Privacy-Preserving Knowledge Distillation for Robust and Data-Efficient MRI Reconstruction

Data privacy restrictions limit the sharing of raw MRI data across multiple clinical sites, creating generalization challenges when a model encounters unseen data distributions. To address this, we propose a knowledge distillation method that facilitates collaborative training while preserving patient data confidentiality. Each site trains a local reconstruction model on its raw MRI data and, instead of sharing local data, transfers images reconstructed from a publicly accessible dataset to a central server. The server builds a global model to generate these reconstructed images from the public dataset and then sends back the final reconstructions. This iterative process enables sites to leverage insights from other sites' data without directly accessing private information. Experimental results on multi-site MRI datasets show that our method achieves superior generalization performance compared to models trained solely on local data or through site-average models.