A New Method of Detecting Microcalcification Clusters for Computer Aided Digital Mammography

Abstract

Digital mammograms are processed for detecting microcalcification clusters (MCCs) and prompting radiologists on their locations without specifying their type, i.e., benign, normal or malignant. The method includes image segmentation using SUSAN edge detector followed by the shape filters. Then the objects are classified with a four-level feed-forward Neural Network with four input features comprising perimeter and other three characterizing foreground-background relation. MCCs are found using the distance and the object count spatial filters. This simple yet robust system is capable to detect MC clusters with 98.4% of true positives at no false positive cases. The trial is performed on 118 mammograms from the DDSM database. It is shown in the paper that the reported performance is achieved due to the outstanding property of the edge detector to capture objects in a closed contour fashion; an efficient classifier, and significant features characterizing MCCs' geometry and intensities with respect to the background.