A Novel Partitioning Approach for Multimodal Brain Tumor Segmentation for Federated Learning

Traditional Machine Learning approaches always require centralizing the training data on a central server or one location. Due to the privacy of medical data, it is often infeasible to collect and share patient data over one centralised server or location. Federated learning helps owners of patient data to share only model weight updates. A high-performing model can be achieved by carefully aggregating the model updates. In this work, we present a new partitioning or clustering methodology for the FeTS dataset based on mutual information scores and apply a federated learning method for effective weight aggregation We have used the well-known U-Net architecture for our task. We have reported results on Training and Validation FeTS 2021 dataset.