Automated Multiclass Artifact Detection in Diffusion MRI Volumes via 3D Residual Squeeze-and-Excitation Convolutional Neural Networks

Abstract

Diffusion MRI (dMRI) is widely used to investigate neuronal and structural development of brain. dMRI data is often contaminated with various types of artifacts. Hence, artifact type identification in dMRI volumes is an essential pre-processing step prior to carrying out any further analysis. Manual artifact identification amongst a large pool of dMRI data is a highly labor-intensive task. Previous attempts at automating this process are often limited to a binary classification (“poor” vs. “good” quality) of the dMRI volumes or focus on detecting a single type of artifact (e.g., motion, Eddy currents, etc.). In this work, we propose a deep learning-based automated multiclass artifact classifier for dMRI volumes. Our proposed framework operates in 2 steps. In the first step, the model predicts labels associated with 3D mutually exclusive collectively exhaustive (MECE) sub-volumes or “slabs” extracted from whole dMRI volumes. In the second step, through a voting process, the model outputs the artifact class present in the whole volume under investigation. We used two different datasets for training and evaluating our model. Specifically, we utilized 2,494 poor-quality dMRI volumes from the Adolescent Brain Cognitive Development (ABCD) and 4,226 from the Healthy Brain Network (HBN) dataset. Our results demonstrate accurate multiclass volume-level main artifact type prediction with 96.61 and 97.52% average accuracies on the ABCD and HBN test sets, respectively. Finally, in order to demonstrate the effectiveness of the proposed framework in dMRI pre-processing pipelines, we conducted a proof-of-concept dMRI analysis exploring the relationship between whole-brain fractional anisotropy (FA) and participant age, to test whether the use of our model improves the brain-age association.

Review of Image Classification Algorithms Based on Graph Convolutional Networks

“…The algorithm allows the network to recalibrate features. Since the SE block is relatively simple, it has been applied to many CNNs to improve the performance of the model [58][59][60][61]. Although the SE block is currently less used on GCNs, it may be an option to consider from the perspective of incorporating attention mechanisms without adding too much to the model complexity.…”