Motion Correction and Super-Resolution for Multi-slice Cardiac Magnetic Resonance Imaging via a Multi-stage Deep Learning Approach

Accurate reconstruction of high-resolution 3D volumes of the heart in cardiac magnetic resonance (CMR) images is important for accurate assessments of heart anatomy and function; however, CMR data is usually acquired as a stack of short-axis (SAX) 2D slices. The reconstruction of 3D volumes from the segmentation contours in 2D slices is challenging due to (1) the presence of inter-slice misalignment caused by cardiac and respiratory motion and (2) the data sparsity from the large gaps between SAX slices. Therefore, motion correction and super-resolution are required to address these two challenges respectively. While existing deep learning (DL) approaches have tried performing motion correction and super-resolution in a single-stage model, we found that such a scheme may compromise reconstruction accuracy. In this study, we propose a novel three-stage DL approach that performs motion correction and super-resolution sequentially and reconstructs more accurate high-resolution 3D volumes of left ventricle blood-pool and myocardium in both a simulation study and in the real-world Sunnybrook Cardiac Dataset compared with the existing single-stage approaches.