Pleural Invasion of Peripheral cT1 Lung Cancer by Deep Learning Analysis of Thoracoscopic Images: A Retrospective Pilot Study

This research developed a deep learning algorithm to predict pathological pleural invasion (pPL) from thoracoscopic imagery in patients with cT1 tumors receiving sublobar resections. The study included 80 patients, with the learning model trained on 64 cases and validated on 16. The algorithm's predictive performance was evaluated against surgeons' intraoperative judgments using McNemar's test. The tumor recognition component achieved 78 percent image-level accuracy, while the pPL prediction model demonstrated 69 percent patient-level accuracy compared to the surgeons' 75 percent (P=0.32). While the researchers acknowledge that clinical implementation would require at least 90 percent accuracy, these findings suggest the AI approach shows promise, achieving performance similar to that of surgeons despite the modest dataset size.