Comparative Analysis on Data Balancing and Augmentation in Skin Cancer Image Classification Using Multiple Datasets with Explainable AI

Abstract

Skin cancer is one of the most common types of cancer worldwide, with a continuously increasing incidence rate. Early detection and accurate classification of skin lesions are crucial for improving patient survival, particularly for melanoma. This study presents a comparative analysis of data-balancing strategies for skin cancer image classification across multiple publicly available datasets, including ISIC 2024, ISIC 2019, ISIC 2020, HAM10000, and PROVe-AI. The ConvNeXt-Tiny architecture is employed as the classification model and evaluated under three dataset configurations, namely RAW without balancing, Undersampling, and Oversampling via data augmentation, combined with two Learning Rate (LR) settings of 0.001 and 0.0001. In addition to reporting classification performance, this study emphasizes a comparative evaluation of accuracy, training stability, and computational efficiency across different balancing strategies. The experimental results indicate that the RAW dataset with an LR=0.0001 provides the best trade-off between performance and efficiency, achieving a validation accuracy of 99.88% and an F1-score of 0.9988. Oversampling via augmentation achieves the highest performance, with a validation accuracy of 99.93% and an F1-score of 0.9993, but requires substantially higher computational resources. Undersampling enables faster training with lower resource consumption, although it results in a slight performance degradation. Furthermore, Explainable Artificial Intelligence techniques, including Grad-CAM and LIME, demonstrate that models trained with an appropriate learning rate consistently focus on clinically relevant lesion regions, thereby improving interpretability and trustworthiness.