IOP-FL: Inside-Outside Personalization for Federated Medical Image Segmentation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Federated learning (FL) enables collaborative model training across institutions without data centralization.
• Global FL models struggle with performance optimization for individual clients due to data heterogeneity (e.g., scanners, demographics).
• Deploying FL models to unseen clients with novel data distributions presents significant challenges.

Purpose of the Study:

• To propose a unified framework, Inside and Outside Personalization in FL (IOP-FL), for optimizing individual client prediction accuracy in medical imaging.
• To enhance the performance of FL models for both existing and new clients within and beyond the federated network.
• To address the limitations of generic global models in heterogeneous medical imaging environments.

Main Methods:

• Developed a lightweight, gradient-based inside personalization approach using both global and local gradients for client-specific optimization.
• Created a routing space using local personalized models and the global model for outside FL client adaptation.
• Implemented a novel test-time routing scheme with consistency loss and shape constraints for dynamic model incorporation based on test data distribution.

Main Results:

• IOP-FL demonstrated significant improvements over state-of-the-art methods for both inside and outside personalization.
• Achieved enhanced prediction accuracy for individual clients in medical image segmentation tasks.
• Validated the framework's effectiveness on two distinct medical imaging datasets.

Conclusions:

• The IOP-FL framework offers a robust solution for personalized FL in medical imaging.
• The proposed methods effectively address data heterogeneity and deployment to unseen clients.
• IOP-FL shows strong potential for practical clinical applications, improving AI model utility.