Around the globe, pneumonia is the leading cause of morbidity and mortality, especially in areas with poor access to medical expertise. This study proposes a robust deep learning framework for automated pneumonia detection using chest X-ray (CXR) images using three clinically diverse and publicly available Kaggle datasets. Three convolutional neural network models were developed and compared: a lightweight CNN based on MobileNetV2 trained on Paulo Breviglieri’s Imbalanced Pneumonia X-ray Images dataset; a ResNet50-based transfer learning model trained on Tolga’s Chest X-ray dataset; and a DenseNet121-based model trained on Paul Mooney’s Chest X-ray Images (Pneumonia) dataset. The highest test accuracy of 94.87% and a strong ROC AUC score of 0.99 were achieved by the ResNet50 model, which was followed by DenseNet121 at 86.38% and MobileNetV2 at 84.94%. All models underwent standardized preprocessing, including augmentation, normalization, and resizing. For interpretability, all models underwent Grad-CAM visualizations, which effectively emphasized the pneumonia-affected lung regions. These results demonstrate the deep learning models’ effectiveness and generalizability for classifying pneumonia as well as their potential as scalable, interpretable, and cost-effective diagnostic tools in clinical and resource-constrained settings.