A hybrid deep feature optimized ensemble framework for multiclass skin disease classification

Skin diseases constitute one of the most prevalent dermatological disorders worldwide and their early and precise diagnosis is essential for reducing diagnostic delays as well minimizing disease associated mortality. Motivated by this clinical need this work proposed a novel hybrid skin lesion classification framework termed ResBR-XGBR is proposed, which integrates deep feature extraction using ResNet50, hybrid feature optimization and ensemble decision fusion using Random Forest and XGBoost classifiers for the multiclass skin lesion classification using HAM10000 dataset. To capture rich hierarchical features, two pretrained deep convolutional architectures, ResNet50 and EfficientNet-B0 were employed. Subsequently, a multi-stage feature optimization strategy incorporating: Boruta, Chi-square, and Recursive Feature Elimination (RFE) was employed to remove redundant and irrelevant feature while retaining only the most discriminative, statistically, and diagnostically features. The selected features were further evaluated on multiple classifiers using an 80:20 training testing split with Grid Search Cross validation to perform hyperparameter tuning and identify the optimal learning for each model. In scenario 1, the feature subsets obtained through Boruta and Recursive Feature Elimination (RFE) demonstrated superior compatibility with Random Forest, XGBoost, and their ensemble classifier, where the best accuracy of 91.4%±1.2% was achieved using ResNet50 features. Motivated by these findings, scenario 2 involved the fusion of Boruta and RFE selected features, followed by re-evaluation across individual and ensemble classifiers. The experimental results revealed that the proposed hybrid ensemble framework (ResBR-XGBR) consistently outperformed all standalone learners and attained the highest test classification accuracy of 95.2%±0.5%.