Multicenter Privacy-Preserving Model Training for Deep Learning Brain Metastases Autosegmentation
🤖 AI Summary
This study addresses the data silos and privacy compliance challenges in multi-center MRI segmentation of brain metastases by proposing the first differentially private federated learning framework tailored to this task. Methodologically, it integrates FedAvg with gradient clipping, Gaussian noise injection, and a 3D U-Net architecture, augmented by multi-center MR image standardization preprocessing; it supports heterogeneous data distributions and dynamic participant institutions. Evaluated on real clinical data from six hospitals, the framework achieves a Dice coefficient of 0.82 ± 0.04—12.3% higher than single-center training—while reducing communication overhead by 37% and strictly complying with GDPR and HIPAA privacy requirements. Its core contribution lies in being the first systematic application of a rigorously differentially private federated learning system to brain metastasis segmentation, successfully balancing segmentation accuracy, robustness across sites, and clinical deployability.
Application Category
🔎 Similar Papers
No similar papers found.
