Artificial Intelligence-Assisted Lung Ultrasound for Pneumothorax: Diagnostic Accuracy Compared with CT in Emergency and Critical Care

Abstract

Simple Summary Pneumothorax is a life-threatening condition that requires rapid and accurate diagnosis, especially in emergency and critical care settings. Although lung ultrasound (LUS) offers a fast and radiation-free diagnostic option, its accuracy can vary depending on the operator's experience. This study evaluated the potential of artificial intelligence (AI) to assist clinicians by automatically detecting pneumothorax on LUS images and videos. Using transformer-based deep learning models, we compared the diagnostic performance of Vision Transformer (ViT), DINOv2, and Video Vision Transformer (ViViT) architectures. When tested on data from different patients, the DINOv2 model achieved 90% accuracy, demonstrating reliable generalization without overfitting. Furthermore, when video sequences were analyzed, both Random Forest and eXtreme Gradient Boosting classifiers trained on ViViT-derived features achieved 90% accuracy, showing that AI can effectively interpret dynamic pleural motion. These results indicate that transformer-based AI can enhance pneumothorax diagnosis by improving consistency and reducing operator dependence, supporting broader use of lung ultrasound in emergency and point-of-care environments.Abstract Background: Pneumothorax (PTX) requires rapid recognition in emergency and critical care. Lung ultrasound (LUS) offers a fast, radiation-free alternative to computed tomography (CT), but its accuracy is limited by operator dependence. Artificial intelligence (AI) may standardize interpretation and improve performance. Methods: This retrospective single-center study included 46 patients (23 with CT-confirmed PTX and 23 controls). Sixty B-mode and M-mode frames per patient were extracted using a Clarius C3 HD3 wireless device, yielding 2760 images. CT served as the diagnostic reference. Experimental studies were conducted within the framework of three scenarios. Transformer-based models, Vision Transformer (ViT) and DINOv2, were trained and tested under two scenarios: random frame split and patient-level split. Also, Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) classifiers were trained on the feature maps extracted by using Video Vision Transformer (ViViT) for ultrasound video sequences in Scenario 3. Model performance was evaluated using accuracy, sensitivity, specificity, F1-score, and area under the ROC curve (AUC). Results: Both transformers achieved high diagnostic accuracy, with B-mode images outperforming M-mode inputs in the first two scenarios. In Scenario 1, ViT reached 99.1% accuracy, while DINOv2 achieved 97.3%. In Scenario 2, which avoided data leakage, DINOv2 performed best in the B-mode region (90% accuracy, 80% sensitivity, 100% specificity, F1-score 88.9%). ROC analysis confirmed strong discriminative ability, with AUC values of 0.973 for DINOv2 and 0.964 for ViT on B-mode images. Also, both RF and XGBoost classifiers trained on the ViViT feature maps reached 90% accuracy on the video sequences. Conclusions: AI-assisted LUS substantially improves PTX detection, with transformers-particularly DINOv2-achieving near-expert accuracy. Larger multicenter datasets are required for validation and clinical integration.