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

Authors

Abstract

Quantitative Susceptibility Mapping (QSM) is a magnetic resonance imaging
technique that quantifies tissue magnetic susceptibility by solving an ill-posed inverse
problem from measured magnetic field perturbations. Its accuracy is fundamentally limited
by the non-local dipole kernel and the lack of a universally accurate and generalizable prior.
Approach. We propose Flow-QSM, a physics-guided conditional flow-matching framework
for efficient and accurate QSM reconstruction. The method first learns a generative prior of
susceptibility maps via unconditional flow matching. It then performs a physics-guided
reverse sampling process, where the generative prior is conditioned and corrected by the
physical forward model through a predicted flow velocity field. This establishes a
probabilistic bridge between the prior and posterior distributions, ensuring an optimal
balance between data fidelity and learned anatomical constraints. Specifically, we design a
customized architecture featuring (i) a patch-wise positional encoding mechanism to
maintain global spatial coherence across partitioned sub-volumes, and (ii) a dual branch
skip-backbone modulation strategy to adaptively fuse spatially enriched and
frequency-refined features for high-fidelity detail recovery. Comprehensive
evaluations on in vivo human brain datasets suggest that Flow-QSM achieves improved
accuracy and artifact suppression compared with representative QSM reconstruction
methods. Its performance is further examined on high-resolution data, clinical data, and
out-of-domain cases, where it demonstrates consistent behavior while maintaining efficient
inference time. Flow-QSM provides a unified probabilistic framework that
integrates learned generative priors with physics-based constraints for susceptibility
mapping. The proposed approach offers a flexible and principled strategy for addressing
ill-posed inverse problems in computational imaging.

Tags

Topics