Federated Visual Autoregressive Transformers for Collaborative Model Training in MRI Reconstruction

Motivation: Federated learning (FL) offers a privacy-preserving framework for multi-site training of generalizable models in MRI reconstruction. Yet, existing FL methods require all sites to use a fixed model architecture, preventing site-level architecture selection. Goal(s): Our goal was to enable collaborative model training across multiple sites with distinct architectural preferences. Approach: We introduced a novel FL method (FedVAT) that builds a multi-site image prior based on visual autoregressive transformers, and uses synthetic MRI data generated by the VAT prior to train local reconstruction models. Results: FedVAT enhances flexibility in collaborative training of MRI reconstruction models, and outperforms state-of-the-art personalized FL methods in generalization. Impact: High-fidelity image generation achieved by FedVAT enables imaging sites to collaboratively train MRI reconstruction models with divergent architectures. Avoidance of architectural constraints combined with reliable generalization can facilitate applications that suffer from data scarcity, such as assessment of rare diseases.