Domain-Adapted Foundation Models for Single-Click Surgical Instrument Segmentation in Spinal Endoscopy

Accurate segmentation of surgical instruments in endoscopic images is essential for developing computer-assisted surgical systems, yet creating annotated training datasets remains labor-intensive. This study aimed to evaluate the efficiency of point-based interactive segmentation using foundation models for surgical instrument annotation in spinal endoscopy.

We conducted a retrospective study comparing point-based segmentation performance between Segment Anything Model (SAM) 2.0, SAM 2.1, and domain-adapted MedSAM 2.1 for surgical instrument segmentation in spinal endoscopy. A test set of 308 images from 10 patients undergoing full-endoscopic lumbar decompression/discectomy or biportal endoscopic decompression/discectomy was evaluated. Models were assessed using standardized point prompts placed at the ground truth mask centroid. Primary outcomes included Dice similarity coefficient at single-click and success rate (Dice≥0.85). Secondary analysis evaluated click-to-quality metrics across iterative refinement up to 10 clicks.

At single-click, MedSAM 2.1 achieved significantly higher Dice scores than SAM 2.1 (0.937±0.074 vs. 0.844±0.222, p<0.001; 95% confidence interval [CI] for difference, 0.070-0.116) and SAM 2.0 (0.937±0.074 vs. 0.824±0.236, p<0.001; 95% CI for difference, 0.088-0.138), with a 3-fold reduction in variability (standard deviation: 0.074 vs. 0.222). External validation on the publicly available Spine Endoscopic Atlas dataset (202 images from 10 patients) confirmed generalizability, with MedSAM 2.1 achieving Dice 0.918±0.083, compared to 0.724±0.281 for SAM 2.1 and 0.708±0.290 for SAM 2.0 (both p<0.001). Interrater agreement between 2 independent annotators demonstrated excellent reliability (intraclass correlation coefficient, 0.991; 95% CI, 0.986-0.994; mean pairwise Dice, 0.953±0.051). The MedSAM 2.1-assisted workflow reduced annotation time by 75.8% compared to manual polygon annotation (7.7±13.3 seconds vs. 31.8±18.2 seconds per image; p<0.001; 4.1-fold speedup). SAM 2.1 improved progressively with additional clicks (Dice 0.844 to 0.943 at 10 clicks), while MedSAM 2.1 showed nonmonotonic progression (0.937 to 0.947), indicating near-optimal first-click predictions.

Domain-adapted MedSAM 2.1 improved single-click annotation accuracy and efficiency for surgical instruments in spinal endoscopy, suggesting that low-resource domain adaptation may facilitate institutional dataset construction for surgical artificial intelligence research.