Physics-driven deep training of dictionary-based algorithms for MR image reconstruction

Techniques involving learned dictionaries can outperform conventional approaches involving (nontrained) analytical sparsifying models for MR image reconstruction. Inspired by iterative dictionary learning-based reconstruction methods, we propose a novel efficient image reconstruction framework involving multiple iterations (or layers). Each layer involves applying a transformation to image patches, thresholding, and then reconstructing the patches in a dictionary, followed by an update of the image using observed k-space measurements. We train the transforms, thresholds, and dictionaries within the multi-layer algorithm to minimize reconstruction errors. Our experiments demonstrate that for highly undersampled k-space data, such trained reconstruction algorithms provide high quality results.