Enhanced Skin Lesion Classification via Swin Transformer and Random Data Augmentation

Early diagnosis of skin cancer is critical for improving the survival rate of patients with malignant melanoma. However, conventional dermoscopic examination heavily relies on dermatologists’ expertise and is prone to subjective bias. Traditional computer-aided diagnosis (CAD) methods alleviate some subjectivity but depend on manual feature engineering and suffer from limited receptive fields of convolutional operations, making it difficult to capture the global heterogeneity of skin lesions. To address these challenges, we propose a skin lesion classification method that integrates Swin Transformer with random data augmentation techniques. A set of augmentation strategies is employed to increase training sample diversity and reduce overfitting. Swin Transformer serves as the backbone network, leveraging hierarchical feature extraction, window-based self-attention, and multi-scale fusion to enhance representation learning and classification performance. Additionally, the model is optimized using a hybrid loss function that combines crossentropy and focal loss, improving the performance of hard-to-classify samples. Experimental results demonstrate that the proposed method achieves an accuracy of 82.8%, outperforming conventional CNN-based models such as ResNet34 and ConvNeXt. Furthermore, the window-based self-attention mechanism effectively addresses the limitations of convolutional operations by dynamically modeling long-range dependencies and capturing global contextual information of skin lesions.