SAM strategies for laryngeal SCC segmentation in multicentric endoscopic frames

Laryngeal squamous cell carcinoma (LSCC) is a common malignancy of the upper aerodigestive tract, and its precise margin delineation is essential for improving surgical outcomes. Although deep learning methods have shown promising results for automated LSCC segmentation from endoscopic images, their generalization across clinical settings is limited by small and heterogeneous datasets. This may also explain why, despite the rise of foundation models in other fields of medical imaging, their use in LSCC segmentation remains sporadic. Within an international collaboration, we here leverage a large proprietary multicentric dataset of 3929 annotated white light and narrow-band frames from 602 patients to investigate the potential of the Segment Anything Model (SAM) and its medical variants, exploring a large variety of experiments, including full fine-tuning, low-rank adaptation strategies, and different prompting configurations. From our analysis, performance strongly depended on the prompting strategy. SAM-based models achieved their strongest performance under oracle bounding-box prompting, reaching Dice similarity coefficient values of 0.90–0.92 and representing an idealized upper-bound setting for surgeon-guided, semi-automatic segmentation. While outperforming convolutional methods in this setting, the comparison reflects the additional spatial information provided by the bounding-box prompt. Prompt-free and point-prompted configurations achieved lower performance, highlighting the importance of spatial guidance. Overall, these findings support the potential of SAM-based foundation models as clinician-in-the-loop tools for LSCC segmentation and motivate further research toward robust automatic prompting and intraoperative decision-support systems.