Comparative Deep Learning and Explainable AI Approaches for Automated Detection of Arsenic Skin Lesions

作者：Md. Siam Uddin Molla Antor, Nitun Kumar Podder, Nakib Aman, Md. Raihanul Haque, Shahriar Siddique Arjon, Tamanna Yasmin · 年份：2025 · DOI：10.1109/compas67506.2025.11381857 · 被引用次数：4 · 研究领域：Cutaneous Melanoma Detection and Management、Arsenic contamination and mitigation、Acne and Rosacea Treatments and Effects

Arsenic contamination continues to be a significant public health threat in some parts of South Asia, leading to severe skin lesions, as well as an increased likelihood of skin cancer, from long-term exposure. This study presents a deep learning framework for the automated detection of arsenic skin lesions with the use of a modified VGG16 architecture, including Gradient-weighted Class Activation Mapping (Grad-CAM) for explainability. The proposed framework was developed and evaluated on an open-access dataset of arsenic-induced skin lesions, achieving validation accuracy of 99.50%, which is better accuracy compared to lightweight baseline (i.e., MobileNetV2, SqueezeNet) models. To bolster interpretability and build a sense of clinical trust, additional visualization techniques, including LIME and t-SNE, were deployed within this study to confirm that the model was detecting relevant areas in the lesions. The experimental results presented in this study demonstrate excellent classification performance that was achieved with little overfitting, highlighting the framework as promising for successful real-world screening and potential for early intervention in austere clinical settings. In the end, this work establishes the foundation for strong and explainable AI solutions to assist in the early diagnosis of environmental dermatosis.