Machine Learning for Large-Scale Quality Control of 3D Shape Models in Neuroimaging

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Anatomy

Background:

• Human quality assessment of MRI-derived data is a bottleneck in large-scale neuroimaging studies.
• Automated quality control methods using machine learning are underdeveloped for complex neuroimaging phenotypes.

Purpose of the Study:

• To optimize predictive models for assessing the quality of meshes representing deep brain structure shapes.
• To develop machine learning classifiers capable of detecting low-quality neuroimaging data.

Main Methods:

• Utilized vertex-wise and global shape features from meshes across 19 cohorts and over 7500 subjects.
• Trained kernelized Support Vector Machine and Gradient Boosted Decision Trees classifiers.
• Validated model generalization across diverse datasets and disease conditions.

Main Results:

• Achieved significant reduction in human workload for quality assessment, ranging from 30-70%.
• Models demonstrated generalization across datasets and diseases.
• Recall rates for detecting failing quality meshes approached inter-rater reliability.

Conclusions:

• Machine learning models can effectively automate neuroimaging data quality control.
• Optimized models reduce manual assessment time and maintain high accuracy.
• This approach addresses a critical bottleneck in large-scale neuroimaging research.