Feed-forward weakly supervised deep learning models for breast cancer diagnosis from histological images

Breast cancer screening has become one of the top priorities to reduce the number of deaths from breast cancer and treat it as soon as possible. When making a diagnosis for breast cancer, the gold standard is histological image analysis. However, these images are very large and require both time and expertise to be correctly interpreted. Research in computer vision algorithms has shown great potential to assist practitioners in making faster and more accurate diagnoses. However, to use these kinds of algorithms, large amounts of data with high quality annotations are often required. In this thesis, we propose to approach this problem from the perspective of weakly supervised learning. In this scenario, we consider that we only have weak annotations such as image-level labels for each sample. With these weak annotations, we seek to train accurate models to accurately predict both image-level (classification) and pixel-level labels (segmentation). As a first contribution, we identify, analyze and evaluate several weakly supervised techniques proposed for natural images with the intent of identifying which one are the most promising for our application. As a second contribution, we generalize a technique proposed for binary classification to the more general settings of multi-class and multi-label classification. We show that this technique scales well to widely used natural image datasets and further evaluate it on histological image datasets. As a third contribution, we study the impact of adding size constraints to train a model. This corresponds to a scenario where an annotator would only estimate the size of the cancerous regions in histological images, reducing the time needed to annotate images when doing a segmentation. We propose a formulation to take advantage of the size information and show that it significantly improves the segmentation performance compared to training with image-level labels only.