Skin Cancer Classification Using Gaussian-Weighted Swin Transformer

چکیده مقاله

Skin cancer is considered one of the most common and, at the same time, one of the deadliest types of cancer worldwide; therefore, early and accurate diagnosis plays a vital role in successful treatment. In this study, we propose a novel deep learning-based method for skin lesion classification using the Swin Transformer architecture enhanced with Gaussian-based attention weighting. The Swin Transformer is renowned for its hierarchical structure and its shifted window attention mechanism. In this study, we add Gaussian weighting to the attention maps to help the model focus more on the middle areas of each window. This is a good design choice because lesions are usually in the middle of dermoscopic images. The proposed model is trained on the publicly available HAM10000 and ISIC-2018 datasets. The model employs techniques such as focal loss and dynamic learning rate scheduling to expedite the learning process. Also, incremental random oversampling is used to remedy class imbalance and make the samples from the minority class more diverse. The results of the evaluation show that the proposed method can accurately classify a wide range of lesion types and is competitive with the best methods available.