MRI super-resolution reconstruction using efficient diffusion probabilistic model with residual shifting.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Magnetic resonance imaging (MRI) offers excellent soft-tissue contrast but suffers from long acquisition times, causing patient discomfort and motion artifacts.
• Current deep learning super-resolution (SR) methods for MRI reconstruction require extensive sampling, hindering real-time applications.

Purpose of the Study:

• To develop a novel diffusion-based SR framework, Res-SRDiff, that accelerates MRI reconstruction by reducing sampling steps.
• To maintain high-fidelity anatomical details and improve computational efficiency in MRI.

Main Methods:

• Introduced Res-SRDiff, a diffusion-based SR framework with a residual error-shifting mechanism integrated into the forward diffusion process.
• Evaluated Res-SRDiff on brain T1 MP2RAGE and prostate T2-weighted images, comparing it against established SR methods.
• Utilized quantitative metrics (PSNR, SSIM, GMSD) and qualitative assessments, including an ablation study and Likert-scale image quality evaluation.

Main Results:

• Res-SRDiff significantly outperformed comparison methods in PSNR, SSIM, and GMSD for both datasets (p<0.05).
• Achieved high-fidelity reconstruction with only four sampling steps, reducing reconstruction time to under one second per slice.
• Qualitative analysis confirmed preservation of fine anatomical details and lesion morphologies; Likert scores were highest for Res-SRDiff.

Conclusions:

• Res-SRDiff demonstrates superior efficiency and accuracy in MRI reconstruction, significantly improving computational speed and image quality.
• The residual error-shifting mechanism enhances diffusion-based SR, enabling rapid and robust high-resolution image reconstruction.
• This advancement holds potential for improving clinical MRI workflows and accelerating medical imaging research.