Establishment Of Artificial Intelligence Model For Diagnosis And Extent Evaluation Of Vitiligo And Clinical Application Of AI Model

Objective:This study aimed at developing an objective and convenient tool to diagnose vitiligo and assess the extent of vitiligo by using an artificial intelligence model.Methods:1)Based on Convolutional Neural Network Yolo V3 architecture and digital photos of vitiligo,we built an AI model of vitiligo lesion detection.Digital photos data were collected,selected and labeled,and the database were prepared to train the Yolo V3.The parameters of the Yolo V3 model were optimized.The reliability of the model was evaluated.Bsed on digital photos of visible light and wood light and the image processing methods,we developed a model to diagnosis of vitiligo.The reliability of the model was evaluated.We compared the performances of AI model and dermatologists using 97 digital photos of visible light and wood light,including 50 vitiligo,47 nonvitiligo.2)Based on CNN Unet++ and digital photos,we developed an AI model of segmentation of vitiligo lesion.Digital photo data were collected,selected and segmented by dermatologists.The database were prepared to train model.The Unet++ was fine tuned with the training portion of the database,and the trained model was validated with the testing portion of the database.3)On the base of the AI model of diagnosis of vitiligo and the AI model of segmentation of vitiligo lesion,we established the methods of image processing for extent assessment of vitiligo,and the model was validated.4)The artificial intelligence model of extent assessment of vitiligo was used as an objective evaluation tool in the clinical trial of the efficacy and safety of excimer laser in the treatment of vitiligo.The changes of lesion area and chroma were measured before and after the treatment of excimer light and excimer laser.5)The artificial intelligence model of detection of vitiligo was used in the wechat program to realize the application of the model in the mobile phone.Results:1)The sensitivity of AI model of vitiligo lesion detection are 92.81%;The accuracy and sensitivity of AI model of diagnosis are 88.66%and 88%,slightly lower than those of dermatologists;the specificity of AI model is 89.36%,similar to dermatologists.2)The AI model of segmentation of vitiligo lesion achieved a mIoU of 79%and Dice of 88.42%.3)Measurement of the relative area:the area of vitiligo lesions was measured before and after treatment in 20 groups,and there was no statistical difference between the evaluation results of the model and doctors and the measurement results of Photoshop;measurement of the actual area:the actual area of 27 vitiligo lesions was measured,and there was no statistical difference between the measurement results of the model and Photoshop;measurement of the chromaticity:the vitiligo lesions were measured before and after treatment in 20 groups.There was no statistical difference between the model and the doctor’s evaluation results.4)The application of the evaluation model in the clinical trial of observation of efficacy and safety of excimer light in the treatment of vitiligo in facial stable stage:the effective rate of the treatment of excimer light evaluated by model is 70%,the effective rate of the treatment of excimer laser evaluated by model is 60%,there is no statistical difference.5)Application of detection model in mobile terminal:the accuracy of detection of vitiligo in images collected by mobile terminal is 86.49%.Conclusions:This study developed an objective and convenient tool to diagnose vitiligo and assess the extent of vitiligo by using an artificial intelligence model.