From September 9th to 12th, 2023, the World Conference on Lung Cancer (WCLC) was held in Singapore. Oncology Frontier deeply engaged in the conference, capturing the latest advancements and witnessing the progress and strides made by China in the international oncology arena. Dr. Liang Hengrui of the First Affiliated Hospital of Guangzhou Medical University presented the results from Chinese researchers in the field of “Artificial Intelligence Application in Thoracoscopic Lobectomy for Lung Cancer” in a brief oral report at the 2023 WCLC.
Dr. Liang Hengrui
He is a member of the International Association for the Study of Lung Cancer (IASLC), a member of the CSCO Smart Medical Committee, a youth committee member of the World Chinese Physicians Association, and a youth committee member of the Guangdong Pharmaceutical Association’s Thoracic Surgery Branch.
MA11.04 Video-Based Artificial Intelligence in Thoracoscopic Lobectomy for Lung Cancer: Surgical Structures Segmentation and Phase Recognition
Introduction : Intelligent surgery, encompassing deep learning-based automatic surgical structure recognition, has emerged as a significant area of research. AI-driven surgical structure recognition using video data has demonstrated promising outcomes in several classic and relatively uncomplicated procedures. However, no studies have been reported regarding the incorporation of AI into video-assisted thoracoscopic surgery (VATS) lobectomy..
Method : The LungSurg system comprised two integrated networks (Fig1): (1) a segmentation network for the identification of intrathoracic anatomy and instruments (SurgSeg); and (2) a classification network for surgical phase recognition (SurgClass). The surgical videos were prospectively collected and annotated by two groups of surgeons as data sources for training and testing the AI model.
Results : This study encompassed 152 VATS lobectomy videos, 26,799 surgical structure annotations, and over one million frames containing surgical phase information. The segmentation network exhibited a mean average precision of 0.920 in identifying five types of instruments and 0.675 in recognizing eight distinct anatomical structures. The classification network attained a top-1 accuracy of 0.757 and a top-3 accuracy of 0.925 in discerning 14 surgical phases. Our novel algorithms further enhanced the performance of existing models in both segmentation and classification tasks by incorporating image filling and co-occurrence loss calculation for SurgSeg, as well as utilizing four types of data for SurgClass. For better visualization, the model was integrated into a designed software (Fig2).
Conclusion : The deep learning model exhibited high accuracy in identifying five categories of instruments, eight distinct anatomical structures, and 14 specific surgical phases during VATS lobectomy for lung cancer treatment. The findings of this study underscore the feasibility and potential of AI integration in VATS lobectomy, representing a substantial advancement in the field of thoracic surgical intelligence.
Dr. Liang Hengrui’s Commentary:
A revolution is unfolding in the medical field. Intelligent surgery, particularly automatic surgical structure recognition based on deep learning, has become a hot research frontier. Although it has achieved encouraging results in conventional and relatively simple surgeries, its application in Video-Assisted Thoracoscopic Surgery (VATS) lobectomy has rarely been reported.
At WCLC , we introduce the LungSurg system, which integrates two powerful neural networks – one for segmenting intrathoracic anatomical structures and surgical instruments (SurgSeg), and the other for identifying surgical stages (SurgClass). The research team proactively collected 152 videos of VATS lobectomy, covering over a million frames. These images were meticulously annotated with surgical structure and phase information.
Excitingly, the segmentation network showed remarkable performance in accurately identifying five different types of surgical instruments with an average accuracy of 0.920 and an accuracy of 0.675 in recognizing eight different anatomical structures. Meanwhile, the classification network also achieved significant success in distinguishing 14 specific surgical stages, with a Top-1 accuracy of 0.757 and a Top-3 accuracy of 0.925.
The research introduced innovative algorithms, further enhancing the performance of existing models. For better visualization and operation, the team integrated this model into specially designed software.
This groundbreaking study has brought unprecedented precision to VATS lobectomy. AI not only accurately identify surgical instruments and anatomical structures but also precisely differentiate various surgical stages. This marks a massive advancement in thoracic surgery intelligence, paving the way for endless possibilities in future surgeries.
