Deep Learning for computer-aided detection and diagnosis of clustered microcalcifications on digital mammograms

Breast cancer is one of the most common cause of cancer death in women worldwide. In most western countries, screening programs are organized in order to detect breast cancers at an early stage. To improve breast cancer detection, many radiologists use computer-aided detection and diagnosis (CAD) systems which are able to detect and characterize mammographic signs of malignancy such as clustered microcalcifications and masses through computerized image analysis. Even though effective in terms of sensitivity, these systems produce a too high number of false alarms, which potentially limits the benefit they can provide. This thesis addresses the problem of accurately detecting and classifying clustered microcalcifications in full field digital mammograms. The goal is to reduce the gap between CAD systems and radiologists in terms of false alarms while maintaining the high sensitivity typical of the commercial CAD systems. To this end, three main contributions are proposed by exploiting innovations and advantages of novel machine learning algorithms, based on deep learning convolutional neural networks (CNNs) : (i) a new proposal of combination of a deep cascade of boosting classifiers and a CNN to deal with the high-imbalance problem of classifying individual pixels in a mammogram as belonging to a microcalcification or not; (ii) a novel method for detecting individual calcifications that provides for the use of multiple-depth CNNs, to exploit both the local features and the surrounding context of MCs; and (iii) a novel end-to-end system able to combine both detection and classification of malignant cluster, by additionally segmenting individual calcifications. Along with these contributions, experimental comparisons with other existing methods in the literature are provided and show significant reduction in the number of false alarms. Moreover a novel end-to-end model that combines detection and classification steps is presented, by showing a significant improvement with respect to single-task systems. When applied to a clinical setting, this would help the radiologists to reduce the number of unnecessarily recalled women with microcalcification clusters, thus improving the effectiveness of screening and diagnosis processes.