SACNet: A Multiscale Diffeomorphic Convolutional Registration Network with Prior Neuroanatomical Constraints for Flexible Susceptibility Artifact Correction in Echo Planar Imaging

Abstract

Susceptibility artifacts (SAs), inevitable in brain diffusion MR (dMRI) scans acquired using single-shot echo planar imaging (EPI), severely compromise the accurate detection of human brain structure. Existing SA correction (SAC) methods offer inadequate correction quality and limited applicability across diverse datasets with varied acquisition protocols. To address these challenges, we proposed SACNet, a SAC framework based on unsupervised registration convolutional networks, featuring: i) a novel diffeomorphism regularization function to avoid unnatural SAC warps, modified from a potential well function; ii) an integration with prior neuroanatomical constraints and coarse-to-fine processing strategy to enables multi-scale geometric and intensity recoveries in severe distorted areas; iii) a unified registration framework that incorporates multiple phase-encoding (PE) EPI images and structural images, ensuring compatibility with both single- and inverse-PE protocols, with or without field maps. Utilizing simulated dMRI images and over 2000 brain scans from neonatal, child, adult and traveling participants, our method consistently demonstrates state-of-the-art correction performance. Notably, SACNet effectively reduces SAs-related multicenter effects compared to existing methods. We have developed user-friendly tools using containerization techniques, hope to facilitate SAC correction quality across extensive neuroimaging studies.