基于深度学习的恒牙牙菌斑智能判读系统的建立及其相关影响因素分析

Deep learning-based dental plaque detection on permanent teeth and the influenced factors

目的建立基于人工智能技术的牙菌斑智能判读系统,分析影响其判读结果的相关因素。方法2018年10月至2019年6月用家用口腔内镜(1280×960像素,东莞立港医疗器材有限公司)拍摄北京大学口腔医学院的25名研究生志愿者[男性12名,女性13名,年龄(23±3)岁]口内牙齿唇颊侧照片,使用菌斑染色剂(Cimedical,日本)进行菌斑染色后,用同样拍摄方法再次拍摄照片,共收集符合纳入标准的549组恒牙牙菌斑染色前后的牙齿照片。将549组图像采用随机数字表法分为440组训练数据集和109组测试数据集。设计并实现基于DeepLab的深度学习模型,采用LabelMe软件(Windows版3.2.1,MIT,美国)进行标注,对标注后的440组训练数据集进行训练,并据此建立恒牙牙菌斑智能判读系统,使用平均交并比(mean intersection over union,MIoU)作为评估该算法识别准确性的量化指标,用建立的恒牙牙菌斑智能判读系统对109组测试照片进行判读。使用Matlab软件(Windows版R2017a,MathWorks,美国)提取109组照片的牙菌斑边缘线,计算菌斑边缘线像素点个数,以此衡量牙菌斑边缘的复杂性,并计算牙菌斑面积百分比。基于多元线性回归分析牙位、牙菌斑面积百分比、牙菌斑边缘线像素点个数、镜头光源光斑是否位于菌斑区域对于恒牙牙菌斑判读准确性的影响,通过方差检验比较模型调整后的决定系数R2选择拟合效果更优的模型。结果恒牙牙菌斑智能判读系统对测试组判读的MIoU值为0.700±0.191。当牙菌斑面积百分比、牙菌斑边缘线像素点个数进入回归模型时,R2值为0.491,高于只有牙菌斑百分比进入模型时的R2,牙菌斑面积百分比、牙菌斑边缘线像素点个数对恒牙牙菌斑判读准确性有显著影响(P<0.05)。牙菌斑边缘线像素点个数的标准化系数为-0.289,菌斑面积百分比的标准化系数为-0.551。结论本研究通过家用口腔内镜采集的恒牙牙面图像构建了恒牙牙菌斑智能判读系统,该系统可以较准确地判断牙菌斑的附着情况;牙菌斑边缘线越复杂,牙菌斑面积百分比越高,菌斑识别的准确性越低。

Objective To develop an artificial intelligence system for detecting dental plaque on permanent teeth and find the influenced factors.Methods Photos of the labial or buccal surfaces of the permanent teeth were taken by using an intraoral camera(1280×960 pixels;TPC Ligang,Shenzhen,China)before and after applying the plaque-disclosing agent(Cimedical,Japan)in 25 volunteers[12 males,13 femals,aged(23±3)years]recruided in accordance with the inclusion criteria from the students of Peking University School of Stomatology from October 2018 to June 2019.A total of 549 groups of photos were captured and then divided into a training dataset containing 440 groups of photos and a test dataset including 109 groups of photos.The scopes of teeth and dental plaque on photos were labeled using LabelMe(Windows 3.2.1,MIT,U S A).A DeepLab based deep learning system was designed for the intelligent detection of dental plaque on permanent teeth.The mean intersection over union(MIoU)was employed to indicate the detection accuracy.Matlab(Windows R2017a,MathWorks,U S A)was used to extract the plaque edge line of 109 groups of photos and to calculate the number of pixels for the measurement of the complexity of the plaque edge line.The percentage of dental plaque area was calculated.Multivariate linear regression was used to explore whether tooth site,plaque percentage,number of plaque edge line pixels and lens light spot location would influence the detection accuracy,of which P<0.05 was considered statistically significant.Results The MIoU of the permanent tooth model was 0.700±0.191 when 440 photos were used for training and 109 photos were used for testing.In the regression model of significance test(P<0.05),the percentage of plaque and the number of pixels on the edge of plaque had significant influence on the accuracy of dental plaque detection.The standardized coefficient of the number of pixels of the plaque edge line is-0.289,and the standardized coefficient of the percentage of plaque is-0.551.Conclusions In the present study,an artificial intelligence system was built to detect dental plaque area on tooth photos collected by family intraoral camera.The system showed the ability to detect the dental plaque of permanent teeth.The more complex the marginal line of dental plaque and higher the percentage of dental plaque are,the lower the accuracy of plaque recognition is.