On-going, large-scale neuroimaging initiatives have produced many MRI datasets with hundreds, even thousands, of individual participants and scans. These databases can aid in uncovering neurobiological causes and correlates of poor mental health, disease pathology, and many other important factors. While volumetric quantification of brain structures can be completed by expert hand-tracing, automated segmentations are becoming the only truly tractable approach for particularly large datasets.
Here, we assessed the spatial and numerical reliability for newly-deployed automated segmentation of hippocampal subfields and amygdala nuclei in FreeSurfer. In a sample of participants with repeated structural imaging scans (N=118), we found numerical reliability (as assessed by intraclass correlations) to be generally high, with 92% of the subregions having ICCs above 0.90 and the remainder still above 0.75. Spatial reliability was lower with only 11% of regions having Dice coefficients above 0.90, but 70% with Dice coefficients above 0.75. Of particular concern, three regions, the hippocampal fissure, the anterior amygdaloid area, and the paralaminar nucleus, had only moderate spatial reliability (0.50-0.75). We also examined correlations between spatial reliability and person-level factors (e.g., age, inter-scan interval, and difference in image quality). For these factors, interscan interval and image quality were related to variations in spatial reliability. Examined collectively, our work suggests strong numerical and spatial reliability for the majority of hippocampal and amygdala subdivisions; however, caution should be exercised for a few regions with more variable reliability.
bioRxiv Subject Collection: Neuroscience