Comparsion of CNN and Transformer Models with Transfer Learning in Monkeypox Diagnosis

Abstract

In this study, five different CNN-based and Transformer-based deep learning models, namely ResNet50, EfficientNet-B3, ConvNeXt-Tiny, ViT-Base, Swin-Tiny, are comparatively evaluated on a six-class dataset consisting of Monkeypox, HFMD, Chickenpox, Varicella, Measles, Cattlepox and healthy skin images. The models trained by transfer learning on the open source dataset using data augmentation techniques were tested with 5-fold cross-validation and analyzed in terms of metrics such as accuracy, precision, recall, F1-score and training time for each fold. According to the results, the ConvNeXt-Tiny model showed the highest classification performance with an accuracy of 82.30% and an F1-score average of 81.36%, and it also outperformed the other models in terms of computational efficiency by training in less time. While EfficientNet-B3 and ResNet50 models offer balanced performance with low training time, ViT-Base and Swin-Tiny models are found to be disadvantageous for limited resource systems due to their long training time and high hardware requirement despite their high representativeness. In this context, the ConvNeXt-Tiny model is proposed as a suitable and feasible model for medical image classification on small datasets with class imbalance. © 2025 IEEE.