Potential improvement of computerized classification for malignant versus benign mammographic microcalcification clusters with the use of special views

We are investigating the use of special mammographic views, i.e. magnification and spot compression views, into computerized classification schemes for malignant versus benign microcalcification clusters. Radiologists often recall patients with suspicious lesions for additional views. We expect that the fusion, or combination, of information extracted from special views and conventional mammograms will improve the performances of computer-aided diagnosis (CAD) schemes for classification of malignant versus benign microcalcification clusters. It has been shown that reading with CAD improves radiologists' performances. The CAD scheme is applied separately to conventional mammograms and special views. The scheme consists of segmentation of manually-identified microcalcifications, followed by extraction of microcalcification-based geometrical and textural features in addition to cluster-based features. Linear discriminant analysis (LDA) is then applied to each image to classify a cluster as malignant or benign. The leave-one-out technique is used for training, and testing the classifier. The resulting likelihoods of malignancy output from the LDA applied separately to the conventional mammograms and special views are combined using the maximum classifier output. We applied the proposed technique to a database of 75 biopsy-proven patients (31 malignant and 44 benign). The case-based performances for classification of malignant versus benign microcalcification clusters resulted in an area under the receiver operating characteristic (ROC) curves, A(z), of. 0.771 on conventional mammograms, 0.845 on special views, and 0.908 when merging likelihoods of malignancy from conventional mammograms and special views. These preliminary results indicate that the proposed technique of combining information from special views with that from conventional mammograms can improve computerized classification of microcalcification clusters.