基于卷积神经网络的下颌第三磨牙和下颌神经管关系诊断

doi:10.13591/j.cnki.kqyx.2025.08.006

摘要/Abstract

摘要：

目的开发一种自动化系统,能够从全景片中准确判断下颌第三磨牙和下颌神经管之间的关系。方法选取600张口腔全景影像组成数据集,精确标注了下颌第三磨牙与下颌神经管的位置。将研究设计的模型TI-YOLOv5与PANet、Faster R-CNN、Mask R-CNN、ResNeSt-101以及原始YOLOv5(you only look once version 5)在图像分割任务上进行对比,评价指标为平均查准率(average precision,AP)和AP₅₀。结果 TI-YOLOv5 达到 AP 54.0%,AP₅₀ 94.9%,较原始 YOLOv5(AP 49.1%,AP₅₀ 88.2%)分别提升4.9%和6.7%,并超越Mask R-CNN(AP 45.1%,AP₅₀ 84.2%)等其他SOTA方法。结论 TI-YOLOv5在下颌第三磨牙与神经管的自动定位与关系分类上显著优于主流网络,具有较高的检测精度和判别准确率,可为下颌第三磨牙拔除术前风险评估提供可靠的技术支持。

关键词: 深度学习, YOLOv5, 下颌第三磨牙, 下颌神经管

Abstract:

Objective To develop an automated system that can accurately determine the relationship between the mandibular third molar and the mandibular nerve canal from panoramic images. Methods A dataset consisting of 600 panoramic images of the oral cavity was selected, and the positions of the mandibular third molar and the mandibular nerve canal were accurately labeled. We compared the research designed TI-YOLOv5 with PANet, Faster R-CNN, Mask R-CNN, ResNeSt-101, and the original YOLOv5 in image segmentation tasks, with evaluation metrics of AP and AP₅₀. Results TI-YOLOv5 achieved AP(average precision) 54.0% and AP₅₀ 94.9%, an increase of 4.9 and 6.7 percentage points respectively compared to the original YOLOv5 (AP 49.1%, AP₅₀ 88.2%), and surpassed other SOTA methods such as Mask R-CNN (AP 45.1%, AP₅₀ 84.2%). Conclusion TI-YOLOv5 is significantly superior to mainstream networks in automatic positioning and relationship classification of mandibular wisdom teeth and neural tubes, with high detection accuracy and discrimination accuracy, and can provide reliable technical support for preoperative risk assessment of mandibular wisdom tooth extraction.

Key words: deep learning, YOLOv5, mandibular third molar, mandibular nerve canal

中图分类号:

R782.1

张金萍, 虞仙, 陈熠名, 王泽辉, 陶瑜, 韦一, 李碧榕, 朱炳震, 张娟. 基于卷积神经网络的下颌第三磨牙和下颌神经管关系诊断[J]. 口腔医学, 2025, 45(8): 596-602.

ZHANG Jinping, YU Xian, CHEN Yiming, WANG Zehui, TAO Yu, WEI Yi, LI Birong, ZHU Bingzhen, ZHANG Juan. Convolutional neural network-based diagnosis of the relationship between mandibular third molar and mandibular nerve canal[J]. Stomatology, 2025, 45(8): 596-602.

