Bilateral Proxy Federated Domain Generalization for Privacy-Preserving Medical Image Diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Centralized domain generalization methods struggle with private, multi-source medical data.
• Existing federated methods face challenges balancing privacy and generalization.
• High privacy of medical data limits centralized training paradigms.

Purpose of the Study:

• To propose a novel framework, the Bilateral Proxy Framework (BPF), for privacy-preserving federated domain generalization.
• To enhance the generalization ability and stability of medical image diagnosis models in federated settings.
• To address the challenge of achieving both strict privacy and robust generalization on out-of-distribution data.

Main Methods:

• Leveraging client-side proxies for privacy-preserving communication and mutual distillation.
• Implementing a server-side proxy with a distance-based strategy and parameter moving average.
• Ensuring smoother convergence of local models and stability of the global model.

Main Results:

• The Bilateral Proxy Framework (BPF) demonstrates superior performance over state-of-the-art methods.
• BPF enhances the stability and robustness of the global model, preventing overfitting.
• The framework achieves accurate and reliable medical image diagnosis in federated settings.

Conclusions:

• The Bilateral Proxy Framework effectively tackles privacy and generalization challenges in federated medical image analysis.
• BPF offers a promising solution for developing robust and privacy-preserving AI diagnostic tools.
• The method shows significant improvements on both simulated and real-world medical image datasets.