Validity of Computer Vision in Surgical Tool Segmentation in Transoral Resection

Introduction: Surgical video datasets for transoral robotic surgery (TORS) require time intensive instrument masks segmentation for model development, yet. We evaluated commonly used segmentation models and explored a practical workflow for initial mask generation and downstream training.

Methods: Four segmentation approaches—SAM2, YOLOv26-seg, Mask R-CNN (ResNet-50), and a MONAI U-Net— were applied on 200 randomly sampled frames from TORS procedure. Manually annotated of surgical instruments were designated as reference standard. SAM2 was evaluated without task-specific training), YOLOv26 and R-CNN were evaluated using publicly available endoscopic weights/fine-tuning, and U-Net was done with built in endotracheal weights. Model outputs were compared to manual annotations using overlap-based matching; summarized with precision precision–recall curves (AUPRC), and mAP@0.5.

Results: Without task-specific training, SAM2 achieved high precision for most tools: Bovie 0.995, Maryland 0.995, needle driver 0.995, suction 0.953, endotracheal tube 0.995, retractor 0.763 with overall precision 0.949. YOLOv26achieved class-wise precision of 0.943 Bovie, 0.904 Maryland, 0.995 needle driver, 0.803 retractor, 0.741 suction, and 0.865 endotracheal tube, with overall 0.875. R-CNN achieved 0.948 Bovie, 0.862 Maryland, 0.772 retraction, 0.805 suction, and 0.812 endotracheal tube, with overall 0.700; endotracheal tube 0.995, with lower precision for the retractor 0.763. Overall precision across all classes was 0.949.U-Net had the lowest precision 0.850 Bovie, 0.816 Maryland, 0.761 suction, endotracheal tube 0.836 with overall precision 0.820 .

Conclusions: SAM2 demonstrated strong baseline instrument segmentation without task-specific training and may reduce annotation burden by generating initial masks for rapid human correction and label propagation. In a multi-stage workflow, corrected masks can support training task-specific models. Future work should report complete AUPRC results across models and evaluate generalization across pathology and physiology of the oropharyngeal airway.