Robust diffusion parametric mapping of motion-corrupted data with a three-dimensional convolutional neural network

TL;DR

This paper introduces a 3D convolutional neural network-based method to robustly recover diffusion MRI measures from motion-corrupted data, significantly reducing motion effects and improving data utilization in diffusion imaging studies.

Contribution

The study extends a hierarchical 3D CNN approach to handle motion-contaminated diffusion MRI data, incorporating motion assessment and corrupted volume rejection for improved accuracy.

Findings

Diffusion measures from the new pipeline are minimally affected by motion.

The method retains as few as eight volumes from contaminated data with comparable accuracy.

It effectively reduces spurious group differences caused by head motion.

Abstract

Head motion is inevitable in the acquisition of diffusion-weighted images, especially for certain motion-prone subjects and for data gathering of advanced diffusion models with prolonged scan times. Deficient accuracy of motion correction cause deterioration in the quality of diffusion model reconstruction, thus affecting the derived measures. This results in either loss of data, or introducing bias in outcomes from data of different motion levels, or both. Hence minimizing motion effects and reutilizing motion-contaminated data becomes vital to quantitative studies. We have previously developed a 3-dimensional hierarchical convolution neural network (3D H-CNN) for robust diffusion kurtosis mapping from under-sampled data. In this study, we propose to extend this method to motion-contaminated data for robust recovery of diffusion model-derived measures with a process of motion assessment and corrupted volume rejection. We validate the proposed pipeline in two in-vivo datasets. Results from the first dataset of individual subjects show that all the diffusion tensor and kurtosis tensor-derived measures from the new pipeline are minimally sensitive to motion effects, and are comparable to the motion-free reference with as few as eight volumes retained from the motion-contaminated data. Results from the second dataset of a group of children with attention deficit hyperactivity disorder demonstrate the ability of our approach in ameliorating spurious group differences due to head motion. This method shows great potential for exploiting some valuable but motion-corrupted DWI data which are likely to be discarded otherwise, and applying to data with different motion level thus improving their utilization and statistic power.