REGION SPECIFIC AUTOMATIC QUALITY ASSURANCE FOR MRI-DERIVED CORTICAL SEGMENTATIONS.

Area of Science:

• Biomedical imaging
• Neuroscience
• Machine learning

Background:

• Quality control (QC) is crucial for biomedical data analysis, especially in medical imaging.
• Automated brain MRI segmentation tools exist but may produce anatomically incorrect labels.
• Traditional QC methods focusing on statistical outliers may incorrectly flag accurate data.

Purpose of the Study:

• To develop and evaluate machine learning models for assessing the anatomical accuracy of human brain MRI parcellations.
• To improve the reliability of automated segmentation in large-scale neuroimaging studies.

Main Methods:

• Utilized a database of over 12,000 human brain MRIs with 68 cortical parcellations.
• Human raters assessed the anatomical accuracy of each parcellation.
• Trained three machine learning models to classify parcellations as 'pass' (anatomically accurate) or 'fail'.
• Tested model performance on an independent dataset.

Main Results:

• Machine learning models demonstrated good performance in classifying the anatomical accuracy of most labeled brain regions.
• The models successfully distinguished between accurate and inaccurate segmentations.

Conclusions:

• Machine learning offers a robust approach for quality control in human brain MRI segmentation.
• This method enhances the anatomical accuracy of large-scale, multi-site neuroimaging research.