Finding intersections of curvi-linear structures in mammograms using oriented local energy feature detection

The increased attenuation at the intersections of curvilinear structures CLS in mammograms leads to speck-like responses which - on a micro-scale resemble microcalcifications. In this paper, we present a novel technique for distinguishing between these CLS intersections and true microcalcifications. Our method is based on a multiresolution, oriented local energy analysis. Local energy enables the detection of features of several different kinds in a unified framework using local phase to distinguish between the different types. Since CLS occur over a wide range of sizes, we decompose the signal in a multiresolution framework which not only helps detect CLS over a range of scales but also lets us estimate at each location the local width of a CLS. Orientation information computed from steerable filters is used in a clustering algorithm to distinguish between curvilinear structures and other responses. By combining scale, phase and orientation information we can distinguish the CLS from non- CLS locally linear features and thus identify positively CLS in a mammogram. Location where these CLS intersect can then be used to validate the responses of a calcification detector.