Motivation: Deep learning-based motion estimation/correction and reconstruction for dynamic MRI often require pre-training models, which leads to limitations in their application and raises concerns about domain shift. Goal(s): To develop and evaluate a zero-shot self-supervised learning framework that can perform both motion estimation and image reconstruction for dynamic MRI without pre-training. Approach: The proposed AIM-ZS iteratively reconstructs and estimates motion from dynamic undersampled k-space data. It combines reconstruction and motion estimation modules with an attention mechanism. Results: AIM-ZS demonstrated superior performance compared to conventional methods, maintaining high image quality even at high acceleration factors (AF=18), with significantly better PSNR values. Impact: This pre-training-free approach enables the widespread adoption of motion-corrected dynamic MRI without requiring large training datasets. It potentially expands clinical applications requiring high temporal resolution imaging and motion analysis and improving accessibility across different imaging protocols.
Fujita et al. (Tue,) studied this question.