基于卷积神经网络深度学习的颈椎骨龄分期模型的建立  

Developing a cervical vertebral staging model using convolutional neural network deep learning

作　　者：周正涛 吴奔[2] 谷岩[1] Zhou Zhengtao;Wu Ben;Gu Yan(Department of Orthodontics,Peking University School and Hospital of Stomatology&National Center for Stomatology&National Clinical Research Center for Oral Diseases&National Engineering Research Center of Oral Biomaterials and Digital Medical Devices,Beijing 100081,China;School of Statistics,Renmin University of China,Beijing 100872,China)

机构地区：[1]北京大学口腔医学院·口腔医院口腔正畸科、国家口腔医学中心、国家口腔疾病临床医学研究中心、口腔生物材料和数字诊疗装备国家工程研究中心,北京100081 [2]中国人民大学统计学院,北京100872

出　　处：《中华口腔正畸学杂志》2024年第4期181-187,共7页Chinese Journal of Orthodontics

摘　　要：目的本研究应用基于深度学习中的卷积神经网络(convolutional neural networks,CNN)开发出自动化颈椎骨龄分期的特征融合模型,以辅助正畸学的临床诊断及研究。方法本研究回顾性纳入了北京大学口腔医院正畸科7~17岁患者的头颅侧位片1797张,这些图像由正畸医师分为6个颈椎骨龄分期(CS1~CS6),最终纳入1655例(男774例,女881例),再通过人工标注头颅侧位片第二至第四颈椎上的标志点组成数据集。构建由关键点检测、特征融合和分类器模块组成的特征融合模型,然后将数据集输入模型进行训练、验证和测试,得到各样本的预测分期。最后计算模型的分期性能指标,并比较随机森林和人工神经网络(artificial neural network,ANN)两种分类器的分期性能。结果两种分类器的特征融合模型中,ANN分类器的分期准确率达到82.18%,加权Kappa值达到0.8919,显著高于随机森林分类器(准确率78.85%、加权Kappa值为0.8715)。结论特征融合模型不仅实现了全自动的颈椎骨龄分期,而且提高了模型准确率,可以在临床实践和研究中帮助正畸医师提供方便、快速和可靠的颈椎骨龄分期。Objective The purpose of this study is to develop a feature fusion model for an automated cervical vertebral maturation classification method based on convolutional neural networks in deep learning,aiming at assisting clinical diagnosis and research in orthodontics.Methods This study included a total of 1797 lateral cephalometric radiograph of orthodontic patients aged 7 to 17 from the Department of Orthodontics,Peking University School and Hospital of Stomatology.These images were classified into six cervical vertebral stages(CS1~CS6)by orthodontists,and 1655 cases(774 males and 881 females)were finally included.Then the landmarks on the second to fourth cervical vertebrae were manually identified on the lateral cephalometric images to form the dataset.A feature fusion model consisting of key points detection,feature fusion and classifier module was constructed,after which the dataset was input into the model for training,verification and testing to obtain the prediction stage for each sample.Finally,we calculated the model's staging performance index and compared the performance of two classifiers.Results In the two classifiers of the feature fusion model,the staging accuracy of the ANN classifier reached 82.18%,and the weighted Kappa value reached 0.8919,both significantly higher than those of the random forest classifier(accuracy=78.85%,wκ=0.8715).Conclusions The feature fusion model not only realizes the automatic cervical vertebral staging,but also improves the model accuracy.It provides orthodontists with convenient,rapid and reliable cervical vertebral stage predictions for clinical practice and research.

关 键 词：人工智能 卷积神经网络 头颅侧位片 颈椎骨龄分期法 特征融合 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] R783.5[自动化与计算机技术—控制科学与工程] TP391.41[医药卫生—口腔医学]