Advancing Quantitative Susceptibility Mapping With 2.5D Diffusion Models for Rapid Intracranial Hemorrhage Quantification.

purposeTo develop a generative diffusion model-based approach for robust and efficient quantitative susceptibility mapping (QSM) reconstruction in intracranial hemorrhage (ICH), applicable to both standard gradient echo (GRE) and rapid echo planar imaging (EPI) acquisitions.

methodsQSMDiff, an unsupervised diffusion model for 3D QSM dipole inversion, was proposed. Three volumetric partitioning strategies including 2D slices, 3D patches, and 2.5D slabs were evaluated, and the memory-efficient 2.5D slab approach was adopted to balance accuracy and efficiency while maintaining anatomical fidelity. A conditional sampling mechanism ensured consistency with measured local fields, and a three-stage training data-generation strategy combining public dataset, synthetic QSM, and ICH lesions simulated from in vivo patients was implemented to overcome data scarcity.

resultsQSMDiff achieved the best overall performance in simulation studies, with SSIM of 0.97 ± 0.07, RMSE of 0.04 ± 0.03, and HFEN of 4.49 ± 0.83, demonstrating superior structural fidelity and noise suppression. For in vivo ICH patients scanned with rapid EPI, QSMDiff showed strong agreement with SWI-QSM references (R

conclusionQSMDiff achieves high-quality and accurate QSM reconstruction from both standard GRE and rapid EPI scans for ICH assessment. By integrating a 2.5D training strategy with synthetic ICH augmentation, it delivers accurate and reliable susceptibility maps even from lower-quality acquisitions, offering a practical solution for fast and robust ICH assessment.