Flow-QSM: bridging learned priors and physical models for quantitative susceptibility mapping.

Area of Science:

• Medical Imaging
• Computational Physics
• Machine Learning

Background:

• Quantitative Susceptibility Mapping (QSM) quantifies tissue magnetic susceptibility using MRI.
• Current QSM methods face limitations due to ill-posed inverse problems and lack of generalizable priors.
• Accuracy is fundamentally limited by non-local dipole kernels and prior knowledge.

Purpose of the Study:

• To introduce Flow-QSM, a physics-guided conditional flow-matching framework for efficient and accurate QSM reconstruction.
• To develop a method that balances data fidelity with learned anatomical constraints for improved QSM.
• To address the challenges of ill-posed inverse problems in computational imaging.

Main Methods:

• Learned a generative prior of susceptibility maps via unconditional flow matching.
• Employed a physics-guided reverse sampling process conditioning the prior with the physical forward model.
• Utilized a customized architecture with patch-wise positional encoding and dual branch skip-backbone modulation.

Main Results:

• Flow-QSM demonstrated improved accuracy and artifact suppression compared to existing QSM methods on in vivo human brain datasets.
• Consistent performance was observed across high-resolution, clinical, and out-of-domain datasets.
• The method maintains efficient inference times.

Conclusions:

• Flow-QSM offers a unified probabilistic framework integrating learned priors with physics-based constraints for susceptibility mapping.
• The approach provides a flexible and principled strategy for solving ill-posed inverse problems in imaging.
• This method advances the field of quantitative susceptibility mapping with improved accuracy and efficiency.