Alignment-Guided Forward-Distortion Model for Deep Unsupervised Correction of Susceptibility Artifacts in EPI

In magnetic resonance imaging (MRI), susceptibility-induced distortions pose a significant challenge for images acquired using echo planar imaging (EPI). Classical methods use two EPI images acquired in reverse phaseencoding (PE) directions to correct susceptibility artifacts. However, these methods suffer from long computation times, making them impractical for clinical usage. Recently, deep learning-based approaches have been proposed to enable a leap in computation efficiency for EPI susceptibility artifact correction. A vital consideration in reverse-PE-based correction is the need to take into account any potential subject motion between reversed-PE acquisitions. In this work, we propose an alignment-guided forward distortion network (agFD-Net) that accounts for subject motion during correction of susceptibility artifacts. Similar to its predecessor FD-Net, agFD-Net is trained in a physics-driven unsupervised fashion to estimate a single corrected image and a displacement field. In agFD-Net, a new pre-trained alignment network called AlignNet is plugged into the network architecture to facilitate motion correction. The results on experimental NIH dataset featuring realistic levels of motion demonstrate that agFD-Net provides rapid and high-fidelity artifact correction, while successfully accounting for subject motion.Clinical Relevance-EPI is the most commonly used sequence for diffusion MRI and functional MRI. While susceptibility artifacts in EPI require correction before any downstream evaluation, the long computation time of classical correction methods make them impractical for use in clinical settings. The proposed agFD-Net provides more than two orders of magnitude speed up in computational efficiency, making it a highly promising approach for use in clinical settings.