Features of Prospectively Overlooked Computer-Aided Detection Marks on Prior Screening Digital Mammograms in Women With Breast Cancer

Transcription

1 Women s Imaging Original Research Women s Imaging Original Research WOMEN S IMAGING Nariya Cho 1 Seung Ja Kim Hye Young Choi Chae Yeon Lyou Woo Kyung Moon Cho N, Kim SJ, Choi HY, Lyou CY, Moon WK Keywords: breast cancer, computer-aided detection, digital images, full-field digital mammography, screening mammography DOI: /AJR Received February 23, 2010; accepted after revision April 12, This study was supported by the National R&D Program for Cancer Control, Ministry of Health and Welfare (grant A01185) and by the Innovative Research Institute for Cell Therapy (grant A062260), Republic of Korea. 1 All authors: Department of Radiology and Clinical Research Institute, Seoul National University Hospital, and Institute of Radiation Medicine, Seoul National University Medical Research Center, 101 Daehangno, Jongno-gu, Seoul , Republic of Korea. Address correspondence to W. K. Moon (moonwk@radcom.snu.ac.kr). AJR 2010; 195: X/10/ American Roentgen Ray Society Features of Prospectively Overlooked Computer-Aided Detection Marks on Prior Screening Digital Mammograms in Women With Breast Cancer OBJECTIVE. The purpose of this article is to describe the features of prospectively overlooked computer-aided detection (CAD) marks on prior screening digital mammograms for women with breast cancer. SUBJECTS AND METHODS. A CAD system embedded in a digital mammography system was prospectively applied to 50,100 screening mammograms between December 2003 and December Each mammogram was originally interpreted by one of five radiologists using the CAD information. Seventy-five mammogram pairs of prior negative screening mammograms and subsequent mammograms of developed cancers were collected. Visible findings and their actionability were determined by three blinded radiologists. All CAD marks, both true-positive and false-positive, and the number of marked views for the visible findings on prior mammograms were analyzed. RESULTS. Of the 75 areas where cancer later developed, 61% (46/75) of mammograms had visible findings (21 masses, 17 microcalcifications, and eight masses with microcalcifications). Of these visible findings, 46% (21/46) were determined to be actionable, and 54% (25/46) were underthreshold. The CAD system had correctly depicted 74% (34/46) of the visible findings 52% (11/21) of masses, 94% (16/17) of microcalcifications, and 88% (7/8) of masses with microcalcifications. Actionable findings showed higher CAD sensitivity than did underthreshold findings (90% [19/21] vs 60% [15/25]; p = 0.04) and were more often marked on both views (58% [11/19] vs 27% [4/15]; p = 0.09). The average number of false-positive marks per case was CONCLUSION. On prior screening digital mammograms, the CAD system had correctly marked 74% (34/46) of visible findings and 90% (19/21) of actionable findings. The actionable findings showed significantly higher CAD sensitivity and were marked on both mammographic views more often than the underthreshold findings were. C omputer-aided detection (CAD) has the potential to decrease the false-negative rate and increase the cancer detection rate in the interpretation of screening mammograms. In earlier retrospective review studies of problematic mammograms, CAD was shown to have correctly marked the missed cancers in 65 77% of prior mammograms [1 4]. In addition, several prospective clinical studies have shown increased cancer detection rates of %, with increased recall rates of % [5 9]. In one study comparing a radiologist s performance before and after the implementation of CAD in clinical practice, CAD was shown to provide increased cancer detection rates of 16.22%, with a 7.79% increased recall rate [10]. However, there are as many differing results on whether the CAD system is truly beneficial in the real clinical setting. One large academic institutional study found that there were no significant differences in the cancer detection rate or recall rate before and after CAD implementation [11]. Moreover, a recent large community-based study found that the use of CAD may actually be harmful because it was associated with decreased specificity ( %), lower positive predictive value ( %), and increased recall and biopsy rates without significantly improved sensitivity ( %; p = 0.32) [12]. However, limited radiologist experience and ignoring of the expected learning curve with CAD were suggested as possible explanations for the low performances [13]. Indeed, in another study 1276 AJR:195, November 2010

