口腔医学研究

口腔医学研究 ›› 2025, Vol. 41 ›› Issue (8): 695-699.DOI: 10.13701/j.cnki.kqyxyj.2025.08.010

• 牙周病学研究 • 上一篇    下一篇

基于深度学习的三维牙槽骨吸收评估及牙周炎分期

邱岳1,2, 韩阳平1,2, 邓冠红2,3, 李菁2,3*   

  1. 1.厦门医学院附属口腔医院牙周病二科 福建 厦门 361000;
    2.厦门市口腔疾病诊疗重点实验室 福建 厦门 361000;
    3.厦门医学院附属口腔医院口腔黏膜病科 福建 厦门 361000
  • 收稿日期:2025-03-21 出版日期:2025-08-28 发布日期:2025-08-15
  • 通讯作者: *李菁,E-mail:787956094@qq.com
  • 作者简介:邱岳(1989~ ),男,福建人,硕士,主治医师,研究方向:牙周病学与人工智能。
  • 基金资助:
    厦门市自然科学基金联合项目(编号:3502Z20227405)

Three-Dimensional Alveolar Bone Resorption Assessment and Periodontitis Staging Based on Deep Learning

QIU Yue1,2, HAN Yangping1,2, DENG Guanghong2,3, LI Jing2,3*   

  1. 1. The Second Department of Periodontology, Stomatological Hospital of Xiamen Medical College, Xiamen 361000, China;
    2. Xiamen Key Laboratory of Stomatological Disease Diagnosis and Treatment, Xiamen 361000, China;
    3. Department of Oral Mucosal Diseases, Stomatological Hospital of Xiamen Medical College, Xiamen 361000, China
  • Received:2025-03-21 Online:2025-08-28 Published:2025-08-15

PDF (PC)

79

摘要： 目的: 本研究旨在构建一种基于多阶段深度学习的三维牙槽骨吸收率自动量化评估模型,为牙周炎分期提供智能化指标。方法: 收集来自牙周炎患者的100例锥形束CT(cone beam computed tomography,CBCT)影像资料,采用基于体积卷积神经网络(fully convolutional neural network for volumetric medical image segmentation,V-Net)与改进的3D医学图像分割Transformer网络(transformer-based encoder-decoder network for efficient and accurate 3D medical image,UNETR++)相结合的多任务分割策略,分别对牙齿、牙槽骨及釉牙骨质界(cemento-enamel junction,CEJ)进行高精度分割。通过膨胀操作与接触面积比值计算相结合,并辅以CEJ方向聚类,实现牙槽骨吸收率的三维自动化量化分析,根据2018年世界牙周病新分类对牙周炎进行初步分期。结果: 模型对牙齿、牙槽骨和CEJ进行自动化分割,Dice相似系数(Dice similarity coefficient,DSC)分别达到95.7%、91.5%和87.9%,且在分割的平均表面距离(average surface distance,ASD)、豪斯多夫距离(Hausdorff distance,HD)和灵敏度(sensitivity,SEN)等方面表现出高精度与稳定性。模型实现了近远中、颊舌向等多维度的骨吸收分析,与牙周专科医师评估结果相比,本模型在牙周炎分期上的总体准确率为85%,并取得较高的一致性(Kappa=0.773)。模型对重度牙槽骨吸收(Ⅲ期或Ⅳ期)的识别准确(F1分数=1.00)。模型较人工评估显著提升效率(配对t检验,P<0.01)。结论: 本方法可快速、准确地测量三维牙槽骨吸收率,为牙周炎精准分期和临床诊断提供重要参考。

关键词: 牙周炎, 牙槽骨吸收率, 锥形束CT, 深度学习, 牙周病新分类

Abstract: Objective: To develop a multi-stage deep learning-based model for automatic quantitative assessment of three-dimensional alveolar bone resorption, providing intelligent quantitative indicators for periodontitis staging. Methods: Cone beam computed tomography (CBCT) scans from 100 periodontitis patients were collected. A multi-task segmentation strategy combining fully Convolutional Neural Network for volumetric medical image segmentation (V-Net) and transformer-based encoder-decoder network for efficient and accurate 3D medical image (UNETR++) was proposed to achieve high-precision segmentation of teeth, alveolar bone, and the cemento-enamel junction (CEJ). By combining dilation operations and contact-area ratio calculations, complemented by clustering in the CEJ direction, the three-dimensional alveolar bone resorption was quantitatively measured. Preliminary determination of periodontitis staging was acquired based on the 2018 new classification of periodontal diseases. Results: The model automatically segmented the teeth, alveolar bone, and CEJ, achieving Dice similarity coefficient (DSC) of 95.7%, 91.5%, and 87.9%. The model also demonstrated high accuracy and stability in terms of average surface distance (ASD), hausdorff distance (HD), and sensitivity (SEN). The model enabled multidimensional analysis of bone resorption in the mesiodistal and buccolingual directions. Compared with periodontists’ evaluations, the overall accuracy of periodontitis staging by this model was 85%, with a high level of agreement (Kappa=0.773). The model achieved an F1 score of 1.00 in identifying severe alveolar bone resorption (Stage Ⅲ or Ⅳ). A paired t-test (P<0.01) indicated a significant improvement in efficiency over manual assessment. Conclusion: This method can measure three-dimensional alveolar bone resorption quickly and accurately, providing a valuable reference for precise periodontitis staging and clinical diagnosis.

Key words: periodontitis, alveolar bone resorption rate, CBCT, deep learning, new classification of periodontal diseases