Boosting Skin Cancer Detection Accuracy: A Hierarchical Ensemble Approach with Feature Fusion

Accurate skin lesion classification is crucial for early detection and treatment of skin cancer. However, achieving high diagnostic accuracy remains challenging due to inter-lesion variability, class imbalance, and the need for robust feature extraction. To address these challenges, we propose a hierarchical ensemble approach that integrates lesion masking, feature extraction from multiple pre-trained models, feature fusion, and class imbalance mitigation. Our methodology significantly improved skin lesion classification accuracy compared to single-model architectures. Balancing class distributions further enhances performance, with the fusion of ResNet50, DenseNet121, and AlexNet models achieving an accuracy of 91% on the HAM10000 dataset. The ensemble models demonstrated varying robustness across datasets, highlighting the importance of dataset-specific tuning for optimal performance. Feature quantity showed a moderate positive correlation with precision, suggesting a potential link between representational capacity and precise predictions. These findings pave the way for robust and accurate skin lesion diagnosis, potentially supporting early skin cancer detection and clinical decision-making.