HPD-Net: A Hybrid CNN–Transformer Architecture for Pneumonia Detection from Chest X-ray Images

Abstract

Pneumonia remains a major global health concern and a leading cause of respiratory-related mortality. Automated analysis of chest X-ray (CXR) images using deep learning has emerged as a promising tool for assisting clinical diagnosis. This thesis presents HPD-Net (Hybrid Pneumonia Detection Network), a hybrid CNN–Transformer architecture designed for binary pneumonia classification from chest radiographs. The proposed model combines EfficientNetV2-S, which extracts fine-grained local features, with Swin Transformer Tiny, which captures global contextual information. To effectively integrate these complementary representations, a learnable gated fusion module is introduced. Experiments were conducted on the Kermany chest X-ray dataset using a custom stratified 70/15/15 train-validation-test split. The performance of HPD-Net was compared against two baseline models: EfficientNetV2-S and Swin Transformer Tiny. The proposed model achieved an AUC-ROC of 0.9947 and a precision of 0.9919, demonstrating competitive performance for pneumonia detection. Furthermore, Grad-CAM++ was employed to provide visual explanations of model predictions by highlighting diagnostically relevant lung regions. The results demonstrate the potential of hybrid CNN–Transformer architectures for accurate and interpretable medical image analysis.