Explainable ensemble deep learning with semantic searching for skin disease classification: a web-based system

Skin diseases are becoming more common around the world. Diagnosing them at an early stage is challenging because of the similarity in shape of different types of lesions and because imaging conditions can vary. This study proposes a unified and lightweight framework for skin lesion classification using the International Skin Imaging Collaboration 2019 dataset by integrating deep learning, stacking ensemble learning, feature embedding-based semantic image retrieval, explainable AI (Grad-CAM and Grad-CAM++), and web-based deployment. The experimental results show that CNN and hybrid CNN–LSTM achieved an accuracy score of 0.9700 and 0.9500, with AUC scores of 0.9987 and 0.9888, respectively. The stacking ensemble achieved high-performance using a XGBoost meta-model, with an accuracy score of 0.9733 and an AUC of 0.9968. While this improvement is modest, the ensemble reduces variance and improves stability and generalization. The main contribution is the integration of classification, case-based semantic retrieval, visual explanation, and practical deployment in a single framework to enable the retrieval of similar images that share the same lesion features as supporting evidence for predictions. Semantic-based retrieval achieved a mAP@5 of 0.981, demonstrating strong feature representations. Since the lightweight architecture has 489,096 parameters, it enables real-time CPU inference and web-based deployment without GPU requirements. Overall, the proposed architecture provides an interpretable, evidence-supported, and practical system for skin lesion classification.