Neural posterior probabilities for microcalcification detection in breast cancer diagnoses

We apply the a Posteriori Probability Model Selection (PPMS) algorithm with the help of Generalized Softmax Perceptron (GSP) neural architecture in order to obtain estimates of the posterior class probabilities at its outputs, in the binary problem of microcalcification detection in a hospital digitalized mammogram database. We first detect windowed images with high probability to belong to the class microcalcification is present, then we locally segment the shape of the calcifications, and finally show the segmented microcalcifications to the radiologist. The segmented images together with the posterior probabilities for each window image can be employed as a valuable information to help predicting a breast diagnosis, in order to distinguish between benignant calcium deposit and malignant accumulation, that is, breast carcinoma.