Cyclic Self-Supervised Diffusion for Ultra Low-Field to High-Field MRI Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Low-field MRI offers cost and safety benefits but yields low-resolution images.
• Synthesizing high-field MRI from low-field data is crucial but faces a clinical fidelity gap.
• Existing methods struggle with anatomical fidelity, fine structural details, and image contrast differences.

Purpose of the Study:

• To develop a novel framework for synthesizing high-field MRI from low-field data.
• To address the clinical fidelity gap in MRI synthesis.
• To improve anatomical accuracy and structural detail in synthesized MRI images.

Main Methods:

• Proposed a cyclic self-supervised diffusion (CSS-Diff) framework.
• Incorporated cycle-consistent constraints for anatomical preservation.
• Introduced slice-wise gap perception and local structure correction networks.

Main Results:

• Achieved state-of-the-art performance in cross-field MRI synthesis (PSNR, SSIM, LPIPS).
• Significantly improved preservation of fine-grained anatomical structures (e.g., white matter, cortex).
• Demonstrated quantitative reliability and anatomical consistency in synthesized images.

Conclusions:

• CSS-Diff effectively synthesizes high-field MRI from low-field data.
• The framework bridges the clinical fidelity gap, enhancing image quality and anatomical accuracy.
• The method holds potential for reducing reliance on costly MRI acquisitions and expanding data availability.