Diffusion Modeling with Unrolled Transformers for Self-Supervised MRI Reconstruction

Motivation: Diffusion models can reconstruct high-quality MR images, but their training neglects physical constraints and requires supervision via ground-truth images derived from fully-sampled acquisitions. Goal(s): Our goal was to devise a diffusion-based method that incorporates physical constraints and that can be trained using undersampled acquisitions. Approach: We introduced a novel diffusion model (SSDiffRecon) based on a physics-driven unrolled transformer architecture; and self-supervised training was achieved by predicting held-out subsets of acquired k-space data from remaining subsets. Results: SSDiffRecon achieved superior reconstructions to alternative self-supervised methods, and performed on par with a supervised benchmark trained on fully-sampled acquisitions. Impact: The improvement in image quality and acquisition speed through SSDiffRecon, combined with the ability to train on undersampled acquisitions, may facilitate adoption of AI-based reconstruction for comprehensive MRI exams in many applications, particularly in pediatric and elderly populations.