Purpose: Analysis of the correspondence of the locations and features of lesion candidates on two mammographic views is a promising method for improving computerized lesion detection. In this study, we investigated the use of joint CC and MLO view information for detection of microcalcification clusters on mammograms, and compared the accuracy of this two-view detection method to our single-view detection method.
Methods and Materials: Candidates of microcalcification clusters were located using our previously developed prescreening algorithm. In single-view detection, false-postives (FPs) among these candidates were reduced by using a classifier based on features extracted by a neural network, as well as the morphological and texture features extracted from the cluster. For two-view detection, object pairs from the CC and MLO view candidates were formed by using their radial distances from the nipple. The pairs were classified using a correspondence classifier, which analyzed the similarity between features in a pair. The correspondence classifier scored each pair as to its likelihood of being a true positive pair. These scores were fused with the single-view detection scores. The algorithm was trained using 108 mammogram pairs from our institution, and tested using an independent public data set from the University of South Florida (USF). The test set included 232 mammograms (116 two-view cases) containing 254 microcalcification clusters. Nine of the microcalcifications were seen only on one view. The FP rate was measured by applying the algorithm to 152 normal mammograms (76 two-view cases) from the USF database.
Results: The prescreening algorithm detected 89% (226/254) of the clusters with an average of 3.5 FPs/image (539/152) on the normal mammograms. After FP reduction, the single-view detection algorithm had a film-based sensitivity of 86% at 0.6 FPs/image. At the same sensitivity, the two-view detection algorithm produced 0.4 FPs/image. The sensitivity of the single-view and two-view detection algorithms was 79% and 83%, respectively, at 0.1 FPs/image. If correct detection was defined as marking a malignant cluster on at least one view, the two-view detection algorithm achieved a sensitivity of 90% at 0.1 FPs/image.
Conclusion: The correspondence of geometric, morphological, textural and neural network features of cluster candidates on two different views provides valuable information for improving the accuracy of computerized microcalcification detection.