Generalizable deep mri reconstruction with cross-site data synthesis

Abstract

Deep learning techniques have enabled leaps in MRI reconstruction from undersampled acquisitions. While they yields high performance when tested on data from sites that the training data originates, they suffer from performance losses when tested on separate sites. In this work, we proposed a novel learning technique to improve generalization in deep MRI reconstruction. The proposed method employs cross-site data synthesis to benefit from multi-site data without introducing patient privacy risks. First, MRI priors are captured via generative adversarial models trained at each site independently. These priors are shared across sites, and then used to synthesize data from multiple sites. Afterwards, MRI reconstruction models are trained using these synthetic data. Experiments indicate that the proposed method attains higher generalization against single-site models, and higher site-specific performance against site-average models.