Attention U-Net with differential privacy in federated learning framework for brain stroke lesion segmentation

Data privacy considerations and data fragmentation between healthcare institutions is causing segmentation of ischemic stroke lesions from neuroimaging to be hindered. The centralized approach may conflict with HIPAA and GDPR, and the federated learning approach does not have proper privacy guarantees nor is it able to capture lesion detail. This work prposes Fed-AttUNet-DP,a three major contributions: a spatial attention mechanism to integrate federated learning for better stroke lesion detection; differential privacy at the client level and secure MPCC at the server level for dual-layered privacy preservation; adaptive federated optimization for non-IID medical data and faster training speed and better performance. The proposed framework is an augmentation to the federated averaging algorithm which exploits attention U-Net and antennas attention (differential privacy i.e. gradient clipping and Gaussian noise). Ten simulated health care institutions were involved in 150 rounds of communication, at a rate of [Formula: see text] participation of the clients. Privacy budget was configured set at [Formula: see text], [Formula: see text] and a multiplier of noise [Formula: see text] as well as gradient clipping threshold [Formula: see text]. The distance between data heterogeneity was measured with Earth Mover (EMD [Formula: see text]). Experiments are conducted on the BRISC2025 dataset. The Dice similarity coefficient and IoU for Fed-AttUNet-DP is 0.930 and 0.890 respectively, sensitivity is 0.941 and specificity is 0.982. It beats all federated baselines by only [Formula: see text] accuracy drop compared to centralized training as well as [Formula: see text] compared to local-only training. The inference per volume is 0.8 s, converges the system with 5.18 GB total communication cost with 150 rounds. It is verified in the works of the ablation research that attention gates ([Formula: see text] DSC), differential privacy ([Formula: see text] DSC trade-off), and secure aggregation individually contribute to it. Our work provides formal privacy guarantees and advances segmentation accuracy, by mitigating non-IID heterogeneity across multiple medical image institutions and respecting HIPAA/GDPR regulations.