2 [14] comparing double and single reading by experienced radiologists with CAD, increased sensitivity was reported ( %). Thus, to realize the potential benefits of the CAD system in clinical practice, the features of CAD marks overlooked by radiologists must be described so that potentially actionable CAD marks can be distinguished from many of the false-positive results identified and acted on prospectively. For the past several years, our institution has used a CAD-embedded digital mammography system for screening examinations, and we have found and herein describe prior mammograms in which CAD had correctly marked a lesion that was missed by the radiologist. To the best of our knowledge, there have been no reports regarding the description of these prospectively applied missed CAD marks. Thus, the purpose of our study was to describe the features of prospectively overlooked CAD marks on prior screening digital mammograms in women with breast cancer. Subjects and Methods Case Selection Institutional review board approval was obtained for this study, and informed consent was not required. Between December 2003 and December 2006, a commercially available CAD system (ImageChecker M1000-DM, version 3.1; R 2 Technology) developed for full-field digital mammography was prospectively applied to 50,100 consecutive mammographic examinations in a screening group of patients. Among them, 230 consecutive patients with biopsy-proven breast cancers were identified and treated at our institution. These patients with prospectively applied CAD information were assessed to determine whether their prior screening mammograms were initially interpreted as negative findings (BI-RADS category 1 or 2) and were retrievable on the PACS. When more than one prior mammogram was available, the most recently obtained mammogram was used. A total of 75 cases of breast cancer had available prior screening digital mammograms that had been interpreted as BI-RADS category 1 or 2 with retrievable CAD information, and those patients comprised our study population. Of the 75 patients with breast cancer (mean age, 52 years; range years), seven (9.3%) had homogeneously fatty breasts (BI-RADS 1 density), 12 (16.0%) had scattered fibroglandular tissues in fatty breasts (BI-RADS 2 density), 40 (53.3%) had heterogeneously dense breasts (BI-RADS 3 density), and 16 (21.3%) had extremely dense breasts (BI-RADS 4 density). The most common histopathologic findings among these cases were invasive ductal carcinoma (71% [53/75]), followed by ductal carcinoma in situ (DCIS; 27% [20/75]), and invasive lobular carcinoma (3% [2/75]). The mean histologic size of the tumors was 1.5 cm (range, cm) for the invasive ductal carcinoma component and 2.8 cm ( cm) for the DCIS component. All prior negative screening mammograms were obtained 5 34 months (median, 12 months) before the mammograms that led to the diagnosis of cancer. Sixty-three percent (47/75) of cancers were detected during subsequent screening and 37% (28/75) were detected before the next scheduled screening round mainly because of palpability (i.e., interval cancers). All cancers were surgically excised at our institution. Mammographic Findings and CAD Application All mammograms were obtained using a CAD-embedded digital mammography system (Senographe 2000D, GE Healthcare). Each patient underwent a standard two-view examination (craniocaudal and mediolateral oblique mammography) of each breast performed by one of two experienced radiologic technologists who had 5 and 10 years of experience, respectively. Then, one of five board-certified radiologists who had 3 6 years of experience in breast imaging interpreted the mammograms after reviewing the CAD marks in a review workstation. Predefined hanging protocols for the mammograms were provided for the radiologist. The automated detection and display process took one second in the workstation. An asterisk marked by the CAD program indicated a pattern suggestive of a mass or an area of architectural distortion, and a solid triangle indicated an area of clustered bright spots suggestive of microcalcifications. A final assessment category was assigned for each examination according to the American College of Radiology BI-RADS. Screencapture images with CAD marks were saved and transferred to a PACS. These images were used later for CAD mark analysis. Liquid crystal display monitors were calibrated according to the manufacturers specifications for mammograms. Mammograms that led to a diagnosis of cancer were referred to as current mammograms, and those mammograms preceding the current ones were referred to as prior mammograms, as in a previous similar study [1]. Current mammograms showing cancers were assessed by a radiologist who marked the location of the histologically proven cancers by superimposing an annotation on the image. Prior mammograms were compared with the corresponding current mammograms with the provided annotations. Visibility of the subsequently diagnosed cancer on prior mammograms was determined by the same radiologist. Of the 75 prior mammograms, 61% (46/75) were