Explainable Skin Cancer Classification Using Deep Feature Fusion and Ensemble Learning

Skin cancer is one of the most dangerous and rapidly increasing forms of cancer worldwide, where early melanoma detection significantly improves survival rates. However, accurate diagnosis using dermoscopic images remains challenging due to class imbalance, visual similarity among lesions, and limited annotated medical datasets. This study proposes an explainable hybrid skin cancer classification framework integrating deep feature fusion with ensemble machine learning techniques for melanoma detection using the HAM10000 dataset. Pre-trained convolutional neural networks, namely ResNet50 and MobileNetV2, are employed as deep feature extractors, while Support Vector Machine (SVM) and XGBoost classifiers are utilized for classification. Synthetic Minority Oversampling Technique (SMOTE), image augmentation, and transfer learning strategies are incorporated to improve robustness and generalization. Furthermore, explainable artificial intelligence (XAI) techniques, including LIME and SHAP, are integrated to enhance interpretability and clinical reliability. Experimental results demonstrate that the SVM classifier achieved the highest classification accuracy of 80.09%, specificity of 86.97%, and AUROC of 0.8641, while XGBoost achieved superior sensitivity and precision-recall performance. The proposed framework can support dermatologists in reliable melanoma screening and can be extended for multiclass skin lesion classification.