Deep learning-based classification of pleural malignancy using medical thoracoscopic images.

Background: Malignant pleural effusion (MPE) is a frequent complication of advanced lung cancer, and rapid and accurate diagnosis is critical for timely therapeutic decision-making. Although medical thoracoscopy (MT) provides direct visualization and targeted biopsy, resulting in high diagnostic yield, the clinical utility of its findings is contingent on operator experience and subsequent confirmation via pathological analysis. Recent advances in deep learning have enabled automated image classification in various fields, but its application in thoracoscopic images remains unexplored. The aim of our study was to develop a deep learning-based model to classify pleural malignancy and to evaluate its diagnostic performance. Methods: We developed a deep learning-based classification model using thoracoscopic images obtained from 426 patients who underwent MT at Incheon St. Mary's Hospital between May 2015 and July 2024. After preprocessing and standardization of 4,932 images (2,093 benign, 2,839 malignant), we trained an InceptionV3-based convolutional neural network using transfer learning and online augmentation during training. Model performance was evaluated according to accuracy, precision, recall, the F1 score, the area under the receiver operating characteristic curve (ROC-AUC), and gradient-weighted class activation mapping (Grad-CAM) visualization. Results: In total, 4,932 thoracoscopic images from 426 patients were used to train and evaluate the model. In the test set, the InceptionV3-based model achieved a classification accuracy of 81.7% [95% confidence interval (CI): ± 3.5%], with a precision of 82.4% (95% CI: ±4.3%), recall of 86.6% (95% CI: ±3.8%), and F1 score of 84.6% (95% CI: ±3.0%). The AUC was 0.90 (95% CI: ±2.6%), indicating excellent discriminative performance. A confusion matrix indicated 165 true positives, 25 false negatives, 29 false positives, and 262 true negatives. Grad-CAM visualizations confirmed that the model consistently focused on visually relevant pleural abnormalities such as nodularity, thickening, and neovascularization. Notably, the model maintained high accuracy even in cases without overt tumor nodules. Conclusions: This study presents the first deep learning-based classification model of pleural malignancy using MT images. The model showed excellent diagnostic performance and has the potential to aid real-time clinical decision-making during thoracoscopy by suggesting malignant targets for biopsy or pleurodesis.