Accelerated dynamic MRI using self expressiveness prior

We introduce a self-expressiveness prior to exploit the redundancies between voxel profiles in dynamic MRI. Specifically, we express the temporal profile of each voxel in the dataset as a sparse linear combination of temporal profiles of other voxels. This scheme can be thought of as a direct approach to exploit the inter-voxel redundancies as opposed to low-rank and dictionary based schemes, which learn dictionaries from the data to represent the signal. The proposed representation may be interpreted as a union of subspaces model or as an analysis transform. The use of this algorithm is observed to considerably improve the recovery of myocardial perfusion MRI data from under sampled measurements.