Artificial Intelligence Tracking of Otologic Instruments in Mastoidectomy Videos.

Abstract

Objective: Develop an artificial intelligence (AI) model to track otologic instruments in mastoidectomy videos.

Study design: Retrospective case series.

Setting: Tertiary care center.

Subjects: Six otolaryngology residents (PGY 3-5) and one senior neurotology attending.

Interventions: Thirteen 30-minute videos of cadaveric mastoidectomies were recorded by residents. The suction irrigator and drill were semi-manually annotated. Videos were split into training (N = 8), validation (N = 3), and test (N = 2) sets. YOLOv8, a state-of-the-art AI computer vision model, was adapted to track the instruments.

Main outcome measures: Precision, recall, and mean average precision using an intersection over union cutoff of 50% (mAP50). Drill speed in two prospectively collected live mastoidectomy videos by a resident and attending surgeon.

Results: The model achieved excellent performance for tracking the drill (precision 0.93, recall 0.89, and mAP50 0.93) and low performance for the suction irrigator (precision 0.67, recall 0.61, and mAP50 0.62) in test videos. Prediction speed was fast (~100 milliseconds per image). Predictions on prospective videos revealed higher mean drill speed (8.6 ± 5.7 versus 7.6 ± 7.4 mm/s, respectively; mean ± SD; p < 0.01) and duration of high drill speed (>15 mm/s; p < 0.05) in attending than resident surgery.

Conclusions: An AI model can track the drill in mastoidectomy videos with high accuracy and near-real-time processing speed. Automated tracking opens the door to analyzing objective metrics of surgical skill without the need for manual annotation and will provide valuable data for future navigation and augmented reality surgical environments.