PoXeptionNet: Accelerating Pox Disease Identification Using a Dual Neural Network Fusing Xception and EfficientNetB0 Architectures

Pox diseases remain prevalent dermatological conditions, and timely detection is essential for mitigating transmission and supporting clinical decision-making. This work introduces PoXeptionNet, a dual-branch convolutional neural network that combines Xception and EfficientNet-B0 through a feature-fusion mechanism to enhance multiclass pox disease recognition. The proposed architecture captures complementary multiscale representations, improving robustness across visually similar lesion categories. Experiments conducted on a six-class skin disease dataset demonstrate that PoXeptionNet achieves a test accuracy of 95.24% and an AUC of 0.95, with precision, recall, and F1-scores all exceeding 95%. These results significantly outperform several established deep learning baselines. We also tested the generalization ability of our model using cross-dataset validation on other two independent external datasets. It obtained 94.3% test accuracy on a binary dataset (Mpox vs others) and 85.5% test accuracy on a four-class of pox-related diseases. Explainable AI techniques such as Grad-CAM, Grad-CAM++, Integrated Gradients, and LIME, were employed to highlight discriminative regions, providing transparency into model behavior and supporting clinical interpretability. A prototype web application was developed to demonstrate real-time deployment feasibility. The results indicate that PoXeptionNet offers an effective and interpretable approach for automated pox disease analysis.