口腔医学研究

口腔医学研究 ›› 2021, Vol. 37 ›› Issue (9): 845-849.DOI: 10.13701/j.cnki.kqyxyj.2021.09.016

• 口腔影像学研究 • 上一篇    下一篇

基于深度学习在曲面体层图像中人工智能辅助诊断系统初步研究

帕克扎提·色依提1, 王铁梅1*, 徐子能2, 白海龙2, 丁鹏2, 刘澍1, 滕跃辉1, 冯英连1, 王蓉1, 单珊1, 钟双泽1   

  1. 1.南京大学医学院附属口腔医院口腔颌面医学影像科 江苏 南京 210008;
    2.北京羽医甘蓝信息技术有限公司 北京 100102
  • 收稿日期:2021-03-01 出版日期:2021-09-28 发布日期:2021-09-16
  • 通讯作者: *王铁梅,E-mail:tiemei106@263.net
  • 作者简介:帕克扎提·色依提(1992~ ),女,新疆乌鲁木齐人,硕士,医师,主要从事口腔颌面医学影像学研究工作。
  • 基金资助:
    江苏省自然科学基金面上项目(编号:BK20150089)南京市医学科技发展资金项目(编号:QRX17079)

Preliminary Study of Common Oral Diseases Recognition in Oral Pantomography with Deep Learning

Pakezhati·SEYITI1, WANG Tiemei1*, XU Zineng2, BAI Hailong2, DING Peng2, LIU Shu1, TENG Yuehui1, FENG Yinglian1, WANG Rong1, SHAN Shan1, ZHONG Shuangze1   

  1. 1. Department of Oral and Maxillofacial Imaging, Nanjing Stomatological Hospital, Medical School of Nanjing University, Nanjing 210008, China;
    2. Beijing DeepCare Information Technology Co., Ltd, Beijing 100102, China
  • Received:2021-03-01 Online:2021-09-28 Published:2021-09-16

PDF (PC)

526

摘要： 目的: 基于深度学习对口腔曲面体层图像分析,开展人工智能在口腔常见疾病辅助诊断系统的研发,挖掘人工智能对曲面体层图像分割及辅助诊断价值。方法: 回顾性纳入2000张口腔曲面体层片建立数据集(训练集1400张、测试集600张,累计标注82042例)。运用基于卷积神经网络的深度学习算法,通过算法设计、模型训练和验证,构建口腔常见疾病智能影像诊断模型PanoNet,利用6个子网络模型分别执行不同口腔疾病的分割及识别。结果: PanoNet在恒牙列识别及龋病、根尖周炎、阻生牙、种植体、牙体修复术后等疾病识别中准确率、敏感度和特异度均高于85%(kappa>0.81);在牙槽骨吸收分级识别中准确率、敏感度、特异度分别为76.50%、75.25%、79.00%(kappa=0.44)。结论: 基于卷积神经网络的深度学习算法建立的口腔曲面体层图像诊断模型PanoNet能有效识别上述口腔常见疾病,体现人工智能在曲面体层片上对口腔常见疾病的影像辅助诊断的应用价值。

关键词: 口腔曲面体层图像, 深度学习, 卷积神经网络, 人工智能

Abstract: Objective: To develop an artificial intelligence aided diagnosis system for common oral diseases based on deep learning algorithm and oral pantomography image analysis. Methods: A total of 2000 oral panoramic radiographs were selected from PACS database of our hospital to establish the data set (1400 for training set and 600 for test set). Using the deep learning algorithm based on convolutional neural network, through the algorithm design, model training, and validation, the intelligent image diagnosis model “PanoNet” for common oral diseases was constructed. The image segmentation and recognition of different oral diseases were performed by six sub network models. Results: The accuracy, sensitivity, and specificity of PanoNet were higher than 85% (kappa>0.81) in the identification of permanent dentition, dental caries, periapical periodontitis, impacted teeth, implants, and postoperative repair of dental defects. In the classification of alveolar bone resorption, the above indexes were 76.50%, 75.25%, and 79.00%, respectively (kappa=0.44). Conclusion: Based on the deep learning algorithm of convolution neural network, PanoNet can effectively identify the above common oral diseases, which reflects the application value of artificial intelligence in the image diagnosis and preliminary screening of common oral diseases.

Key words: oral panoramic radiographs, deep learning, convolutional neural network, artificial intelligence