A Spatio-Temporal Diffusion Model for Cardiac Real-Time Imaging

Real-time imaging of cardiac function is favorable due to shorter scan times and becomes necessary when arrhythmia or inability to hold breath leads to insufficient quality of electrocardiogram (ECG)-gated Cartesian cine. However, comparable spatio-temporal resolution can only be achieved in undersampled settings, which in turn demand performant reconstruction methods. This study investigates image quality improvements using a novel spatio-temporal diffusion-based reconstruction, applied to accelerated spiral real-time acquisitions.

In a clinical study, real-time acquisition was performed using accelerated spiral sampling patterns acquired during breath hold and free-breathing. Retrospective binning enabled calculation of segmented spiral cine images, which were used to train a spatio-temporal diffusion model. Reconstruction of accelerated acquisitions was performed using the proposed model, as well as a 2D spatial diffusion model and compressed sensing-based techniques for comparison. Reconstruction quality was assessed by calculating quantitative image metrics for breath-held data and by means of an expert-reader study for free-breathing scans.

Real-time acquisitions enabled significantly shorter scan durations in comparison to clinical cine, with improved quality for participants with irregular heartbeats. Quantitative image metrics indicate superior image quality of the proposed method compared to the baseline methods. Expert reader scores imply consistent sharpness and reduced apparent noise for the proposed model.

Including temporal information within the diffusion model improved consistency between frames, reduced noise, and preserved sharpness in the reconstructions of undersampled spiral acquisitions. Long reconstruction times and demanding computational burdens are obstacles to overcome.