Interpretable Deep Learning-Based Arsenicosis Classification from Regional Skin Lesion Images Using MobileNetV3-Small

Arsenicosis, a chronic condition resulting from prolonged exposure to arsenic-contaminated groundwater, remains a major public health issue in several regions of Bangladesh. Early detection of arsenicosis through visible skin lesions can significantly aid medical intervention, yet clinical diagnosis is often limited by accessibility and subjectivity. This study proposes a deep learning-based framework for automated arsenicosis classification using digital skin lesion images collected from affected regions. A MobileNetV3-Small architecture, pre-trained on ImageNet, was employed for binary classification of affected versus healthy skin images. The model achieved an impressive 95.42\% classification accuracy with significantly fewer number of trainable parameters. Furthermore, Gradient-weighted Class Activation Mapping (Grad-CAM) was integrated to generate visual explanations, confirming that the network consistently focused on clinically relevant lesion regions during prediction. The lightweight nature of MobileNetV3-Small ensures low computational cost, making the model practical for deployment in mobile or edge-based healthcare systems. Overall, the proposed approach presents a reliable, interpretable, and efficient method for early arsenicosis diagnosis, contributing to scalable, technology-assisted healthcare screening in resource-constrained communities.