Deep Learning Model for Pneumothorax Detection in Chest Radiographs: A Multicenter Retrospective Cross-Sectional Study

Background: Pneumothorax is a common clinical condition characterized by the presence of air within the pleural space, occurring in about half of chest trauma cases. Its clinical presentation ranges from asymptomatic cases to severe conditions causing hemodynamic instability or death. Deep learning models offer transformative potential for both clinical diagnosis and medical education through automated detection and interactive training tools. This study sought to evaluate deep learning models for detecting pneumothorax in Chest Radiographs (CXRs), assessing their diagnostic accuracy and potential to enhance medical education.Methods: This retrospective cross-sectional study was conducted between February 2022 and September 2023 to assess the performance of four deep learning models for pneumothorax detection: Mask Region-based Convolutional Neural Network (Mask R-CNN), Deep Labelling version 3 (DeepLabv3), You Only Look Once version 8 (YOLOv8), and the U-shaped CNN model (U-Net). The evaluation was conducted using 20,000 chest X-ray images sourced from three hospitals in Iran, along with three open source datasets, including PTX-498, PTX-227, and SIIM-ACR-Pneumothorax. Images were labeled by consensus from two radiologists and two traumatologists. Rather than applying a conventional percentage-based split, a tiered data strategy was applied: internal datasets for training and validation, and external datasets (CheXpert and NIH) for independent testing to verify generalizability. Each model was trained to detect pneumothorax by extracting features and performing segmentation. Performance was evaluated using sensitivity, specificity, precision, recall, and F1-score. The outputs were analyzed for integration into virtual learning platforms to train medical students in diagnosing pneumothorax.Results: The YOLOv8 algorithm showed the best performance for detecting and localizing pneumothorax, achieving an F1 score of 0.68. The final model’s precision was 0.79, and a recall of 0.60, and it worked best on chest X-ray images with 1024x1024 resolution, particularly showing greater accuracy in identifying larger pneumothoraces.Conclusion: Integration of YOLOv8 into medical education has the potential to improve diagnostic training via interactive AI-based simulations. However, challenges remain in detecting smaller pneumothoraces, highlighting the need for further optimization.