A Hybrid Swin Transformer–ResNet50 Architecture for Skin Lesion Classification

This study presents a hybrid architecture combining local attention and deep feature extraction for the classification of skin lesions. A hybrid structure was developed by integrating the shifted window-based local attention mechanism of Swin Transformer with the deep feature extraction capability of ResNet50. In experiments conducted on the HAM10000 dataset, the standalone Swin Transformer model achieved a maximum validation accuracy of 88.82%, while the proposed Swin Transformer–ResNet50 hybrid model reached a maximum validation accuracy of 90.01%. When evaluated on the independent test set, the Swin Transformer and hybrid models achieved classification accuracies of 89.0% and 90.0%, respectively. The performance of the proposed model was comprehensively evaluated using ROC curves, confusion matrices, and class-based performance metrics. To improve classification performance in minority classes, focal loss, class-based data augmentation, CutMix, and MixUp techniques were employed. The results demonstrate that the proposed hybrid model provides an effective and reliable solution for skin lesion classification, achieving high accuracy and balanced performance across different lesion categories.