Development and Validation of an Artificial Intelligence Surgical Video Analysis Model for Predicting Visceral Pleural Invasion in Lung Cancer Surgery: A Multicenter Study

Background: Intraoperative diagnosis of visceral pleural invasion (VPI) during video-assisted thoracoscopic surgery (VATS) remains challenging. This study aimed to develop and validate a deep learning-based model to improve diagnostic accuracy and guide surgical decision-making. Methods: Thoracoscopic videos and clinical data from 346 patients (3367 images, 2015–2024) in one hospital were divided into training, validation, and internal-test sets (7:2:1), whereas data from 53 patients (1274 images) in two other hospitals formed the external-test set. A spatial dropout-based Residual Convolutional Neural Network (VPI-Net) was developed for estimating patients’ VPI status and VPI risk score (VPIscore). The model’s performance was compared against intraoperative estimations by surgeons and preoperative assessments by radiologists. Results: The VPI-Net model demonstrated significantly higher area under the curve (AUC, 0.84–0.94) and accuracy (79.67–88.68%,) than two surgeons and one radiologist across all cohorts (p < 0.05). Additionally, the VPI-Net model outperformed human experts in sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) across all cohorts. A lower VPIscore (VPIscoreL) was significantly correlated with longer overall survival (OS), relapse-free survival (RFS), and time to progression (TTP) than a higher VPIscore (VPIscoreH) (all p < 0.001). Similar results were observed in patients who had small tumors, with those who had VPIscoreH exhibiting significantly worse RFS and TTP than those with VPIscoreL (RFS [p = 0.012], TTP [p = 0.035]). The VPIscoreL patients had a significantly longer TTP (p = 0.03) than the VPIscoreH patients after sublobectomy. Conclusion: The proposed model enables satisfactory intraoperative identification of VPI, potentially improving patient outcomes during VATS.