Zero-Shot Dynamic MRI Reconstruction With Global-to-Local Diffusion Model

Diffusion models have emerged as promising tools for tackle the challenges of MRI reconstruction, demonstrating superior performance in sample generation compared to traditional methods. However, their application in dynamic MRI reconstruction remains relatively underexplored, primarily owing to the substantial demand for fully sampled training data, which is challenging to obtain because of the spatiotemporal complexity and high acquisition costs associated with dynamic MRI. To address this challenge, this paper proposes a zero-shot learning framework for accurate dynamic MR image reconstruction from undersampled k-space data directly. Specifically, a unique time-interleaved acquisition scheme is employed to merge undersampled k-space data from adjacent temporal frames, thereby constructing pseudo fully encoded reference data. Moreover, while merging all the frames enhances the signal-to-noise ratio (SNR), it also reduces interframe correlation. In contrast, merging only local adjacent frames preserves interframe uniqueness but decreases the SNR. Therefore, a two-stage refinement strategy is applied during the diffusion process to learn the global-to-local prior, ensuring the diffusion model effectively captures the data distribution for zero-shot reconstruction. Extensive experiments demonstrate that the proposed method performs well in terms of noise reduction and detail preservation, achieving reconstruction quality comparable to that of supervised approaches.