Published online before print March 15, 2005, 10.1148/radiol.2352040899
(Radiology 2005;235:385-390.)
© RSNA, 2005
Accuracy of Segmentation of a Commercial Computer-aided Detection System for Mammography1
Jay A. Baker, MD,
Eric L. Rosen, MD,
Michele M. Crockett, MD and
Joseph Y. Lo, PhD
1 From the Department of Radiology, Duke University Medical Center, Box 3800, Erwin Rd, Durham, NC 27710. Received May 19, 2004; revision requested June 30; revision received August 26; accepted October 1. Address correspondence to J.A.B. (e-mail: jay.baker@duke.edu).
 |
ABSTRACT
|
|---|
PURPOSE: To assess the accuracy of segmentation in a commercially available computer-aided detection (CAD) system.
MATERIALS AND METHODS: Approval for this study was obtained from the authors’ institutional review board. Informed consent was not required by the board for this review, as data were stripped of patient identifiers. Two thousand twenty mammograms from 507 women were analyzed with the hardware and software of a commercial CAD system. The accuracy of the segmentation process was determined semiquantitatively and categorized as near perfect if the skin line of the breast was accurately detected, acceptable if only subcutaneous fat was excluded, or unacceptable if any breast parenchyma was excluded from consideration. The accuracy of segmentation was compared for different breast densities and film sizes by using logistic regression (P < .05).
RESULTS: Overall, segmentation was near perfect or acceptable in almost 96.8% of images. However, segmentation defects were significantly more common in mammograms with heterogeneously dense breast tissue (8% unacceptable) than in those with fatty replaced (0% unacceptable), scattered (1.2% unacceptable), or extremely dense (1.8% unacceptable) breast parenchyma (P < .05). For images with unacceptable segmentation, the average percentage of breast parenchyma excluded was almost 25% (range, 5%–100%), with no significant differences among breast densities.
CONCLUSION: For one commercial CAD system, segmentation was usually near perfect or acceptable but was unacceptable more than five times more frequently for mammograms of breasts with heterogeneously dense parenchyma than for those with all other breast densities. On average, one-quarter of the breast parenchyma was excluded from CAD analysis for images with unacceptable segmentation.
© RSNA, 2005
|
INTRODUCTION
|
|---|
Computer-aided detection (CAD) is now used in some clinical practices in the United States to assist radiologists in the interpretation of screening mammograms. Studies have documented that the use of CAD improves sensitivity for cancer detection (1–3). Computer-aided classification systems that differentiate benign from malignant lesions have also been developed, and in research studies, these systems have demonstrated a potential to improve sensitivity and reduce unnecessary biopsies with benign results (4–7). For both CAD and computer-aided classification systems, one of the first critical steps in image analysis—after a digital image is generated—is the separation or segmentation of the entire breast from nonanatomic background regions (8–11). By excluding background regions from further analysis, segmentation improves the accuracy of image analysis and reduces the computation time required.
While segmentation may seem to be of only theoretical importance, this stage can directly influence the success of clinicians’ use of CAD. Subsequent stages of CAD analysis disregard those regions of the image categorized as background during the segmentationstep; therefore, a malignant lesion in a region of the breast mistakenly excluded by the segmentation process will not be identified with the CAD system, regardless of the conspicuity of the lesion. Radiologists essentially receive no second reading benefit from the CAD system for excluded fibroglandular tissue. Therefore, because segmentation accuracy can directly affect the sensitivity of a CAD system, our study was performed to assess the accuracy of segmentation in a commercially available CAD system.
|
MATERIALS AND METHODS
|
|---|
Study Population: Images
Approval for this study was obtained from our institutional review board. The board did not require informed consent for this review, as data were stripped of patient identifiers. Compliance with the rules of the Health Insurance Portability and Accountability Act was maintained.
In July and August 2003, 507 consecutive women who had no breast symptoms underwent routine screening mammography at our facility. Individual patient age was not available because patient identification was removed. Routine bilateral four-view mammographic examinations (ie, craniocaudal and mediolateral oblique views of each breast) were performed in 503 patients, and two-view unilateral examinations were performed in four patients who had undergone mastectomy. In total, 2020 mammographic views were obtained and constitute the study population for this investigation. No cases were excluded from consideration.
