Hybrid Deep Learning Model for Skin Cancer Classification: Integrating PSO-Tuned CNN with Stacking Ensemble Learning

Skin cancer remains a leading cause of morbidity worldwide, necessitating early detection for effective treatment. This paper presents optimized EfficientNet-PSO ensemble model (OEP-SEM), a deep learning framework for automated skin cancer classification utilizing the HAM10000 dataset. EfficientNetB1 serves as a feature extractor, while the final classification model, referred to as Optimized PSO-CNN classifier (OPC-CNN), is optimized using Particle Swarm Optimization (PSO) for hyper parameter selection. The extracted features are not directly classified by a CNN, but instead, they are used to train OPC-CNN, which is further optimized and integrated into a stacking ensemble comprising a CNN wrapper, support vector machine (SVM), and XGBoost, with logistic regression as the meta-classifier to enhance predictive accuracy. Furthermore, a real-time prediction system is integrated to facilitate practical usability. Extensive experimentation confirms that OEP-SEM surpasses conventional CNN-based models, providing an efficient and interpretable dermatological diagnosis solution. The inclusion of PSO enhances the adaptability of OPC-CNN, while ensemble learning improves classification accuracy compared to standalone CNN-based models. Our results indicate a significant improvement in classification performance, robustness, and generalization, making this approach suitable for real-world medical applications.