Microcalcification cluster detection in mammograms with sparse based breast tissue removal and local density dependent multiple classifier

Abstract

Breast cancer is the most common cancer among women worldwide. While breast cancer has high mortality rates, studies have shown that the early detection of breast cancer can improve the chances of recovery. One widely recognized early indication of cancer is clustered microcalcifications. So far, the most effectively used diagnostic modality for microcalcification detection is Mammograms. In an effort to detect the early signs of cancer, recently screening mammography has been adopted in many countries increasing the demand on specialists. Therefore, many research efforts have been focusing on designing computer-aided detection (CAD) to help guide specialists towards suspicious microcalcification clusters by (1) detecting and (2) classifying them. However in raw mammograms, the low contrast between the microcalcification and the surrounding tissue, especially in dense tissue regions, have made it difficult to distinguish the subtle microcalcification particles. Added to that, at the detection stage in the state of the art systems, the low contrast between the microcalcification and the surrounding tissue causes a large number of falsely detected regions. To that end, a CAD system that exploits the difference in texture between normal tissue and microcalcifications is proposed to assure the detection of the most subtle microcalcifications. In the detection stage, the proposed approach adopts sparse representation to estimate normal breast tissue texture only; such that the difference between estimated image and the original image can emphasize subtle microcalcifications. Furthermore, given the importance of the surrounding tissue in classifying malignant microcalcification clusters, the proposed CAD system relies on the texture features in classifying the detected microcalcification cluster regions. Therefore, the inner-class variations due to the surrounding tissue have to be accounted for in the design. Accordingly, local density multiple classifiers are prop...