Joint Motion Estimation with Geometric Deformation Correction for Fetal Echo Planar Images via Deep Learning

Abstract

In this paper, we introduce a novel end-to-end predictive model for efficient fetal motion correction using deep neural networks. Diverging from conventional methods that estimate fetal brain motions and geometric distortions separately, our approach introduces a newly developed joint learning framework that not only reliably estimates various degrees of rigid movements, but also effectively corrects local geometric distortions of fetal brain images. Specifically, we first develop a method to learn rigid motion through a closed-form update integrated into network training. Subsequently, we incorporate a diffeomorphic deformation estimation model to guide the motion correction network, particularly in regions where local distortions and deformations occur. To the best of our knowledge, our study is the first to simultaneously track fetal motion and address geometric deformations in fetal echo-planar images. We validated our model using real fetal functional magnetic resonance imaging data with simulated and real motions. Our method demonstrates significant practical value to measure, track, and correct fetal motion in fetal MRI.