Performance of an artificial intelligence–based software in detecting pneumothorax on supine chest radiographs: a retrospective study

Abstract

Purpose: To evaluate the diagnostic performance of artificial intelligence (AI)-based software for pneumothorax detection on supine radiographs and its impact on physicians’ interpretation. Methods: This single-center retrospective study analyzed 114 hemithoraces with pneumothorax and 340 without, using computed tomography as the reference standard. We evaluated the performance of CXR-AID, an AI-based software, for pneumothorax detection on supine chest radiographs, and conducted a reader study to assess the utility of AI assistance. Sensitivities and specificities were adjusted for clustering within patients. Results: Sensitivity and specificity of the AI in detecting overall pneumothorax on a supine chest radiograph were 61.0% (95% confidence interval [CI], 50.6%–70.4%) and 94.3% (95% CI, 90.8%–96.5%), respectively. Sensitivity of the AI in detecting a large pneumothorax with a maximum radial interpleural distance > 35 mm was 97.4% (95% CI, 83.6%–99.6%). Sensitivity was significantly higher in the upper lung zone than in the lower lung zone (69.5% [95% CI, 59.3%–78.1%] vs. 37.5% [95% CI, 27.3%–48.8%]). In the reader study, the AI significantly improved resident sensitivity (46.8% to 57.3%, P < 0.001). For experts, the AI did not improve sensitivity significantly (P = 0.32) but significantly improved specificity (90.9% to 95.6%, P = 0.02). Conclusions: The AI demonstrated high sensitivity for detecting large pneumothoraces on supine radiographs, helping identify patients requiring tube thoracotomy. It may serve as a diagnostic safety net for residents by increasing sensitivity and enhances experts’ diagnostic confidence by improving specificity. However, pneumothorax detection in the lower lung zone remains challenging.