Generative Diffusion Model for Dynamic MRI Reconstruction with Temproal State Space Representation

Motivation: In accelerated dynamic MR imaging, exploiting the temporal redundancy in the acquired data can play a crucial role in enhancing the precision and quality of image reconstruction. Goal(s): To develop a denoising diffusion probabilistic model that leverages the temporal redundancy of dynamic MRI to improve the performance of reconstruction. Approach: In this work, we propose a Mamba-based temporal block which can be easily plugged into the backbone of a diffusion model for exploiting the spatiotemporal redundancy within dynamic MRI. Results: Qualitative and quantitative results show that our methods significantly improve reconstruction performance in multiple acceleration factors. Impact: Our work will enable faster and higher-quality dynamic CMR imaging for improving the MR imaging workflow and aiding in clinical diagnosis.