Deep learning algorithm for automatic prediction of visceral pleural invasion of lung cancer based on surgical video

e20578 Background: Visceral pleural invasion (VPI) is an important prognostic indicator in lung cancer that is difficult to predict. We aimed to develop a deep learning algorithm that predicts VPI using videos of lesions. Methods: We included 346 patients who underwent thoracoscopic surgery at our department from 2015 to 2021, corresponding to 3800 images. The patient data and image data were divided into training set, validation set and testing set. The algorithm was developed based on the ResNet architecture used for image feature extraction and classification. All images were labeled for lesions and processed to fit the algorithm. We recruited two thoracic attendings and one radiologist to identify VPI using intraoperative video and preoperative CT, respectively. The performance of the algorithm was evaluated by comparison with the performance of the attendings and the radiologist. Results: At the patient level, the algorithm had satisfactory AUC values in the training set (0.85), validation set (0.83) and testing set (0.81). The DeLong test showed statistic difference between the algorithm and two attendings and a radiologist in the training set (p < 0.001, p < 0.001, p < 0.001), validation set ( p= 0.006, p= 0.002, p= 0.026) and testing set ( p= 0.008, p= 0.012, p= 0.007). At the image level, the algorithm also outperformed two attendings and a radiologist with significantly great AUC values in the training set (0.83 vs. 0.52 vs. 0.53 vs. 0.65; p< 0.001, p< 0.001, p< 0.001), validation set (0.78 vs. 0.50 vs. 0.54. vs. 0.68; p< 0.001, p< 0.001, p= 0.048) and testing set (0.79 vs. 0.53 vs. 0.60 vs. 0.58; p< 0.001, p< 0.001, p< 0.001). Conclusions: The video-based deep learning algorithm could satisfactorily predict VPI and achieved expert-level performance better than thoracic attending surgeons and the radiologist we recruited.