基于深度学习方法建立颞下颌关节核磁共振影像的自动分割模型

doi:10.13591/j.cnki.kqyx.2025.06.009

摘要/Abstract

摘要：

目的开发一种基于深度学习的自动分割模型,用于颞下颌关节(TMJ)核磁共振成像(MRI)的分割。方法收集104例研究对象的颞下颌关节MRI数据,并对关节盘、髁突及关节窝进行标注。利用自适应U-Net框架(nnU-Net),构建分割模型,并对该模型进行定量和定性评估。结果所建立的分割模型在分割准确性方面表现出色。在不同关节结构的分割上,关节盘的Dice相似系数(Dice)达0.77,髁突的Dice达0.85,关节窝的Dice达0.66。模型在处理开口和闭口状态下的MRI图像时,展现了相近的分割性能。结论本研究开发了一种基于深度学习的颞下颌关节MRI自动分割模型,该模型有助于临床医生对颞下颌关节盘移位做出快速判断。

关键词: 颞下颌关节, 核磁共振, 深度学习, 自动分割

Abstract:

Objective To build an automatic segmentation model of temporomandibular joint(TMJ) based on magnetic resonance imaging(MRI) using deep learning method. Methods The MRI data of TMJ of 104 subjects were collected, with the articular disc, condyle and glenoid fossa marked. The adaptive U-Net framework(nnU-Net) was used to construct a segmentation model, which was subjected to both quantitative and qualitative assessments. Results The segmentation model demonstrated excellent accuracy in segmentation. In the segmentation of different joint structures, the model achieved Dice of 0.77 for the articular disc, 0.85 for the condyle, and 0.66 for the glenoid fossa. The model showed similar segmentation performance when processing MRI images in both open-mouth and closed-mouth states. Conclusion This study developed an automatic segmentation model for TMJ MRI based on deep learning, which can assist clinicians in diagnosing anterior displacement of the TMJ disc.

Key words: temporomandibular joint, magnetic resonance imaging, deep learning, automatic segmentation

中图分类号:

R782.63

刘飞, 张久楼, 金若帆, 张楠, 周薇娜. 基于深度学习方法建立颞下颌关节核磁共振影像的自动分割模型[J]. 口腔医学, 2025, 45(6): 445-452.

LIU Fei, ZHANG Jiulou, JIN Ruofan, ZHANG Nan, ZHOU Weina. The automatic segmentation of the temporomandibular joint based on MRI using deep learning method[J]. Stomatology, 2025, 45(6): 445-452.

