Shifted windowing vision transformer-based skin cancer classification via transfer learning

Skin cancer is widely recognized as one of the most perilous diseases on a global scale. Early identification of skin lesions can significantly enhance the treatment effects by aiding in clinical decision-making, hence mitigating the risk of disease progression and metastasis. Unfortunately, the skin images used for training are usually limited and imbalanced. To tackle the aforementioned challenges, a considerable number of deep learning models have been introduced for distinguishing skin cancer. And promising findings have been achieved by these algorithms. Nevertheless, the cross-domain adaptability and resilience of these models remain unresolved for actual clinical applications. One limitation of convolutional neural network-based deep models is their difficulty in finding the embeddings of global relationships for an image. This issue arises from the restricted receptive field of the convolutional operator, which primarily focuses on local information. To address this disparity, the present work introduces an innovative approach based on a vision transformer, with the aim of achieving skin cancer categorization. The presented model leverages the attention mechanism. Furthermore, transfer learning is employed to enhance the weighting parameters of the proposed model. Experimental results show the superior performance of the proposed approach compared to traditional approaches, indicating the efficacy of the proposed vision transformer architecture. The promising results demonstrate that the proposed approach is a potentially valuable instrument for skin cancer identification.