基于单步神经网络的牙根与下颌管关系自动检测

周炎锜; 戴修斌; 王东苗; 朱书进; 冒添逸

doi:10.15953/j.ctta.2022.083

基于单步神经网络的牙根与下颌管关系自动检测

doi: 10.15953/j.ctta.2022.083

1.
南京邮电大学地理与生物信息学院, 南京210023
2.
南京医科大学附属口腔医学院, 南京211166

基金项目: 江苏省自然科学基金（基于多尺度细粒度网络和二值自编码模型的病理图像快速检索研究（BK20200745））；江苏省卫生健康委面上项目（基于深度学习模型精准评估下颌智齿拔除术并发下牙槽神经损伤的临床研究（M2020021））；江苏省高等学校自然科学研究项目（基于动态变尺度栈式二值自编码的病理图像实时检索研究（20KJB510022））。

详细信息

作者简介:
周炎锜：男，南京邮电大学地理与生物信息学院硕士研究生，主要从事医学图像处理，E-mail：1020173024@njupt.edu.cn

戴修斌：男，南京邮电大学地理与生物信息学院副教授、研究生导师，主要从事医学图像处理，E-mail：daixb@njupt.edu.cn

通讯作者:
戴修斌*，

中图分类号: O 242；R 814
计量
- 文章访问数: 79
- HTML全文浏览量: 35
- PDF下载量: 12
- 被引次数: 0
出版历程
- 收稿日期: 2022-05-07
- 修回日期: 2022-09-25
- 录用日期: 2022-10-27
- 网络出版日期: 2022-11-04
- 刊出日期: 2023-03-31

Automatic Identification of Relationship between Tooth Root and Mandibular Canal Based on One Step Deep Neural Network

1.
School of Geographic and Biologic Information, Nanjing University of Posts and Telecommunications, Nanjing 210023, China
2.
School of Stomatology, Nanjing Medical University, Nanjing 211166, China

摘要

摘要: 为了提高曲面体层片中下颌阻生智齿牙根与下颌管位置关系的识别精度和效率，提出一种基于深度卷积神经网络的自动检测方法。该方法将下颌阻生智齿牙根与下颌管位置关系的自动检测视为回归任务与分类任务的结合，以YOLOv5网络为框架构建可同时完成分类和定位任务的深度卷积神经网络，将对应锥形束CT图像中获取的空间位置关系信息作为分类金标准，训练其学习曲面体层片图像特征与接触下颌管的智齿牙根之间的非线性关系。将新获得的曲面体层片输入到训练好的网络模型后，即可获得该曲面体层片下颌阻生智齿牙根与下颌管相互接触的概率值，同时预测出存在牙根与下颌管相互接触情况的区域。实验结果表明，本文方法能准确地判断出下颌阻生智齿牙根与下颌管是否接触，并能预测出存在牙根与下颌管相互接触情况的区域；与人工判读和其他方法相比，能获得更准确的检测结果。
- 深度卷积神经网络 /
- 曲面体层片 /
- 下颌阻生智齿 /
- 位置关系
Abstract: To improve the accuracy and efficiency of identifying the relationship between the root of the impacted mandibular third molar (M3M) and the mandibular canal in panoramic radiographs, we proposed an automatic method based on a deep convolutional neural network. This method treats the automatic identification of the relationship between the root of the M3M and the mandibular canal as a combination of regression and classification tasks. It uses the YOLOv5 (You Only Look Once) network as a framework for constructing a deep convolutional neural network that can accomplish detection and classification tasks simultaneously. This network, which takes the spatial relationship information extracted from the corresponding cone-beam CT images as the ground-truth, was trained to learn the nonlinear relationship between image features and the root of the M3M contacting the mandibular canal. When inputting a newly acquired panoramic radiograph into the trained network, the network will output the probability value for the root of the M3M contacting the mandibular canal. In the meantime, the region that includes the root of the M3M contacting the mandibular canal can be predicted. The experimental results show that the proposed method can provide an accurate judgment of whether the roots of impacted mandibular wisdom teeth in the panoramic radiographs are in contact with the mandibular canal and the location of regions in which the roots of the M3M are in contact with the mandibular canals; compared to manual diagnosis and the other methods, the proposed method can obtain more accurate results.
- deep convolutional neural network /
- panoramic radiograph /
- impacted mandibular third molar /
- position relationship

