DeepSepSTI-R2*: R2-Free Anisotropic Susceptibility Source Separation in Susceptibility Tensor Imaging with Deep Learning

Motivation: Tissue magnetic susceptibility originates from multiple sources, including paramagnetic iron and diamagnetic myelin. Current techniques for source separation in susceptibility imaging assume the susceptibility to be isotropic and require time-consuming R2 measurements. Goal(s): To develop a method for anisotropic susceptibility source separation without requiring R2 measurements. Approach: We developed DeepSepSTI-R2*, a deep learning method that jointly estimates paramagnetic susceptibility scalar (expected from isotropic susceptibility of tissue iron), diamagnetic susceptibility tensor (expected from anisotropic susceptibility of myelin sheath), and R2 maps from GRE phase and R2* measurements. Results: DeepSepSTI-R2* shows comparable results to R2-based counterparts, while being more efficient and more practical clinically. Impact: DeepSepSTI-R2* enhances anisotropic susceptibility source separation by eliminating the need for extra R2 measurement, i.e., using only R2* and phase maps derived solely from gradient echo data, thereby significantly reducing scan time.