🤖 AI Summary
This study addresses the challenge of automatic multi-center brain metastasis segmentation while overcoming medical data silos and ensuring privacy compliance. We propose the first framework that deeply integrates differential privacy with federated learning, featuring a dual privacy-preserving mechanism: gradient clipping combined with noise injection during local training, and secure model aggregation across sites—enabling collaborative training without sharing raw medical images. Leveraging a 3D U-Net architecture and a heterogeneous data alignment strategy, our method achieves a Dice score of 0.892 ± 0.021 on data from six hospitals—representing a 12.3% improvement over single-center baselines. Under a privacy budget of ε = 2.0, the framework rigorously satisfies HIPAA and GDPR requirements. This work establishes a verifiable, deployable paradigm for privacy-sensitive, cross-institutional AI modeling in clinical neuro-oncology.