Mammographic imaging was performed with the screen-film technique (MinR-2000; Eastman Kodak, Rochester, NY), and mammographic imaging systems (Mammomat 3000; Siemens Medical Systems, Erlangen, Germany) were approved according to the Mammography Quality Standards Act (MQSA). The mammograms employed complied with all of the manufacturer’s recommended quality parameters and those required by the MQSA. Each mammogram was subsequently digitized at 50-µm resolution and 12-bit gray scale and analyzed with the hardware and software of a commercial computer-aided detection system (ImageChecker M1000, version 3.2; R2 Technologies, Sunnyvale, Calif).
Image Evaluation
Two dedicated breast imaging radiologists (J.A.B., E.L.R.) prospectively reviewed each case of two or four mammographic views, with consensus. Each observer was qualified to evaluate mammograms according to MQSA rules and had 6 or 7 years of breast imaging experience. The observers first evaluated each case to determine the overall density of the breast parenchyma. The density was categorized according to the four density categories in the Breast Imaging Reporting and Data System mammography lexicon published by the American College of Radiology (12). These categories were as follows: almost entirely fat (<25% glandular), scattered fibroglandular densities (approximately 25%–50% glandular), heterogeneously dense (approximately 51%–75% glandular), and extremely dense (>75% glandular). After establishing the density of the parenchyma, the observers assessed the accuracy of segmentation of the CAD system. For the CAD system used in our study, the area of the mammographic film determined to represent the breast was brighter on the display than was the area determined to represent background (Fig 1). The detection algorithms of the CAD system are not applied to areas determined to represent background structure, and no CAD marks are placed on these background regions.
View larger version (76K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 1. Bilateral CAD images in a 62-year-old woman demonstrate near-perfect segmentation of breast tissue. The breast outline is accurately highlighted by the CAD system. Left: Craniocaudal CAD image. Right: Medioloateral oblique CAD image.
|
|
The observers assessed each mammogram and the corresponding CAD output to determine whether the area highlighted by the CAD system accurately reflected the area of the breast on the mammogram, with the mammogram as the reference standard. Each segmentation outcome was assigned one of three accuracy ratings, which were similar to those of grading systems used by other authors in prior studies (10,13): near perfect, acceptable, or unacceptable segmentation.
Near-perfect segmentation indicates that the segmentation algorithm accurately identified the contour of the skin and all breast parenchyma, and virtually all subcutaneous fat of the breast was highlighted by the CAD segmentation algorithm (Fig 1). Acceptable segmentation indicates that although 100% of the breast parenchyma was included in the segmented image, less than 100% of the nonparenchymal adipose tissue was included by the segmentation process. The excluded regions included subcutaneous adipose tissue, a wedge of adipose tissue superior to the breast parenchyma on the mediolateral oblique image, or adipose tissue medial or lateral to the breast parenchyma on the craniocaudal image (Fig 2). While it is not ideal to exclude this tissue, breast cancers rarely occur in the subcutaneous fat, and its exclusion from CAD consideration is unlikely to affect the sensitivity of the system. Segmentation was deemed unacceptable if more than 5% of the actual breast parenchyma was excluded by the segmentation algorithm.
View larger version (81K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2. Bilateral CAD images in a 47-year-old woman with acceptable segmentation of breast parenchyma. Left: Craniocaudal CAD image shows wedge of adipose tissue (thick arrow) that was excluded from the lateral aspect of the right breast. Right craniocaudal view shows location of a possible mass (*), which was marked by the CAD system. Right: Mediolateral oblique CAD image shows wedge of adipose tissue (thin arrows) that was excluded from the superior aspect of both breasts. All breast parenchyma was included in highlighted tissue.
|
|
The observers determined the area of excluded breast tissue semiquantitatively with the use of templates with decile and quartile markings. These templates were overlays in the shape of mammographic images with lines indicating the breast divided into tenths (deciles) and quarters (Fig 3). The overlays were compared with the actual CAD images to determine the area of parenchyma excluded in the segmentation process.
Statistical Analysis
Segmentation accuracy was determined for all 2020 images. Breast parenchyma density and mammographic film size—large (24 x 30 cm) or small (18 x 24 cm)—were also recorded. The accuracy of segmentation was compared among breast density categories and between film sizes by means of logistic regression analyses. Differences among groups were determined from the logistic model (SAS Software System, version 8.2, released 2001; SAS, Cary, NC) in consultation with a statistician, with a significance level of P < .05.
