Real-time multicenter semantic segmentation for laparoscopic gastrectomy: development, external validation and...

Area of Science:

• Surgical Technology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Laparoscopic gastrectomy is a complex procedure where limited situational awareness can lead to errors.
• Existing artificial intelligence (AI)-based computer vision for intraoperative use has limited evidence for generalizable, real-time applications.

Purpose of the Study:

• To develop and validate a real-time AI navigation system for laparoscopic gastrectomy.
• To assess the system's performance, feasibility for bedside use, and impact on surgeon workload and usability.

Main Methods:

• A You Only Look Once version 10 (YOLOv10)-based AI model was trained on laparoscopic gastrectomy videos from four hospitals to recognize 18 categories (instruments and organs).
• The model was integrated into a user interface for real-time bedside deployment, with performance evaluated using metrics like mean average precision (mAP) and F1 score.
• Surgeon cognitive workload (NASA-TLX) and system usability (SUS) were assessed.

Main Results:

• The AI system achieved strong performance in external validation (mAP@50 0.832 for boxes, 0.836 for masks) and operated in real-time on a laptop-grade GPU.
• The system demonstrated good usability (SUS score 78.33±5.20) and low cognitive burden (NASA-TLX score 26.72±6.51) for surgeons.
• High-performing classes included stapler, ultrasonic scalpel, and forceps; lower performance was noted for omentum and pancreas.

Conclusions:

• A real-time AI navigation system for laparoscopic gastrectomy was successfully developed, externally validated across multiple centers, and deployed live.
• The system is feasible for bedside use, showing robust instrument and organ detection/segmentation.
• Further prospective trials are recommended to evaluate the system's impact on intraoperative safety, efficiency, and surgical training.