A Lightweight Hybrid CNN Model for Classification of Arsenic-Induced Skin Lesions

Chronic arsenic poisoning, largely attributable to sustained intake of contaminated groundwater, represents a critical public health concern, particularly within South Asian populations. It leads to severe medical conditions such as skin lesions, cancer, and cardiovascular ailments. This study introduces an advanced deep learning framework designed to automatically detect and classify arsenic-induced skin lesions using images captured via smartphones in Bangladesh. The developed hybrid convolutional neural network (CNN) architecture combines parallel CNN pathways, Dense and Residual blocks for efficient feature extraction, and lightweight Fire modules to optimize computational performance. Preprocessing of images involved resizing, normalization, and specialized augmentation strategies to mitigate class imbalances and enhance the training efficacy. The model achieved outstanding accuracy of 98.33%, surpassing the performance of standard CNNs (DenseNet121, ResNet50V2) and other hybrid configurations. Detailed performance assessments confirmed robust predictive power, yielding precision of 0.9900, recall of 0.9867, F1-score of 0.9883, specificity of 0.9867, Cohen’s kappa of 0.9767, and Matthews correlation coefficient of 0.9766. Further validation through interpretability techniques such as Grad-CAM and Grad-CAM++ illustrated the model's precise identification of clinically significant lesion areas, thereby enhancing confidence and interpretability of predictions. Its compactness, computational efficiency, and superior accuracy demonstrate significant potential for real-time diagnostics, especially beneficial in resource-limited healthcare environments, facilitating early diagnosis and intervention for arsenic-related conditions.