A Novel Framework for Fat, Glandular Tissue, Pectoral Muscle and Nipple Segmentation in Full Field Digital Mammograms

Automated segmentation of mammograms is an important initial step in a wide range of applications including breast density and texture analysis and computer aided detection of abnormalities. In this paper, we propose a unified machine learning framework that enables simultaneous segmentation of the breast region, fatty tissue, glandular tissue, pectoral muscle and nipple region in full field digital mammograms. We calculate both a multi-label segmentation mask and a probability map associated with each of the segmented classes. The probability map facilitates interpretation of the segmentation mask prior to further analysis. The method is evaluated using left or right MLO views from 100 women in a 5-fold cross validation manner. Our framework is shown to be robust and accurate, achieving sensitivity/specificity from 82.7% to 98.5% at the equal-error-rate point of the ROC curves and area under the ROC curve values from 0.9220 to 0.9998 for the corresponding segmentations.