基于多源半监督学习的牙齿结构分割

doi:10.11871/jfdc.issn.2096-742X.2025.02.017

摘要/Abstract

摘要：

【目的】现有的牙齿结构分割研究中，牙齿数据具有成本高昂和来源复杂的特点，研究人员普遍选择在多源数据集上使用半监督方法构建深度学习网络，而从多源未标注数据中学习差异成为新的挑战。【方法】本研究以半监督算法Mean-Teacher为核心，提出一种多源半监督算法，并使用新的损失函数计算一致性损失，旨在更好地帮助半监督模型进行训练。【结果】在真实临床场景中收集的测试数据集上评估了这一方法的分割性能，结果表明，本文提出的方法在整体牙齿上的分割结果为平均Dice系数81.13%，平均Jaccard系数82.96%，平均召回率82.78%，平均精确度82.91%，与其他方法相比获得了更准确的分割性能。【结论】比起直接将未标注数据统一投入训练的主流方法，基于多源半监督学习的牙齿分割模型Multi-Student可以更好地帮助半监督模型提高精度。

关键词: 人工智能, 医学影像, 迁移学习, 深度学习, 牙齿分割

Abstract:

[Objective] Current research in dental structure segmentation faces two primary constraints: the prohibitively high costs of annotated dental data and the inherent complexity of multi-source datasets. While semi-supervised learning approaches have been widely adopted for constructing deep learning networks using multi-source data, effectively extracting discriminative features from heterogeneous unlabeled data remains a significant challenge. [Methods] Building upon the Mean-Teacher framework, we propose a novel Multi-Source Semi-Supervised Learning algorithm that incorporates, a domain-adaptive consistency regularization mechanism, and a hybrid loss function combining structural similarity and uncertainty-aware constraints to optimize model training. [Results] Comprehensive evaluation on a clinically acquired dataset demonstrated superior segmentation performance, achieving mean scores of: Dice 81.13%, Jaccard 82.96%, recall 82.78%, and precision 82.91%. Statistical analysis confirmed significant improvements over other methods. [Conclusions] Compared to mainstream methods that directly incorporate unlabeled data into training, the Multi-Student model, based on multi-source semi-supervised learning, can better enhance the accuracy of semi-supervised tooth segmentation models.

Key words: artificial intelligence, medical imaging, transfer learning, deep learning, tooth segmentation

贾子昂. 基于多源半监督学习的牙齿结构分割[J]. 数据与计算发展前沿, 2025, 7(2): 175-185.

JIA Ziang. Teeth Structure Segmentation Based on Multi-Source Semi-Supervised Learning[J]. Frontiers of Data and Computing, 2025, 7(2): 175-185, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2025.02.017.

图/表 11

图1

表1

图2

表2

CBAM-Unet基础模型设定"

Layer	Output Shape	Operation Type	Configuration
Input	(C, H, W)	Input Image	-
enc1	(16, H, W)	2×Conv+BN+ReLU	$C i n = 3, C o u t = 16$
enc2	(32, H/2, W/2)	2×Conv+BN+ReLU	$C i n = 16, C o u t = 32$
enc3	(64, H/4, W/4)	2×Conv+BN+ReLU	$C i n = 32, C o u t = 64$
enc4	(128, H/8, W/8)	2×Conv+BN+ReLU	$C i n = 64, C o u t = 128$
center	(256, H/16,W/16)	2×Conv+BN+ReLU	$C i n = 128, C o u t = 256$
dec4	(128, H/8, W/8)	2×Conv+BN+ReLU	$C i n = 256, C o u t = 128$
dec3	(64, H/4, W/4)	2×Conv+BN+ReLU	$C i n = 128, C o u t = 64$
dec2	(32, H/2, W/2)	2×Conv+BN+ReLU	$C i n = 64, C o u t = 32$
dec1	(16, H, W)	2×Conv+BN+ReLU	$C i n = 32, C o u t = 16$
Output	(8, H, W)	1×Conv	$C i n = 16, C o u t = 8$

