LungSurg: A Generative AI System for Segmentation and Phase Classification in Thoracoscopic Lobectomy

Area of Science:

• Thoracic Surgery
• Artificial Intelligence
• Medical Imaging

Background:

• Surgical practices are increasingly integrating artificial intelligence (AI) for enhanced precision.
• Video-assisted thoracoscopic surgery (VATS) lobectomy for lung cancer presents opportunities for AI-driven advancements.

Purpose of the Study:

• To assess the potential of AI in VATS lobectomy by developing and evaluating an AI system named LungSurg.
• LungSurg aims to improve the analysis of intrathoracic anatomy, surgical instruments, and surgical phases during VATS lobectomy.

Main Methods:

• Development of LungSurg, featuring interconnected segmentation and classification networks.
• Prospective collection of 222 VATS lobectomy videos from eight centers, with extensive annotations and frame-level phase information.
• External validation of segmentation and classification network performance, alongside comparative studies with senior surgeons and an educational assessment with surgical residents.

Main Results:

• The segmentation network achieved high mean Average precision scores for lung and instrument identification (0.745 for left lung, 0.726 for right lung).
• The classification network demonstrated strong Top-1 (71.5%) and Top-3 (88.0%) accuracies in identifying 14 distinct surgical phases.
• LungSurg showed comparable anatomical identification to senior surgeons and superior sensitivity, with significant skill improvement in residents trained using the system.

Conclusions:

• LungSurg accurately analyzes VATS lobectomy procedures, demonstrating the feasibility of AI tools in thoracic surgery.
• AI-driven systems like LungSurg have the potential to enhance surgical precision, training, and overall thoracic surgical practices.