Auto-prompting Foundation Models for Clinical Segmentation: The Case of Pathological Scapula

Automatic segmentation of the scapula in CT volumes is critical for surgical planning and other clinically relevant morphometric tasks, but especially pathological cases pose significant challenges due to high anatomical variability, complex geometry, and proximity to adjacent structures. Standard deep learning methods, including large-scale segmentation models such as TotalSegmentator, often fail on this task due to the scarcity of high-quality, annotated pathological data. Conversely, general-purpose foundation models such as SAM2 require manual prompting, precluding full automation. In this work, we introduce a novel, fully automatic, and prompt-free segmentation framework. Our key idea is to cascade existing tools, repurposing the imperfect output of an automatic segmentation model (TotalSegmentator) not as a final result, but as a mechanism to generate robust 3D prompts for a SAM-based foundation model. This approach creates a fully automatic pipeline that requires no user interaction or task-specific fine-tuning. We evaluated our framework on a challenging dataset of about 40 expert-annotated pathological scapulae. Quantitative and qualitative results demonstrate that our method significantly outperforms both the baseline automatic segmentation model and prompted general-purpose medical foundation models (MedSAM2), reaching an average Dice Score of 92.40%, compared to 81.24% of TotalSegmentator and 82.41% of MedSAM2. Our auto-prompting strategy offers a powerful and data-efficient paradigm for tackling complex segmentation tasks in clinically realistic, low-data scenarios, bridging the gap between large-scale models and specialized clinical needs.