Segmented rapid magnetic resonance imaging using structured sparse representations

This paper presents a novel approach for segmented rapid magnetic resonance (MRI) imaging using structured sparse representations. In recent years, MR imaging speed has been improved through techniques relying on reduced data acquisition either to allow less-perceivable artifacts or to exploit the data-redundancy. But, these techniques require design of randomized sampling patterns which prove to be sub-optimal from the point of attaining high acceleration gains. The proposed approach treats the imaging problem as a segmented imaging problem constrained to structured dictionaries learned from a small training dataset. The solution to which, is further relaxed to allow recovery of segmented component images from the under-sampled Fourier space. The under-sampling corresponds to traversing few commonly used sampling trajectories only (such as: spirals and radials), i.e., no special sampling pattern design is needed. This highly reduces the number of trajectory acquisitions needed per image formation and, reconstruction in learned dictionaries leads to recovery of automatic segmented images. Experimental results on a real-patient cardiac cine-MR data-set recover high quality segmented images at an acceleration factor of R=13.5 using only 2-spiral sampling trajectories.