Lightweight CNN-SVM Framework Optimized with HMO for Robust and Explainable Skin Cancer Diagnosis

Background: Skin cancer, particularly melanoma, poses a significant global health challenge, necessitating accurate and early detection methods. Although deep learning-based systems have demonstrated potential in dermoscopic image analysis, challenges such as limited interpretability, lack of robustness, and poor generalization across datasets impede their clinical adoption. Methods: We propose a lightweight, interpretable, and uncertainty-aware framework for automated skin cancer detection. The system combines a Harmonic Mean Optimizer (HMO)-tuned Convolutional Neural Network (CNN) with a Support Vector Machine (SVM) classifier. It also features Monte Carlo Dropout-based uncertainty estimation and Grad-CAM for visual explainability. The model was trained on the American Cancer Society (ACS) dataset and validated without retraining on the ISIC 2018 dataset. Its robustness was evaluated under synthetic perturbations, including noise, brightness variation, and rotation. Results: The proposed HMO-CNN+SVM achieved 96.1% accuracy, 96.8% sensitivity, and 94.7% specificity on the ACS dataset. On the ISIC 2018 dataset, it maintained 94.8% accuracy, demonstrating strong cross-dataset generalization. Accuracy remained above 93.7% under all perturbation scenarios. Grad-CAM heatmaps aligned with expert-annotated lesion regions, and entropy-based confidence scores effectively distinguished uncertain predictions. Conclusions: The proposed framework combines accuracy, interpretability, uncertainty handling, and robustness in a lightweight design, making it suitable for deployment in both high- and low-resource clinical settings. It addresses key limitations of current AI-based diagnostic tools and enhances the reliability of automated melanoma screening systems.