that CAD always should be used according to this procedure. That is, the radiologist first reads the study without the benefit of CAD and then engages CAD and reviews the prompts. The additional reading time per mammogram is only a few seconds on average, and this method prevents the radiologist from using CAD as a diagnostic tool or for reassurance. The radiologist must make a decision before engaging CAD and use CAD only to initiate, never to cancel, a patient recall. Findings of the prospective study Main outcome measures of the study included patient recall rates, detection rates, PPV, cost, and time. For 12,860 consecutive screening mammograms, the WDC recall rate without CAD would have been 6.5%. With the use of CAD, the recall rate rose to 7.7%, which represented a 20% increase in absolute numbers of patients. Regarding detection rates, had CAD not been used, 41 cancers would have been detected. With CAD, 49 cancers were detected, which represented a 19.5% increase. All cancers detected with CAD were stage 0 or stage 1. Without CAD, the PPV for biopsy would have been 38%. With the additional biopsy procedures performed because of CAD results, the PPV for biopsy remained at 38%. This lack of change (lack of decrease) indicated that the decision to perform a biopsy was based solely on the merits of the lesion, not on whether CAD had prompted recall. If the decision to perform a biopsy had been based on only CAD prompting, the PPV for biopsy would have decreased. The cost per each diagnosis of breast cancer before and after CAD was investigated. Research is ongoing, but preliminary data suggest that the cost per diagnosis decreased with CAD. This appears to be a result of the additional cancers detected. A parameter measured but not included in the original publication by Freer and Ulissey was the extra time it took to read mammograms with CAD. CAD increased reading time by approximately 20%; however, this additional time added an average of only about 12 min/day in the screening room. Both the radiologist and CAD prospectively detected most of the cancers. Although CAD detected 8 cancers that the radiologists did not, the radiologists detected 9 cancers that CAD did not; this observation reinforced the concept that CAD should not be used for diagnosis. Although this comparison is interesting, it has limited clinical relevance. How CAD performs alone is not important because CAD never should be used alone. What matters is how the radiologist's detection rate improves by using CAD. In reviewing how CAD contributed to the early detection of breast cancer at the WDC, an interesting point emerged: Seven of the 8 cancers that CAD identified and the radiologist missed were related to microcalcifications, and all 9 of the cancers the radiologist detected and CAD missed were masses. In addition, CAD prospectively marked all cancers related to microcalcification. This performance may be specific to only the WDC. However, when these results analyzed in conjunction with the false-negative mammograms analyzed with CAD (Burhenne, 2000), the WDC modified its clinical practice. The center now applies double reading (with 2 radiologists) plus CAD, which effectively provides triple reading. The first radiologist focuses on detecting only masses, and the second radiologist uses CAD to detect subtle microcalcifications. To maximize the benefit of CAD, each center that implements CAD should periodically determine how CAD is helping and then adjust its clinical practice accordingly. Ongoing data collection will determine whether the practice is effective. Conclusions of the prospective study With CAD, 20% more women were recalled, 20% more cancers were detected, and 20% more time was spent in the screening room than without CAD. The patient recall rate increased from 6.5% to 7.7%. The PPV for biopsy (38%) did not change, and all additional cancers detected as a result of CAD were stage 0 or stage 1. The average of 12 additional min/day in the screening room to find 20% more breast cancers seems to be a fair trade. CAD may have lowered the cost per diagnosis of each cancer, and it may have lowered the medical-legal exposure of the radiologists by helping to detect the cancers a year or more earlier than they would have been. In addition, the increased detection rate with CAD exceeded rates previously noted in traditional double-read studies. SUMMARYSection 7 of 9
Although current results with CAD are encouraging, the general medical community needs additional information before the true clinical utility of CAD can be determined. First, the reproducibility of the WDC experience must be confirmed at other sites. The WDC is also investigating whether their first-year success will continue in subsequent years. Second, because of the relatively low annual incidence of breast cancer, researchers must analyze hundreds of thousands of mammograms to find enough breast cancers to achieve statistically high confidence intervals. A multicenter investigation would be ideal for this study. In 2001, the WDC had prospectively analyzed more than 30,000 studies with CAD. By now, this number has increased exponentially. Detailed records of the results have been kept, and to continually improve the quality of their clinical work, WDC radiologists periodically review the data. CAD clearly marks most breast cancers and appears to perform as well as the radiologist in its ability to detect breast cancer. Although the mass algorithm seems to have some limitations, it should improve over time, and so far, the newer version of CAD software seems to have an excellent ability to mark malignant calcium in at least 1 view. In the WDC study, CAD marked 40 of 49 cancers, and the radiologist marked 41 of 49 cancers. However, when results of these 2 approaches were combined, 8 more cancers were detected; these would have been missed that year's round of screening. Remember, how CAD operates alone is not important because it never should be used alone. What is important is how much better the radiologist can perform when he or she uses CAD. In the WDC experience, CAD has proven to be a useful, reliable, and cost-effective tool for the early detection of breast malignancies. Should other experiences parallel those at the WDC, CAD may prove to be the single most significant technologic advance in the detection of early stage breast cancer in the last 25 years. For excellent patient education resources, visit eMedicine's Imaging Center, Cancer and Tumors Center, and Women's Health Center. Also, see eMedicine's patient education articles Mammogram, Breast Cancer, Breast Lumps and Pain, and Breast Self-Exam. MULTIMEDIASection 8 of 9
REFERENCESSection 9 of 9
Mammography - Computer-Aided Detection excerpt Article Last Updated: Jan 26, 2005 | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
