Adaptive deep dictionary learning for MRI reconstruction

Abstract

This work addresses the well known problem of reconstructing magnetic resonance images from their partially samples K-space. Compressed sensing (CS) based techniques have been used rampantly for the said problem. Later studies, instead of employing a fixed basis (like DCT, wavelet etc. as used in CS), learnt the basis adaptively from the image itself. Such studies, loosely dubbed as dictionary learning (DL) showed marked improvement over CS. This work proposes deep dictionary learning based inversion. Instead of learning a single level of basis, we learn multiple levels adaptively from the image, while reconstructing it. The results show marked improvement over all previously known techniques.