Model-based federated learning for accurate MR image reconstruction from undersampled k-space data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning excels in Magnetic Resonance (MR) image reconstruction but requires large, multi-center datasets, raising privacy issues.
• Federated learning (FL) enables multi-center training without data sharing, but struggles with data heterogeneity and simple averaging aggregation.
• Existing FL methods show limited reconstruction and generalization capabilities due to data variability and basic aggregation techniques.

Purpose of the Study:

• To introduce ModFed, a novel Model-based Federated learning framework for robust MR image reconstruction.
• To overcome limitations of existing FL methods in handling data heterogeneity and improving model performance.
• To enhance MR image reconstruction accuracy and generalization using privacy-preserving techniques.

Main Methods:

• Developed attention-assisted model-based neural networks to reduce data dependency on individual clients.
• Implemented an adaptive dynamic aggregation scheme to mitigate data heterogeneity and boost model robustness.
• Incorporated a spatial Laplacian attention mechanism and personalized client-side loss for detailed information capture.

Main Results:

• ModFed demonstrated superior MR image reconstruction performance compared to six state-of-the-art FL approaches.
• The framework achieved enhanced generalization capabilities on three in-vivo datasets.
• Experimental results validate the effectiveness of ModFed in addressing data heterogeneity and improving reconstruction quality.

Conclusions:

• ModFed offers an effective solution for privacy-preserving MR image reconstruction using federated learning.
• The proposed adaptive aggregation and attention mechanisms significantly improve model performance and generalization.
• ModFed advances the field by enabling robust and accurate MR image reconstruction from heterogeneous multi-center data.