口腔医学研究

口腔医学研究 ›› 2026, Vol. 42 ›› Issue (3): 242-246.DOI: 10.13701/j.cnki.kqyxyj.2026.03.012

• 口腔正畸学研究 • 上一篇    下一篇

基于YOLOv7深度学习模型的正畸患者口内照片牙菌斑目标检测研究

尹清, 唐倩, 包幸福*   

  1. 吉林大学口腔医院正畸科 吉林 长春 130021
  • 收稿日期:2025-10-16 发布日期:2026-03-26
  • 通讯作者: * 包幸福,E-mail:baoxf@jlu.edu.cn
  • 作者简介:尹清(2000~ ),女,河北人,硕士在读,研究方向:深度学习在口腔中的应用。
  • 基金资助:
    吉林省口腔正畸临床医学研究中心(编号:YDZJ202402073CXJD)

YOLOv7-based Deep Learning Model for Dental Plaque Object Detection in Intraoral Photographs of Orthodontic Patients

YIN Qing, TANG Qian, BAO Xingfu*   

  1. Department of Orthodontics, Stomatological Hospital of Jilin University, Changchun 130021, China
  • Received:2025-10-16 Published:2026-03-26

PDF (PC)

35

摘要： 目的:接受固定矫治器的正畸患者常因口腔内托槽、弓丝等装置的存在而难以观察牙菌斑,易导致清洁不足,进而引发釉质脱矿、龋齿及牙周疾病等问题。本研究旨在系统评估YOLOv7模型通过固定矫治患者使用智能手机拍摄的口腔照片识别牙菌斑的性能与可行性。方法:本研究于2024年3月~2025年6月期间,从吉林大学口腔医院正畸科收集23例固定矫治患者的口腔内牙菌斑染色图像,将其划分为训练集、验证集和测试集3组。采用配备图形处理器(graphics processing unit,GPU)的计算机,安装Python、PyTorch、Torchvision等必要软件包构建YOLOv7运行环境。通过训练集训练模型、验证集验证模型、测试集测试模型,确定最佳学习权重,使算法能够识别正畸患者口腔照片中的牙菌斑。结果:实验结果表明,该模型在接受固定正畸治疗的患者中牙菌斑识别精确率(precision,P)为82.91%,回归率(recall,R)为78.00%,平均精度均值(mean average precision,mAP50)为79.60%。该模型检测成功率优于患者和正畸医师对牙菌斑图像的视觉识别。结论:上述结果表明基于YOLOv7的深度学习算法已达到临床应用水平。本研究使固定矫治患者仅需使用智能手机拍摄口腔照片即可获得实时牙菌斑检测结果。未来将该模型集成至软件系统,将显著提升固定矫治患者日常口腔卫生管理的可行性和有效性。

关键词: 菌斑识别, 正畸, 口腔健康, 口腔健康保健, 深度学习

Abstract: Objective: To systematically evaluate the performance and feasibility of the YOLOv7 model in identifying dental plaque from intraoral photographs taken by fixed orthodontic patients using their smartphones. Methods: The study collected intraoral plaque staining images from 23 patients undergoing fixed orthodontic treatment at the Orthodontics Department of Jilin University Stomatological Hospital between March 2024 and June 2025. They were categorized into three sets: training, validation, and test sets. A GPU-equipped computer was used, with Python, PyTorch, Torchvision, and the other necessary software packages installed to configure the environment for YOLOv7. The model was trained on the training set, validated on the validation set, and tested on the test set to determine the optimal learning weights, enabling the algorithm to identify plaque in intraoral photographs of orthodontic patients. Results: Experimental results showed that the model achieved a plaque identification precision of 82.91% for patients undergoing fixed orthodontic treatment, a Recall of 78.00%, and a mAP50 of 79.60%. The model’s detection success rate outperformed the visual identification of plaque images by both patients and orthodontists. Conclusion: The deep learning algorithm based on YOLOv7 achieved a level suitable for clinical application. This study enables fixed orthodontic patients to obtain real-time plaque detection results simply by taking intraoral photos with a smartphone. In the future, integrating the model into software will significantly enhance the feasibility and effectiveness of daily oral hygiene management for fixed orthodontic patients.

Key words: plaque detection, orthodontics, oral health, oral health care, deep learning