1 Academy of Medical Engineering and Translational Medicine, Tianjin University, Tianjin 300354, China
2 School of Computer and Network Security, Chengdu University of Technology, Chengdu 610000, China
3 School of Mechanical Engineering, Anhui University of Technology, Maanshan 243002, China
Show less
文章历史+
Received
Published
Revised
2025-08-30
2026-08-28
2025-12-16
Online Date
2026-09-04
PDF (10264K)
摘要
3D牙颌数据中的牙齿准确分割对于诊断、治疗计划至关重要,针对已有算法在牙齿边界不清晰、拥挤、错位和缺失等情况下分类分割精度低、高显存消耗的痛点,提出一种采用注意力机制增强特征提取的二阶段牙齿分类分割模型。由于3D点云模型的输入较大,导致计算量较大,为降低显存消耗,定位牙齿采用2D图像进行检测,将3D点云投影成2D图像,针对现有方法在缺牙错位时分类准确率低的问题,在现有算法作为主干网络基础上,增加了由Transformer构建的牙齿特征增强损失模块(feature enhancement loss module,FELM)强化各个牙齿区域数据点之间的相关性学习,以提高检测的分类精度。第二阶段,需要对检测到 2D 牙齿区域之后裁切得到3D牙颌模型上对应的单个牙齿点云数据进行分割。针对 EdgeConv只能学习局部特征的问题,增加了基于注意力机制的全局通道特征关联(global channel relation feature,GCRF)模块,增强了对全局特征进行融合学习,得到EdgeConv-GCRF分割模型,对单颗牙齿进行分割。通过在 Teeth3DS 数据集实验,所提方法得到的平均准确率(mean accuracy,mAcc)、平均交并比(mean intersection over union,mIOU)和平均戴斯值(mean Dice,mDice)分别为0.860 8、0.824 9、0.866 6,相比现有其他方法,显著提升了分类分割精度,能够高效完成分割任务,极具经济及社会价值。
Abstract
Accurate tooth segmentation in 3D dental jaw data is crucial for diagnosis and treatment planning. Aiming at the pain points of existing algorithms, such as low classification and segmentation accuracy and high video memory consumption under conditions of unclear tooth boundaries, crowding, misalignment, and tooth loss, a two-stage tooth classification and segmentation model with an attention mechanism to enhance feature extraction was proposed Due to the large input of 3D point cloud models, which led to a large amount of computation, to reduce video memory consumption, teeth were located by detecting 2D images and projects 3D point clouds into 2D images. To address the problem of low classification accuracy of existing methods in cases of tooth loss and misalignment, a tooth feature enhancement loss module (FELM) constructed by Transformer was added on the basis of the existing algorithm as the backbone network to strengthen the correlation learning between data points in each tooth region, so as to improved the classification accuracy of detection. In the second stage, after detecting the 2D tooth region, the corresponding single-tooth point cloud data on the cropped 3D dental jaw model needed to be segmented. To address the problem that EdgeConv can only learn local features, a global channel relation feature (GCRF) module based on the attention mechanism was added to enhance the fusion learning of global features, and an EdgeConv-GCRF segmentation model was obtained to segment a single tooth. Through experiments on the Teeth3DS dataset, the proposed method achieves the mean accuracy (mAcc), mean intersection over union (mIOU), and mean Dice score (mDice) of 0.860 8, 0.824 9, and 0.866 6, respectively. Compared with other existing methods, it significantly improves the classification and segmentation accuracy, can efficiently complete the segmentation task, and has great economic and social value.
DAI Chao, YAO Guang-le, XU Min-peng, WANG Yang.
High-precision 3D Tooth Classification and Segmentation Method with Feature Enhancement[J].
