Supporting data for " Physics-informed deep learning reconstruction in medical imaging"

Using simulated data, one deep-learning QSM pipeline consisting of a background field removal network (POCSnet1) and a dipole inversion network (POCSnet2) was proposed. On COSMOS (N=1), results from both models were improved compared to those from other models evaluated, and the evaluations on clinical data were also performed. This pipeline was further developed into a deep learning-regularized, single-step QSM quantification model, SS-POCSnet. A data fidelity term based on a single-step model was iteratively applied with SS-POCSnet regularized susceptibility maps. On synthetic datasets (N=10), SS-POCSnet showed the best performance and also reduced the underestimation of susceptibility values in deep brain nuclei compared with other models considered. Furthermore, this model was sensitive to cerebral microbleed /calcification (N=24) and multiple sclerosis lesions (N=10), demonstrating its clinical applicability.