determined to have visible lesions. These 46 digital mammograms with visible findings were saved as DICOM files for later review. To determine the actionability of the 46 visible findings on prior mammograms, 46 DICOM files of visible prior mammograms were randomly ordered and mixed with 100 DICOM files of normal mammograms assessed as BI-RADS category 1 or 2. The 100 normal mammograms were obtained between November 1, 2003, and November 15, 2003, and 2-year follow-ups revealed no mammographic or clinical abnormalities. A total of 146 mammograms were reviewed independently by three radiologists (with 5, 7, and 10 years, respectively, of experience in breast imaging), without information regarding current mammograms, histopathologic results, and CAD marks, using a 5-megapixel (2,560 2,048) liquid crystal display system (ME511 L; Totoku Electric). The radiologists assessed the BI-RADS assessment category, location, and type of the finding. If the area of the visible finding on prior mammogram was correctly detected and categorized as 0, 4, or 5 by two or more radiologists, the finding was determined to be an actionable finding. If the visible finding on prior mammograms was detected by no or only one radiologist, the finding was determined to be an underthreshold finding. CAD Mark Evaluation Two other radiologists who had 5 and 3 years of experience in breast imaging determined in consensus whether the CAD marks prompted the correct location of the lesion that later developed into cancer on the basis of all available information of magnification mammography, ultrasound, MR images, and surgical histologic analysis reports of the biopsy-proven cancers. A true-positive mark was defined as when the asterisk was placed on a location of a known malignant mass and when a triangle was placed on a corresponding location of known malignant calcifications on at least one view. Lesion mark types (i.e., mass or microcalcification) also had to agree with the mammographic characteristics (i.e., mass or microcalcification) of the known malignant lesion. If both mass and microcalcification features were noted for the lesion, then either CAD mark type was considered correct. All CAD marks that did not mark the known malignancy or did not agree with the lesion characteristics were defined as false-positive marks. In addition, the number of visible views and the number of views with CAD marks were analyzed. When a finding was visible only on either mediolateral oblique or craniocaudal view, the AJR:195, November

3 A C Fig year-old woman with actionable finding and both-view marks on prior mammograms. A, Screening digital mammogram with mediolateral oblique (left) and craniocaudal (right) views was initially reported as negative with computer-aided detection (CAD) information. Focal asymmetry (arrows) in left lower inner quadrant was determined to be actionable because all three readers assessed it as recallable. B, Screen-capture images of CAD display in mediolateral oblique (left) and craniocaudal (right) views show mass marks (asterisks) in both views. C, Screening digital mammogram of mediolateral oblique (left) view obtained 16 months later shows ill-defined mass (arrow). Maximum intensity projection image (right) of contrast-enhanced MR image of left breast shows irregular enhancing mass (arrow). Surgical histologic analysis revealed 1.5-cm ductal carcinoma in situ. B number of visible views was one. When a finding was visible on both the mediolateral oblique and craniocaudal views, the number of visible views was two. When a CAD mark having the same radial distance from the nipple on both mediolateral oblique and craniocaudal views was prompted on the finding with similar appearance on both views, it was defined to be a both-view mark (Fig. 1). When a CAD mark was prompted on only a mediolateral oblique or craniocaudal view, the number of views with CAD marks was one AJR:195, November 2010

4 A C Statistical Analysis In the prior digital mammograms with visible findings, total mass mark number, microcalcification mark number, number of true-positive marks, and number of false-positive marks were calculated. True-positive and false-positive marks were also assessed for masses and microcalcifications, respectively. The average number of true-positive and false-positive marks per patient was calculated according to the correct marking of at least one true-positive lesion in either view (craniocaudal, mediolateral oblique, or both) and were presented with range and median value. The number of views with CAD marks was compared between actionable and underthreshold findings using the chi-square test. All statistical analyses were performed using SPSS for Windows (version 12.0, SPSS). A p value of less Fig year-old woman with underthreshold finding and one-view mark on prior mammograms. A, Screening digital mammogram