图/表 15

表1

表2

图1

图2

图3

图4

图5

图6

图7

表3

表4

表5

图8

表6

图9

参考文献 23

[1]	谢丛蔓. 中国大陆地区颞下颌关节紊乱病流行趋势的系统评价及Meta分析[D]. 重庆: 重庆医科大学, 2020.
[2]	Lei J, Yap AU, Li Y, et al. Clinical protocol for managing acute disc displacement without reduction: A magnetic resonance imaging evaluation[J]. Int J Oral Maxillofac Surg, 2020, 49(3): 361-368.
[3]	Spitzer WJ, Lenz M, Sauter R. Imaging the articular disk of the temporomandibular joint using nuclear magnetic resonance: Preliminary report[J]. Dtsch Zahnarztl Z, 1986, 41(7): 693-696. pmid: 3461994
[4]	傅开元, 雷杰. 颞下颌关节紊乱病的分类、诊断及治疗进展[J]. 口腔医学, 2024, 44(1): 6-10.
[5]	张倩, 李俊, 马丽霞, 等. 颞下颌关节张闭口位磁共振成像在颞下颌关节盘前移位中的诊断价值[J]. 影像研究与医学应用, 2021, 5(14): 26-28.
[6]	Widmalm SE, Brooks SL, Sano T, et al. Limitation of the diagnostic value of MR images for diagnosing temporomandibular joint disorders[J]. Dentomaxillofac Radiol, 2006, 35(5): 334-338.
[7]	Nebbe B, Brooks SL, Hatcher D, et al. Magnetic resonance imaging of the temporomandibular joint: Interobserver agreement in subjective classification of disk status[J]. Oral Surg Oral Med Oral Pathol Oral Radiol Endod, 2000, 90(1): 102-107.
[8]	Alkhader M, Ohbayashi N, Tetsumura A, et al. Diagnostic performance of magnetic resonance imaging for detecting osseous abnormalities of the temporomandibular joint and its correlation with cone beam computed tomography[J]. Dentomaxillofac Radiol, 2010, 39(5): 270-276.
[9]	LeCun Y, Bengio Y, Hinton G. Deep learning[J]. Nature, 2015, 521(7553): 436-444.
[10]	殷晓航, 王永才, 李德英. 基于U-Net结构改进的医学影像分割技术综述[J]. 软件学报, 2021, 32(2): 519-550.
[11]	Savjani R. nnU-Net: Further automating biomedical image autosegmentation[J]. Radiol Imaging Cancer, 2021, 3(1): e209039.
[12]	Isensee F, Jaeger PF, Kohl SAA, et al. nnU-Net: A self-configuring method for deep learning-based biomedical image segmentation[J]. Nat Methods, 2021, 18(2): 203-211. doi: 10.1038/s41592-020-01008-z pmid: 33288961
[13]	Zwijnen AW, Watzema L, Ridwan Y, et al. Self-adaptive deep learning-based segmentation for universal and functional clinical and preclinical CT image analysis[J]. Comput Biol Med, 2024, 179: 108853.
[14]	Kang H, Witanto JN, Pratama K, et al. Fully automated MRI segmentation and volumetric measurement of intracranial meningioma using deep learning[J]. J Magn Reson Imaging, 2023, 57(3): 871-881.
[15]	Bareja R, Ismail M, Martin D, et al. nnU-net-based segmentation of tumor subcompartments in pediatric medulloblastoma using multiparametric MRI: A multi-institutional study[J]. Radiol Artif Intell, 2024, 6(5): e230115.
[16]	方媛媛, 毛伟玉, 王中振, 等. 基于深度学习的颞下颌关节骨关节病锥形束CT自动诊断及分类[C]// 中华口腔医学会颞下颌关节病学及牙合学专委会. 第20次全国颞下颌关节病学及牙合学研讨会暨第七届亚洲颞下颌关节学术大会论文汇编. 北京,2023:284-285.
[17]	Ito S, Mine Y, Yoshimi Y, et al. Automated segmentation of articular disc of the temporomandibular joint on magnetic resonance images using deep learning[J]. Sci Rep, 2022, 12(1): 221. doi: 10.1038/s41598-021-04354-w pmid: 34997167
[18]	Lin BL, Cheng MS, Wang SZ, et al. Automatic detection of anteriorly displaced temporomandibular joint discs on magnetic resonance images using a deep learning algorithm[J]. Dentomaxillofac Radiol, 2022, 51(3): 20210341.
[19]	Nozawa M, Ito H, Ariji Y, et al. Automatic segmentation of the temporomandibular joint disc on magnetic resonance images using a deep learning technique[J]. Dentomaxillofac Radiol, 2022, 51(1): 20210185.
[20]	Emshoff R, Brandlmaier I, Bertram S, et al. Relative odds of temporomandibular joint pain as a function of magnetic resonance imaging findings of internal derangement, osteoarthrosis, effusion, and bone marrow edema[J]. Oral Surg Oral Med Oral Pathol Oral Radiol Endod, 2003, 95(4): 437-445.
[21]	Chan HP, Samala RK, Hadjiiski LM, et al. Deep learning in medical image analysis[J]. Adv Exp Med Biol, 2020, 1213: 3-21.
[22]	Li MX, Punithakumar K, Major PW, et al. Temporomandibular joint segmentation in MRI images using deep learning[J]. J Dent, 2022, 127: 104345.
[23]	Soldati E, Rossi F, Vicente J, et al. Survey of MRI usefulness for the clinical assessment of bone microstructure[J]. Int J Mol Sci, 2021, 22(5): 2509.

关节盘位置	左侧	右侧
正常	43	39
ADDwR	25	23
ADDwoR	36	42
总计	104	104

关节盘位置	左侧	右侧
正常	6	3
ADDwR	6	6
ADDwoR	8	11
总计	20	20

模型	图像类别	关节组织	Dice	召回率	精确率	F₁分数
2D模型	闭口位	关节盘	0.775 911	0.761 949	0.802 372	0.781 638
		髁突	0.861 552	0.858 067	0.880 847	0.869 308
		关节窝	0.661 321	0.613 758	0.717 471	0.661 574
	开口位	关节盘	0.790 665	0.788 616	0.807 662	0.798 026
		髁突	0.899 456	0.901 339	0.908 778	0.905 043
		关节窝	0.662 487	0.624 911	0.711 179	0.665 260
3D模型	闭口位	关节盘	0.778 488	0.748 471	0.821 614	0.783 339
		髁突	0.869 087	0.871 318	0.882 744	0.876 994
		关节窝	0.680 833	0.643 324	0.722 766	0.680 735
	开口位	关节盘	0.796 855	0.781 431	0.827 924	0.804 006
		髁突	0.902 941	0.906 160	0.912 518	0.909 328
		关节窝	0.680 003	0.650 240	0.716 696	0.681 853

关节组织		平方和	自由度	均方	F	P
关节盘	组间	0.006	3	0.002	0.350	0.789
	组内	0.429	76	0.006
	总计	0.435	79
髁突	组间	0.026	3	0.009	1.998	0.121
	组内	0.335	76	0.004
	总计	0.362	79
关节窝	组间	0.007	3	0.002	0.376	0.771
	组内	0.263	57	0.005
	总计	0.471	79

模型		平方和	自由度	均方	F	P
2D/闭口位	组间	0.404	2	0.202	34.933	<0.001
	组内	0.329	57	0.006
	总计	0.733	59
3D/闭口位	组间	0.355	2	0.177	30.568	<0.001
	组内	0.331	57	0.006
	总计	0.685	59
2D/开口位	组间	0.563	2	0.281	52.560	<0.001
	组内	0.305	57	0.005
	总计	0.868	59
3D/开口位	组间	0.497	2	0.249	53.872	<0.001
	组内	0.263	57	0.005
	总计	0.761	59