Explainable benchmarking of U-Net variants for lung segmentation on chest radiographs

Accurate lung field segmentation in chest radiographs is essential for reliable computer-aided diagnosis of pulmonary diseases. This study aimed to systematically benchmark baseline, lightweight, and attention-based U-Net architectures while integrating explainable artificial intelligence (AI) to evaluate both segmentation performance and anatomical focus.

We conducted a comparative evaluation of three U-Net variants-baseline U-Net, attention U-Net, and shallow U-Net-using the chest X-ray Masks and Labels dataset. All models were trained under identical conditions with five-fold cross-validation and evaluated using accuracy, Intersection over Union (IoU), and Dice coefficient. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to visualize model attention and assess whether network activations were anatomically localized within lung regions. All architectures were implemented in TensorFlow.

U-Net and attention U-Net achieved the highest segmentation performance (Dice ≈ 0.97, IoU ≈ 0.94), with Grad-CAM activations consistently localized to lung fields, indicating reliable anatomical focus. The shallow U-Net showed slightly lower accuracy (Dice= 0.96, IoU= 0.92) but demonstrated faster inference and broader sensitivity to internal parenchymal structures, which may facilitate future disease-focused pulmonary analysis.

This study highlights the trade-offs between segmentation accuracy, computational efficiency, and model interpretability across U-Net variants. By combining quantitative benchmarking with explainable AI-based analysis, our results provide performance insights, supporting the development of trustworthy AI tools for chest radiograph analysis.