Automatic tip detection of surgical instruments in biportal endoscopic spine surgery

Recent advances in robotics and deep learning can be used endoscopic surgeries provide numerous advantages by freeing one of the surgeon's hands. This study aims to automatically detect tip instrument, localize a point, evaluate detection accuracy biportal spine surgery (BESS). The could serve as preliminary for development vision intelligence robotic endoscopy. dataset contains 2310 frames from 9 BESS videos with x y coordinates annotated an expert. We trained two state-of-the-art detectors, RetinaNet YOLOv2, bounding boxes centered around annotations specific margin sizes determine optimal size detecting instrument localizing point. calculated recall, precision, F1-score fixed box both ground truth predicted midpoints compare performance models different boxes. For RetinaNet, 150 pixels was recall 1.000, precision 0.733, 0.846. 0.864, 0.808, 0.835. Also, cross-validate its overall robustness. resulting mean were 1.000 ± 0.000, 0.767 0.033, 0.868 0.022, respectively. In this study, we evaluated automatic method surgical instruments surgery, compared algorithms, validated robustness cross-validation. applied types endoscopy detection.