Deep Learning Approaches for Pneumonia Detection and Analysis with XAI

Artificial intelligence (AI) has become increasingly popular for assistance in accurate reading of X-ray images. A convolutional neural network (CNN) is a deep learning AI model designed to process and analyze image data. In this study, we investigate four high-performance CNN architectures DenseNet201, Xception, EfficientNetB3, and ResNet-50 for automatic classification of pediatric chest X-rays for Pneumonia detection. Wasserstein Generative Adversarial Network-Gradient Penalty (WGAN-GP) is employed to address class imbalance and enrich the training distribution. All CNNs are implemented in a transfer-learning setting with frozen ImageNet-pretrained backbones, a standardized preprocessing pipeline (resizing, normalization, Gaussian filtering), and lightweight task-specific classification heads. In addition, a hybrid soft-voting ensemble of the three best-performing models is constructed to further improve robustness. Experiments on the chest X-ray images dataset show that the individual models achieve strong performance, while the hybrid ensemble attains the best results with 0.9727 accuracy, 0.9798 precision, 0.9828 recall, and 0.9813 F1-score on the test set. To generate visual explanations, Gradient-weighted Class Activation Mapping (Grad-CAM) is used. The proposed WGAN-GP–augmented, ensemble-based, and explainable framework constitutes a practical and more transparent decision-support tool for pneumonia detection in chest radiography.