Abstract/Summary

The thalamus is a brain region formed from functionally distinct nuclei, which contribute in important ways to various cognitive processes. Yet, much of the human neuroscience literature treats the thalamus as one homogeneous region, and consequently the unique contribution of specific nuclei to behaviour remains under-appreciated. This is likely due in part to the technical challenge of dissociating nuclei using conventional structural imaging approaches. Yet, multiple algorithms exist in the neuroimaging literature for the automated segmentation of thalamic nuclei. One recent approach developed by Iglesias and colleagues (2018) generates segmentations by applying a probabilistic atlas to subject-space anatomical images using the FreeSurfer software. Here, we systematically validate the efficacy of this segmentation approach in delineating thalamic nuclei using Human Connectome Project data. We provide several metrics quantifying the quality of segmentations relative to the Morel stereotaxic atlas, a widely accepted anatomical atlas based on cyto- and myeloarchitecture. The automated segmentation approach generated boundaries between the anterior, lateral, posterior, and medial divisions of the thalamus. Segmentation efficacy, as measured by metrics of dissimilarity (Average Hausdorff Distance) and overlap (DICE coefficient) within groups was mixed. Regions were better delineated in anterior, lateral and medial thalamus than the posterior thalamus, however all the volumes for all segmented nuclei were significantly different to the corresponding region of the Morel atlas. These mixed results suggest users should exercise care when using this approach to study the structural or functional relevance of a given thalamic nucleus.