Deep learning-based real-time intraoperative detection of thoracic duct.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Surgical Technology

Background:

• Thoracic duct (TD) injury during esophageal cancer surgery can lead to fatal chylothorax.
• Accurate intraoperative identification of the TD is crucial for patient safety.
• Deep learning (DL) offers potential for real-time image analysis to aid surgeons.

Purpose of the Study:

• To develop and validate a DL-based system for real-time intraoperative TD detection.
• To assess the AI system's performance against surgical expertise.

Main Methods:

• A YOLOv5-seg model was trained on 2,500 images from 30 thoracoscopic esophagectomy videos.
• The dataset included 1,400 TD annotations and 1,100 background images.
• Performance was evaluated using Dice coefficient and Intersection over Union (IoU), compared to resident and attending surgeons.

Main Results:

• The AI system achieved 92.5% accuracy, with a Dice coefficient of 0.677 and IoU of 0.569.
• AI performance was superior to resident surgeons (Dice: 0.677 vs. 0.577) but inferior to attending surgeons (Dice: 0.677 vs. 0.797).
• The AI demonstrated a significant performance advantage over less-experienced surgeons.

Conclusions:

• A DL-based real-time TD segmentation system is technically feasible.
• The AI system shows potential as a foundation for future AI-guided intraoperative applications.
• The AI's performance is superior to residents, suggesting its utility in enhancing surgical safety.