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

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Image Analysis

Background:

• Diffusion MRI (dMRI) is crucial for studying brain development but is susceptible to artifacts.
• Manual artifact identification in dMRI is time-consuming and previous automated methods are limited.
• Accurate artifact classification is essential for reliable dMRI data analysis.

Purpose of the Study:

• To develop a deep learning-based automated multiclass artifact classifier for dMRI volumes.
• To improve the efficiency and accuracy of dMRI pre-processing pipelines.
• To evaluate the model's performance on large-scale datasets.

Main Methods:

• A two-step deep learning framework was proposed for artifact classification.
• The model predicts artifact labels on 3D sub-volumes (slabs) and uses a voting process for whole-volume classification.
• Trained and evaluated on 2,494 (ABCD) and 4,226 (HBN) poor-quality dMRI volumes.

Main Results:

• Achieved high average accuracies of 96.61% (ABCD) and 97.52% (HBN) for multiclass artifact prediction.
• Demonstrated the model's effectiveness in a proof-of-concept dMRI analysis.
• The proposed framework significantly improves dMRI pre-processing.

Conclusions:

• The developed deep learning model offers accurate and automated multiclass artifact identification in dMRI.
• This approach enhances the reliability of neuroimaging studies by improving data quality.
• The framework has the potential to streamline dMRI pre-processing workflows.