Hybrid Deep Learning for Skin Cancer Classification Using Dermatoscopic Images With Explainable AI

Abstract

Skin cancer is one of the most prevalent and life-threatening diseases worldwide, with melanoma representing the most aggressive form due to its high metastatic potential and mortality rate. Early and accurate detection is essential for improving treatment effectiveness and patient survival outcomes. However, conventional diagnostic approaches based on visual inspection and dermatoscopy remain highly subjective and dependent on clinician expertise, which can lead to variability and misdiagnosis. To overcome these limitations, this study proposes an explainable hybrid deep learning framework that integrates Convolutional Neural Networks (CNNs) and Gated Recurrent Units (GRU) for automated skin cancer classification using dermatoscopic images. The proposed CNN–GRU model leverages the CNN’s capability to extract fine-grained spatial features such as texture, color variation, and lesion boundaries, while the GRU component captures broader contextual dependencies across feature representations to enhance classification robustness. The model was trained and evaluated using publicly available datasets, including the ISIC Archive and HAM10000, which contain high-quality annotated dermatoscopic images of benign and malignant skin lesions. Standard preprocessing techniques, including image resizing, normalization, and data augmentation, were applied to improve dataset diversity and reduce overfitting. An explainable AI framework was incorporated to improve clinical interpretability, utilizing Gradient-weighted Class Activation Mapping (Grad-CAM) for visualizing CNN feature importance and attention-based visualization for interpreting GRU outputs. These methods provide transparent insights into the regions of interest influencing model predictions, thereby increasing clinical trust and interpretability. Experimental results show that the proposed CNN–GRU hybrid model achieved an overall classification accuracy of 85%, outperforming the standalone CNN model, which achieved 84%. The hybrid model also recorded improved macro and weighted average precision, recall, and F1-scores of 0.85 compared to 0.84 for the CNN baseline, indicating more balanced and reliable performance. At the class level, the model achieved a precision of 0.87 and recall of 0.84 for benign lesions, and a precision of 0.81 and recall of 0.86 for malignant lesions, demonstrating strong sensitivity in detecting malignant cases, which is critical for reducing false negatives in clinical diagnosis. Hence, the results confirm that the integration of CNN and GRU architectures enhances feature representation by combining local spatial feature extraction with global contextual learning. The hybrid model demonstrates improved generalization, stable convergence, and reduced misclassification across classes. When combined with explainable AI techniques, the framework not only delivers high classification performance but also provides interpretable decision support, making it a reliable and clinically applicable tool for early skin cancer detection and diagnosis.