Undersampled dynamic magnetic resonance imaging using patch-based spatiotemporal dictionaries

Dynamic magnetic resonance imaging (dMRI) requires high spatial and temporal resolutions, which is challenging due to the low imaging speed. To reduce the imaging time, a patch-based spatiotemporal dictionary learning (DL) model is proposed for compressed-sensing reconstruction of dynamic images from undersampled data. Specifically, the dynamic image sequence is divided into overlapping patches along both the spatial and temporal directions. These patches are expected to be sparsely represented over a set of temporal-dependent spatiotemporal dictionaries. The images are then reconstructed from the undersampled data in (k,t) space under such sparseness constraints, where the dictionaries are learned iteratively. Alternating optimization is applied to solve the problem. Simulation results show that the proposed method is capable of preserving details in both spatial and temporal directions.