浅谈皮肤图像质量在AI研究中的价值  

The Discussion on the Value of Skin Image Quality in the Research of AI

作　　者：孟如松[1] 李艺鹏 胡博[1] 刘海军[1] 杨世飞[1] MENG Rusong;LI Yipeng;HU Bo;LIU Haijun;YANG Shifei(Airmy Institute of Dermatology, Dermatosis Screenage Diagnostic Center of Dermatology Hospital, Air Force General Hospital, Beijing 100142, China)

机构地区：[1]全军皮肤病研究所,空军总医院皮肤病医院皮肤病影像诊断中心,北京100142

出　　处：《皮肤科学通报》2018年第2期229-237,共9页Dermatology Bulletin

摘　　要：近年,以深度学习技术为代表的人工智能(AI)正席卷各行各业,而AI框架有多种,多数采用深度卷积神经网络(CNN)技术结合迁移学习进行训练,虽然在皮肤AI研究中取得长足进展,但其研究结果未能真正走出实验室进入临床应用。制约这些因素主要是缺乏高质量的皮肤疾病图像的大型数据集。本文针对皮肤科常见的图像采集方法,包括临床摄影图像、皮肤镜图像、反射式共聚焦激光扫描显微镜(RCM)图像、皮肤B超图像和组织病理图像的质量要素进行探讨和述评,希望对解决因皮肤图像质量的问题而影响AI研究进展的瓶颈问题能有所帮助。The artificial intelligence （AI）, on behalf of deep learning of the technology, has swept across various walks of life in recent year. Among the multiple frames of Al,the technology of deep convolution neural network （CNN）,in combination with transfer- ring learning, has been most adopted for training. Although great progress has been made in the research of AI for dermatology,the research findings have failed to take clinical application into practice. The lack of large data set for high-quality Dermatosis images has been the main cause. In this article, we discuss and review the quality factors of all the images acquired by the common image-acquisition methods in dermatology, including ordinary camera, dermoscopy, reflectance confocal miscroscopy （RCM） , skin uhrasound, and histopathology. Hupefully, it can help solving the bottleneck problem about the image quality and to promote AI.

关 键 词：皮肤图像质量 人工智能 皮肤摄影图像 皮肤镜图像 反射式共聚焦激光扫描显微镜图像 皮肤超声图像 组织病理图像 

分 类 号：R445[医药卫生—影像医学与核医学] R751[医药卫生—诊断学]