A Hybrid Deep Learning Model for Multiclass Skin Cancer Classification Using ConvNeXtV2 and Separable Self-Attention Mechanisms

Abstract

Objectives: Skin cancer is among the most prevalent and life-threatening cancers worldwide, making accurate and automated diagnostic approaches essential. This study aims to develop an efficient deep learning–based framework for multiclass skin lesion classification to enhance early detection and support clinical decision-making. 

Methodology: The study proposes a hybrid deep learning architecture that integrates ConvNeXtV2 convolutional blocks with separable self-attention mechanisms, enabling simultaneous learning of fine-grained local features and long-range contextual dependencies. Transfer learning, extensive data augmentation, and class balancing techniques were employed to improve model generalization and robustness. The proposed model was trained and evaluated on the HAM10000 dataset using five-fold stratified cross-validation to ensure reliable performance assessment. 

Results: Experimental results demonstrate that the proposed framework outperforms conventional convolutional neural networks and Vision Transformer-based models. The model achieved an average accuracy of 93.52%, precision of 93.17%, recall of 91.24%, and an F1-score of 92.18%, along with a ROC-AUC value of 0.957, indicating strong discriminative capability across multiple skin lesion classes. 

Conclusion: The findings confirm that combining ConvNeXtV2 with effective attention mechanisms provides a computationally efficient and highly accurate solution for automated skin cancer detection. Future work will focus on multimodal data integration, edge deployment, and clinical validation to enhance real-world applicability and translational impact in healthcare settings.