AG-UNet: Attention-Gated Pneumothorax Segmentation and Severity Classification

Pneumothorax is a serious medical condition that arises when air leaks into pleural space which is area separating lungs from the chest cavity. This study presents a customized AG-UNet architecture, specifically designed for the semantic segmentation of pneumothorax regions in chest radiograph (CXR) images. The SIIM-ACR dataset is used for implementing the proposed model. The proposed architecture follows a symmetric encoder-decoder structure based on U-Net and is enhanced with attention-gated decoder blocks to improve focus on pneumothorax-affected regions. The encoder path extracts hierarchical features, while the decoder path reconstructs the segmentation map with spatial precision using attention mechanisms and skip connections. The proposed architecture enhances the traditional U-Net by incorporating attention-gated decoder blocks, enabling the model to focus on clinically significant regions and minimize false positives. The model achieves a Dice coefficient of 71.66 %, testing accuracy of 79.72 %, precision of 79.72 %, recall of 80.07 %, and an F 1 -score of 79.39 %, demonstrating robust performance in pixel-wise segmentation. Furthermore, a severity classification module is integrated to categorize pneumothorax into three grades: no pneumothorax, small pneumothorax, and large pneumothorax. While the model shows reliable identification of normal and mild cases, it exhibits underestimation in severe cases, highlighting the challenge of detecting subtle features in complex radiographic images.