DARQ

DARQ evaluates stereotaxic registrations of T1-weighted brain MRI scans using a deep neural network to automate quality control of linear registration in neuroimaging workflows.

Key Features:

• Automation: Automates QC of stereotaxic registration for T1-weighted brain MRI and can replace manual human assessment of registration quality.
• Deep Learning-Based Approach: Employs a deep neural network trained on 9,325 MRI scans and 64,476 registrations from publicly available sources to detect registration failures.
• High Accuracy and Reliability: Demonstrated 89% accuracy and an 85% true negative rate (15% false positive rate) in balanced cross-validation, and 96.1% accuracy with a 95.5% true negative rate (4.5% false positive rate) on an independent multiple sclerosis dataset, comparable to manual QC test-retest accuracy of 93%.
• Robustness and Generalizability: Validated across multiple datasets, including patients with multiple sclerosis, indicating consistent performance across different data sources.

Scientific Applications:

• Registration QC in Preprocessing: Quality control of linear registration to stereotaxic space as an initial step in automated neuroimaging pipelines.
• Large-Scale Neuroimaging Studies: Automated evaluation of registration quality for extensive datasets where manual QC is impractical.

Methodology:

Trains a deep neural network on labels from human raters who assessed registrations produced by seven different linear registration tools, with performance validated using balanced cross-validation and an independent multiple sclerosis dataset.