Research on Dermoscopic Lesion Classification Based on Deformable Transition Layer and Multi-path Hybrid Attention Fusion

As the largest organ of the human body, the skin undertakes crucial physiological functions such as protecting the organism and regulating metabolism, and its health status is directly related to the quality of human life. Skin lesions are diverse in types and similar in morphology; failure to accurately identify and intervene in a timely manner can easily lead to delayed diagnosis and treatment, posing a serious threat to patients' lives and health. However, dermoscopic images are commonly plagued by irregular lesion morphology, blurred boundaries and low discrimination of fine-grained features, which restrict the classification performance of traditional deep learning models. To address the above pain points, an improved DenseNet121 architecture integrating multi-path hybrid attention and deformable transition layer (HADT-DenseNet) is proposed. Based on DenseNet121, this architecture introduces deformable convolution to replace the traditional convolution in the transition layer to realize adaptive sampling of irregular lesion features. Meanwhile, a multi-path branch structure is designed, with hybrid attention blocks inserted between key dense blocks to enhance attention features in both channel and spatial dimensions. Through a hierarchical feature fusion strategy, the basic dense features are organically combined with attention-enhanced features to improve the multi-level nature of feature expression. Experimental results on the public ISIC 2017 dataset show that HADT-DenseNet achieves a classification accuracy of 92.3%, a precision of 88.2%, an AUC of 96.1% and an F1-score of 91.8%, which is significantly superior to the baseline model DenseNet121 and other improved variants based on common attention mechanisms. This study provides an effective and feasible new approach for the intelligent auxiliary diagnosis of skin lesions.