PD08.06. AI-Based Surgical Recognition System Developed by an Esophageal Surgeon: Recurrent Laryngeal Nerve Identification and Traction Detection

Abstract

Esophageal Cancer: Surgical Treatment of Esophageal Cancer Artificial intelligence (AI) has rapidly advanced in medical imaging and diagnostics, and its application in surgery is expanding toward intraoperative guidance and skill assessment. For clinically meaningful implementation, surgeons must actively participate in development based on real-world needs. In particular, recurrent laryngeal nerve (RLN) injury remains a critical complication in esophagectomy. We developed an AI-based surgical recognition system designed by esophageal surgeons to identify bilateral recurrent laryngeal nerves and detect traction-related surgeon behaviors during robot-assisted esophagectomy. The system was trained using annotated intraoperative video datasets and incorporated deep learning–based anatomical segmentation and action recognition modules. A retrospective evaluation was conducted using recorded surgical videos. Participating surgeons reviewed videos either with or without AI-generated overlays indicating RLN location and traction alerts. The primary outcome was nerve identification accuracy. Secondary outcomes included surgeon confidence levels in nerve recognition. The system was designed not as a decision-making replacement but as an intraoperative attention-support tool. From the early development phase, product design considered target users, disease indications, potential limitations, regulatory strategy, cost structure, and impact on standard practice, in collaboration with regulatory, quality assurance, and reimbursement specialists through a university-origin startup. In the evaluation study, AI assistance significantly improved right RLN identification accuracy. Without AI overlays, surgeons correctly identified the nerve in 46.9% of cases, whereas accuracy increased to 81.3% when AI-generated recognition was provided (p = 0.004). The improvement was particularly notable among early-career surgeons. In addition to objective accuracy gains, AI support significantly increased surgeons’ confidence in nerve identification. The system successfully visualized traction-related movements that may contribute to postoperative RLN palsy, providing real-time attention cues. These findings suggest that structured intraoperative information presentation can reduce recognition variability among surgeons. The software does not override surgical judgment but functions as a safety-oriented cognitive support layer. Following clinical performance evaluation, the system obtained regulatory approval in December 2025. Ongoing real-world implementation will further assess clinical utility and external validity across diverse institutional settings. We developed and clinically implemented an AI-based RLN recognition and traction detection system for robot-assisted esophagectomy. Surgeon-driven development significantly improved anatomical recognition accuracy and confidence. This information-support model demonstrates a feasible pathway for socially implementing surgical AI technologies while preserving surgeon autonomy. Further real-world evaluation will clarify its broader clinical impact and generalizability.