Multi-layer Convolutional Sparse Coding Framework for Restoration of Under-sampled MR Images

Abstract

Magnetic Resonance Imaging (MRI) plays an important role in the diagnosis of different pathologies associated with human anatomy. The need to acquire images with higher temporal and spatial resolution require longer scan times resulting in patient fatigue and claustrophobia. In addition to scan times, the induced motion artifacts in acquired MRI images further necessitate the reduction in scan time for better image quality in case the process is repeated. To circumvent the longer scan time problem, Parallel Imaging (PI) and Compressive Sensing (CS) techniques have been proposed for scan time accelerations. The emergence of deep learning-based techniques that rely on fully sampled MR images to learn image priors and key parameters for non-linear mappings between fully sampled and undersampled MR images have enabled compressive sensing-magnetic resonance imaging (CS-MRI) restoration architectures that are much better than the traditional regularization-based restoration techniques. In this article, we propose a multilayer convolutional sparse coding (ML-CSC) based framework utilizing layered basis pursuit for CS-MRI reconstruction and demonstrate its effectiveness on brain MR images with different acceleration factors. The proposed generic architecture is shown to provide successful reconstruction from undersampled MR images that can be further used for clinical interpretations.