Does computer-aided detection increase the accuracy of interpretation of mammograms?

Background Computer-aided detection (CAD) programs that identify suspicious areas on mammograms have been developed to assist radiologists. Objective To assess the clinical benefit of using CAD when performing mammograms. Design and intervention This study paired two sets of data. The first set, on factors that could affect mammogram interpretation, was gathered from surveys mailed to three US mammography facilities and affiliated radiologists in 2002. The second set consisted of Breast Cancer Surveillance Consortium registry data from the same facilities on routine bilateral mammograms and cancer outcomes from women >= 40 years of age who were screened during the period 1998-2002. In 2005, affiliated facilities provided additional information on the use of CAD from 1998 to 2002. Outcome measures The outcome measures were the specificity (true-negative mammogram), sensitivity (truepositive mammogram), and overall accuracy of mammography with or without the use of CAD, and the rates of recall (percentage of positive mammograms) and cancer detection on biopsy (per 1,000 mammograms). Results Forty-three (84%) of the 51 facilities that contributed data to the consortium registries responded to the survey, 7 (16%) of which implemented CAD during the study period. Of 630 NATURE CLINICAL PRACTICE ONCOLOGY the radiologists affiliated to these facilities, 122 (77%) provided complete survey responses. Overall, data were available for 429,345 mammograms, 332,869 (78%) of which were interpreted by participating radiologists and included in the analysis, and 31,186 of which were interpreted with use of CAD. Of a total of 222,135 women screened, 2,351 received a diagnosis of breast cancer within a year after screening; 156 of these diagnoses were made using CAD. Women screened at facilities that did not use CAD during the period of the study were older and had denser breast tissue than women screened at facilities that did use CAD. Specificity was significantly lower and the recall rate significantly higher at facilities that did not use CAD than at those that had adopted but not yet implemented CAD (P<0.001 for both). In the centers that implemented CAD, the sensitivity increased from 80.4% to 84.9% after CAD implementation (P=0.32). When facilities implemented CAD, the biopsy rate increased by 20%, but the cancer detection rate was unchanged (4.20 cases per 1,000, versus 4.15 per 1,000 before implementation of CAD; P=0.03). The rate of detection of ductal carcinoma in situ increased by 34% after the implementation of CAD; however, the rate of detection of invasive breast cancer decreased by 12%. In analyses that adjusted for the characteristics of the patients and radiologists, use of CAD was still associated with lower specificity and slightly higher sensitivity than non-use. The modeled area under the receiver-operating-characteristic curve was 0.919 without the use of CAD and 0.871 with its use (P=0.005), indicating a lower overall accuracy with CAD. Conclusion The use of CAD is associated with reduced accuracy of mammogram interpretation, no improvement in cancer detection rates, and increased recall and biopsy rates, and therefore does not have obvious clinical benefit.