Sparsifying transform learning for Compressed Sensing MRI

Abstract

Compressed Sensing (CS) enables magnetic resonance imaging (MRI) at high undersampling by exploiting the sparsity of MR images in a certain transform domain or dictionary. Recent approaches adapt such dictionaries to data. While adaptive synthesis dictionaries have shown promise in CS based MRI, the idea of learning sparsifying transforms has not received much attention. In this paper, we propose a novel framework for MR image reconstruction that simultaneously adapts the transform and reconstructs the image from highly undersampled k-space measurements. The proposed approach is significantly faster (>10x) than previous approaches involving synthesis dictionaries, while also providing comparable or better reconstruction quality. This makes it more amenable for adoption for clinical use.