The Boundary versus The Core: A Quantitative Comparison of TotalSegmentator and MOOSE Performance in Medical Image Segmentation

The rise of total-body PET/CT imaging necessitates robust automated segmentation tools, fueling a debate between using general-purpose "foundation" models and domain-specific "specialised" models. This paper presents a rigorous, quantitative comparison between TotalSegmentator, a foundation model trained on over 1,200 CT scans, and MOOSE, a specialised tool designed for PET/CT analysis. We evaluated both models on public datasets for spleen and vertebrae segmentation, using a comprehensive suite of metrics covering both volumetric overlap (Dice, Jaccard) and boundary precision (Hausdorff Distance). Our results reveal a critical trade-off: TotalSegmentator delivers significantly superior boundary accuracy for all tested structures, making it the premier choice for precision-critical tasks, such as radiotherapy planning. Conversely, MOOSE achieves higher volumetric overlap for vertebrae, suggesting better performance in capturing the core anatomical volume. However, this comes at the cost of severe and frequent boundary errors. We conclude that TotalSegmentator has better performance in Total-body PET/CT images segmentation. The optimal choice is application-dependent, requiring a careful balance between the need for precise surface delineation and accurate core volume assessment. Our findings provide evidence-based guidance for researchers and clinicians in selecting the appropriate AI tool for their specific quantitative imaging needs.