图/表 6

图1

图2

图3

图4

表1

表2

参考文献 29

[1]	Puciło M, Puciło A, Safranow K, et al. The influence of age, sex, and tooth type on the anatomical relationship between tooth roots and the mandibular canal[J]. Imaging Sci Dent, 2021, 51(4): 373-382. doi: 10.5624/isd.20210031 pmid: 34987997
[2]	Komerik N, Muglali M, Tas B, et al. Difficulty of impacted mandibular third molar tooth removal: Predictive ability of senior surgeons and residents[J]. J Oral Maxillofac Surg, 2014, 72(6): 1062. e1-1062.e6.
[3]	Mendes MLT, do Nascimento-Júnior EM, Reinheimer DM, et al. Efficacy of proteolytic enzyme bromelain on health outcomes after third molar surgery. Systematic review and meta-analysis of randomized clinical trials[J]. Med Oral Patol Oral Cir Bucal, 2019, 24(1): e61-e69.
[4]	Xu GZ, Yang C, Fan XD, et al. Anatomic relationship between impacted third mandibular molar and the mandibular canal as the risk factor of inferior alveolar nerve injury[J]. Br J Oral MaxillofacSurg, 2013, 51(8): e215-219.
[5]	Rytkönen K, Ventä I. Distance between mandibular canal and third molar root among 20-year-old subjects[J]. Clin Oral Investig, 2018, 22(7): 2505-2509.
[6]	Su NC, van Wijk A, Berkhout E, et al. Predictive value of panoramic radiography for injury of inferior alveolar nerve after mandibular third molar surgery[J]. J Oral Maxillofac Surg, 2017, 75(4): 663-679.
[7]	Reia VCB, de Toledo Telles-Araujo G, Peralta-Mamani M, et al. Diagnostic accuracy of CBCT compared to panoramic radiography in predicting IAN exposure: A systematic review and meta-analysis[J]. Clin Oral Investig, 2021, 25(8): 4721-4733.
[8]	Bell GW. Use of dental panoramic tomographs to predict the relation between mandibular third molar teeth and the inferior alveolar nerve. Radiological and surgical findings, and clinical outcome[J]. Br J Oral Maxillofac Surg, 2004, 42(1): 21-27.
[9]	Nyachhyon R, Joshi U, Mainali A, et al. Compression of the inferior alveolar canal by mandibular third molar among images taken from patients visiting dental imaging centres of kathmandu: A descriptive cross-sectional study[J]. JNMA J Nepal Med Assoc, 2022, 60(245): 26-30. doi: 10.31729/jnma.7124 pmid: 35199669
[10]	Lin PL, Huang PY, Huang PW. Automatic methods for alveolar bone loss degree measurement in periodontitis periapical radiographs[J]. Comput Methods Programs Biomed, 2017, 148: 1-11.
[11]	Krois J, Ekert T, Meinhold L, et al. Deep learning for the radiographic detection of periodontal bone loss[J]. Sci Rep, 2019, 9(1): 8495. doi: 10.1038/s41598-019-44839-3 pmid: 31186466
[12]	Bayrakdar IS, Orhan K, Akarsu S, et al. Deep-learning approach for caries detection and segmentation on dental bitewing radiographs[J]. Oral Radiol, 2022, 38(4): 468-479.
[13]	Weng S, Wersing H, Steil JJ, et al. Learning lateral interactions for feature binding and sensory segmentation from prototypic basis interactions[J]. IEEE Trans Neural Netw, 2006, 17(4): 843-862.
[14]	Sukegawa S, Yoshii K, Hara T, et al. Multi-task deep learning model for classification of dental implant brand and treatment stage using dental panoramic radiograph images[J]. Biomolecules, 2021, 11(6): 815.
[15]	Chen H, Zhang KL, Lyu PJ, et al. A deep learning approach to automatic teeth detection and numbering based on object detection in dental periapical films[J]. Sci Rep, 2019, 9(1): 3840. doi: 10.1038/s41598-019-40414-y pmid: 30846758
[16]	Cha JY, Yoon HI, Yeo IS, et al. Peri-implant bone loss measurement using a region-based convolutional neural network on dental periapical radiographs[J]. J Clin Med, 2021, 10(5): 1009.
[17]	Hou L, Chen CH, Wang SJ, et al. Multi-object detection method in construction machinery swarm operations based on the improved YOLOv4 model[J]. Sensors, 2022, 22(19): 7294.
[18]	Ahmad MA, Ourak M, Gruijthuijsen C, et al. Deep learning-based monocular placental pose estimation: Towards collaborative robotics in fetoscopy[J]. Int J Comput Assist Radiol Surg, 2020, 15(9): 1561-1571. doi: 10.1007/s11548-020-02166-3 pmid: 32350788
[19]	Wang SD, Huang L, Jiang D, et al. Improved multi-stream convolutional block attention module for sEMG-based gesture recognition[J]. Front Bioeng Biotechnol, 2022, 10: 909023.
[20]	Zhang Y, Liu CQ, Liu M, et al. Attention is all you need: Utilizing attention in AI-enabled drug discovery[J]. Brief Bioinform, 2023, 25(1): bbad467.
[21]	Yi J, Liu X, Cheng S, et al. Multi-scale window transformer for cervical cytopathology image recognition[J]. Comput Struct Biotechnol J, 2024, 24: 314-321.
[22]	Hu JF, Sun JX, Lin ZH, et al. APANet: Auto-path aggregation for future instance segmentation prediction[J]. IEEE Trans Pattern Anal Mach Intell, 2022, 44(7): 3386-3403.
[23]	Wang Y, Chen XQ, Li JQ, et al. Convolutional block attention module-multimodal feature-fusion action recognition: Enabling miner unsafe action recognition[J]. Sensors (Basel), 2024, 24(14): 4557.
[24]	Papanastasiou G, Dikaios N, Huang JH, et al. Is attention all you need in medical image analysis? a review[J]. IEEE J Biomed Health Inform, 2024, 28(3): 1398-1411.
[25]	Wang WJ, Li XY, Ren HL, et al. Chinese clinical named entity recognition from electronic medical records based on multisemantic features by using robustly optimized bidirectional encoder representation from transformers pretraining approach whole word masking and convolutional neural networks: Model development and validation[J]. JMIR Med Inform, 2023, 11: e44597.
[26]	Li XM, Chen H, Qi XJ, et al. H-DenseUNet: Hybrid densely connected UNet for liver and tumor segmentation from CT volumes[J]. IEEE Trans Med Imaging, 2018, 37(12): 2663-2674.
[27]	Masood M, Masood Y, Newton JT. The clustering effects of surfaces within the tooth and teeth within individuals[J]. J Dent Res, 2015, 94(2): 281-288. doi: 10.1177/0022034514559408 pmid: 25421840
[28]	Buyuk C, Akkaya N, Arsan B, et al. A fused deep learning architecture for the detection of the relationship between the mandibular third molar and the mandibular canal[J]. Diagnostics (Basel), 2022, 12(8): 2018.
[29]	Yang Y, Bao DY, Ni C, et al. Three-dimensional positional relationship between impacted mandibular third molars and the mandibular canal[J]. BMC Oral Health, 2023, 23(1): 831. doi: 10.1186/s12903-023-03548-0 pmid: 37924035

网络结构	AP	AP₅₀
YOLOv5	20.1	88.2
YOLOv5+ Transformer	22.1	92.1
YOLOv5+Transformer+ Fusion	23.4	93.1
YOLOv5+Transformer+Fusion+Involution	25.1	94.1
TI-YOLOv5(上述+CBAM)	29.1	94.9