Hybrid Approach for Multi-Class Skin Cancer Classification with DCNN Feature and Ensemble Techniques

The classification of skin cancer is a critical factor in early diagnosis and treatment. This paper presents a hybrid multi-class skin cancer classification that uses a deep convolutional neural network as the feature extraction model and an ensemble of Histogram Gradient Boosting, Random Forest, and XGBoost classifiers with soft voting for the final prediction. The proposed model achieved a high classification accuracy of 92% on the ISIC dataset and 93% on the PAD-UFES-20 dataset, with corresponding specificity values of 98.92% and 98.75%, respectively. The overall false positive rates were 0.21% and 0.17%, respectively. The precision, recall, and F1-scores also exceeded 91% demonstrating high sensitivity and fairly balanced performance. The false negative rate was also kept relatively low at 1.90% for ISIC and 0.67% for PAD, while the Matthews correlation coefficients(MCC) were quite high at 97.60% for ISIC and 91.85% for PAD. The proposed model is also evaluated on an external dataset, HAM10000, yielding an accuracy of 66.67% confirming generalization capability under a domain shift. This DCNN-ensemble approach to automated skin cancer detection has shown itself to be a solid and scalable solution. Future studies should focus on feature fusion, expanding this approach to more diverse datasets that may encompass different patient populations and analysis, and integrating applications that enable real-time diagnostic support.