Title: Automated motion artifact detection in early pediatric diffusion MRI using a convolutional neural network
Citation: Weaver, J. M., DiPiero, M., Rodrigues, P. G., Cordash, H., Davidson, R. J., Planalp, E. M., & Dean, D. C., 3rd (2023). Automated motion artifact detection in early pediatric diffusion MRI using a convolutional neural network. Imaging neuroscience (Cambridge, Mass.), 1, 10.1162/imag_a_00023. https://doi.org/10.1162/imag_a_00023
Abstract: Diffusion MRI (dMRI) is a widely used method to investigate the microstructure of the brain. Quality control (QC) of dMRI data is an important processing step that is performed prior to analysis using models such as diffusion tensor imaging (DTI) or neurite orientation dispersion and density imaging (NODDI). When processing dMRI data from infants and young children, where intra-scan motion is common, the identification and removal of motion artifacts is of the utmost importance. Manual QC of dMRI data is (1) time-consuming due to the large number of diffusion directions, (2) expensive, and (3) prone to subjective errors and observer variability. Prior techniques for automated dMRI QC have mostly been limited to adults or school-age children. Here, we propose a deep learning-based motion artifact detection tool for dMRI data acquired from infants and toddlers. The proposed framework uses a simple three-dimensional convolutional neural network (3DCNN) trained and tested on an early pediatric dataset of 2,276 dMRI volumes from 121 exams acquired at 1 month and 24 months of age. An average classification accuracy of 95% was achieved following four-fold cross-validation. A second dataset with different acquisition parameters and ages ranging from 2-36 months (consisting of 2,349 dMRI volumes from 26 exams) was used to test network generalizability, achieving 98% classification accuracy. Finally, to demonstrate the importance of motion artifact volume removal in a dMRI processing pipeline, the dMRI data were fit to the DTI and NODDI models and the parameter maps were compared with and without motion artifact removal.
Investigator: Doug Dean III, PhD
About the Lab: Doug’s research in the Developing Brain Imaging Lab, focuses on the development and application of novel quantitative magnetic resonance imaging (MRI) methods to measure and evaluate the brain structure throughout early neurodevelopment and aging. He is particularly experienced in pediatric imaging, scanning more than 800 infants and young children (3 months to 12 years) during his graduate work. His current research is focused on examining how the white matter microstructure of the brain develops across the early development and how these microstructural processes are related to changes in cognition and behavior.