Heuristic Approach to Detect Pectoral Muscle and Nipple in Mammogram for Computer-Aided Diagnosis

Abstract

In computer-aided diagnosis algorithms for mammograms, many preprocessing steps like pectoral muscle detection and the coordinates of the nipple are done automatically before the actual detection and diagnosis process begin. Normally, in medio-lateral oblique (MLO) views, the pectoral muscle should be seen to the tip of the nipple or below. Consequently, straight line estimate of pectoral muscle is parallel to the tangent drawn at nipple. Considering this heuristic anatomical relation, in the proposed work, hybrid combination of textured Hough transform with distant surround suppression was used to detect straight or curved pectoral muscle on MLO view followed by orthogonal Radon transform to detect the nipple position. The proposed algorithm was tested on 266 MLO views from miniMIAS and 74CC views from local hospital. The pectoral muscle was detected with accuracy of 97%, and detected nipple position was acceptable in 93.67% of mammograms with average error of 5.755 mm with location annotated by a radiologist.