P1.245. Real-Time AI-Based Thoracic Duct Recognition During Thoracoscopic Esophagectomy

Abstract

Abstract Topic Esophageal Cancer: Surgical Treatment of Esophageal Cancer – technique Background Thoracic duct injury during thoracoscopic esophagectomy may result in postoperative chylothorax, a serious complication associated with increased morbidity. However, intraoperative identification of the thoracic duct remains challenging because of its small diameter and variable visualization. We aimed to develop and validate a deep learning–based system for real-time thoracic duct recognition to assist surgeons during minimally invasive esophagectomy. Methods Thoracoscopic esophagectomy videos were retrospectively collected to construct a segmentation dataset. A total of 2,500 images, including thoracic duct–positive and background frames, were manually annotated and divided into training and validation sets (4:1). An independent external test set was created from five additional surgical videos. A YOLOv5-seg deep learning model was trained for thoracic duct semantic segmentation. Model performance was evaluated using Dice coefficient, intersection over union (IoU), and detection accuracy. Segmentation performance was compared with resident and attending surgeons who independently evaluated identical test images under blinded conditions. Statistical comparisons were performed with significance defined as P<0.05. Results The AI system achieved an overall accuracy of 92.5% on the independent test set, with a mean Dice coefficient of 0.677 and an IoU of 0.569. Performance demonstrated a hierarchical pattern, with attending surgeons achieving the highest segmentation accuracy (Dice 0.797), followed by AI and resident surgeons (Dice 0.577). The AI model significantly outperformed residents in both Dice and IoU metrics (P<0.05) but remained inferior to attending surgeons. Real-time inference latency reached 17.7 ms per frame, enabling near-real-time analysis of thoracoscopic video streams. Qualitative assessment showed improved discrimination compared with residents in anatomically ambiguous conditions, although performance decreased in low-contrast or poorly exposed thoracic ducts. Conclusion Deep learning–based real-time thoracic duct segmentation during thoracoscopic esophagectomy is technically feasible and demonstrates performance superior to resident surgeons while approaching attending-level recognition. This system may serve as an intraoperative decision-support tool to enhance anatomical identification and potentially improve surgical safety. Prospective multicenter studies are warranted to validate its clinical impact and integration into routine surgical workflows.