Abstract

Motivation: Deep learning methods have achieved superior reconstruction in MR quantitative T1ρ imaging due to their ability to learn the non-linearity relationship between the undersampled k-space data and corresponding quantitative maps. Goal(s): In this study, we investigate the use of DDPM for highly accelerated T1ρ imaging. Approach: The DDPM learns the image properties from fully acquired images during training without the knowledge of the subsampling patterns used for the accelerated scans. This is advantageous to most existing models that need to be retrained every time for a new sampling scheme. Results: Our results demonstrate that DDPM can achieve superior T1ρ-weighted images and T1ρ map. Impact: The proposed DDPM can achieve superior T1ρ map then compressed sensing and other learning methods.