Efficient and explainable skin lesion classification: a single ConvNeXt-Large surpassing ConvNeXt ensembles with Grad-CAM

Abstract

Automatic classification of skin lesions from dermoscopic images plays a crucial role in early cancer detection. However, this task remains challenging due to visual similarities between lesion types, class imbalance, and image artifacts. While recent deep learning approaches, including ConvNeXt-based ensembles, have demonstrated impressive classification accuracy, their clinical reliability is hindered by uneven performance across classes, especially lower sensitivity in detecting high-risk conditions. In this study, we propose an efficient and interpretable framework based on a single ConvNeXt-Large model,trained and evaluated on the ISIC 2019 skin lesion dataset. Through systematic empirical optimization, we achieved a balanced performance that surpasses ensemble methods in clinically critical metrics, attaining a sensitivity of 86.27% and a specificity of 98.08% on the validation set. This was accomplished without compromising overall accuracy, while significantly reducing computational complexity compared to ensemble architectures. To enhance transparency and clinical trust, we further integrated Grad-CAM visual explanations, which reveal that the model consistently focuses on medically relevant lesion regions. Overall, this work contributes a practical, lightweight, and trustworthy deep learning solution for skin lesion classification that balances accuracy, efficiency, and interpretability—making it more suitable for real-world deployment.