Deep Learning Super-Resolution Single-Beat CMR

Overview

Deep learning super-resolution reconstruction is an emerging technique that enhances the resolution of cardiac magnetic resonance (CMR) scans beyond the original acquisition through post-processing. This study investigates whether a deep learning-based single-beat super-resolution CMR protocol-including cine, T2-STIR, and LGE sequences-can provide diagnostic equivalence to a standard segmented CMR protocol. Total scan time, diagnostic confidence, and diagnostic interchangeability are compared between protocols, with particular focus on wall motion abnormalities, myocardial edema, pericardial effusion, late gadolinium enhancement and final diagnosis. The goal is to assess diagnostic interchangeability while improving efficiency and motion robustness.

Cardiac magnetic resonance (CMR) is the gold standard for non-invasive assessment of myocardial diseases, providing comprehensive information through e.g. cine imaging, T2-weighted sequences, and late gadolinium enhancement (LGE). Conventional CMR protocols typically rely on segmented (multi-shot) acquisitions over multiple heartbeats and require repeated breath-holds, which can limit patient comfort and compliance. While these segmented sequences offer high spatial resolution, they are prone to motion and respiratory artifacts-particularly in patients with arrhythmias or dyspnea-and contribute to long total examination times. Recent advances in deep learning (DL) reconstruction techniques have enabled substantial acceleration of segmented CMR sequences, particularly for cine and LGE imaging. These approaches effectively reduce acquisition time but still rely on regular cardiac rhythm and adequate breath-holding capacity, limiting their applicability in more challenging patient populations. In contrast, single beat (or: single-shot) imaging acquires data within a single heartbeat, offering a motion-robust alternative, though at the cost of lower spatial resolution. Efforts to streamline CMR are ongoing, with some studies proposing to reduce comprehensive exam times to 30 minutes or less. In parallel, full DL-based reconstruction MRI protocols are being increasingly explored across MRI domains, including neuroimaging and musculoskeletal imaging. Applying deep learning super-resolution to CMR, particularly in combination with single-beat acquisitions with the option of free-breathing acquisition, may enhance both speed and robustness. This prospective investigates whether a deep learning-based single-beat super-resolution CMR protocol - including single-shot cine, T2-STIR, and LGE sequences in both short- and long-axis views - can provide diagnostic interchangeability to a standard segmented protocol. All participants undergo both protocols during the same exam session. Total scan times are compared between protocols using Student's t-test. Three blinded readers evaluate predefined diagnostic categories including wall motion abnormalities, pericardial effusion, myocardial edema, LGE, and the final CMR diagnosis. Generalized estimating equations with bootstrapped 95% confidence intervals and a predefined equivalence margin of ±5% was used for the interchangeability analysis. Agreement in categorical ratings was evaluated using Cohen's Kappa and Fleiss' Kappa, as appropriate. Diagnostic confidence was rated on a 5-point Likert scale.