TTP-SSFL: Test-Time Personalization Self-Supervised Federated Learning for Accelerating MR Image Reconstruction

Federated learning (FL) has emerged as a promising paradigm for accelerating magnetic resonance (MR) image reconstruction while preserving data privacy in multicenter collaborations. However, existing FL-based reconstruction methods face two major challenges: 1) a heavy reliance on fully sampled k-space datasets for model training, which is often a tricky problem in clinical settings, and 2) significant performance degradation due to distribution shifts between training and test domains. To address these limitations, a test-time personalization self-supervised FL (TTP-SSFL) method is proposed to accelerate MR image reconstruction. In this study, cross-institutional collaboration without any fully sampled data is implemented by introducing a Siamese-based self-supervised strategy with a hybrid loss function at each client. Moreover, a low-rank adaptation (LoRA)-based test-time adaptation (TTA) strategy is proposed to further mitigate domain shift during deployment. By inserting lightweight adapters into the global model and optimizing them using only testing data via self-supervision, the proposed method can achieve efficient model personalization and robust generalization under distribution shifts. Extensive experiments on multicenter datasets show that TTP-SSFL achieves state-of-the-art performance among self-supervised methods and matches the accuracy of supervised personalized FL (PFL) models, providing a practical and privacy-preserving solution for robust MR reconstruction across heterogeneous clinical environments.