A comparative study of segmentation techniques for the quantification of brain subcortical volume

Area of Science:

• Neuroimaging
• Neuropsychiatry
• Medical Image Analysis

Background:

• Manual tracing of MRI is the gold standard for brain structure segmentation in research.
• Automated segmentation methods are increasingly used due to time constraints.
• Systematic investigation of automated segmentation accuracy in large neuropsychiatric samples is lacking.

Purpose of the Study:

• To compare the accuracy of automated, semi-automated, and stereological segmentation techniques against manual tracing.
• To evaluate segmentation accuracy for the caudate (easy) and hippocampus (difficult) structures.
• To assess consistency, agreement, and bias of different segmentation methods in neuropsychiatric MRI.

Main Methods:

• Compared FSL-FIRST, volBrain, FreeSurfer, and Measure® against manual ITK-SNAP segmentation.
• Utilized high-resolution T1-weighted MRI scans from 177 psychiatric patients and 104 healthy controls.
• Analyzed relative consistency (partial correlation), absolute agreement (ICC), and bias (Bland-Altman plots).

Main Results:

• All techniques showed high correlation for caudate segmentation, with moderate ICCs.
• Stereology demonstrated good consistency and ICC for hippocampus segmentation.
• Automated/semi-automated methods had poor ICC but moderate consistency for the hippocampus; FreeSurfer showed least bias for the caudate.

Conclusions:

• Automated segmentation offers good accuracy for easy-to-segment structures like the caudate in neuropsychiatric MRI.
• For the hippocampus, automated methods show moderate correlation but poor absolute agreement.
• Manual or stereological volume estimation is crucial for studies demanding high precision, especially with small sample sizes.