Radial Approach for Dipole Inversion (RADI): A deep learning-based 2D approach to quantitative susceptibility mapping

Motivation: Previous methods are time-consuming or require a lot of computational resources to solve the dipole inversion problem in Quantitative Susceptibility Mapping (QSM). Goal(s): To propose a 2D U-Net-based approach called "Radial Approach for Dipole Inversion (RADI)". Approach: In RADI, planes orthogonal to the B0 direction are radially sampled. A 2D U-Net-based model was trained to output 2D artificially simulated susceptibility from the calculated local field map. The model was used to process planes from brain images. Results: The correlation coefficient between RADI and Morphology Enabled Dipole Inversion (MEDI) was 0.680. In RADI, streak and susceptibility artifacts were suppressed. Impact: We propose a 2D deep learning-based approach to the dipole inversion in QSM called "Radial Approach for Dipole Inversion (RADI)". RADI can accelerate the calculation process of QSM and reduce artifacts that appear in conventional methods.