A robust deep learning method for quantitative susceptibility mapping using diffusion model with a time-travel and resampling refinement module

Motivation: Deep learning-based QSM reconstruction approaches often suffer from generalization problems. Goal(s): To develop a robust deep learning-based method for QSM reconstruction using diffusion models along with a time-travel and resampling refinement strategy. Approach: The diffusion prior is unconditionally trained using high-quality QSM images without explicit knowledge about the measurement. The physical constraint is plugged into the sampling process of the diffusion model by solving an inverse problem. A refinement strategy is proposed to apply the time-reverse and resampling strategy in the latter sampling steps. Results: The proposed method shows high-quality and robust QSM reconstruction results compared with supervised deep learning-based methods. Impact: We introduce a diffusion model-based method for QSM reconstruction by enforcing hard data consistency during inference. We also present a time-travel and resampling refinement module in the latter steps to enhance performance. Our approach enables robust and high-quality QSM reconstruction.