An Ensemble Multi-CNN Framework for Enhanced Skin Cancer Classification

Skin cancer is a rapidly growing disease that demands initial and precise detection for effective treatment. Automatic classification, however, is hindered by issues like class imbalance, intra-class variation, and differences in lesion characteristics. To overcome these obstacles, this study introduces an ensemble learning framework combining three advanced Convolutional Neural Networks (CNNs)—ResNet50, EfficientNetB4, and DenseNet201. Each model contributes distinct advantages: DenseNet201’s dense connectivity, EfficientNetB4’s efficient scaling, and ResNet50’s residual learning. These models work together to extract strong, complementary features from dermatoscopic images, which are then classified using ensemble methods like XGBoost. The approach achieves impressive results, with 97.23% and 95% accuracy on the ISIC 2017 and HAM100000 datasets, respectively. By tackling challenges like feature redundancy and class overlap, this method significantly enhances multiclass skin cancer classification, providing a promising and efficient solution for clinical applications.