DermPox: A Skin Lesion Classification System Using State-of-the-Art Deep Learning Models with Explainable AI

Abstract

The global emergence of mpox has prompted efforts to strengthen guidelines for diagnosis, treatment, and prevention to help healthcare providers differentiate it from other diseases with similar clinical presentations. This study develops a deep learning-based system to classify mpox and distinguish it from similar skin lesions (chickenpox, cowpox, and measles) using convolutional neural networks. Four state-of-the-art architectures, ResNet50, MobileNetV3Large, EfficientNetV2L, and ConvNeXtBase, were fine-tuned via transfer learning and evaluated using stratified five-fold cross-validation. To enhance generalization and mitigate skin color bias, two augmentation strategies were applied: standard transformations and a color-based method adjusting HSV channels. ConvNeXtBase achieved the highest performance score across all metrics and was deployed as both a web application (with Grad-CAM and LIME interpretability) and an offline, lightweight mobile version (TensorFlow Lite). Clinical validation by a dermatologist on 40 images showed 100% concordance with expert diagnoses. Explainable AI revealed that model decisions aligned with clinically relevant lesion features, improving transparency. This system offers a rapid, cost-effective alternative to PCR testing, particularly valuable in resource-limited settings and limits exposure.