Real-time multicenter semantic segmentation for laparoscopic gastrectomy: development, external validation and operating room deployment of an AI navigation system.

Abstract

Background: Laparoscopic gastrectomy is technically demanding and errors often arise from limited situational awareness. Artificial intelligence (AI)-based computer vision may enhance intraoperative recognition of instruments and anatomy, but evidence for generalizable, real-time deployment remains limited.

Methods: We conducted a multicenter study using laparoscopic gastrectomy videos from 4 tertiary hospitals. Data were split by institution into training, internal test, and external validation sets. A You Only Look Once version 10 (YOLOv10)-based model was developed to recognize 18 categories (11 instruments and 7 organs) and integrated into a lightweight user interface for bedside use. Mean average precision at Intersection over Union (IoU) 0.50 (mAP@50) and AP@[0.50-0.95] for boxes and masks, recall, and F1 score were calculated. The feasibility of live operating-room deployment was assessed. Surgeon cognitive workload was assessed using the National Aeronautics and Space Administration Task Load Index (NASA-TLX), and system usability was evaluated with the System Usability Scale (SUS).

Results: The training set comprised 4308 frames (41,066 labels); the internal testing set, 1803 frames (17,211 labels); and the external validation set, 721 frames (6685 labels). Training performance reached box precision 0.896, recall 0.891, mAP@50 0.937, and AP@[0.50-0.95] 0.846; mask AP@[0.50-0.95] was 0.809. On the internal testing set, box AP@[0.50-0.95] was 0.670 and mask AP@[0.50-0.95] 0.642. External validation achieved box precision 0.798, recall 0.773, F1 score 0.785, mAP@50 0.832, and AP@[0.50-0.95] 0.684; mask mAP@50 was 0.836 and AP@[0.50-0.95] 0.658. High-performing classes included stapler, ultrasonic scalpel, and forceps, whereas omentum and pancreas were relatively lower. The system operated in real time during laparoscopic gastrectomy on a laptop-grade GPU, providing on-demand overlays without workflow disruption. The NASA-TLX score was 26.72±6.51, indicating low cognitive burden, and the SUS score was 78.33±5.20, demonstrating good usability.

Conclusions: We present development, external validation, and live deployment of a real-time AI navigation system for laparoscopic gastrectomy across multiple centers. The model demonstrated robust detection and segmentation of instruments and organs and was feasible for bedside use. Prospective trials are warranted to evaluate effects on intraoperative safety, efficiency, and training.