Specificity-Preserving Federated Learning for MR Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Federated learning (FL) offers privacy-preserving collaboration for magnetic resonance (MR) image reconstruction.
• Domain shift from varying MR imaging protocols degrades FL model performance.
• Existing FL methods prioritize global model generalization, neglecting valuable domain-specific features.

Purpose of the Study:

• To introduce FedMRI, a federated learning algorithm for MR image reconstruction that preserves domain-specific properties.
• To address the performance degradation caused by domain shifts in collaborative MR imaging.
• To improve the accuracy and efficiency of MR image reconstruction in multi-institutional settings.

Main Methods:

• Proposed FedMRI algorithm with a globally shared encoder and client-specific decoders.
• Implementation in both frequency and image spaces for simultaneous generalized and specific feature extraction.
• Introduction of weighted contrastive regularization to enhance global encoder convergence amidst domain shifts.

Main Results:

• FedMRI achieved reconstructed MR images closest to ground-truth across multi-institutional data.
• The method demonstrated superior performance compared to state-of-the-art federated learning techniques.
• Preservation of domain-specific features led to improved local reconstruction accuracy.

Conclusions:

• FedMRI effectively balances global generalization and local specificity in federated MR image reconstruction.
• The proposed approach mitigates performance degradation due to domain shifts.
• FedMRI represents a significant advancement for privacy-preserving, collaborative MR imaging.