Abstract
<title>Abstract</title> <p>Pneumonia is a worldwide health problem associated with considerable morbidity and mortality, and prompt diagnosis is vital to decrease its burden. We present an explainable hybrid deep learning framework using ConvNeXt-Base with a channel attention mechanism based on the CBAM approach and Vision Transformer (ViT-B/16) for automated pneumonia detection from chest X-ray images. The divided architecture aims to memorize local spatial features and global contextual dependencies that contribute positively to the classification performance. A combined dataset of 6,590 chest X-ray images, comprising 4,400 pneumonia and 2,190 normal cases.We applied standard preprocessing and augmentation techniques to enhance diversity in the dataset and aid generalization. We pre-trained the model with transfer learning with ImageNet pre-trained weights and optimized it with the AdamW optimizer with cosine annealing scheduling. The hybrid model proposed in this paper improves the accuracy, precision, recall, and F1-score to 94.03%, 95.01%, 92.74%, and 93.62%, respectively, indicating better performance than several other baseline models such as ConvNeXt, MobileNetV2, ResNet152, Vision Transformer, and DenseNet121 by an order of magnitude. For the first time, Grad-CAM-based visualization can visualize the clinically relevant areas for chest X-ray images from different datasets, which greatly improves model interpretability. The proposed framework as a computer-aided diagnostic tool is proven to be reliable and efficient with high potential of help available for the healthcare professionals to diagnose pneumonia at an early stage with more precision.</p>