HTML全文

图 1 本文所提方法的流程图

Figure 1. The pipeline of the proposed method

下载: 全尺寸图片幻灯片

图 2 本文所用网络结构图

Figure 2. The structure of the network used in this paper

下载: 全尺寸图片幻灯片

图 3 FPN+PAN结构示意图

Figure 3. FPN+PAN structure

下载: 全尺寸图片幻灯片

图 4 使用不同模型获得结果对应ROC曲线的对比

Figure 4. The comparison of corresponding ROC curves of the results obtained using the different models

下载: 全尺寸图片幻灯片

图 5 使用本文方法所得结果对应的PR曲线

Figure 5. The PR curves corresponding to the results obtained using the method in this paper

下载: 全尺寸图片幻灯片

图 6 使用本文方法得到的输出结果

每一行左侧图像均为原始图像，右侧是本文方法输出结果图。结果图中的棕色字符contact代表下颌阻生智齿牙根与下颌管的位置关系是接触状态，蓝色字符contactless代表两者没有接触；棕色数字是预测为接触的置信度，蓝色数字是预测为不接触的置信度。

Figure 6. Demonstration of the output results obtained from the proposed method in this paper

下载: 全尺寸图片幻灯片

图 7 待检测物体所在区域被不同方向移动后图像的部分测试结果

Figure 7. Some examples of predicted results, after the regions including the objects to be detected in panoramic images, have been shifted in different directions

下载: 全尺寸图片幻灯片

表 1 Backbone网络涉及的主要参数

Table 1. The main parameters of the backbone network

模块名称	数量	卷积核	尺寸	步长	输入尺寸	输出尺寸
Conv	1	80	3⊆3	2	608⊆608⊆3	304⊆304⊆80
Conv	1	160	3⊆3	2	304⊆304⊆80	152⊆152⊆160
CSP1_4	4	160	－	－	152⊆152⊆160	152⊆152⊆160
Conv	1	320	3⊆3	2	152⊆152⊆160	76⊆76⊆320
CSP1_8	8	320	－	－	76⊆76⊆320	76⊆76⊆320
Conv	1	640	3⊆3	2	76⊆76⊆320	38⊆38⊆640
CSP1_12	12	640	－	－	38⊆38⊆640	38⊆38⊆640
Conv	1	1280	3⊆3	2	38⊆38⊆640	19⊆19⊆1280
CSP1_4	4	1280	－	－	19⊆19⊆1280	19⊆19⊆1280
SPPF	1	1280	－	－	19⊆19⊆1280	19⊆19⊆1280

下载: 导出CSV

表 2 本文方法与其他方法及人工判读所得预测结果对应的分类性能评价指标的对比

Table 2. The comparison of classification performance for the proposed method, manual diagnosis, and the other models

方法	准确率	灵敏度	特异度	精确度
人工判读	0.845	0.741	0.892	0.759
AlexNet	0.778	0.506	0.919	0.764
GoogLeNet	0.770	0.434	0.944	0.800
VGG-16	0.737	0.422	0.900	0.686
ResNet-50	0.831	0.663	0.919	0.809
本文方法	0.881	0.819	0.913	0.829

下载: 导出CSV

表 3 部分概率阈值对应的分类性能评价指标值

Table 3. The measurements of classification performance for different thresholds

概率阈值	准确率	灵敏度	特异度	精确度
0.60	0.881	0.819	0.913	0.829
0.65	0.881	0.795	0.925	0.846
0.70	0.868	0.747	0.931	0.849
0.75	0.860	0.699	0.944	0.866

下载: 导出CSV

表 4 使用本文方法时，不同训练迭代次数、批大小、学习率以及优化器参数对应的各分类性能指标值

Table 4. The measurements of classification performance for different iterations, epochs, learning rates, and parameters of the optimizer in the proposed method

参数名称	参数取值	准确率	灵敏度	特异度	精确度
训练迭代次数	800	0.877	0.807	0.913	0.827
	1200	0.881	0.819	0.913	0.829
	1600	0.835	0.614	0.95	0.864
批大小	4	0.840	0.602	0.963	0.893
批大小	6	0.881	0.819	0.913	0.829
学习率	0.0022	0.889	0.759	0.956	0.900
	0.0032	0.881	0.819	0.913	0.829
	0.0042	0.823	0.590	0.944	0.845
优化器参数β₁（β₂=0.999）	0.743	0.868	0.723	0.944	0.870
	0.843	0.881	0.819	0.913	0.829
	0.943	0.864	0.747	0.925	0.838
优化器参数β₂（β₁=0.843）	0.9	0.856	0.675	0.950	0.875
	0.99	0.868	0.759	0.925	0.840
	0.999	0.881	0.819	0.913	0.829

