LAUSR

Compressed sensing (CS) has shown great potential for fast magnetic resonance imaging (fastMRI). Traditional CS methods model the inverse problem by leveraging the sparsity prior to guarantee the success of signal recovery, which is not rich enough to capture the detailed features of MRI modality. The other challenge is computational complexity in CS methods which often include an iterative optimization-based solver, hindering the growth and development of modern high resolution MRI. Inspired by existing researches in machine vision tasks, two novel network blocks are presented here which respectively leverage a) the spatial correlations and b) data consistency prior, and a novel multi-level densely connected framework is devised to improve the model capacity for removing aliasing artifacts from the under-sampled MR images and recovering missing anatomical information in high resolution MRIs. It is demonstrated that the framework produces more realistic and faithful structures and textural details, providing superior reconstructions in terms of less visual artifacts and relevant metrics.