表2

图3

表3

图4

图5

表4

图6

表5

参考文献 30

[1]	RAJPURKAR P, LUNGREN M P. The Current and Future State of AI Interpretation of Medical Images[J]. New England Journal of Medicine, 2023, 388(21): 1981-1990.
[2]	BRADY A. The vanishing radiologist—an unse-en danger, and a danger of being unseen[J]. Eu-ropean Radiology, 2021, 31(8): 5998-6000.
[3]	HOLM P, VIGILD M, NITSCHKE I, et al. De-ntal care for aging populations in Denmark, Sw-eden, Norway, United Kingdom, and Germany[J]. Journal of Dental Education, 2005, 69(9): 987-997.
[4]	GAO H, CHAE O. Individual tooth segmentatio-n from CT images using level set method with shape and intensity prior[J]. Pattern Recognition, 2010, 43(7): 2406-2417.
[5]	GAN Y, XIA Z, XIONG J, et al. Toward accu-ratee tooth segmentation from computed tomography images using a hybrid level set model[J]. Medical Physics, 2015, 42(1): 14-27.
[6]	KEYHANINEJAD S, ZOROOFI A, SETAREH-DAN K, et al. Automated segmentation of teeth in multi-slice CT images[C]. IET International Conference on Visual Information Engineering, St-evenage, Herts, UK. 2006: 339-344.
[7]	LECUN Y, BENGIO Y, HINTON G. Deep lear-ning[J]. Nature, 2015, 521(7553): 436-444.
[8]	LECUN Y, BOTTOU L, BENGIO Y, et al. Gr-adient-based learning applied to document recog-nition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.
[9]	CHO K, MERRIËNBOER B, GULCEHRE C, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translate-on[J]. arXiv Preprint arXiv:1406.1078, 2014.
[10]	RONNEBERGER O, FISCHER P, BROX T. U-net: Convolutional networks for biomedical ima-ge segmentation[C]// Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015, Munich, DE. 2015: 234-241.
[11]	ÇIÇEK Ö, ABDULKADIR A, LIENKAMP S, et al. 3D U-Net: learning dense volumetric segm-entation from sparse annotation[C]// Medical Im-age Computing and Computer-Assisted Intervent-ion-MICCAI 2016. Athens, GR, 2016: 424-432.
[12]	MILLETARI F, NAVAB N, AHMADI A. V-net: Fully convolutional neural networks for volumet-ric medical image segmentation[C]// 2016 Fourth International Conference on 3D Vision, San Francisco, CA, USA. 2016: 565-571.
[13]	RAJPURKAR P, CHEN E, BANERJEE O, et al. AI in health and medicine[J]. Nature Medicine, 2022, 28(1): 31-38.
[14]	SUN M, LI K, QI X, et al. Contextual informa-tion enhanced convolutional neural networks for retinal vessel segmentation in color fundus ima-ges[J]. Journal of Visual Communication and I-mage Representation, 2021, 77: 103134.
[15]	HUANG G, LIU Z, MAATEN L, et al. Densely connected convolutional networks[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Hawaii, HI, USA, 2017: 4700-4708.
[16]	SUN M, ZHOU W, QI X, et al. Prediction of BAP1 expression in uveal melanoma using dens-ely-connected deep classification networks[J]. Cancers, 2019, 11(10): 1579.
[17]	TARVAINEN A, VALPOLA H. Mean teachers a-re better role models: Weight-averaged consiste-ncy tar-gets improve semi-supervised deep learn-ing results[J]. Advances in Neural Information Processing Systems, 2017, 30.
[18]	JIAO R, ZHANG Y, DING L, et al. Learning with limited annotations: a survey on deep semi-supervised learning for medical image segment-ation[J]. Computers in Biology and Medicine, 2024, 169: 107840.
[19]	LI D, YANG J, KREIS K, et al. Semantic seg-mentation with generative models: Semi-supervis-ed learning and strong out-of-domain generalizat-ion[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021: 8300-8311.
[20]	XU D, WANG Z. Semi-supervised semantic seg-mentation using an improved generative adversa-rial network[J]. Journal of Intelligent & Fuzzy S-ystems, 2021, 40(5): 9709-9719.
[21]	MITTAL S, TATARCHENKO M, BROX T. Se-mi-supervised semantic segmentation with high a-nd low-level consistency[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 43(4): 1369-1379.
[22]	YU L, WANG S, LI X, et al. Uncertainty-aware self-ensembling model for semi-supervised 3D l-eft atrium segmentation[C]// Medical Image Co-mputing and Computer Assisted Intervention-MICCAI 2019, Shenzhen, 2019: 605-613.
[23]	OUALI Y, HUDELOT C, TAMI M. Semi-super-vised semantic segmentation with cross-consiste-ncy training[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re-cognition, Seattle, WA, USA, 2020: 12674-12684.
[24]	WANG G, YE C, MAN B. Deep learning for t-omographic image reconstruction[J]. Nature Mac-hine Intelligence, 2020, 2(12): 737-748.
[25]	CUI Z, FANG Y, MEI L, et al. A fully autom-atic AI system for tooth and alveolar bone seg-mentation from cone-beam CT images[J]. Nature Communications, 2022, 13(1): 2096.
[26]	LI K, TAN M, XIAO D, et al. Research on ro-ad extraction from high-resolution remote sensin-g images based on improved UNet++[J]. IEEE Access, 2024, 12: 50300-50309.
[27]	HAO J, ZHU Y, HE L, et al. T-Mamba: A uni-fied framework with Long-Range Dependency in dual-domain for 2D & 3D Tooth Segmentation[J]. arXiv preprint arXiv:2404.01065, 2024.
[28]	CHEN S, ZHUO J, LI X, et al. CMT: Co-train-ing Mean-Teacher for Unsupervised Domain Ad-aptation on 3D Object Detection[C]// Proceedings of the 32nd ACM International Conference on Multimedia, Melbourne, VIC, AUS, 2024: 4738-4747.
[29]	LI X, YU L, CHEN H, et al. Transformation c-onsistent self-ensembling model for semisupervis-ed medical image segmentation[J]. IEEE Transa-ctions on Neural Networks and Learning Systems, 2020, 32(2): 523-534.
[30]	MENDEL R, RAUBER D, SOUZA L, et al. E-rror-correcting mean-teacher: Corrections instead of consistency-targets applied to semi-supervised medical image segmentation[J]. Computers in B-iology and Medicine, 2023, 154(C): 106585.

	已标注数据	未标注数据
X1	150	0
X2	150	330
X3	0	330

	Dice(%)	Jaccard(%)	Recall(%)	Precision(%)
CCT^[28]	80.16	82.54	80.28	82.43
TCSM^[29]	80.10	81.29	81.04	80.96
ECMT^[30]	80.97	82.34	82.10	81.63
Multi-Student (Ours)	81.13	82.96	82.78	82.91

	Dice(%)	Jaccard(%)	Recall(%)	Precision(%)
Mean-Teacher	75.34	74.61	73.17	73.90
Multi-Student+MSE	79.84	79.19	78.33	80.87
Multi-Student (Ours)	81.13	82.96	82.78	82.91