Explainable Deep Learning for Early Melanoma Detection Using Dermoscopic Images

Abstract: This paper presents an early melanoma detection system using an Explainable Artificial Intelligence (XAI) deep learning model applied to skin lesion images. The proposed system integrates three convolutional neural networks: ResNet-50, VGG-16, and SqueezeNet. LIME (Local Interpretable Model-Agnostic Explanations) is used to enhance both classification accuracy and model interpretability. Experiments were conducted on two benchmark dermoscopic image datasets: the skin lesion melanoma dataset and the HAM10000 dataset (Human Against Machine, with 10,000 training images). On the first dataset, ResNet50, VGG16, and SqueezeNet achieved classification accuracies of 85.64%, 88.08%, and 90.50%, respectively. On the HAM10000 dataset, the models achieved accuracies of 95.15%, 96.06%, and 97.80%, respectively. Comparative analysis indicates that SqueezeNet consistently delivers superior performance across both datasets. By combining high-performing deep learning models with explainable AI techniques, the proposed framework offers a reliable and transparent solution for early melanoma detection, supporting clinical decision-making and improving diagnostic confidence.