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

U-Net and attention U-Net achieved Dice ≈0.97 for lung segmentation on chest radiographs, with Grad-CAM showing anatomical focus; shallow U-Net had Dice 0.96 but faster inference and broader parenchymal sensitivity. Retrospective, single dataset.

• Comparative evaluation of baseline, attention, and shallow U-Net on the Chest X-ray Masks and Labels dataset using 5-fold cross-validation (n not reported).
• Attention U-Net and baseline U-Net had top segmentation (Dice ≈0.97, IoU ≈0.94); shallow U-Net was slightly lower (Dice 0.96, IoU 0.92).
• Limitation: single public dataset, no external validation or clinical outcome correlation.

RadPigeon summaries are original and for information only. They are not clinical advice.