A deep framework for enhancement of diagnostic information in CSMRI reconstruction

Abstract

In compressed sensing (CS)-based magnetic resonance imaging (MRI), it is very challenging to maintain the diagnostic quality due to limited measurements. Diagnostically critical information, like fine anatomical details, edges, and boundaries are distorted due to the leakage of energy and artifacts during CS-reconstruction. In this paper, we have proposed a deep learning architecture to reconstruct high-resolution diagnostically enhanced MR images from comprehensively sensed k-space data with reduced scan time. Proposed network is based on deep back-projection and generative adversarial network (GAN) architectures. For the performance evaluation, we have considered a pathological dataset having 3064 MR images with three types of brain tumors. The performance of the proposed technique is also compared with some of the well-known image super-resolution techniques. It has been observed that the proposed technique outperforms some of the recent image super-resolution techniques, both quantitatively and qualitatively.