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

Manual tracing of magnetic resonance imaging (MRI) represents the gold standard for segmentation in clinical neuropsychiatric research studies, however automated approaches are increasingly used due to its time limitations. The accuracy of segmentation techniques for subcortical structures has not been systematically investigated in large samples. We compared the accuracy of fully automated [(i) model-based: FSL-FIRST; (ii) patch-based: volBrain], semi–automated (FreeSurfer) and stereological (Measure®) segmentation techniques with manual tracing (ITK-SNAP) for delineating volumes of the caudate (easy-to-segment) and the hippocampus (difficult-to-segment). High resolution 1.5 T T1-weighted MR images were obtained from 177 patients with major psychiatric disorders and 104 healthy participants. The relative consistency (partial correlation), absolute agreement (intraclass correlation coefficient, ICC) and potential technique bias (Bland–Altman plots) of each technique was compared with manual segmentation. Each technique yielded high correlations (0.77–0.87, p < 0.0001) and moderate ICC’s (0.28–0.49) relative to manual segmentation for the caudate. For the hippocampus, stereology yielded good consistency (0.52–0.55, p < 0.0001) and ICC (0.47–0.49), whereas automated and semi-automated techniques yielded poor ICC (0.07–0.10) and moderate consistency (0.35–0.62, p < 0.0001). Bias was least using stereology for segmentation of the hippocampus and using FreeSurfer for segmentation of the caudate. In a typical neuropsychiatric MRI dataset, automated segmentation techniques provide good accuracy for an easy-to-segment structure such as the caudate, whereas for the hippocampus, a reasonable correlation with volume but poor absolute agreement was demonstrated. This indicates manual or stereological volume estimation should be considered for studies that require high levels of precision such as those with small sample size.