Stacking-based Deep Ensemble for Skin Cancer Detection

One of the most common and deadly cancers in the world is skin cancer, and early detection greatly enhances the prognosis for patients. This study presents a comprehensive deep learning- based framework for the classification and staging of skin cancer using dermoscopic images from the HAM10000 dataset. To address class imbalance, resampling and augmentation techniques were employed, followed by training three state-of-the-art convolutional neural network (CNN) architectures—ResNet50, DenseNet121, and EfficientNetV2B0. A stacking ensemble strategy was implemented, where the predictions of the base CNNs served as metafeatures for a Random Forest classifier, yielding improved generalization and robustness. Additionally, lesion staging was performed through image processing techniques, where contour-based diameter estimation combined with anatomical localization provided rule-based stage determination. Experimental evaluation demonstrated that the stacking ensemble achieved superior classification performance compared to individual CNNs.