Analysis of microcalcification features for pathological classification of mammograms

In this study, we aim to analyze the quantitative features for characterizing microcalcification clusters on mammograms for pathological classification into benign and malignant classes. Our database includes 101 cases: 48 cases of benign and 53 cases of malignant with biopsy proven. Two views of mammogram images, Cranial Caudal (CC) view and MedioLateral Oblique (MLO) view were used in our experiments to extract the promising features. A total of 72 features were extracted from each view for analysis. Performances of the considered features are evaluated with an application to mammogram classification. A deep neural network (DNN) classifier was used as a supervised learning in the classification. Performances of the features are evaluated by three measures: Sensitivity, Specificity, and Accuracy. The results showed that for the CC view, Mean of Areas of microcalcification spots, Mean of Major Axis Lengths of microcalcification spots performed better than the other features, followed by Mean of Minor Axis Lengths of microcalcification spots and Mean Diameter of all microcalcifications; for the MLO view, Mean of Perimeters of microcalcification spots, and Mean Diameter of all microcalcifications performed better than the other features, followed by Mean of Contrasts of microcalcification spots, and Diffuseness (standard deviation of inter-distances of all microcalcifications in the ROI). The Mean Diameter of all microcalcification spots is the good feature for both CC and MLO views.