dDiMo: Domain-conditioned and Temporal-guided Diffusion Modeling for Accelerated Dynamic MRI

Motivation: Diffusion probabilistic methods synthesize realistic images from Gaussian noise but can yield suboptimal performance in reconstructing accelerated MRI data with additional time-resolved dimensions, causing temporal misalignment. Goal(s): This paper aims to develop a novel generative AI approach based on diffusion modeling that incorporates temporal features from time-resolved dimensions, like dynamic cardiac motion, to generate high-quality images. Approach: Our method introduces a domain-conditioned, temporal-guided diffusion model that leverages spatiotemporal correlations and self-consistent k-t priors to guide the diffusion process in the native data domain. Results: The proposed method shows significant promise for multi-coil cardiac MRI reconstruction at high acceleration factors. Impact: This work demonstrates the feasibility of a novel generative AI method for rapid dynamic MRI by leveraging temporal information and self-consistent k-t priors. Beyond its immediate applications, the method shows potential for generalization, adapting to inverse problems across various domains.