with mediolateral oblique (left) and craniocaudal (right) views was initially reported as negative with computer-aided detection (CAD) information. Focal asymmetry (arrow) in left upper quadrant was determined to be underthreshold because all three readers assessed it as negative. B, Screen-capture images of CAD display in mediolateral oblique (left) and craniocaudal (right) views show mass mark (asterisks) in craniocaudal view only. Mediolateral oblique view showed no mark in subtle asymmetry in left upper area. One calcification mark (triangles) in periareolar area was false-positive. C, Screening digital mammograms obtained 13 months later show ill-defined mass (arrows) on both mediolateral oblique (left) and craniocaudal (right) views. Surgical histologic analysis revealed 1.0-cm invasive ductal carcinoma. than 0.05 was considered to indicate statistical significance. Results CAD Marks and False-Positive Marks A total of 130 marks were placed by the CAD system, with 60 mass and 70 microcalcification marks. Of the 60 mass marks, 23 were true-positives and 37 were false-posi- B AJR:195, November

5 TABLE 1: Mass and Microcalcification Detection Performance of the Computer-Aided Detection System Based on Actionability Findings tives. Of the 70 microcalcification marks, 33 were true-positives and 37 were false-positives. The false-positive mark rate per patient was 1.61 (range, 0 8 marks; median, 0 mark) with 0.80 mass mark (range, 0 3 marks; median, 0.5 mark) and 0.80 microcalcification mark (range, 0 8 marks; median, 0 mark). Sensitivity of Masses (n = 21) Sensitivity of Microcalcifications (n = 17) Craniocaudal Mediolateral Oblique Either Craniocaudal Mediolateral Oblique Either Sensitivity of Masses With Microcalcifications (n = 8) Craniocaudal Mediolateral Oblique Actionable findings (n = 21) 60 (3/5) 40 (2/5) 60 (3/5) 91 (10/11) 82 (9/11) 100 (11/11) 80 (4/5) 60 (3/5) 100 (5/5) 90 (19/21) Underthreshold findings (n = 25) 25 (4/16) 31 (5/16) 50 (8/16) 83 (5/6) 33 (2/6) 83 (5/6) 33 (1/3) 33 (1/3) 67 (2/3) 60 (15/25) Total (n = 46) 33 (7/21) 33 (7/21) 52 (11/21) 88 (15/17) 65 (11/17) 94 (16/17) 63 (5/8) 50 (4/8) 88 (7/8) 74 (34/46) Note Data are percentage (no. of findings/total). TABLE 2: Sensitivity of the Computer-Aided Detection (CAD) System According to the Histopathologic and Mammographic Appearances of Visible Lesions on Prior Mammograms in Women With Breast Cancer Histopathologic Appearance of Tumor at Time of Diagnosis Mass (n = 21) Microcalcification (n = 17) Mass With Microcalcification (n = 8) Total (n = 46) Invasive ductal carcinoma (mean, 1.5 cm; range, cm) a Total no. of lesions (% of 31 patients) 16 (52) 8 (26) 7 (23) 31 CAD marked (no. of lesions) CAD sensitivity (%) Invasive lobular carcinoma (1.7 cm) a Total no. of lesions (% of 1 patient) 1 (100) CAD marked (no. of lesions) CAD sensitivity (%) 0 NA NA 0 Ductal carcinoma in situ (mean, 2.8 cm, range cm) a Total no. of lesions (% of 14 patients) 4 (29) 9 (64) 1 (7) 14 CAD marked (no. of lesions) CAD sensitivity (%) Note Percentages have been rounded. NA = not applicable. a The histologic lesion size is defined as the greatest diameter of the lesion on surgical histologic analysis. TABLE 3: Comparison of Number of s With Computer-Aided Detection (CAD) Marks Between Actionable Lesions and Underthreshold Lesions No. of s With CAD Marks Actionable Lesions (n = 21) Underthreshold Lesions (n = 25) Three Readers (n = 9) Two Readers (n = 12) One Reader (n = 8) No Reader (n = 17) 0 (no mark) 0 2 (17) 2 (25) 8 (47) 1 (one view) 4 (44) 4 (33) 4 (50) 7 (41) 2 (both views) 5 (56) 6 (50) 2 (25) 2 (12) Note Data are no. (%) of lesions. Percentages have been rounded. Either Total CAD Performance: Visible Findings on Prior Mammograms Of the 46 visible findings on prior mammograms, 21 findings presented as masses, 17 as microcalcifications, and eight as masses with microcalcifications. The CAD system had correctly marked 74% (34/46) of the visible findings on prior mammograms. According to the mammographic appearance, 52% (11/21) of lesions presented as mass, 94% (16/17) of lesions presented as microcalcification, and 88% (7/8) of lesions presented as mass with microcalcification and were correctly marked (Table 1). The system showed a significantly higher sensitivity for calcified findings than for mass only findings on prior mammograms (92% [23/25] vs 52% [11/21]; p = 0.006). According to the histopathologic analysis at the time of cancer diagnosis, 71% (22/31) of the invasive ductal carcinomas, 0% (0/1) of the invasive lobular carcinoma, and 86% (12/14) of the DCIS lesions were correctly marked on prior mammograms (Table 2). CAD Performance: Actionable Versus Underthreshold Lesions Of the 46 visible findings on prior mammograms, 46% (21/46) were determined to be actionable, and 54% (25/46) were underthreshold. Of the 21 actionable findings, there were five masses, 11 microcalcifications, and five masses with microcalcifications. Of the 25 underthreshold findings, there were 16 masses, six microcalcifications, and three masses with microcalcifications. Calcified findings were more likely to be determined as actionable findings than mass only findings (64% [16/25] vs 24% [5/21]; p = 0.009). Actionable findings on prior mammograms showed higher CAD sensitivity than 1280 AJR:195, November 2010

