Federated Deep Learning for Brain Tumor Segmentation: Comparative Analysis of U-Net Architectures

Medical image segmentation is essential for precise diagnosis and treatment planning, particularly in brain tumor analysis where accurate boundary delineation directly impacts surgical outcomes and therapeutic decisions. However, privacy concerns and regulatory constraints limit data sharing across healthcare institutions, restricting the development of robust deep learning models that require large, diverse training datasets. We propose a federated deep learning framework that enables collaborative brain tumor segmentation model training without compromising patient data privacy. Our approach implements three U-Net architectural variants (standard U-Net, Residual U-Net, and Attention U-Net) across multi-institutional BraTS datasets (2019-2021) using the Federated Averaging algorithm. Performance evaluation on three tumor sub-regions (whole tumor, tumor core, enhancing tumor) demonstrates that federated Attention U-Net achieves Dice scores of 0.905, 0.859, and 0.811 respectively on BraTS 2021, representing only 0.8% degradation compared to centralized training while enabling privacy-preserving collaboration. Our federated approach preserves data privacy while achieving comparable accuracy to centralized methods, demonstrating comprehensive evaluation of attention-based federated learning for multi-institutional brain tumor segmentation with statistical significance validation across three consecutive challenge datasets.