A generic registration-assisted framework for dynamic magnetic resonance imaging super-resolution with misaligned...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Deep learning for dynamic MRI (dMRI) super-resolution faces challenges with motion-induced misalignment between low- and high-resolution (LR-HR) training pairs.
• Previous methods simulating LR from HR images create a domain gap, hindering real-world performance.

Purpose of the Study:

• Propose a generic registration-assisted super-resolution framework (RegSR) for direct supervised learning on misaligned clinical dMRI data.
• Enable robust training on real-world, motion-corrupted dMRI datasets.

Main Methods:

• RegSR employs a synergistic approach where super-resolution and registration mutually enhance each other.
• Introduces a multi-scale recursive registration network (MRReg) for precise spatial correction of LR images.
• Utilizes a dual-coordinate training scheme to ensure specialized, interference-free training for super-resolution and registration modules.

Main Results:

• RegSR significantly outperforms state-of-the-art methods on abdominal and cardiac dMRI datasets.
• Achieved quantitative improvements: 8.15% reduction in MAE, 3.47% increase in SSIM, 2.48% increase in PSNR.
• Demonstrated superior performance in perceptual quality metrics (LPIPS) and ranked highly in NIQE.

Conclusions:

• RegSR offers a robust and generalizable solution for supervised super-resolution training on real-world dMRI.
• Effectively addresses motion-induced misalignment and enhances overall reconstruction quality.