DenseNet201SA++: Enhanced Melanoma Recognition in Dermoscopy Images via Soft Attention Guided Feature Learning

ABSTRACT As the first line of defense in the human immune system, the skin is highly susceptible to environmental toxins. Melanoma, the most lethal type of skin cancer, is characterized by high mortality and a strong tendency for metastasis. It can sometimes originate from pre‐existing nevi, particularly dysplastic nevi. Early identification is crucial for improving patient survival rates. However, traditional skin lesion detection faces challenges due to image quality limitations, dataset imperfections, and the complexity of lesion features. This study proposes the DenseNet201SA++ model, which uses image augmentation techniques and the soft attention mechanism to optimize dermoscopy image quality and automatically capture critical features. Experiments on the HAM10000 dataset with 10,015 dermoscopic images, focusing on binary classification (melanoma vs. nevus), show that the DenseNet201SA++ model achieves significant performance gains, with improvements in precision, recall, F1‐score, and accuracy of at least 7.2%, 14.7%, 12.7%, and 14.7% compared to baseline networks. The proposed soft attention‐guided feature fusion in DenseNet201SA++ addresses feature redundancy in traditional attention mechanisms, achieving superior performance in distinguishing Mel from Nv, while the DenseNet201 backbone shows distinct advantages. Ablation studies confirm the significant role of data augmentation. The integrated DenseNet201SA++ model achieves robust results with precision, recall, F1‐score, and accuracy all reaching 0.983, complemented by an AUC of 0.993. These metrics demonstrate the model's exceptional balance between discriminative power and generalization capability, validating the effectiveness of our proposed architecture.