Hybrid EfficientNet-U-Net Architecture for Automated Pneumothorax Localization and Diagnosis from Chest Radiographs

Abstract

Pneumothorax is a critical thoracic condition characterized by air accumulation in the pleural cavity, which can be life-threatening if not promptly diagnosed and treated. However, conventional chest radiograph diagnosis heavily relies on radiologists' experience, leading to diagnostic variability and potential delays in treatment. This research proposes an innovative deep learning model integrating EfficientNet and U-Net architectures to accurately detect pneumothorax from chest X-ray images. The proposed hybrid EfficientNet-UNet model leverages EfficientNet's robust feature extraction capabilities combined with U-Net’s effectiveness in segmentation tasks, significantly enhancing detection precision. Comprehensive training and validation were conducted using publicly available datasets of chest radiographs. Experimental results demonstrated promising performance, with the model achieving a high ROC_AUC score of 0.86, improved mean Intersection over Union (mIoU) of 0.20, and mean Dice coefficient (mDice) of 0.30. Additionally, sensitivity reached approximately 0.40, while specificity remained stable around 0.89, indicating the model’s robust discriminatory ability. These outcomes highlight the proposed model’s effectiveness in automating pneumothorax diagnosis, offering a powerful tool to assist clinical decision-making, reduce diagnostic errors, and improve patient outcomes.