Autoregressive Masked Diffusion Transformer for Consistent Zero-shot Angular Space Super-Resolution in Diffusion MRI

Motivation: High angular resolution diffusion MRI acquisitions are often clinically infeasible due to prolonged imaging times and patient discomfort. Current methods struggle to reconstruct high-dimensional diffusion data effectively from undersampled data with arbitrary sampling schemes. Goal(s): Our aim is to enable accurate reconstruction of high-resolution diffusion signals from undersampled dMRI data, adaptable to variable sampling schemes without retraining. Approach: We propose the Masked Diffusion Transformer (MDiT), integrating a random masking strategy for zero-shot generalization across sampling patterns and a DiT backbone with autoregressive inference for high-fidelity reconstruction. Results: Our model achieved better results in the angular super resolution task compared to current state-of-the-art methods. Impact: MDiT's ability to accurately reconstruct diffusion data under various sampling schemes without retraining could advance clinical dMRI applications, supporting efficient, high-quality dMRI imaging even under practical clinical constraints.