Early Screening Of Periodontal Disease Based On Deep Learning Of Oral Digital Image

Objective:Periodontal disease is the leading cause of tooth loss.The diagnosis of this disease usually depends on periodontal examination,history inquiry and imaging interpretation of periodontal specialists.However,due to the shortage of periodontal specialists and the difference in medical resources between primary hospitals and grade A hospitals,it is difficult to cope with the high incidence of periodontal disease in China.If periodontal disease is not diagnosed early and treated in time,it may lead to tooth loss eventually,which will bring serious psychological and economic burden to patients.Artificial intelligence is a technology that uses computers to simulate human behavior in order to perform tasks more efficiently.Deep learning,represented by Convolutional Neural Network(CNN),can simulate human brain for data processing and has excellent performance in image recognition.In recent years,CNN has been widely used in the field of medical research,and related reports have also been reported in the field of stomatology.In this study,an intelligent analysis model of intra oral digital images was constructed based on deep learning technology,and the early screening of periodontal diseases could be effectively carried out by identifying the non-healthy lesions in periodontal images.Methods:The present study retrospectively collected oral digital images and clinical data from patients and periodontal healthy people who visited the Stomatology Diagnosis and Treatment Center of the Second Affiliated Hospital of Nanchang University from September 2019 to October 2021.After the informed consent of the subjects,periodontal specialists were used to diagnose and classify the oral digital images into periodontal healthy images and periodontal non-healthy images,and the corresponding database was established,and the included data were divided into training sets and test sets.The training set was used to construct an intelligent analysis model of oral digital image based on CNN structure.The performance of CNN model was evaluated according to the results of the test set,and the image features extracted from the early screening model of periodontal disease were displayed through deconvolution visualization analysis.Results:A total of 3869 oral digital images were collected from 578 subjects,including2230 periodontal healthy images and 1639 periodontal unhealthy images.CNN training was carried out on digital images of nine grids,orthotopic digital images of occlusal mouth,and orthotopic digital images of occlusal mouth after removing invalid background.Three kinds of image intelligent analysis models of digital images of occlusal mouth were established in this study.Among them,the AUC value of oral digital image classification model(training set 3153,test set 716)was 0.651.The AUC value of digital image classification model(445 training sets and 133 test sets)was 0.767.The AUC value of digital image classification model of orthotopic occlusal mouth after removing invalid background was 0.784.The accuracy of the three CNN models were 66.62%,64.66% and 77.44%,the accuracy was 0.4375,0.4824 and 0.6944,the recall rate was 0.5805,0.9318 and 0.5682,the F1 value was0.4990 and 0.6357,respectively.0.6250.The deconvolution visualization analysis shows that the three CNN models have different abilities to extract image features when they classify different data sets.The highlight area of feature extraction obtained by the CNN model based on intra-oral digital image after removing invalid background is the most consistent with clinical expectations.Conclusion:(1)The intelligent early screening model of periodontal disease based on CNN structure can quickly,objectively and efficiently recognize the intra-oral digital image.The prediction results can be obtained only by inputting images into the model,providing an effective method for the early screening of periodontal disease.(2)There are differences in the performance of CNN models constructed according to different databases,and the model constructed by intra-oral digital image data after removing invalid background has the best performance;(3)The inverse convolution visualization analysis in this study can identify the possible affected areas and provide new ideas for the standardized treatment of periodontal disease.