Automated subcortical segmentation using FIRST: test-retest reliability, interscanner reliability, and comparison to...

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Automated segmentation of magnetic resonance images (MRIs) is crucial for quantitative analysis.
• Assessing the reliability of automated tools (test-retest, interscanner, manual comparison) is essential for clinical and research applications.
• The FSL/FIRST tool is a widely used automated subcortical segmentation method.

Purpose of the Study:

• To evaluate the reliability and consistency of the FSL/FIRST automated subcortical segmentation tool.
• To assess performance across different scanning platforms, sessions, and in patient populations.
• To compare automated segmentation with manual segmentation for the hippocampus.

Main Methods:

• Retrospective analysis of 190 MRI scans from 87 subjects (mood/anxiety disorders and healthy volunteers).
• Scans were acquired on the same or different MRI platforms.
• Comparison with manual segmentation of the hippocampus was performed on a subset of scans.

Main Results:

• High reliability (Intraclass Correlation Coefficient > 0.83) for thalamus, caudate, putamen, hippocampus, and pallidum across sessions on the same scanner.
• Inter-platform reliability varied (0.527 < ICC < 0.953), with lower values for hippocampus between different coil types.
• Accumbens and amygdala segmentation showed generally unreliable results.
• Acceptable ICC (> 0.7) for hippocampal volumes between automated and manual segmentation; significant gender differences detected by both.
• FIRST segmentations showed consistency with manual segmentations in the left caudate and bilateral putamen.

Conclusions:

• FSL/FIRST demonstrates good reliability for most subcortical structures in multisite or meta-analysis settings, particularly within the same scanner.
• Reliability is reduced between different scanning platforms and for specific regions like the hippocampus with certain coil configurations.
• Segmentation of the accumbens and amygdala requires caution due to observed unreliability.
• The tool shows acceptable agreement with manual segmentation and can detect clinically relevant differences, validating its use in patient studies.