Conditional Diffusion Model with K-space Guidance for Highly-Accelerated Cardiac Cine MRI Reconstruction

Motivation: To develop a new deep-learning framework for highly accelerated cardiac cine reconstruction with sharp details and high SNR. Goal(s): To reconstruct high-quality cine MRI with high undersampling rates using denoising diffusion probabilistic frameworks. Approach: A diffusion model conditioned on slice information is trained to generate images of different phases. Data consistency enforced by k-space alignment controls the phase generation. Results: The diffusion model reconstructs high-quality cine MRI from highly undersampled data, validated by radiologists' evaluation. Impact: A new framework with diffusion models is proposed for cardiac cine reconstruction. It utilizes high-quality reconstruction of generative models and provides reliable results by data consistency control. The method is applicable to other dynamic MRI reconstructions in highly undersampled scenarios.