Pneumothorax, the abnormal collection of air in the pleural space, poses significant diagnostic challenges. Manual interpretation of chest X-rays is time-consuming and error-prone, especially in resource-limited settings. This study introduces a hybrid deep learning approach using an EfficientNet-UNet architecture for precise and efficient pneumothorax detection. Leveraging EfficientNet’s depth-wise feature extraction and U-Net’s segmentation capability, the model provides high-resolution localization of thoracic abnormalities. Trained on publicly available annotated datasets, the model achieved an AUC of 0.86, a mean Dice coefficient of 0.30, and sensitivity of 0.40. These results underscore the model's ability to support automated radiographic diagnosis with high specificity and clarity.
The proposed model shows excellent potential for real-time, AI-driven thoracic diagnostics. By combining segmentation and classification strengths, the system offers both visual explanations and clinical reliability, thereby supporting faster and more accurate pneumothorax management.
One of the key limitations lies in the moderate sensitivity, indicating that small or subtle pneumothorax regions might be missed. Additionally, the system’s performance may vary across datasets from different radiographic equipment or hospitals.
Further enhancements will focus on integrating a multi-class output to detect other thoracic abnormalities such as pleural effusion and consolidation. Real-time deployment using mobile-based applications and federated learning for privacy-preserving training are also future priorities.