AN EXPLAINABLE DEEP LEARNING APPROACH FOR PNEUMONIA DETECTION USING CHEST X-RAY IMAGES

Abstract

Pneumonia is still one of the leading causes of death in children under five, and the burden is worst in low-income regions where trained radiologists are scarce. Chest X-ray imaging is usually the first diagnostic step, and deep learning classifiers have gotten very good at reading these images, with benchmark accuracies now above 95%. The problem is that a lot of these models are not actually looking at the disease, they often key in on scanner artifacts or other spurious patterns instead of real pathology, a failure known as shortcut learning, which makes them risky to trust clinically no matter how high the reported accuracy is. This study compares four explainable AI methods side by side under identical conditions, Layer-wise Relevance Propagation (LRP), Grad-CAM, FGSM-based Adversarial Training, and a Spatial Attention Mechanism (SAM), all built on the same ResNet-50 backbone and tested on a public pediatric chest X-ray dataset of 5,856 labeled images. A stricter preprocessing pipeline, a sensitivity-targeted decision threshold, and a new composite Trade-off Score capturing accuracy and explanation quality together were also introduced. The results point to a genuine trade-off. LRP came out on top for interpretability, reaching a Mean Relevance Score of 0.9078 while holding accuracy at 93.91%, the same as the baseline. Grad-CAM gave the best classification numbers, 95.03% accuracy and 98.72% AUC-ROC, with a still-decent interpretability score of 0.7963. All four variants beat the previously reported baseline, including a 12.30-point jump in specificity for Grad-CAM, a 10.45-point gain in AUC-ROC for Adversarial Training, and a 0.5481 improvement in MRS for SAM. This shows accuracy and interpretability need not be a trade-off accepted blindly, giving clinician’s real evidence to base that choice on.