口腔医学研究

口腔医学研究 ›› 2022, Vol. 38 ›› Issue (10): 959-962.DOI: 10.13701/j.cnki.kqyxyj.2022.10.012

• 口腔影像学研究 • 上一篇    下一篇

基于卷积神经网络的移位型牙根折裂人工智能诊断的初步研究

胡燕妮1, 曹丹彤1, 王柏鑫2, 陈颖2, 林梓桐1*   

  1. 1.南京大学医学院附属口腔医院,南京市口腔医院口腔颌面医学影像科 江苏 南京 210008;
    2.南京大学电子科学与工程学院 江苏 南京 210023
  • 收稿日期:2022-02-09 出版日期:2022-10-28 发布日期:2022-10-20
  • 通讯作者: *林梓桐,E-mail:linzitong710@163.com
  • 作者简介:胡燕妮(1997~ ),女,安徽人,硕士在读,主要从事口腔医学的研究工作。
  • 基金资助:
    江苏省南京市卫生科技发展项目(编号:YKK19090);江苏省医学会科研专项资金资助[编号:SYH-3201150-0007(2021002)];江苏省卫生健康委面上项目(编号:M2021077)

Preliminary Exploration of Artificial Intelligence in Diagnosis of Displaced Root Fracture Based on Convolutional Neural Network

HU Yanni1, CAO Dantong1, WANG Baixin2, CHEN Yin2, LIN Zitong1*   

  1. 1. Department of Oral and Dentomaxilofacial Radiology, Nanjing Stomatological Hospital, Medical School of Nanjing University, Nanjing 210008, China;
    2. School of Electronic Science and Engineering, Nanjing University, Nanjing 210023, China
  • Received:2022-02-09 Online:2022-10-28 Published:2022-10-20

PDF (PC)

140

摘要： 目的: 探讨卷积神经网络(convolutional neural network,CNN)在移位型牙根折裂(displaced root fracture, DRF)锥形束CT(cone beam computed tomography,CBCT)图像诊断中的应用价值。方法: 筛查我院就诊的286例DRF患者的CBCT图像,分别采用自动截图和人工截图的方法获得兴趣区并人工标注阳性或阴性,使用ResNet 50分别进行训练并建立诊断模型,并计算两种截图方式所建立的诊断模型的准确率、特异度和灵敏度。结果: 286名患者中,164名患者的CBCT图像用于自动截图,122名患者的CBCT图像用于人工截图。自动截图组中,训练集包括阳性图片184幅和阴性图片186幅,测试集包括阳性图片64幅和阴性图片66幅;人工截图组中,训练集包括阳性图片96幅和阴性图片114幅,测试集包括阳性图片32幅和阴性图片36幅。人工截图组模型诊断的准确率为98.5%、特异度为100%,灵敏度为96.9%,自动截图组模型诊断的准确率为90.0%,特异度为95.5%,灵敏度为84.4%。结论: 卷积神经网络ResNet 50对于移位型牙根折裂具有良好的诊断效能。

关键词: 移位型牙根折裂, 卷积神经网络, 锥形束CT

Abstract: Objective: To explore the value of convolutional neural network in the diagnosis of displaced root fracture (DRF) based on cone-beam CT (CBCT). Methods: The CBCT images of 286 patients with DRF were collected in our hospital. The regions of interest were obtained by automatic and manual cropping, and labeled as positive or negative manually. ResNet 50 was used for training and establishing diagnosis models for DRF. The diagnosis efficiency of two models was calculated (accuracy, specificity, and sensitivity). Results: One hundred and sixty-four patients' images were used for automatic cropping and 122 patients' images were used for manual cropping. In the automatic cropping group, there were 184 positive images and 186 negative images in the training set; 64 positive images and 66 negative images in the test set. While in the manual cropping group, there were 96 positive images and 114 negative images in the training set; 32 positive images and 36 negative images in the test set. The accuracy, specificity, and sensitivity were 98.5%, 100%, and 96.9% with manual cropping; while they were 90.0%, 95.5%, and 84.4% with automatic cropping. Conclusion: Convolutional neural network ResNet 50 could be used to detect displaced root fractures with high diagnosis efficiency.

Key words: displaced root fractures, convolutional neural network, cone beam CT