The percentage of breast parenchyma excluded in the segmentation process was recorded for all images categorized as having unacceptable segmentation. These percentages were compared between breast tissue density categories by means of a t test, with a significance level of P < .05.
|
RESULTS
|
|---|
Breast Imaging Reporting and Data System Category and Film Size
The Breast Imaging Reporting and Data System category for breast tissue density was fatty replaced for 77 patients (302 [15.0%] of 2020 images), scattered fibroglandular for 275 patients (1098 [54.4%] of 2020 images), heterogeneously dense for 134 patients (536 [26.5%] of 2020 images), and extremely dense for 21 patients (84 [4.2%] of 2020 images). Small mammography film (18 x 24 cm) was used for 1418 (70.2%) of the 2020 images, and large film (24 x 30 cm) for the remaining 602 (29.8%) images.
Segmentation Accuracy
Overall, segmentation was near perfect or acceptable for 1956 (96.8%) of the 2020 images analyzed and unacceptable for the remaining 64 (3.2%) mammograms. For mammograms categorized as fatty replaced, segmentation was near perfect or acceptable for all mammographic images analyzed. For mammograms with scattered fibroglandular densities, segmentation was near perfect or acceptable in 98.2% (1078 of 1098) of images and unacceptable in 1.8% (20 of 1098). For those with extremely dense breast parenchyma, segmentation was near perfect or acceptable in 98.8% (83 of 84) of images and unacceptable in 1.2% (one of 84). In contrast, only 92.0% (493 of 536) of images categorized as heterogeneously dense demonstrated near-perfect or acceptable segmentation, and 8.0% (43 of 536) were categorized as unacceptable. This level of unacceptable segmentation for heterogeneously dense breasts (8%) is over five times the 1.4% unacceptable rate (21 of 1484 images) for the other three categories of breast density combined. With the logistic regression model, mammograms of heterogeneously dense breasts were significantly more likely to result in unacceptable segmentation than were mammograms of breasts with other density categories (P < .001).
The area of breast parenchyma excluded for images categorized with unacceptable segmentation ranged from 5% to almost 100% (Figs 4–7). Overall, the area of excluded tissue in these images averaged 24.7% ± 22.2 (standard deviation). For mammograms with scattered breast density and unacceptable segmentation, the area of tissue excluded averaged 24.3% ± 20.4, compared with 25.0% ± 23.5 for mammograms with heterogeneously dense breasts. For the single image with extremely dense breast parenchyma and unacceptable segmentation, 20% of breast parenchyma was excluded in the segmentation process. Analysis with a t test showed no significant differences in the areas of excluded tissue according to breast parenchyma density (scattered vs heterogeneously dense) (t = 0.12, P > .90).
View larger version (91K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 4. Bilateral CAD images in a 54-year-old woman demonstrate unacceptable segmentation of breast parenchyma. Left: Craniocaudal CAD image. CAD segmentation excluded 50% of breast parenchyma (arrows). Right: Mediolateral oblique CAD image demonstrates near-perfect segmentation.
|
|
View larger version (129K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 5. Bilateral craniocaudal CAD images in a 44-year-old woman demonstrate unacceptable segmentation of breast parenchyma. Left: CAD segmentation excluded medial half of breast parenchyma (arrowhead) on right view. Right: CAD segmentation excluded 75% of breast parenchyma (arrows) on left view.
|
|
View larger version (93K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 6. Bilateral CAD images in a 71-year-old woman demonstrate unacceptable segmentation of breast parenchyma. Left: Craniocaudal CAD image. The medial 25% of breast parenchyma (arrows) on bilateral views was excluded. Right: Mediolateral oblique CAD image. CAD segmentation excluded almost all breast parenchyma (arrowheads) on the left view, as shown by the interface of highlighted tissue and darker excluded regions.
|
|
View larger version (80K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 7a. Images in a 59-year-old woman with invasive ductal carcinoma of the right breast. (a) Mediolateral oblique mammogram of the right breast shows subtle architectural distortion (arrow) in the superior aspect of the breast. (b) Radiograph of lumpectomy tissue demonstrates distortion that represents carcinoma surrounding barb of localization wire and biopsy marking clip. (c) Right mediolateral oblique CAD image demonstrates unacceptable segmentation (arrowheads) and biopsy-proved cancer missed by the CAD system. (d) Repeated right mediolateral oblique CAD image demonstrates acceptable segmentation. Site of malignancy is correctly identified by the CAD system, as indicated (*).
