Learning Representations for Medical Image Diagnosis: Impact of Curriculum Training and Architectural Design

This thesis investigates two key aspects of learning deep-based image representations for medical diagnosis. The two are confronted with common challenges of medical image databases, namely, the limited number of samples, the presence of unreliable annotations and class-imbalance; as well as, domain shift and data privacy constraints for collaborative learning across institutions. The first part of this thesis concerns the architectural design of deep learning approaches. We explore the importance of localizing the region of interest in the image prior to the classification and the implicit capsule networks' approach to model spatial information. We verify the importance of localization as a preliminary step to the classification, provide a sensitivity analysis of the size of the region of interest, and discuss image retrieval as a clinical use case. We also validate that capsules create equivariance, thus requiring to see fewer viewpoints of the object of interest. The second part of the thesis focuses on easing the optimization of the deep network parameters by gradually increasing the difficulty of the training samples. This gradual increase is based on the concept of curriculum learning and achieved with a data scheduler that controls the order and pace of the samples. We validate the beneficial effect of the curriculum data schedulers in two scenarios. First, we leveraged prior knowledge and uncertainty for the fine-grained classification of proximal femur fractures. In this case, we demonstrated the benefits of our proposed curriculum method under controlled scenarios: with limited amounts of data, under class-imbalance, and in the presence of label noise. Second, we verified the positive effect of the curriculum data scheduler for multi-site breast cancer classification in a federated learning setup.