Federated Learning for organ segmentation

Training deep learning models for medical imaging requires access to large volumes of sensitive patient data. To this end, the models are generally trained on centralised, deidentified databases that are hard to collect because of privacy requirements. Federated Learning proposes an alternative approach, in which a coalition of hospitals collaboratively trains a central model without exchanging any clinical data. This thesis explores the combination of Federated Learning with U-Net models, and applies it to the task of image segmentation of the heart, the lungs and the oesophagus. The experiments are made locally, using three distinct datasets to simulate different hospitals, with the inter-variability that it involves. A variant of Federated Learning referred to as Federated Equal-Chances that improves segmentation performance on unbalanced datasets is introduced as well. The necessary steps to implement it in a real-life scenario are detailed and the potential of Federated Learning to improve medical image segmentation is proven.