Lung region complexity affects Grand-CAM and Deep Taylor outcomes in pneumonia predictions

Accurate and timely pneumonia diagnosis is critical for patient care, and cоnvolutional neural networks (CNNs) have shown significant effectiveness and high accuracy in automating this process. However, CNNs act as black boxes, which limits their interpretability in clinical settings. Explainable AI (XAI) techniques are designed to address this limitation by offering visual or quantitative explanations for model predictions. This study investigated two widely used XAI techniques—Gradient-weighted Class Activation Mapping (Grad-CAM) and Deep Taylor Decomposition (DTD)—for detecting pneumonia in chest X-ray images. Our analysis focuses on how different anatomical regions of the lungs affect the accuracy of explanations provided by these methods. We hypothesized that XAI methods will show statistically significant differences in activation and relevance across lung regions, with the lower regions exhibiting the highest importance values. Because pneumonia most commonly affects the lower lobes, we expect both methods to highlight these areas more strongly than the upper or central regions. Since CNNs learn increasingly complex features, subtle pneumonia patterns in the lower lobes or near the heart can be difficult to interpret. XAI techniques such as Grad-CAM and DTD help visualize these abstract features. Statistical analyses revealed consistent regional differences in how each method emphasized specific lung areas, with one region showing particularly strong relevance in both methods. These findings suggest that aligning interpretability techniques with clinically relevant anatomical regions can improve diagnostic support in medical imaging by helping identify subtle or complex abnormalities more effectively.