Super-MoCo-MoDL: A combined super-resolution and motion-corrected undersampled deep-learning reconstruction framework for 3D whole-heart cardiac MRI.
Authors
Affiliations (6)
Affiliations (6)
- School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom.
- School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom; Department of Electrical and Electronic Engineering, Imperial College London, London, United Kingdom.
- Institute for Computational Imaging and AI in Medicine, Technical University of Munich, Munich, Germany.
- Department of Cardiology, Aarhus University Hospital, Palle Juul Boulevard 99, 8200, Aarhus N, Denmark; Department of Clinical Medicine, Aarhus University, Aarhus, Denmark.
- School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom; School of Engineering, Pontificia Universidad Católica de Chile, Santiago, Chile; Millennium Institute for Intelligent Healthcare Engineering, Santiago, Chile.
- School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom; School of Engineering, Pontificia Universidad Católica de Chile, Santiago, Chile; Millennium Institute for Intelligent Healthcare Engineering, Santiago, Chile; Institute for Biological and Medical Engineering, Pontificia Universidad Católica de Chile, Santiago, Chile; Institute of Advanced Study, Technical University of Munich, Munich, Germany.
Abstract
Cardiac magnetic resonance (CMR) is a well-established imaging modality for the assessment of cardiovascular diseases. However, attainable image resolution remains lower than that of X-ray computed tomography (CT) due to long scan times and the need for respiratory motion correction. In this work, we combine a previously proposed motion-corrected model-based deep-learning reconstruction for undersampled 3D whole-heart CMR with data-consistent super-resolution to enable high-resolution 3D whole-heart CMR from significantly shortened scans. Our proposed framework, Super-MoCo-MoDL, utilises two neural networks; the first estimates non-rigid respiratory motion from zero-padded and zero-filled bin images, the second applies these fields in an iterative motion-corrected model-based ADMM (alternating direction method of multipliers) reconstruction which alternates between applying a super-resolving U-Net and imposing data-consistency in the acquired centre of k-space. The framework was trained using 156 isotropic-resolution free-breathing 3D datasets. It was subsequently applied to prospective anisotropic low-resolution free-breathing 3D data acquired in a cohort of congenital heart disease (CHD) patients, and to prospective undersampled and low-resolution data acquired in a cohort of patients with suspected coronary artery disease (CAD). Isotropic resolution whole-heart 3D images were reconstructed from ~ 0.8- and ~ 2.1-minute scans, for CHD patients at 1.5-mm resolution and suspected-CAD patients at 0.9-mm resolution, respectively, representing an overall scan acceleration of ~ 18-fold in each case. Visual inspection, expert image quality scores and rankings, and quantitative vessel sharpness measurements demonstrated that the Super-MoCo-MoDL reconstructions produced sharp high-quality images that were comparable with high-resolution acquisitions. For patients with suspected CAD, comparison was made with computed tomography coronoary angiography (CTCA), demonstrating that coronary plaque visualisation was possible with the Super-MoCo-MoDL technique. Super-MoCo-MoDL is able to reconstruct high-resolution 3D whole-heart images from low-resolution and undersampled anisotropic acquisitions.