下载: 导出CSV

表 5 本文所用深度网络与其他模型涉及参数数量的对比

Table 5. The comparison of the number of parameters used in our network and the others

网络模型	AlexNet	GoogLeNet	VGG-16	ResNet-50	本文方法
参数量/M	61.0	7.0	138.4	25.5	87.3

下载: 导出CSV

表 6 待检测物体所在区域朝不同方向移动前和移动后，本文方法对测试图像所得预测结果的分类性能评价指标对比

Table 6. The comparison of classification performance for the predicted results between the cases with or without the regions including the targets shifted in different directions

方法	准确度	灵敏度	特异度	精确度
移动前	0.881	0.819	0.913	0.829
移动后	0.864	0.723	0.938	0.857

下载: 导出CSV

参考文献(16)

[1]	王东苗, 金致纯, 丁旭, 等. 锥形束CT评估下颌阻生智齿拔除术后下牙槽神经损伤的风险[J]. 南京医科大学学报(自然科学版), 2016,36(10): 1263−1266.
[2]	WANG D M, HE X T, WANG Y L, et al. Topographic relationship between root apex of mesially and horizontally impacted mandibular third molar and lingual plate: Cross-sectional analysis using CBCT[J]. Scientific Reports, 2016, 6(1): 39268−39278. doi: 10.1038/srep39268
[3]	EKERT T, KROIS J, MEINHOLD L, et al. Deep learning for the radiographic detection of apical lesions[J]. Journal of Endodontics, 2019, 45(7): 917−922. doi: 10.1016/j.joen.2019.03.016
[4]	CHANG H J, LEE S J, YONG T H, et al. Deep learning hybrid method to automatically diagnose periodontal bone loss and stage periodontitis[J]. Scientific Reports, 2020, 10(1): 7531−7539. doi: 10.1038/s41598-020-64509-z
[5]	ARIJI Y, YANASHITA Y, KUTSUNA S, et al. Automatic detection and classification of radiolucent lesions in the mandible on panoramic radiographs using a deep learning object detection technique[J]. Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 2019, 128(4): 424−430. doi: 10.1016/j.oooo.2019.05.014
[6]	VINAYAHALINGAM S, TONG X, BERGÉ S, et al. Automated detection of third molars and mandibular nerve by deep learning[J]. Scientific Reports, 2019, 9(1): 9007−9014. doi: 10.1038/s41598-019-45487-3
[7]	LEE J, KIM D, JEONG S. Diagnosis of cystic lesions using panoramic and cone beam computed tomographic images based on deep learning neural network[J]. Oral Diseases, 2020, 26(1): 152−158. doi: 10.1111/odi.13223
[8]	FUKUDA M, ARIJI Y, KISE Y, et al. Comparison of 3 deep learning neural networks for classifying the relationship between the mandibular third molar and the mandibular canal on panoramic radiographs[J]. Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 2020, 130(3): 336−343. doi: 10.1016/j.oooo.2020.04.005
[9]	CHOI E, LEE S, JEONG E, et al. Artificial intelligence in positioning between mandibular third molar and inferior alveolar nerve on panoramic radiography[J]. Scientific Reports, 2022, 12(1): 2456−2463. doi: 10.1038/s41598-022-06483-2
[10]	ZWA B, LJA B, SHUAI W. Apple stem/calyx real-time recognition using YOLO-v5 algorithm for fruit automatic loading system[J]. Postharvest Biology and Technology, 2022, 185(2): 111808−111815.
[11]	BOCHKOVSKIY A, WANG C Y, LIAO H. YOLOv4: Optimal speed and accuracy of object detection[J]. arXiv Preprint, 2020, arXiv: 2004.10934v1.
[12]	LIN T Y, DOLLAR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]//IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017: 2117-2125.
[13]	LI H, XIONG P, AN J, et al. Pyramid attention network for semantic segmentation[J]. arXiv Preprint, 2018, arXiv:1805.10180.
[14]	REZATOFIGHI H, TSOI N, GWAK J Y. Generalized Intersection over Union: A metric and a loss for bounding box regression[C]//IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2019: 658-666.
[15]	COOK N R. Use and misuse of the receiver operating characteristic curve in risk prediction[J]. Circulation, 2007, 115(7): 928−35. doi: 10.1161/CIRCULATIONAHA.106.672402
[16]	BUCKLAND M K, GEY F C. The relationship between recall and precision[J]. Journal of the Association for Information Science & Technology, 2010, 45(1): 12−19.