上海口腔医学

上海口腔医学 ›› 2026, Vol. 35 ›› Issue (3): 233-237.doi: 10.19439/j.sjos.2026.03.002

• 论著 • 上一篇    下一篇

基于深度学习的搜索识别网络辅助上颌磨牙根管口定位的准确性评价

沈栖云1, 傅裕杰1, 周珂帆2, 陈宇飞2,*, 张旗1,*   

  1. 1.上海市同济口腔医院 牙体牙髓病科,同济大学口腔医学院,上海牙组织修复与再生工程技术研究中心,同济大学口腔医学研究所,上海 200072;
    2.同济大学 计算机科学与技术学院,上海 201804
  • 收稿日期:2025-05-06 修回日期:2025-08-03 发布日期:2026-07-02
  • 通讯作者: 张旗,E-mail：qizhang@tongji.edu.cn;陈宇飞,E-mail：yufeichen@tongji.edu.cn。*共同通信作者
  • 作者简介:沈栖云（1998—）,女,硕士,住院医师,E-mail：shenqiyun@tongji.edu.cn
  • 基金资助:
    国家自然科学基金（82170945、82370950）; 上海申康医院发展中心临床三年行动计划（SHDC2020CR3058B）; 人工智能促进科研范式改革赋能学科跃升计划

Deep learning-based SINet-V2 for localization of root canal orifices in maxillary molars

Shen Qiyun1, Fu Yujie1, Zhou Kefan2, Chen Yufei2, Zhang Qi1   

  1. 1. Department of Endodontics, Shanghai Tongji Stomatological Hospital and Dental School, Tongji University; Shanghai Engineering Research Center of Tooth Restoration and Regeneration & Tongji Research Institute of Stomatology. Shanghai 200072;
    2. School of Computer Science and Technology, Tongji University. Shanghai 201804, China
  • Received:2025-05-06 Revised:2025-08-03 Published:2026-07-02

PDF (PC)

9

摘要： 目的: 基于深度学习构建根管口定位模型,为临床高效精准定位根管口提供参考。方法: 收集80颗离体上颌磨牙,开髓并清理髓室后,插入根管锉以确认根管口位置,采集髓室底显微照片,建立常规根管口数据集。采用随机函数将其中64组数据纳入训练集,16组纳入测试集。将一名牙体牙髓病专家采用MITK软件人工标注髓室底显微影像中的根管口作为参考标准。将训练集输入搜索识别网络2.0版（search and identification network version 2, SINet-V2）,以构建常规根管口定位模型（normal canal orifice localization model,NCO-LM）。将模型测试结果与参考标准进行数目、形态和准确性评估,采用平均交并比（mean intersection over union,MIoU）、根管口数目预测准确率（num equal rate,NER）、平均绝对误差（mean absolute error,MAE）、平均连通域中心距离差（mean distance,MDis）作为评估指标。随后,在上述80颗磨牙的根管口处覆盖光固化复合树脂,再次获取模拟钙化髓室底照片,建立钙化根管口数据集,按NCO-LM的方法训练并测试钙化根管口定位模型（calcified canal orifice localization model,CCO-LM）。结果: NCO-LM与CCO-LM的预测结果与参考标准相似性高,定位距离偏差较小。NCO-LM的MIoU为（87.42±5.47）%,MAE为（0.000 594± 0.000 351）,MDis为(1.63±1.40）,NER为98.44%。CCO-LM的MIoU为（91.87±4.03）%,MAE为（0.000 801± 0.000 543）,MDis为（1.91±1.31）,NER为96.25%。结论: NCO-LM与CCO-LM可精准定位髓室底显微影像中常规及模拟钙化根管口,为人工智能在临床实时辅助根管口定位提供前期理论基础。

关键词: 深度学习, 根管口定位, 钙化根管口, 搜索识别网络2.0版

Abstract: PURPOSE: To construct a deep learning-based model for root canal orifice localization, aiming to provide a reference for clinically efficient and precise root canal orifice identification. METHODS: Eighty maxillary molars were collected. After pulp chamber opening and cleaning, K-files were inserted to confirm root canal orifice positions, and microphotographs of the pulp chamber floor were captured to establish a conventional root canal orifice dataset. A random function was used to allocate 64 samples to the training set and 16 samples to the test set. Manual annotations of root canal orifices in pulp chamber floor images by an endodontic expert using MITK software served as the reference standard. The training set was input into the search identification network version 2 (SINet-V2) to develop a conventional root canal orifice localization model (NCO-LM). Model performance was evaluated against the reference standard using the following metrics: mean intersection over union (MIoU), root canal number prediction accuracy (NER), mean absolute error (MAE), and average connected domain center distance difference(MDis). Subsequently, light-cured composite resin was applied to cover the root canal orifices of the same 80 molars to simulate calcified pulp chambers. Photographs of the simulated calcified pulp chamber floors were acquired to create a calcified root canal orifice dataset. A calcified root canal orifice localization model (CCO-LM) was trained and tested following the NCO-LM methodology. RESULTS: Both NCO-LM and CCO-LM demonstrated high similarity to the reference standard with minimal localization deviations. NCO-LM achieved an MIoU of (87.42±5.47)%, MAE of (0.000 594±0.000 351), MDis of (1.63±1.40) pixels, and NER of 98.44%. CCO-LM showed an MIoU of (91.87±4.03) %, MAE of (0.000 801±0.000 543), MDis of (1.91±1.31) pixels, and NER of 96.25%. CONCLUSIONS: NCO-LM and CCO-LM models based on the SINet-V2 network, which are capable of accurately locating both conventional root canal orifices and hidden orifices covered by simulated calcifications in pulp chamber floor images. These findings provide a preliminary theoretical foundation for artificial intelligence-assisted real-time clinical root canal orifice localization.

Key words: Deep learning, Root canal orifices localization, Calcified root canal orifices, SINet-V2

中图分类号: