Unsupervised Susceptibility Artifact Correction in DTI Using a Deep Learning Forward-Distortion Model

Diffusion weighted imaging (DWI) requires correction of susceptibility artifacts before conducting quantitative analyses. Correction is typically performed by acquiring DWI images in reversed phase-encode directions, which are used to estimate and correct for the effects of susceptibility-induced field. In this work, we propose a Forward-Distortion Network (FD-Net) for correcting susceptibility artifacts at multiple b-values. We evaluate the quality of the corrected DWI images and Diffusion Tensor Imaging (DTI) metrics, using FSL’s TOPUP as a reference classical method. In addition to rapid execution times, FD-Net exhibits high-fidelity performance for both DWI images and DTI metrics.