Abstract: As the most lethal type of skin cancer, early and accurate diagnosis of melanoma is essential to improve the survival rate of patients. In recent years, deep learning technology has shown great potential in the field of melanoma classification and diagnosis, providing new technical support for clinical diagnosis. This paper systematically reviews the research progress of deep learning in melanoma classification, focusing on the technical evolution and clinical application of core methods such as convolutional neural networks, Transformers, generative adversarial networks and recurrent neural networks. Firstly, the characteristics of authoritative datasets such as HAM10000, ISIC, and PH2 and their value in algorithm development are summarized, and the preprocessing methods and enhancement strategies of different datasets are analyzed in detail, which provides a high-quality data basis for model training. Secondly, the improvement strategies of different deep learning models are deeply analyzed, including network architecture optimization, multimodal feature fusion, and data imbalance processing. In addition, the role of multiple learning strategies such as transfer learning and ensemble learning in improving model performance is also discussed. Finally, the limitations of current technology are summarized, and future research directions are prospected, including the application prospects of multimodal large models, federated learning and lightweight technology.

Key words: melanoma, deep learning, convolutional neural networks, Transformer；transfer learning

摘要： 黑色素瘤作为皮肤癌中最具致命性的类型，其早期准确诊断对提高患者生存率至关重要。近年来，深度学习技术在黑色素瘤分类诊断领域展现出巨大潜力，为临床诊断提供了新的技术支撑。系统回顾了深度学习在黑色素瘤分类中的研究进展，重点关注卷积神经网络、Transformer、生成对抗网络和循环神经网络等核心方法的技术演进及其临床应用。归纳了HAM10000、ISIC、PH2等权威数据集的特性及其在算法开发中的价值，详细分析了不同数据集的预处理方法和增强策略，为模型训练提供了高质量的数据基础；深入分析了不同深度学习模型的改进策略，包括网络架构优化、多模态特征融合及数据不平衡处理等；还探讨了迁移学习、集成学习等多元学习策略在提升模型性能中的作用。总结了当前技术的局限性，并对未来研究方向进行了展望，重点包括多模态大模型、联邦学习及轻量化技术的应用前景。

关键词: 黑色素瘤, 深度学习, 卷积神经网络, Transformer, 迁移学习