Prompt Guided Adaptation of the Segment Anything Model Enhances Small Organ Segmentation In Head and Neck Radiotherapy
Abstract
Purpose
Accurate segmentation of organs at risk (OARs) is essential for high precision head and neck (H&N) radiotherapy, particularly for small and low contrast anatomical structures. Although nnUNet demonstrates strong overall segmentation performance, it typically requires large training datasets and substantial computational resources. In contrast, the foundation Segment Anything Model (SAM) enables flexible prompt based segmentation but shows limited accuracy for small and anatomically complex structures. This study proposes a task specific adaptation of a 3D SAM framework (SAM-FT) that incorporates ROI guided tiny organ enhancement prompting to improve clinically relevant H&N OAR segmentation.
Methods
Our institutional cohort of anonymized H&N CT datasets with expert defined delineations of 24 OARs and target volumes was used for model fine-tuning and evaluation. SAM-FT was adapted using ROI guided tiny organ enhancement strategies, including localized prompt initialization and anatomy aware calibration, to improve prediction stability in challenging anatomical regions. Segmentation performance was compared with SAM-Med3D and nnUNet using overlap and boundary metrics.
Results
For large OARs, SAM-FT achieved Dice comparable to nnUNet (average Dice: 0.802 vs. 0.788, p = 0.0020) and significantly higher than SAM-Med3D (0.605, p < 10⁻⁴). For small organs, SAM-FT achieved a substantial Dice improvement over SAM-MED3D (ΔDice = +0.355, p = 0.0156), while remaining close to nnUNet (ΔDice = −0.066). Across all OARs, SAM-FT demonstrated significantly higher recall than both SAM-Med3D (p = 0.0011) and nnUNet (p = 0.0315).
Conclusion
Task specific fine tuning combined with interactive, anatomy guided prompting is associated with improved segmentation performance for SAM based models, particularly for small and anatomically complex OARs. SAM-FT mitigates key limitations observed in the baseline SAM-Med3D framework and achieves robustness comparable to nnUNet, supporting the potential of foundation model adaptation for clinically meaningful auto contouring in H&N radiotherapy.