|
|
View larger version (125K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 7b. Images in a 59-year-old woman with invasive ductal carcinoma of the right breast. (a) Mediolateral oblique mammogram of the right breast shows subtle architectural distortion (arrow) in the superior aspect of the breast. (b) Radiograph of lumpectomy tissue demonstrates distortion that represents carcinoma surrounding barb of localization wire and biopsy marking clip. (c) Right mediolateral oblique CAD image demonstrates unacceptable segmentation (arrowheads) and biopsy-proved cancer missed by the CAD system. (d) Repeated right mediolateral oblique CAD image demonstrates acceptable segmentation. Site of malignancy is correctly identified by the CAD system, as indicated (*).
|
|
View larger version (95K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 7c. Images in a 59-year-old woman with invasive ductal carcinoma of the right breast. (a) Mediolateral oblique mammogram of the right breast shows subtle architectural distortion (arrow) in the superior aspect of the breast. (b) Radiograph of lumpectomy tissue demonstrates distortion that represents carcinoma surrounding barb of localization wire and biopsy marking clip. (c) Right mediolateral oblique CAD image demonstrates unacceptable segmentation (arrowheads) and biopsy-proved cancer missed by the CAD system. (d) Repeated right mediolateral oblique CAD image demonstrates acceptable segmentation. Site of malignancy is correctly identified by the CAD system, as indicated (*).
|
|
View larger version (78K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 7d. Images in a 59-year-old woman with invasive ductal carcinoma of the right breast. (a) Mediolateral oblique mammogram of the right breast shows subtle architectural distortion (arrow) in the superior aspect of the breast. (b) Radiograph of lumpectomy tissue demonstrates distortion that represents carcinoma surrounding barb of localization wire and biopsy marking clip. (c) Right mediolateral oblique CAD image demonstrates unacceptable segmentation (arrowheads) and biopsy-proved cancer missed by the CAD system. (d) Repeated right mediolateral oblique CAD image demonstrates acceptable segmentation. Site of malignancy is correctly identified by the CAD system, as indicated (*).
|
|
Segmentation was near perfect or acceptable for 96.6% (1370 of 1418) of mammograms obtained with small mammographic film and 97.3% (586 of 602) for those obtained with large film. The logistic regression model analysis results revealed no significant differences in segmentation accuracy according to size of film (P > .3).
|
DISCUSSION
|
|---|
Segmentation of mammograms into breast and background regions has been accomplished with a number of techniques (10,11), including global histogram analysis (14,15), analysis of local gray-value range (10), and gradient analysis (9,16). The accuracy of the segmentation process has been previously evaluated in several studies about nonclinical research protocols. Segmentation was reported to be near perfect or acceptable in 89% (9), 95% (11), 97% (10), and 98% (13) of mammographic images. These values are similar to our overall value of 96.8% acceptable or near perfect for what we believe is the first study of a CAD system in routine clinical use. To our knowledge, the role of breast tissue density has not previously been considered in evaluations of the accuracy of the segmentation process.
Segmentation of the breast from the surrounding background offers two primary benefits in CAD systems. The first benefit is an overall improvement in the accuracy of CAD analysis (11). Identification of breast regions allows subsequent detection algorithms and false-positive reduction techniques to be applied only to breast tissue and not to background regions. If the background region is also analyzed, numerous false-positive regions—such as left and right markers, the patient-identifying nameplate, and imaging artifacts in the background region—may be incorrectly identified as potential lesions. Because segmentation reduces the number of false-positive sites identified by the CAD system, the overall specificity of the system is improved. The segmentation process can also indirectly improve the sensitivity of some CAD systems. For those that are programmed to limit the total number of sites identified per image, true-positive lesions may be below the detection threshold due to large numbers of false-positive marks (17). In this setting, actual breast cancers may be detected but not marked on the CAD output due to the number of more conspicuous false-positive sites.
The second important benefit of segmentation is to reduce the overall area of the image that requires further analysis with the CAD algorithms (11). Digitized mammograms provide a large data set; the reduction of computing effort because of image segmentation allows reduction in the time necessary for CAD analysis, which can require several minutes per case. The exact time savings is not known for any of the commercial systems because of their proprietary nature. Similar benefits are achieved when a segmentation step is included in computer-aided classification systems.
Beyond the engineering benefits of reduced false-positive marks and improved computation time, breast segmentation should be recognized by practicing radiologists as directly affecting their clinical use of a CAD system. Because CAD systems disregard regions of the image excluded in the segmentation process, breast tissue erroneously excluded by this step is not evaluated by the CAD detection algorithm. Therefore, regardless of lesion conspicuity, breast cancers cannot be marked in regions of the breast excluded from CAD evaluation by faulty segmentation. These excluded regions of breast parenchyma do not benefit from the potential advantage of double reading by a human and by a computer system.
