文章摘要

文章信息/Info

Title:: Research on artificial intelligence diagnosis method of psoriasis vulgaris based on deep learning

作者:: 陈宇翔1; 2; 王磊3; 闫峻3; 周冬梅2△; 1. 北京中医药大学，北京 100029；2. 首都医科大学附属北京中医医院，北京 100010；3. 医渡云（北京）技术有限公司，北京 100083

Author(s):: CHEN Yuxiang1; 2; WANG Lei3; YAN Jun3; ZHOU Dongmei2; 1.Beijing University of Chinese Medicine，Beijing 100029，China；2.Beijing Hospital of Traditional Chinese Medicine Affiliated to Capital Medical University，Beijing 100010；3.Yidu Cloud (Beijing) Technology Co.，Ltd.，Beijing 100083

Keywords:: Psoriasis vulgaris; Skin disease; Skin lesion segmentation; Deep learning; Convolutional neural network

摘要:: 寻常型银屑病是一种常见的慢性皮肤病，由于病程长，需要进行长期的疾病管理。为了给寻常型银屑病在基层的诊治提供切实可行、低成本的方案，本研究将是否进行皮损分割作为条件，利用来自北京中医医院皮肤科的830例患者共计13 409张皮损病历照片，训练了两个深度学习模型，探讨不使用皮损分割进行智能诊断的可行性。测试结果表明，使用已进行皮损分割数据建立的模型，准确率为87.812%，AUC值0.94；进行全图识别的模型，准确率为84.098%，AUC值0.91，不进行皮损分割对寻常型银屑病进行识别，能够取得良好的效果，同时能够节省大量的标注、算法适配时间，进一步提升皮肤病智能诊断系统的开发效率。

Abstract:: Psoriasis vulgaris is a kind of common chronic skin disease. Due to the long course of disease, long-term disease management is required. Therefore, it is necessary to provide practicable, low-cost, and repeatable treatment management solutions for psoriasis vulgaris. In the study, a total of 13 409 skin lesion medical records of 830 patients from the Department of Dermatology of Beijing Hospital of Traditional Chinese Medicine were used to learn the data set using the DenseNet network.We trained two deep learning models with the variable of whether the skin lesion was marked or not. The test results show that the model from labelled group has an accuracy rate of 87.812% and an AUC value of 0.94; while the model from full-image recognition has an accuracy rate of 84.098% and an AUC value of 0.91. Recognizing psoriasis without segmentation of skin lesions is able to achieve good results, and can save a lot of time of labelling and algorithm adapting, which is expected to further improve the development efficiency of the artificial diagnosis system for skin diseases.