[Identification model of tooth number abnormalities on pediatric panoramic radiographs based on deep learning].

Objective: To identify tooth number abnormalities on pediatric panoramic radiographs based on deep learning. Methods: Eight hundred panoramic radiographs of children aged 4 to 11 years meeting the inclusion and exclusion criteria were selected and randomly assigned by writing programs in Python (version 3.9) to the training set (480 images), verification set (160 images) and internal test set (160 images), taken in Department of Pediatric Dentistry, Peking University School and Hospital of Stomatology between November 2012 to August 2020. And all panoramic radiographs of children aged 4 to 11 years taken in the First Outpatient Department of Peking University School and Hospital of Stomatology from June 2022 to December 2022 were collected as the external test set (907 images). All of the 1 707 images were obtained by operators to determine the outline and to label the tooth position of each deciduous tooth, permanent tooth, permanent tooth germ and additional tooth. The deep learning model with ResNet-50 as the backbone network was trained on the training set, validated on the verification set, tested on the internal test set and external test set. The images of test sets were divided into two categories according to whether there was abnormality of tooth number, to calculate sensitivity, specificity, positive predictive value and negative predictive value, and then divided into four types of extra teeth and missing permanent teeth both existed, extra teeth existed only, missing permanent teeth existed only, and normal teeth number, to calculate Kappa values. Results: The sensitivity, specificity, positive predictive value and negative predictive value were 98.0%, 98.3%, 99.0% and 96.7% in the internal test set, and 97.1%, 98.4%, 91.9% and 99.5% in the external test set respectively, according to whether there was abnormality of tooth number. While images were divided into four types, the Kappa value obtained in the internal test set was 0.886, and that in the external test set was 0.912. Conclusions: In this study, a deep learning-based model for identifying abnormal tooth number of children was developed, which could identify the position of additional teeth and output the position of missing permanent teeth on the basis of identifying normal deciduous and permanent teeth and permanent tooth germs on panoramic radiographs, so as to assist in diagnosing tooth number abnormalities.目的： 基于深度学习技术识别儿童曲面体层X线片（以下简称曲面体层片）中的牙齿数目异常，提高临床医师工作效率，减少误诊与漏诊。 方法： 从北京大学口腔医学院·口腔医院儿童口腔科2012年11月至2020年8月间拍摄的符合纳入和排除标准的曲面体层片中抽取800张4~11岁儿童的曲面体层片，使用Python（3.9版本）编写程序随机分配为训练集（480张图像）、验证集（160张图像）和内部测试集（160张图像）；并收集北京大学口腔医学院·口腔医院第一门诊部连续半年内拍摄的全部4~11岁儿童曲面体层片，共计1 707张图像由医师阅片确定每颗乳牙、恒牙、恒牙胚和额外牙的轮廓并标识牙位。使用训练集训练以ResNet-50为骨干网络的深度学习模型，在验证集中对模型进行微调，通过内部测试集和外部测试集评估模型性能，根据有无牙齿数目异常分为两类计算灵敏度、特异度、阳性预测值和阴性预测值，再分为同时存在额外牙与恒牙缺失、仅存在额外牙、仅存在恒牙缺失、牙齿数目正常四类计算Kappa值。 结果： 有无牙齿数目异常两类图像在内部测试集中的灵敏度、特异度、阳性预测值和阴性预测值分别为98.0%、98.3%、99.0%、96.7%，外部测试集中的灵敏度、特异度、阳性预测值和阴性预测值分别为97.1%、98.4%、91.9%、99.5%。牙齿数目正常、同时存在额外牙与恒牙缺失、仅存在额外牙、仅存在恒牙缺失四类图像在内部测试集中获得的Kappa值为0.886，在外部测试集中获得的Kappa值为0.912。 结论： 本研究开发了基于深度学习的儿童牙齿数目异常识别模型，其能在识别儿童曲面体层片正常乳恒牙及恒牙胚的基础上，确定额外牙的位置并输出缺失恒牙的牙位，从而辅助诊断有无牙齿数目异常。.