Federated Pseudo Modality Generation for Incomplete Multi-Modal MRI Reconstruction.

Area of Science:

• Medical Imaging
• Machine Learning
• Federated Learning

Background:

• Multi-modal learning is effective for MRI reconstruction but requires paired data, which is scarce in clinical settings.
• Federated learning in medical imaging often encounters clients with missing or single-modal data, hindering standard framework deployment.

Purpose of the Study:

• To propose a novel communication-efficient federated learning framework (Fed-PMG) for multi-modal MRI reconstruction.
• To address the challenge of missing modalities in federated multi-modal MRI reconstruction.

Main Methods:

• Utilized a pseudo modality generation mechanism to recover missing modalities by sharing frequency domain amplitude spectrum distribution.
• Introduced a clustering scheme to compress amplitude spectrum information into centroids, significantly reducing communication costs.

Main Results:

• Fed-PMG effectively recovers missing modalities within acceptable communication overhead.
• The proposed method outperforms existing state-of-the-art approaches.
• Achieved performance comparable to the ideal scenario where all clients possess complete multi-modal data.

Conclusions:

• Fed-PMG offers a viable solution for federated multi-modal MRI reconstruction with missing data.
• The framework balances performance with communication efficiency, making it practical for real-world clinical applications.