Abstract
Brain image segmentation plays a pivotal role in modern healthcare by enabling precise diagnosis and treatment planning. Federated Learning (FL) enables collaborative model training across institutions while safeguarding sensitive patient data. The integration of these technologies holds significant potential for advancing artificial intelligence (AI) in healthcare. However, medical institutions frequently encounter data imbalances, where some have limited annotated brain imaging data, whereas others possess larger datasets and more diverse cases. Such data exhibit non-independent, non-identically distributed characteristics, which adversely affect segmentation accuracy and generalizability. To address these issues, this paper proposes a client-side brain tumor image segmentation model utilizing Virtual Adversarial Training (VAT) integrated into a 3D U-Net to improve model performance under conditions of limited datasets, effectively addressing data scarcity and imbalance within the federated learning environment by optimizing the use of brain tumor image data held by each client. FedHG introduces an effective federated model aggregation strategy that leverages key parameters, specifically the 'weights' derived from a public validation dataset. Additionally, instance normalization parameters are incorporated into client models during training. These strategies collectively enhance the generalizability of the federated model. Empirical experiments validate the proposed algorithm, showing a 2.2% improvement in the Dice Similarity Coefficient (DSC) for brain tumor segmentation over the baseline federated learning algorithm, with a marginal 3% reduction in performance compared to centralized training, highlighting its practical applicability.