Deep Learning in mammography

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

Breast cancer is the most common cancer among women. Despite great progress in patient care and early screening, it remains a major health issue. Mammography screening is now widely used, but it faces significant problems such as a high rate of false positives and overdiagnosis. To alleviate the work load induced by this widespread strategy, computer-aided diagnosis has been introduced in the 1970s, but it still encounters reluctance from senologists. Deep Learning solutions for mammography assessment are reaching maturity and are now being used in healthcare facilities. However, these approaches are very data-intensive, making their deployment complex for medical tasks. We suggest using Active Learning to address these concerns, which allows the model to request specific annotations from an expert and achieve better performance. First of all, this master thesis aims to assess three Deep Learning approaches to diagnose benign and malignant masses in mammograms. We use the publicly available dataset CBIS-DDSM to train our models and evaluate our work. We show that a whole-image classifier based on ResNet has difficulties identifying the right area. Then, a patch approach is investigated, proving that Deep Learning models can discriminate masses with a specificity of 0.9 at 0.1 sensitivity. Afterward, we implement a Faster R-CNN network with a sensitivity of 0.463 at 0.5 false positives average per image. Furthermore, we provide a comprehensive overview of Active Learning techniques for Deep Learning. We evaluate several strategies from the literature with the same experimental parameters to make the results comparable. The entropy measure, core-set approach, ensemble method, Monte Carlo dropout, learning loss, and discriminative Active Learning are studied on the MNIST dataset. Finally, we apply some of these methods to mammography and investigate the relevance of using uncertainty, diversity, and a combination of both measures to annotate regions of interest intelligently.