Detecting Clusters of Microcalcifications in High-Resolution Mammograms Using Support Vector Machines

This paper presents a new method for detecting clusters of microcalcifications in highresolution digital mammograms. Using cluster analysis, we have designed a descriptive set of mammogram image features which enables precise recognition of microcalcifications. These features are fed into the Support Vector Machine classifier trained to discriminate between normal image occlusions and deposits of calcium in breast tissue. Initial candidates for microcalcifications, i.e. suspicious regions on a mammogram image, are selected by means of a discrete wavelet transform, image filtering and morphological operations. Once microcalcifications are detected, our algorithm assesses whether they form groups (clusters) and for each such group verifies its diagnostic significance. This verification is performed by employing another, appropriately trained, Support Vector Machine classifier. Accuracy of our system has been evaluated on the Breast Cancer Research Program (BCRP) volumes of the DDSM database. On this largest publicly available databases of mammograms our system achieved a sensitivity of 85.1% with average number of 5.0 false positive detections per image. Such an accuracy is competitive with other published results obtained on the same dataset.