Multi-institutional Collaborations for Improving Deep Learning-based Magnetic Resonance Image Reconstruction Using...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Deep learning excels at magnetic resonance (MR) image reconstruction from undersampled data.
• Data scarcity and privacy concerns hinder the development of robust deep learning models for MR imaging.
• Federated learning (FL) offers a solution by enabling collaborative model training across institutions without data sharing.

Purpose of the Study:

• To develop a federated learning framework for privacy-preserving MR image reconstruction.
• To address the challenge of domain shift in FL for MR imaging.
• To improve the generalizability and performance of MR reconstruction models trained across multiple institutions.

Main Methods:

• Proposed a federated learning (FL) approach for MR image reconstruction.
• Introduced a cross-site modeling technique to align latent features between source and target sites.
• Conducted extensive experiments to evaluate the proposed framework.

Main Results:

• The proposed FL framework effectively utilizes multi-institutional data while preserving patient privacy.
• Cross-site modeling significantly improved the generalizability of MR reconstruction models.
• Experimental results demonstrated the framework's promise for enhanced MR image reconstruction.

Conclusions:

• Federated learning combined with cross-site modeling is a viable strategy for privacy-preserving MR image reconstruction.
• The approach mitigates domain shift issues inherent in multi-institutional FL.
• This work paves the way for improved clinical applications of MR imaging through collaborative AI.