Comparison of standard and double reading and computer-aided detection (CAD) of interval cancers at prior negative screening mammograms: blind review

The study evaluates the role of computer-aided detection (CAD) in improving the detection of interval cancers as compared to conventional single (CONV) or double reading (DOUBLE). With this purpose, a set of 89 negative cases was seeded with 31 mammograms reported as negative and developing interval cancer in the following 2-year interval (false negative (FN)=11, minimal signs (MS)=20). A total of radiologists read the set with CONV and then with CAD. Overall, there were 589 cancer and 1691 noncancer readings with both CONV and CAD. Double reading was simulated by combining conventional readings in all 171 possible combinations of 19 radiologists, resulting in a total of 5301 cancer and 15 219 noncancer readings. Conventional single, DOUBLE and CAD readings were compared in terms of sensitivity and recall rate. Considering all 19 readings, cancer was identified in 190 or 248 of 589 readings (32.2 vs 42.1%, χ2=11.80, df=1, P<0.01) and recalls were 287 or 405 of 1691 readings (16.9 vs 23.9%, χ2=24.87, df=1, P<0.01) at CONV or CAD, respectively. When considering FN and MS cases separately, sensitivity at CONV or CAD was 50.2 or 62.6% (χ2=6.98, df=1, P=0.01) for FN and 22.3 or 30.7% (χ2=6.47, df=1, P=0.01) for MS cases, respectively. Computer-aided detection (average of 19 readings) was slightly and not significantly less sensitive (sensitivity: 42.1 vs 46.1%, χ2=3.24, df=1, P=0.07) but more specific (recall rate 23.9 vs 26.1%, χ2=3.8, df=1, P=0.04) as compared to DOUBLE (average of 171 readings). Average sensitivity for FN cases only was 62.6% for CAD and 64.8% for DOUBLE (χ2=0.32, df=1, P=0.58). Corresponding values for MS cases were 30.7% for CAD and 35.7% for DOUBLE (χ2=3.53, df=1, P=0.06). Compared to CONV, CAD allowed for improved sensitivity, though with reduced specificity, both effects being statistically significant. Computer-aided detection was almost as sensitive as DOUBLE but significantly more specific. Computer-aided detection might be used in the current practice to improve sensitivity of conventional single reading. Based on estimates of screening sensitivity and FN/MS interval cancer expected frequency, the absolute increase of screening sensitivity expected by introducing CAD-assisted reading may be estimated around 0.9%. The use of CAD as a possible surrogate to conventional DOUBLE needs to be confirmed by further studies, which should include a cost-effective analysis.