6 did underthreshold findings (90% [19/21] vs 60% [15/25]; p = 0.04). Of the 34 correctly marked findings, 15 findings were marked on both views and 19 findings were marked on only one view (Table 3). Actionable findings were more likely to show a both-view mark than underthreshold findings (58% [11/19] vs 27% [4/15]), but this difference was not statistically significant (p = 0.09) (Figs. 1 and 2). When more readers determined a finding as actionable, the finding tended to have CAD marks and to have both-view marks (p = 0.081) (Table 3). Discussion According to the results of our study, the CAD system had correctly marked 74% (34/46) of visible findings and 90% (19/21) of actionable findings, with an overall falsepositive mark rate of 1.61 per woman on prior screening digital mammograms obtained in women with breast cancer, which were prospectively overlooked by radiologists. Our results are similar to those of previous studies that have reported that the CAD system marked 42 77% of areas that subsequently developed into cancers on prior mammograms that had been initially interpreted without CAD information [1 4]. Our study differs from those studies, however, in that the CAD system in our study was prospectively applied and the CAD marks were overlooked by the radiologists even with the given CAD information. Indeed, many researchers have pointed out that correctly marked cancers on CAD of prior mammograms does not translate into increased cancer detection, mainly as a result of the high false-positive rates of CAD, which lead radiologists to overlook the correct CAD marks [13]. We believe that understanding the features of true-positive CAD marks on actionable findings that had been overlooked by radiologists may help radiologists differentiate false-positive marks and true-positive marks in clinical practice. An interesting observation in our study was that actionable findings showed significantly higher CAD sensitivity (90% [19/21] vs 60% [15/25]; p = 0.04) and were more likely to be both-view marks (58% [11/19] vs 27% [4/15]; p = 0.09) than were underthreshold findings. Therefore, we suggest that, when a CAD mark is present on both views, more careful attention should be paid. As for nonactionable lesions, in a retrospective blinded review study regarding the CAD application of prior mammograms [4], the authors suggested that there were nonspecific findings that were perceptible on prior mammograms and subsequently developed into cancer but were considered as normal or benign by three or more radiologists among five radiologists. In that study, the CAD system had marked 42% of the 172 subtle findings; hence, the authors suggested that failure to act on nonspecific but CAD-marked findings prospectively was not below the acceptable standard of care. The subtle findings in the previous study are analogous to the underthreshold findings in our study. We found that both-view marks were a possible indicator of actionability for missed lesions on CAD. However, current CAD systems independently detect abnormal findings in separate mammographic images, whereas human readers have used two view mammographic interpretations in screening mammography for years because it improves the detection of breast cancers [14, 15]. Recent studies have also reported that a fusion of two-view information may improve the performance of CAD systems for microcalcifications as well as for masses [16, 17]. In one study, cluster candidates were paired using their radial distances from the nipple and a similarity classifier from the information of both views [16]. For the malignant microcalcification clusters, at a mammogram-based sensitivity of 80%, the false-positive rate was decreased from 0.35 with single-view method to 0.18 with the two-view fusion. For the malignant masses, detection sensitivity of the CAD system was increased from 62% to 73% at a false-positive rate of 1 per image [17]. Our study results also suggest that a correspondence of lesion candidates on two different views provides valuable additional information in distinguishing true-positive actionable marks from underthreshold marks. With regard to invasive lobular carcinomas, previous research has shown that the sensitivities of CAD were 91 95% [18, 19]; however, none of the invasive lobular carcinomas were marked by CAD in our study, although there was only one case. This fact may be related to