A generic registration‐assisted framework for dynamic magnetic resonance imaging super‐resolution with misaligned training data

BACKGROUND: Deep learning-based super-resolution is a promising solution for restoring dynamic magnetic resonance imaging (dMRI) from undersampled acquisitions; however, physiological motion precludes obtaining perfectly aligned low- and high-resolution (LR-HR) training pairs. Previous studies have circumvented this limitation by retrospectively simulating LR from HR images, introducing a domain gap that impairs performance on real data. PURPOSE: This study proposes a generic registration-assisted super-resolution framework (RegSR) that enables direct supervised learning on clinically-acquired, misaligned LR-HR pairs. METHODS: Inspired by our finding that improved registration accuracy directly enhances super-resolution fidelity, we incorporate three improvements into our framework. First, RegSR capitalizes on a synergistic interplay where super-resolution outputs reduce image-quality discrepancies and registration dynamically corrects spatial offsets, enabling pixel-level supervision for both tasks to enhance their performance mutually. Furthermore, we introduce a multi-scale recursive registration network (MRReg) that estimates deformation fields in a deep-to-shallow manner over feature maps, yielding precise spatial corrections amidst noise and artifacts in LR images. Finally, a dual-coordinate training scheme is designed to decouple super-resolution and registration, ensuring that each module specializes exclusively in its role without functional interference. RESULTS: Evaluations using an abdominal four-dimensional MRI dataset (20 training, six validation cases) and a cardiac cine MRI dataset (100 training, 50 validation cases) show that RegSR significantly outperforms state-of-the-art methods in both structural fidelity and visual realism. Quantitatively, RegSR reduces the MAE mean absolute error by 8.15%, increases the structural similarity index by 3.47% and the peak signal-to-noise ratio PSNR by 2.48%, achieves the best learned perceptual image patch similarity LPIPS score, and ranks second in the natural image quality evaluator. Moreover, RegSR is compatible with diverse super-resolution backbones, consistently improving their performance under misaligned training conditions. CONCLUSIONS: RegSR provides a robust, generalizable solution for supervised super-resolution training on real-world dMRI, effectively addressing motion-induced misalignment while enhancing reconstruction quality.