皮肤疾病智能识别技术研究综述：从传统CAD到多模态大模型的演进、挑战与展望

A Review of Intelligent Recognition Technologies for Skin Disorders: Evolution, Challenges, and Prospects from Traditional CAD to Multimodal Foundation Models

摘要: 皮肤疾病是全球非致命性疾病负担的主要来源之一，但皮肤科医疗资源分布不均，尤其是在发展中国家及偏远地区，智能识别技术与生物医学工程的交叉融合为解决这一困境提供了创新路径。本文系统综述了智能识别技术在皮肤健康管理领域的研究进展、技术方法与应用现状。首先，梳理了从20世纪90年代传统计算机辅助诊断、2012年后深度学习突破、2017~2019年多模态融合探索，到2020年以来多模态大模型驱动的智能化健康管理的完整技术演进历程。其次，深入分析了多模态智能识别、跨模态图像生成(如RGB至UV图像转换)及智能诊断系统的核心原理与典型应用场景，阐述了其在皮肤疾病诊断、肤质分析及美容护肤中的重要作用。在此基础上，对比了国内外研究现状，指出我国在跨模态生成、辅助诊断系统及消费级AI测肤方面取得显著突破，但仍面临数据质量与标准化不足、算法可解释性差、泛化能力弱、跨种族适用性与公平性等挑战。最后，展望了多模态大模型、边缘计算与便携式设备、联邦学习及动态病程监测等未来发展方向。本综述旨在为生物医学工程领域的研究人员提供系统性的技术参考，推动智能识别技术在皮肤健康管理中的深入应用与发展。

Abstract: Skin disorders represent one of the major sources of non-fatal disease burden globally. However, the distribution of dermatological healthcare resources is uneven, particularly in developing countries and remote areas. The interdisciplinary integration of intelligent recognition technologies with biomedical engineering offers an innovative pathway to address this challenge. This paper provides a systematic review of the research progress, technical methods, and application status of intelligent recognition technologies in the field of skin health management. First, it outlines the complete technological evolution from traditional computer-aided diagnosis in the 1990s, the breakthrough of deep learning after 2012, the exploration of multimodal fusion during 2017~2019, to the multimodal foundation model-driven intelligent health management since 2020. Second, it analyzes in depth the core principles and typical application scenarios of multimodal intelligent recognition, cross-modal image generation (e.g., RGB-to-UV conversion), and intelligent diagnosis systems, and elaborates on their important roles in skin disease diagnosis, skin quality analysis, and cosmetic skincare. On this basis, it compares the current research status at home and abroad, noting that China has made significant breakthroughs in cross-modal generation, assisted diagnosis systems, and consumer-grade AI skin analysis, yet still faces challenges such as insufficient data quality and standardization, poor algorithmic interpretability, weak generalization ability, and cross-ethnic applicability and fairness. Finally, it discusses future directions including multimodal foundation models, edge computing and portable devices, federated learning, and dynamic disease progression monitoring. This review aims to provide researchers in biomedical engineering with systematic technical references and to promote the further development and application of intelligent recognition technologies in skin health management.