Robust Cardiac Cine MRI Reconstruction With Spatiotemporal Diffusion Model

Abstract

Accelerated dynamic magnetic resonance imaging (MRI) is highly expected in clinical applications. However, its reconstruction remains challenging due to the inherently high dimensionality and spatiotemporal complexity. While diffusion models have demonstrated robust performance in spatial imaging, their application to spatiotemporal data has been underexplored. To address this gap, we propose a novel spatiotemporal diffusion model (STDM) specifically designed for robust dynamic MRI reconstruction. Our approach decomposes the complex 3D diffusion process into manageable sub-problems by focusing on 2D spatiotemporal images, thereby reducing dimensionality and enhancing computational efficiency. Each 2D image is treated independently, allowing for a parallel reverse diffusion process guided by data consistency to ensure measurement alignment. To further improve the image quality, we introduce a dual-directional diffusion framework (dSTDM), which simultaneously performs reverse diffusion along two orthogonal directions, effectively capturing the full 3D data distribution. Comprehensive experiments on cardiac cine MRI datasets demonstrate that our approach achieves state-of-the-art performance in highly accelerated reconstruction. Additionally, it exhibits preliminary robustness across various undersampling scenarios and unseen datasets, including patient data, non-Cartesian radial sampling, and different anatomies.