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

Area of Science:

• Medical Imaging
• Computer-Assisted Surgery
• Artificial Intelligence in Medicine

Background:

• Accurate segmentation of surgical instruments in endoscopic images is crucial for computer-assisted surgery.
• Manual annotation of training datasets is time-consuming and labor-intensive.

Purpose of the Study:

• To evaluate the efficiency of point-based interactive segmentation using foundation models for surgical instrument annotation in spinal endoscopy.
• To compare the performance of Segment Anything Model (SAM) 2.0, SAM 2.1, and domain-adapted MedSAM 2.1.

Main Methods:

• Retrospective study comparing SAM 2.0, SAM 2.1, and MedSAM 2.1 on 308 spinal endoscopy images.
• Models assessed using point prompts at the ground truth mask centroid.
• Primary outcomes: Dice similarity coefficient (DSC) at single-click and success rate (DSC≥0.85).

Main Results:

• MedSAM 2.1 achieved significantly higher single-click DSC (0.937±0.074) than SAM 2.1 (0.844±0.222) and SAM 2.0 (0.824±0.236).
• MedSAM 2.1 demonstrated a 75.8% reduction in annotation time compared to manual methods (4.1-fold speedup).
• External validation confirmed MedSAM 2.1's generalizability, achieving DSC 0.918±0.083.

Conclusions:

• Domain-adapted MedSAM 2.1 significantly improves single-click annotation accuracy and efficiency for surgical instruments in spinal endoscopy.
• Low-resource domain adaptation can facilitate institutional dataset construction for surgical artificial intelligence research.