Efficient and Scalable Arsenicosis Detection Using Lightweight CNNs and Transformer-Based Models for Resource-Limited Settings

Drinking water contamination with arsenic is still a major global health issue, especially in low-resource settings. Our research proposes a computationally efficient lightweight Convolutional Neural Network (CNN) for automatic detection of arsenicosis and achieves a testing accuracy of 98.75%. It employs modern data augmentation methods, improving robustness and generalizability in various environments. A thorough evaluation against state-of-the-art architectures, including Swin Transformer, Vision Transformer (ViT), and EfficientNetV2-B0, emphasizes the lightweight CNN’s potential to provide competitive accuracy while being amenable to real-time tasks in resourced-constrained settings. Trained on the publicly available ArsenicSkinImageBD dataset, this work should serve as an illustration of how scalable AI-driven diagnostic tools can be developed to tackle global health problems. We hope that by providing a regional solution between state-of-the-art approaches and real-world implementation we can lay the groundwork for a way to bridge the gap on accessible and viable solutions for medical imaging.