2018 年第 4 期 第 13 卷

深度学习方法建立痤疮分级系统的探索

Exploration of acne grading system using deep learning method

作者：吴意平陈海龙张大磊

英文作者：

单位：100730北京医院皮肤科

英文单位：

关键词：痤疮；深度学习；卷积神经网络；临床图片；分级系统

英文关键词：

• 摘要：

• 目的    探索用深度学习方法建立痤疮分级系统。方法    用北京医院皮肤科门诊及互联网收集到有效诊断为痤疮的面部独立图片共9 862张，由培训后皮肤专科医师参照《中国痤疮治疗指南（2014修订版）》分级标准对图片进行分级标注并经高级职称专科医师复核。随机选用其中8 000张为独立训练图片，1 862张为测试图片，使用图形处理器建立了3级痤疮严重程度分级系统（Airdoc-Ac）。由北京医院11名分别具有中（9名）、高（2名）级职称的皮肤科医师对随机选取的30张痤疮患者面部临床图片做分级标注，并用同样30张图片输入Airdoc-Ac系统进行疾病严重程度分级。另取门诊新采集病例36张图片输入Airdoc-Ac系统进行分级并在临床由医师同时分级判定，比较2种方法的准确率。结果    Airdoc-Ac系统的痤疮图片分级识别准确率为85.5%(1 592/1 862)，高于文献报道可以在临床应用的痤疮严重程度分级方法80.0%的准确率；11名皮肤科专科医师的分级识别准确率为60.0%（18/30），2名高级职称专科医师的分级准确率83.3%（25/30），以30张痤疮图片用Airdoc-Ac进行分级准确率为86.7%（26/30），临床新采集图片用Airdoc-Ac分级准确率80.6%（29/36）。Airdoc-Ac痤疮分级系统的准确率与专科医师分级最高准确率差异无统计学意义（χ2=0.112，P=0.738）。结论    Airdoc-Ac系统可以识别痤疮图片中皮肤损害区域及进行疾病严重程度分级，其准确率可达到皮肤专科高级职称医师的水平，且随着病例数的积累，机器学习可能会进一步提高其准确率，为临床皮肤专科医师，基层全科医师和患者提供痤疮的辅助诊断方法，可作为临床医师选择痤疮治疗方法和疗效评估的人工智能辅助诊断工具。

• Objective    To explore a grading system of acne by learning method. Methods    A total of 9 862 facial pictures of acne patients were collected from Department of Dermatology of Beijing Hospital and network resource; 8 000 were independent training pictures and 1 862 were test pictures; the acne severity grading system(Airdoc-Ac) was established by graphics processor units. Nine dermatologists with middle professional title and 2 dermatologists with senior professional title were enrolled to assess the severity of acne in 30 pictures which were also assessed by the Airdoc-Ac system. Thirty-six new diagnosed acne pictures were assessed by both Airdoc-Ac system and dermatologists to evaluate the accuracy rate of the 2 methods. Results   The accuracy rate of Airdoc-Ac system assessing the severity of acne was 85.5%(1 592/1 862), which was higher than that reported by literatures（80.0%）. The accuracy rate of 11 dermatologists was 60.0%(18/30); the accuracy rate of 2 dermatologists with senior professional title was 83.3%(25/30). The accuracy rate of Airdoc-Ac system assessing 30 acne pictures was 86.7%(26/30); the accuracy rate of Airdoc-Ac system assessing 36 new acne pictures was 80.6%(29/36). There was no significant difference of the accuracy rate between Airdoc-Ac system and dermatologists（χ2=0.112, P=0.738）. Conclusion    The Airdoc-Ac grading system shows equal accuracy in assessing the severity of acne as dermatologists with senior professional title.