Optimal image representations for massdetection in digital mammography

This work addresses a two–class classification problem related to one of the leading cause of death among women worldwide, namely breast cancer. The two classes to separate are tumoral masses and normal breast tissue. The proposed approach does not rely on any feature extraction step aimed at finding few measurable quantities characterizing masses. On the contrary, the mammographic regions of interest are passed to the classifier—a Support Vector Machine (SVM)—in their raw form, for instance as vectors of gray–level values. In this sense, the approach adopted is a featureless approach, since no feature is extracted from the region of interest, but its image representation embodies itself all the features to classify. In order to find the optimal image representation, several ones are evaluated by means of Receiver Operating Characteristic (ROC) curve analysis. Image representations explored include pixel–based, wavelet–based, steer–based and ranklet– based ones. In particular, results demonstrate that the best classification performances are achieved by the ranklet–based image representation. Due to its good results, its performances are further explored by applying SVM Recursive Feature Elimination (SVM–RFE), namely recursively eliminating some of the less discriminant ranklets coefficients according to the cost function of SVM. Experiments show good classification performances even after a significant reduction of the number of ranklet coefficients. Finally, the ranklet–based and wavelet–based image representations are practically applied to a real–time working Computer–Aided Detection (CAD) system developed by our group for tumoral mass detection. The classification performances achieved by the proposed algorithm are interesting, with a false–positive rate of 0.5 marks per–image and 77% of cancers marked per–case.