In our study, we evaluated the overall accuracy of breast segmentation for one CAD system and determined that, although segmentation is sufficiently accurate for most examinations, the accuracy of the segmentation process is significantly dependent on the density of the breast parenchyma. The specific causes of faulty segmentation are unknown because CAD algorithms in commercial systems remain proprietary. Findings in prior studies have suggested digitization artifacts and poor mammographic technique as potential causes (10). The mammograms in our study that were most likely to demonstrate faulty segmentation were those with heterogeneous tissue patterns. With more complex parenchymal patterns, the segmentation algorithm may follow interfaces across the breast between dense breast parenchyma and interparenchymal adipose tissue, which may mimic the interface between dense parenchyma and subcutaneous fat.
Although our study demonstrates the accuracy of the segmentation process and potential areas of faulty segmentation, its primary limitation is that it does not directly measure differences in sensitivity or specificity caused by faulty segmentation. Although we evaluated more than 2000 mammographic images, to demonstrate a significant difference in CAD performance caused by differences in segmentation accuracy would require a prohibitively large study, because of the relatively low incidence of breast cancer in the screening population. In addition, in our study, we did not assess the reproducibility or the exact cause of failures in the segmentation process. Our findings, however, alert practicing radiologists to the issue of segmentation and demonstrate the dependence of segmentation on breast tissue density. Further, our study demonstrates that a substantial portion of the breast tissue is often excluded when the segmentation process is inaccurate. In images with faulty segmentation, an average of one-quarter of the breast parenchyma was excluded from further CAD analysis. Given these results, physicians who use CAD systems to assist in the interpretation of mammograms should be aware that they do not provide a double reading for breast parenchyma excluded in the segmentation process, so particular attention should perhaps be paid to these regions.
Although more advanced versions of the commercial CAD system used in our study are now available, the results presented here are relevant because the more innovative systems are not yet in widespread use. Further, the specific results of our study both illustrate the effect of the segmentation technique on CAD analysis and demonstrate the importance of attention to segmentation accuracy during clinical implementation of CAD for all purposes, including interpretation of mammographic, chest computed tomographic (CT), or CT colonographic images. In the future, the segmentation process for breast CAD may improve as images are more commonly acquired directly in full-field digital mammographic format without the need for film digitization. The lack of added noise and added artifacts from digitization and the increased dynamic range of direct full-field digital mammography could theoretically allow more accurate estimation of the breast boundary.
Segmentation is a critical initial step for virtually all CAD systems, including those designed to detect breast cancer on mammograms, lung nodules on chest radiographs (18) or CT scans (8,19), polyps on CT colonographic scans (20), and cerebrovascular accidents on brain CT scans (21). In all such circumstances, the interpreting radiologist must recognize the contribution of the segmentation process to the accuracy of CAD analysis. Incomplete or inaccurate segmentation in any CAD system may contribute to failure to detect a lesion. Developers of CAD systems must continue to improve the accuracy of the segmentation procedure, and practicing radiologists must recognize that the CAD systems on which they rely will fail to detect important lesions in tissue excluded during the segmentation step.
|
FOOTNOTES
|
|---|
Abbreviations: CAD = computer-aided detection,
MQSA = Mammography Quality Standards Act of 1992
Authors stated no financial relationship to disclose.
