Comparison of Multiple View Strategies to Reduce False Positives in Breast Imaging

Abstract

Although state-of-the-art CAD systems for breast imaging have a high sensitivity, their practical applicability is limited by the large number of false positive detections. Recently different multiple view strategies were proposed to increase the specificity. So far it was not possible to compare the performance of these methods, because different validation procedures were used. In this paper we validate all multiple view strategies using the same database and CAD system to make a performance comparison possible. Our results show that the performance difference between different multiple view strategies is small. Asymmetry (difference between left and right breast) and the corresponding view (e.g. if we examine the CC view we use additional information from the MLO view) yields equal performance. The performance is slightly worse using a breast atlas. Multiple view features are weaker than single view features (e.g. CAD score), but it is advantageous that multiple view features provide complementary information.