Segmentation of the breast region with pectoral muscle suppression and automatic breast density classification

Breast cancer is one of the major causes of death among women. Nowadays screening mammography is the most adopted technique to perform an early breast cancer detection ahead other procedures like screen film mammography (SFM) or ultrasound scan. Computer assisted diagnosis (CAD) of mammograms attempts to help radiologists providing an automatic procedure to detect possible cancers in mammograms. Suspicious breast cancers appear as white spots in mammograms, indicating small clusters of micro-calcifications. Mammogram sensitivity decreases due some factors like density of the breast, presence of labels, artifacts or even pectoral muscle. The pre-processing of mammogram images is a very important step in the breast cancer analysis and detection because it might reduce the number of false positives. In this thesis we propose a method to segment and classify automatically mammograms according to their density. We perform several procedures including pre-processing (enhancement of the image, noise reduction, orientation finding or borders removal) and segmentation (separate the breast from the background, labels and pectoral muscle present in the mammograms) in order to increase the sensitivity of our CAD system. The final goal is the classification for diagnosis, in other words, finding the density class for an incoming mammography in order tot determine if more tests are needed to find possible cancers in the image. This functionality will be included in a new clinical imaging annotation system for computer aided breast cancer screening developed by the Communications and Remote Sensing Department at the Université Catholique de Louvain[11]. The source code for the pre-processing and segmentation step has been programmed in C++ using the library of image processing ITK and CMake compiler. The performed method has been applied to medio-lateral oblique-view (MLO) mammograms as well as on caniocauldal mammograms (CC) belonging to different databases. The classification step has been implemented in Matlab. We have tested our pre-processing method obtaining a rate of 100% success removing labels and artifacts from mammograms of mini-MIAS database. The pectoral removal rate has been evaluated subjectively obtaining a rate of good removal of 57.76%. Finally, for the classification step, the best recognition rate that we have obtained was 76.25% using only pixel values, and 77.5% adding texture features, classifying images belonging to mini-MIAS database into 3 different density types. These results can be compared with the actual state of the art in segmentation and classification of biomedical images.