Author contributions: Guarantor of integrity of entire study, J.A.B.; study concepts, J.A.B., J.Y.L.; study design, J.A.B., E.L.R., M.M.C.; literature research, J.A.B., J.Y.L.; experimental studies, J.A.B., E.L.R., M.M.C.; data acquisition, J.A.B., E.L.R., M.M.C.; data analysis/interpretation, J.A.B., E.L.R., J.Y.L.; statistical analysis, J.A.B., J.Y.L.; manuscript preparation, all authors; manuscript definition of intellectual content and editing, J.A.B.; manuscript revision/review, J.A.B., E.L.R., M.M.C.; manuscript final version approval, all authors
|
REFERENCES
|
|---|
- Freer TW, Ulissey MJ. Screening mammography with computer-aided detection: prospective study of 12,860 patients in a community breast center. Radiology 2001; 220:781-786.[Abstract/Free Full Text]
- Warren Burhenne LJ, Wood SA, D’Orsi CJ, et al. Potential contribution of computer-aided detection to the sensitivity of screening mammography. Radiology 2000; 215:554-562.[Abstract/Free Full Text]
- Brem RF, Baum J, Lechner M, et al. Improvement in sensitivity of screening mammography with computer-aided detection: a multiinstitutional trial. AJR Am J Roentgenol 2003; 181:687-693.[Abstract/Free Full Text]
- Leichter I, Lederman R, Bamberger P, Novak B, Fields S, Buchbinder SS. The use of an interactive software program for quantitative characterization of microcalcifications on digitized film-screen mammograms. Invest Radiol 1999; 34:394-400.[CrossRef][Medline]
- Chan HP, Sahiner B, Helvie MA, et al. Improvement of radiologists’ characterization of mammographic masses by using computer-aided diagnosis: an ROC study. Radiology 1999; 212:817-827.[Abstract/Free Full Text]
- Kouskos E, Markopoulos C, Revenas K, Koufopoulos K, Kyriakou V, Gogas J. Computer-aided preoperative diagnosis of microcalcifications on mammograms. Acta Radiol 2003; 44:43-46.[CrossRef][Medline]
- Huo Z, Giger ML, Vyborny CJ, Metz CE. Breast cancer: effectiveness of computer-aided diagnosis—observer study with independent database of mammograms. Radiology 2002; 224:560-568.[Abstract/Free Full Text]
- Armato SG, Altman MB, La Riviere PJ. Automated detection of lung nodules in CT scans: effect of image reconstruction algorithm. Med Phys 2003; 30:461-472.[CrossRef][Medline]
- Mendez AJ, Tahoces PG, Lado MJ, Souto M, Correra JL, Vidal JJ. Automatic detection of breast border and nipple in digital mammograms. Comput Methods Programs Biomed 1996; 49:253-262.[CrossRef][Medline]
- Bick U, Giger ML, Schmidt RA, Nishikawa RM, Wolverton DE, Doi K. Automated segmentation of digitized mammograms. Acad Radiol 1995; 2:1-9.[CrossRef][Medline]
- Ojala T, Nappi J, Nevalainen O. Accurate segmentation of the breast region from digitized mammograms. Comput Med Imaging Graph 2001; 25:47-59.[CrossRef][Medline]
- American College of Radiology. ACR BI-RADS—mammography In: ACR breast imaging reporting and data system, breast imaging atlas. 4th ed. Reston, Va: American College of Radiology, 2003.
- Abdel-Mottaleb M, Carman CS, Hill CR, Vafai S. Locating the boundary between the breast skin edge and the background in digitized mammograms. In: Doi K, Giger ML, eds. Digital mammography ‘96. Amsterdam, the Netherlands: Elsevier, 1996; 467-470.
- Yin FF, Giger ML, Doi K, Metz CE, Vyborny CJ, Schmidt RA. Computerized detection of masses in digital mammograms: analysis of bilateral subtraction images. Med Phys 1991; 18:955-963.[CrossRef][Medline]
- Lau TK. Bischof WF. Automated detection of breast tumors using the asymmetry approach. Comput Biomed Res 1991; 24:273-295.
- Karssemeijer N. Adaptive noise equalization and recognition of microcalcification clusters in mammograms. Int J Pattern Recognit Artif Intell 1993; 7:1357-1376.[CrossRef]
- Zheng B, Leader JK, Abrams G, et al. Computer-aided detection schemes: the effect of limiting the number of cued regions in each case. AJR Am J Roentgenol 2004; 182:579-583.[Abstract/Free Full Text]
- Li Q, Katsuragawa S, Doi K. Computer-aided diagnostic scheme for lung nodule detection in digital chest radiographs by use of a multiple-template matching technique. Med Phys 2001; 28:2070-2076.[CrossRef][Medline]
- Gurcan MN, Sahiner B, Petrick N, et al. Lung nodule detection on thoracic computed tomography images: preliminary evaluation of a computer-aided diagnosis system. Med Phys 2002; 29:2552-2558.[CrossRef][Medline]
- Nappi J, Dachman AH, MacEneaney P, Yoshida H. Automated knowledge-guided segmentation of colonic walls for computerized detection of polyps in CT colonography. J Comput Assist Tomogr 2002; 26:493-504.[CrossRef][Medline]
- Maldjian JA, Chalela J, Kasner SE, Liebeskind D, Detre JA. Automated CT segmentation and analysis for acute middle cerebral artery stroke. AJNR Am J Neuroradiol 2001; 22:1050-1055.[Abstract/Free Full Text]
TOP
ABSTRACT
INTRODUCTION
