Cross-Modal Vertical Federated Learning for MRI Reconstruction.

Area of Science:

• Medical imaging
• Machine learning
• Federated learning

Background:

• Federated learning (FL) allows collaborative model training across hospitals without sharing raw patient data.
• Current FL methods often assume uniform data modalities, which is impractical due to varying hospital imaging guidelines.
• This limitation restricts the utility of FL in real-world multi-institutional medical research.

Purpose of the Study:

• To address the challenge of cross-modal federated learning with heterogeneous data from multiple hospitals.
• To develop a novel framework, Fed-CRFD, for enhancing Magnetic Resonance Imaging (MRI) reconstruction.
• To effectively utilize overlapping samples (same patients, different modalities) and mitigate domain shift issues.

Main Methods:

• Proposed Federated Consistent Regularization constrained Feature Disentanglement (Fed-CRFD) framework.
• Implemented an intra-client feature disentanglement scheme to separate modality-invariant and modality-specific features.
• Introduced a cross-client latent representation consistency constraint for overlapping samples to align features across modalities.

Main Results:

• Fed-CRFD effectively explores overlapping samples to boost MRI reconstruction.
• The framework successfully mitigates the domain shift problem caused by different data modalities.
• Demonstrated superior performance compared to state-of-the-art MRI reconstruction methods on two datasets.

Conclusions:

• Fed-CRFD offers a robust solution for cross-modal federated learning in healthcare.
• The method maximizes the utility of multi-source hospital data while addressing modality discrepancies.
• This approach has significant implications for improving collaborative medical AI development and MRI reconstruction accuracy.