Multiclass skin lesion classification in dermoscopic images using swin transformer model

Automatic skin lesion classification in dermoscopic images is a very challenging task due to the huge intraclass variation, the high degree of interclass visual similarity, low contrast between skin lesion and surrounding normal skin, and the existence of extraneous and intrinsic artifacts. However, existing algorithms for skin lesion classification are developed by leveraging convolutional neural networks (CNNs), and the effectiveness of these algorithms is mostly validated for binary classification of skin lesions. In addition, the relatively low diagnostic sensitivity achieved by these studies demonstrates the uncertainty involved in skin lesion classification. In order to overcome these difficulties, a swin transformer model for multiclass skin lesion classification is proposed by taking advantage of both transformer and CNNs that are based on end-to-end mapping and do not require prior knowledge. Furthermore, the problem of class imbalance is addressed through a weighted cross entropy loss. Moreover, key components of the proposed approach are explored in detail in order to ensure efficient and effective learning process with multiclass data in the skin lesion classification. The proposed method is extensively evaluated on International Skin Imaging Collaboration (ISIC) 2019 Skin Lesion Analysis Towards Melanoma Detection Challenge dataset and achieves a sensitivity, specificity, accuracy, and balanced accuracy value of 82.3%, 97.9%, 97.2%, and 82.3%, respectively. Experimental results demonstrate that the proposed method has the highest balanced accuracy value and outperforms most of the other state-of-the-art methods in multiclass skin lesion classification.