Pneumonia is a major cause of mortality worldwide, necessitating fast and accurate diagnosis. A deep learning architecture using advanced architectures like Xception, EfficientNet B0, and ResNet50, along with an end-to-end architecture based on vision transformers (ViT), is presented. Explainable artificial intelligence (XAI) techniques like Grad-CAM, LIME, Integrated Gradients, and Saliency Maps are used to grant transparency to model prediction. The models are compared to current methods for pneumonia recognition, highlighting the trade-offs between accuracy, generalization, and computational efficiency. The explanation quality of various XAI methods is demonstrated, with ResNet50 and EB0 showing strong accuracy and low overfitting. ViT suggests the ability to manage larger datasets. The XAI techniques provide local insights into pneumonia detection and improved model interpretability, highlighting the need for merging high-performing models with explainability in medical AI applications.