人工智能口腔实景图像识别技术用于龋齿筛查的效能研究

Study on the effectiveness of artificial intelligence image recognition technology to screen dental caries on intraoral photographs

目的基于人工智能(AI)图像识别技术,评估智能手机(Sp)和数字化单反相机(DSLR)两种设备采集的口内实景照片用于龋齿筛查的效能研究。方法建设“益生牙kid”AI龋齿筛查云平台,通过AI图像识别深度学习算法不断迭代学习口腔临床医生筛查龋齿的方法,并用混合矩阵验证其识别效果。使用Sp及DSLR两种设备,同时采集27位受试者口内照片。每位受试者按照采集标准,Sp和DSLR分别获取3张、5张不同牙位照片,两组照片提取上传至龋齿筛查云平台,分别获得359、480个牙位龋齿筛查结果。受试者同时接受口腔医师使用WHO龋齿分级评判标准进行临床龋齿筛查。采用敏感度、特异度及曲线下面积(AUC)等作为AI龋齿图像识别效果的评价指标。结果混淆矩阵结果显示低风险早期龋和高风险成洞龋识别准确度高达90%以上,而中风险釉质轻度缺损龋齿识别准确度仅为32%。口腔医师采用WHO龋齿评判标准与AI筛查龋齿结果比较,Sp拍摄的口内照片诊断龋齿敏感度、特异度及AUC值分别为75.6%、85.8%及0.849;DSLR拍摄的口内照片诊断龋齿敏感度、特异度及AUC值分别为80.0%、86.2%及0.859。结论AI口腔实景图像识别龋齿筛查技术与临床医生龋齿诊断结果接近。Sp拍摄龋齿照片更便捷高效,与临床DSLR相比筛查正确率接近,实际推广应用价值更高。

Objective To evaluate the effectiveness of intraoral photographs collected by smartphone(Sp)and digital single-lens reflex(DSLR)for dental caries screening based on artificial intelligence(AI)image recognition technology.Methods The“YiShengYa kid”AI caries screening cloud platform was constructed.The AI image recognition deep learning algorithm was used to iteratively learn the methods of dental clinicians to screen dental caries,and the confusion matrix was used to verify the recognition effect.Two devices,Sp and DSLR,were used to collect intraoral photographs of 27 subjects at the same time.According to the collection criteria,Sp and DSLR acquired 3 and 5 photos of different tooth positions for each subject,respectively.The two groups of photos were uploaded to the dental caries screening cloud platform to obtain 359 and 480 dental caries screening results.The same subjects were screened for dental caries by dentists with clinical experience using WHO dental caries classification criteria.Sensitivity,specificity and area under the curve(AUC)were used to evaluate the effect of AI dental caries image recognition.Results The results of the confusion matrix showed that the accuracy of identifying low risk early caries and high risk caries with cavities was more than 90%,while the accuracy of identifying moderate risk caries with mild enamel defect was only 32%.Comparing the diagnostic results of dentists under the WHO criteria with AI screening results for caries,Sp images had an AUC value of 0.849 with a sensitivity of 75.6%and specificity of 85.8%.DSLR images had an AUC value of 0.859 with a sensitivity of 80.0%and specificity of 86.2%in diagnosing dental caries.Conclusion AI image recognition technology to screen dental caries is close to the caries diagnosis results of clinicians.Sp is more convenient and efficient to take caries photos,and the screening accuracy is close to that of the DSLR,which has higher practical application value.