Abstract This study addresses the challenge of multiclass skin lesion classification, which is critical in clinical practice due to the high incidence of skin cancer and the need for early detection. However, accurate diagnosis remains difficult due to visual similarity between lesions and the limited availability of expert dermatologists, particularly in resource-constrained settings. While deep learning models have shown promising performance, many existing approaches rely on complex architectures that are not suitable for efficient deployment. To address this limitation, this study proposes a lightweight classification framework based on transfer learning using a modified EfficientNetV2B0 architecture. The experiments were conducted using the ISIC 2019 dataset, focusing on four representative skin lesion classes with imbalanced data distribution. To improve model performance, data balancing through augmentation and structured hyperparameter optimization were applied, including variations in learning rate, dropout, and batch size. In addition, the impact of input image resolution was investigated to enhance feature representation. The results show that the best-performing model, using an input resolution of 448 × 448, achieves a test accuracy of 89.94%, with a macro-average recall of 88.70%, macro-average precision of 88.42%, and a macro-average F1-score of 88.56%. These findings indicate that higher-resolution inputs improve classification performance, although the improvement becomes marginal at higher resolutions. Compared to existing methods, the proposed model provides competitive performance while maintaining parameter efficiency, with only approximately 6.2 million parameters, making it suitable for deployment in resource-constrained environments.
Fathonah et al. (Sun,) studied this question.