The quest for the best: manual, atlas- and spatial prior-based delineation of locus coeruleus

Despite advances in neuromelanin-sensitive brain imaging and a plethora of manual labelling, atlas- and signal intensity-based software solutions, the reliable non-invasive delineation of the locus coeruleus (LC) in the human brain stem remains challenging. We sought to evaluate the spatial accuracy and consistency of atlas-and probabilistic spatial prior-based LC delineation. We acquired neuromelanin-sensitive magnetic resonance imaging data in healthy volunteers (n = 24; mean age 40.0 ± 16.8 years; 42% females). Manual labelling by 9 raters performed twice provided the basis for individual- and group-level comparisons with the automated delineation methods. For the atlas-based labelling, we separately tested seven open-access LC atlases, the averaged output of the manual labelling, and a consensus reference representing the atlases’ overlap. Each one of the atlases served as spatial prior for automated LC delineation in a probabilistic segmentation framework. Manual labelling showed moderate inter-rater agreement (mean Dice = 0.7), with higher delineation variability in the rostral and caudal LC. For the atlas-based labelling we observed a low spatial concordance for all open-access atlases (Dice = 0.2-0.4) with inconsistent boundary accuracy and volume similarity indices relative to the consensus reference. In the same comparison, the averaged manual labelling atlas showed higher spatial overlap (Dice = 0.6). Probabilistic delineation using spatial priors showed the strongest voxel-wise similarity with manual labelling (correlation coefficient r = 0.3) when the averaged manual labelling atlas was used as prior. Principal component analysis confirmed the greater spatial compactness for atlas-based labelling in comparison with spatial prior-based delineation, underscoring method-dependent differences in anatomical organization. Our results highlight the potential advantage of atlas-based labelling for robust and spatially reliable LC identification. The observed variability across methods and atlases calls for harmonised validation strategies and context-sensitive approaches that improve reliability.