A Modified Swin Transformer Architecture for Accurate Melanoma Detection

Melanoma detection typically involves identifying suspicious skin lesions and determining whether they are malignant. Machine learning and deep learning methods have become increasingly popular in detecting melanoma, with Vision Transformer (ViT) playing a major role due to their ability to analyze image data effectively. However, image quality limitations can compromise the robustness of skin image analysis. Therefore, developing an automated intelligent system for skin cancer diagnosis with high detection accuracy is essential. In this study, a Modified Swin Transformer (MST) is proposed by integrating the linear layer from Residual Neural Network (ResNet), and the feature extractor from Swin Transformer on the overall structure of ViT. The performance of this MST is evaluated on dermoscopic images of skin lesions, and its results are compared against state-of-the-art models, including ResNet-101, Swin Transformer, and ViT. The experiments are conducted on a Graphics Processing Unit (GPU) to accelerate the training process. Results show that the proposed MST achieves superior performance compared to ResNet-101, Swin Transformer, and ViT in terms of testing accuracy, precision, recall, and F1-score for melanoma detection. Specifically, the proposed MST achieved the highest testing accuracy of 74.67% and outperformed other models in key metrics.