Preliminary results of a featureless CAD system on FFDM images

A novel featureless approach to the detection of masses and microcalcifications has been adopted, based on a Support Vector Machine (SVM) classifier. This method does not rely on any feature extraction task; on the contrary, the algorithm automatically learns to detect the lesions by the examples presented to it during the training phase. Our technique includes a pre-selection step, in which we select the intra-breast areas that will be analyzed. Those regions are then provided to an SVM classifier, trained to recognize suspect masses or microcalcifications. The CAD performance have been already assessed on digitized mammogram freely available on the net (DDSM USF and Nijmegen databases). In this paper we are going to test the CAD scheme on digital images coming from Giotto Image MD FFDM unit, a mammography system based on an amorphous Selenium detector. Images have a spatial resolution equal to 85 um and 13 bit gray-level resolution and have been collected at two different sites: Maggiore Hospital in Bologna (Italy) and Triemli Hospital in Zurich (Switzerland). Preliminary results are presented on a database gathered at these hospitals. The CAD system marked in the FFDM images 19 cancers out of 23, with a false-positive rate of 0.9 marks per image.