Comparative evaluation of deep learning models for lung segmentation in chest X-rays: applications in infectious disease screening

Abstract

Respiratory infections such as pneumonia and tuberculosis remain the leading contributors to global morbidity and mortality, and early diagnosis is essential for effective treatment. Chest X-ray imaging is widely used for the detection of these conditions; however, reliable interpretation requires accurate delineation of lung fields. Manual annotation is time-consuming and prone to variability, underscoring the need for automated segmentation methods. In this study, we conducted a systematic evaluation of state-of-the-art deep learning architectures for binary lung segmentation, including U-Net, Attention U-Net, Double U-Net, U2-Net, VGG-UNet, UNet++, ResNet-UNet, Dense-UNet, Swin U-Net and HieraSeg Net. The performance of the model was compared using the dice coefficient, the intersection of the union (IoU), the mean absolute error (MAE), the Hausdorff distance and the average symmetric surface distance (ASSD). Among these models, Dense-UNet achieved the best results, yielding a Dice score of 0.9848 and an IoU of 0.9700, with the lowest surface error measures. These findings highlight the potential of Dense-UNet to serve as a robust backbone for AI-assisted diagnostic systems in infectious respiratory diseases, thereby supporting faster, more reliable, and scalable approaches to clinical decision-making.