中华皮肤科杂志 ›› 2021, Vol. 54 ›› Issue (7): 586-589.doi: 10.35541/cjd.20200865

• 论著 • 上一篇    下一篇

面部白癜风人工智能诊断模型的建立及评价

郭丽芳    葛一平    杨寅    林彤   

  1. 中国医学科学院、北京协和医学院皮肤病医院激光科，南京  210042
  • 收稿日期:2020-09-01 修回日期:2021-04-30 发布日期:2021-07-02
  • 通讯作者: 林彤 E-mail:ddlin@hotmail.com

Establishment and evaluation of an artificial intelligence model for the diagnosis of facial vitiligo

Guo Lifang, Ge Yiping, Yang Yin, Lin Tong   

  1. Department of Laser Surgery, Hospital of Dermatology, Chinese Academy of Medical Sciences and Peking Union Medical College, Nanjing 210042, China
  • Received:2020-09-01 Revised:2021-04-30 Published:2021-07-02
  • Contact: Lin Tong E-mail:ddlin@hotmail.com

PDF

4

PDF (Mobile)

0

摘要： 【摘要】 目的 构建面部白癜风人工智能诊断模型，实现面部白癜风的人工智能辅助诊断。方法 利用白癜风皮损单反相机图像和YOLO （You Only Look Once）v3算法建立皮损目标检测模型Vit3，比较Vit3模型的检测结果与皮肤科医生的标注结果，评价Vit3模型的性能。在Vit3模型的基础上，利用面部皮肤人工智能图像采集器拍摄的白癜风及非白癜风皮损的普通光学和紫外光图像，采用图像处理技术测量紫外光图像上皮损区域的灰度值，通过白癜风/非白癜风皮损灰度值阈值鉴别白癜风和非白癜风，建立面部白癜风诊断模型Vit4，通过Cochran′s Q检验比较Vit4模型与皮肤科医生的诊断结果，评价Vit4模型的诊断性能。结果 对于100张白癜风皮损（167处）和100张正常皮肤的单反相机图像，Vit3模型的诊断敏感性为92.81%（155/167）。对于97组（包括50组白癜风、30组白色糠疹、7组无色素痣、10组正常皮肤）面部皮肤图像，Vit4模型的诊断准确率为88.66%（86/97），敏感性为88.00%（44/50），特异性为89.36%（42/47），与皮肤科医生的诊断准确率92.78%（90/97）差异无统计学意义（χ2 = 2.323，P > 0.05）。结论 建立了面部白癜风人工智能诊断模型Vit4，该模型显示出较好的诊断性能，提供了一种较为客观、便捷的面部白癜风辅助诊断方法。

关键词: 白癜风, 人工智能, 诊断, 图像处理, 计算机辅助

Abstract: 【Abstract】 Objective To construct an artificial intelligence model for the diagnosis of facial vitiligo, so as to realize artificial intelligence-assisted diagnosis of facial vitiligo. Methods Based on digital single-lens reflex （SLR） camera images of vitiligo skin lesions and YOLO （You Only Look Once） v3 algorithm, a skin lesion detection model Vit3 was established, and its performance was evaluated by comparing its detection results and labeling results of dermatologists. On the basis of the Vit3 model, both optical and ultraviolet images of vitiligo and non-vitiligo skin lesions were taken by using an artificial intelligence-based facial skin image collector, and the gray values of vitiligo and non-vitiligo skin lesion areas on the ultraviolet images were measured by using an image processing technique. According to the gray-value threshold between vitiligo and non-vitiligo skin lesions, a facial vitiligo diagnosis model Vit4 was established. Cochran′s Q test was used to compare the diagnostic results of the Vit4 model and dermatologists, and the diagnostic performance of the Vit4 model was evaluated. Results For 100 SLR camera images of vitiligo skin lesions （167 lesional sites） and 100 SLR camera images of normal skin, the diagnostic sensitivity of the Vit3 model was 92.81% （155/167）. For 97 pairs of facial skin images （including 50 vitiligo lesions, 30 pityriasis alba lesions, 7 amelanotic nevus leisons, and 10 normal skin tissues）, the diagnostic accuracy rate, sensitivity and specificity of the Vit4 model were 88.66% （86/97）, 88.00% （44/50） and 89.36% （42/47） respectively, and there was no significant difference in the diagnostic accuracy rate between the Vit4 model and dermatologists （92.78% ［90/97］, χ2 = 2.323, P > 0.05）. Conclusion The artificial intelligence model Vit4 was established for the diagnosis of facial vitiligo with favorable diagnostic performance, and could serve as an objective and convenient method for the auxiliary diagnosis of facial vitiligo.

Key words: Vitiligo, Artificial intelligence, Diagnosis, Image processing, computer-assisted