the differences in case selection between the studies. Previous research included cancers detected by screening mammography, whereas our study included prior mammograms that were negatively reported prospectively. Thus, although there were visible findings, they were very subtle. Our study has several limitations. First, the actionability of the findings was determined by three radiologists. The inclusion of more radiologists with varying degrees of experiences could have changed the actionability in our study. Second, the number of views with CAD marks was determined by the radial distance from the nipple and similarity of findings on both mediolateral oblique and craniocaudal views, which might be subjective. Third, the sample size may be too small to reach a solid conclusion. Fourth, the proportion of interval cancers in our study was relatively higher than that in the usual screening population. This fact may be related to the relatively higher proportion of dense breasts in Asian women, which may lead to a lower sensitivity of screening mammog raphy [20]. In conclusion, on prospectively overlooked CAD marks on prior screening digital mammograms obtained in women with breast cancers, the CAD system had prompted 74% (34/46) of visible findings and 90% (19/21) of actionable findings. In addition, actionable findings showed higher CAD sensitivity and were more likely to be marked on both views than underthreshold findings were. Therefore, when a finding on a screening mammogram shows a CAD mark on both craniocaudal and mediolateral oblique views, more attention should be paid. References 1. 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: Brem RF, Baum J, Lechner M, et al. Improvement in sensitivity of screening mammography with computer-aided detection: a multiinstitutional trial. AJR 2003; 181: Birdwell RL, Ikeda DM, O Shaughnessy KF, Sickles EA. Mammographic characteristics of 115 missed cancers later detected with screening mammography and the potential utility of computer-aided detection. Radiology 2001; 219: Ikeda DM, Birdwell RL, O Shaughnessy KF, Sickles EA, Brenner RJ. Computer-aided detection output on 172 subtle findings on normal mammograms previously obtained in women with breast cancer detected at follow-up screening mammography. Radiology 2004; 230: Freer TW, Ulissey MJ. Screening mammography with computer-aided detection: prospective study of 12,860 patients in a community breast center. Radiology 2001; 220: Birdwell RL, Bandodkar P, Ikeda DM. Computer-aided detection with screening mammography in a university hospital setting. Radiology 2005; 236: AJR:195, November

7 7. Morton MJ, Whaley DH, Brandt KR, Amrami KK. aided detection system. J Natl Cancer Inst 2004; 17. Sahiner B, Chan HP, Hadjiiski LM, et al. Joint Screening mammograms: interpretation with com- 96: two-view information for computerized detection puter-aided detection prospective evaluation. Ra- 12. Fenton JJ, Taplin SH, Carney PA, et al. Influence of microcalcifications on mammograms. Med diology 2006; 239: of computer-aided detection on performance of Phys 2006; 33: Dean JC, Ilvento CC. Improved cancer detection screening mammography. N Engl J Med 2007; 18. Evans WP, Warren Burhenne LJ, Laurie L, using computer-aided detection with diagnostic and screening mammography: prospective study of 104 cancers. AJR 2006; 187: Ko JM, Nicholas MJ, Mendel JB, Slanetz PJ. Prospective assessment of computer-aided detection in interpretation of screening mammography. AJR 2006; 187: Cupples TE, Cunningham JE, Reynolds JC. Impact of computer-aided detection in a regional screening mammography program. AJR 2005; 185: Gur D, Sumkin JH, Rockette HE, et al. Changes in breast cancer detection and mammography recall rates after the introduction of a computer- 356: Birdwell RL. The preponderance of evidence supports computer-aided detection for screening mammography. Radiology 2009; 253: Ikeda DM, Sickles EA. Second-screening mammography: one versus two views per breast. Radiology 1988; 168: Sickles EA, Weber WN, Galvin HB, Ominsky SH, Sollitto RA. Baseline screening mammography: one vs two views per breast. AJR 1986; 147: Paquerault S, Petrick N, Chan HP, Sahiner B, Helvie MA. Improvement of computerized mass detection on mammograms: fusion of two-view information. Med Phys 2002; 29: O Shaughnessy KF, Castellino RA. Invasive lobular carcinoma of the breast: mammographic characteristics and computer-aided detection. Radiology 2002; 225: Brem RF, Rapelyea JA, Zisman G, Hoffmeister JW, Desimio MP. Evaluation of breast cancer with a computer-aided detection system by mammographic appearance and histopathology. Cancer 2005; 104: Mandelson MT, Oestreicher N, Porter PL, et al. Breast density as a predictor of mammographic detection: comparison of interval- and screendetected cancers. J Natl Cancer Inst 2000; 92: AJR:195, November 2010