Diffusion-QSM: Diffusion Model With Time-Travel and Resampling Refinement for Quantitative Susceptibility Mapping.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Image Reconstruction
• Computational Imaging

Background:

• Quantitative susceptibility mapping (QSM) is a vital MRI technique for visualizing magnetic susceptibility variations in tissues.
• Current QSM reconstruction methods face challenges with data perturbations and generalization.
• Deep learning (DL) offers potential but often struggles with robustness and out-of-distribution data.

Purpose of the Study:

• To introduce Diffusion-QSM, a robust deep learning-based method for high-quality QSM reconstruction.
• To develop a method that generalizes well across diverse data perturbations.
• To improve the reliability and applicability of QSM in various clinical and research settings.

Main Methods:

• Developed Diffusion-QSM, a diffusion model incorporating a time-travel and resampling refinement module.
• Trained the diffusion prior unconditionally on high-quality QSM images to enhance generalization.
• Integrated physical constraints from the QSM forward model and measurements during inference to guide reconstruction.

Main Results:

• Diffusion-QSM demonstrated superior performance compared to traditional and unsupervised DL methods in simulation, in vivo, and ex vivo data.
• The method exhibited better generalization capabilities than supervised DL methods when processing out-of-distribution data.
• Experimental results confirmed high-quality QSM reconstruction under various perturbations like contrast, resolution, and scan direction.

Conclusions:

• Diffusion-QSM effectively unifies data-driven diffusion priors with subject-specific physics constraints for robust QSM reconstruction.
• The developed method bridges the generalization gap in deep learning for QSM.
• Diffusion-QSM shows significant potential for diverse, realistic applications due to its excellent quality and generalization capabilities.