Enhancing Skin Lesion Classification with Streamlined Self-Attention and Multi-scale Features

Using computer technology for automatic diagnosis of various skin lesions from dermoscopic images has remained a challenging task up to now. Skin cancer, one of the deadliest forms of cancer, makes accurate classification of dermoscopic images particularly crucial. However, existing methods based on Convolutional Neural Networks (CNNs) for automatic skin lesion diagnosis require substantial computational costs and memory demands, and most of these methods cannot effectively run on mobile or embedded devices. To address this issue, we propose EdgeNeXt-ASCEND, an improved lightweight skin disease classification network that achieves a balance between computational load and classification effectiveness by simplifying the attention mechanism and expanding the classifier. Experimental results on the ISIC-UFES dataset, constructed from the ISIC2019 and PAD-UFES-20 datasets, indicate that our proposed method outperforms similar algorithms with only a fraction of the computational requirements and parameters.