A critical assessment of LGN segmentation quality based on volume, contour, and location

Abstract

Reliable segmentation of the lateral geniculate nucleus (LGN), a small thalamic relay nucleus essential for visual processing, is necessary for characterizing its anatomy in healthy populations and structural alterations in neurological disorders. However, accurate delineation remains challenging on routinely acquired T1-weighted images because the LGN is small, deeply located, and has limited contrast relative to surrounding thalamic tissue. Although manual segmentation remains the reference standard, it is time consuming and susceptible to rater variability, limiting its use in large datasets. Automated approaches may therefore provide a more scalable alternative. In this study, we performed a systematic benchmarking of seven automated approaches: FreeSurfer, NextBrain, DeepThalamus, HIPS-THOMAS, MASSP 2.0, BigBrain alignment, and manual novice segmentation against expert-defined ground truth. We applied a multi-metric evaluation incorporating the Dice similarity coefficient, Hausdorff distance and balanced average Hausdorff distance, volumetric similarity, and Euclidean distance. Together, these measures evaluate overlap, contour, volume, and location against expert manual segmentations. Our results indicate that HIPS-THOMAS and DeepThalamus outperformed the other automated methods in all metrics. However, manual segmentation retained high precision despite exhibiting greater inter-subject variance than the top-performing automated segmentation tools. The different metrics do not necessarily report the same order of quality, highlighting the need of using multiple metrics for evaluation segmentation quality.