Science Technology and Engineering, 2026, 26(24): 10506-10515 DOI:10.12404/j.issn.1671-1815.2506283
针对3D牙颌模型点云数据量大导致3D检测模型输入规模过大、计算复杂度高,进而难以部署的问题,提出一种两阶段3D牙齿分割方法,该方法由牙齿检测模块和单颗牙齿局部点云分割模块构成,整体框架如图2所示。牙齿检测模块在2D图上进行检测,大大降低了计算量和显存消耗,包含基础检测模块YOLOv5x和特征增强损失模块(feature enhancement loss module,FELM),得到YOLOv5x-FELM,其输入为3D数据投影到2D平面上融合了光照图和深度图的图片。图2中,黄色区域为YOLOv5x主干网络,灰色区域为FELM模块。通过2D牙齿检测模型可得到每颗牙齿对应的2D检测结果框,
QiC R, SuH, MoK, et al. Pointnet: deep learning on point sets for 3d classification and segmentation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 652-660.
[2]
QiC R, YiL, SuH, et al. Pointnet++: deep hierarchical feature learning on point sets in a metric space[J]. Advances in Neural Information Processing Systems, 2017, 30: 5099-5108.
[3]
WangY, SunY, LiuZ, et al. Dynamic graph CNN for learning on point clouds[J]. ACM Transactions on Graphics, 2019, 38(5): 1-12.
[4]
QiC R, LitanyO, HeK, et al. Deep hough voting for 3D object detection in point clouds[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019: 9277-9286.
[5]
MaX, QinC, YouH, et al. Rethinking network design and local geometry in point cloud: a simple residual MLP framework[J]. arXiv Preprint, arXiv: 2202.07123, 2022.
[6]
ZhaoH, JiangL, JiaJ, et al. Point Transformer[C]// Proceedings of the IEEE/CVF international Conference on Computer Vision. Montreal: IEEE, 2021: 16259-16268.
[7]
LaiX, LiuJ, JiangL, et al. Stratified Transformer for 3D point cloud segmentation[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans: IEEE, 2022: 8500-8509.
[8]
LianC, WangL, WuT H, et al. Deep multi-scale mesh feature learning for automated labeling of raw dental surfaces from 3D intraoral scanners[J]. IEEE Transactions on Medical Imaging, 2020, 39(7): 2440-2450.
[9]
ZhaoY, ZhangL, YangC, et al. 3D Dental model segmentation with graph attentional convolution network[J]. Pattern Recognition Letters, 2021, 152: 79-85.
[10]
JanaA, SubhashH M, MetaxasD. 3D tooth mesh segmentation with simplified mesh cell representation[C]// IEEE 20th International Symposium on Biomedical Imaging (ISBI). Cartagena de Indias: IEEE, 2023: 1-5.
[11]
AlmalkiA, LateckiL J. Self-supervised learning with masked autoencoders for teeth segmentation from intra-oral 3D scans[C]// Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. Seattle: IEEE, 2024: 7820-7830.
[12]
LiJ, ChengB, NiuN, et al. A fine-grained orthodontics segmentation model for 3D intraoral scan data[J]. Computers in Biology and Medicine, 2024, 168: 107821.
[13]
XiongH, LiK, TanK, et al. TFormer: 3D tooth segmentation in mesh scans with geometry guided transformer[J]. arXiv Preprint, arXiv: 2210.16627, 2022.
[14]
ZhangL, ZhaoY, MengD, et al. TSGCNet: discriminative geometric feature learning with two-stream graph convolutional network for 3D dental model segmentation[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Nashville: IEEE, 2021: 6699-6708.
[15]
KrenmayrL, von SchwerinR, SchaudtD, et al. DilatedToothSegNet: tooth segmentation network on 3D dental meshes through increasing receptive vision[J]. Journal of Imaging Informatics in Medicine, 2024, 37: 1846-1862.
[16]
DuanF, ChenL. 3D dental mesh segmentation using semantics-based feature learning with graph-Transformer[C]// International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer, 2023: 456-465.
[17]
LinZ, HeZ, WangX, et al. DBGANet: dual-branch geometric attention network for accurate 3D tooth segmentation[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2023, 33: 5876-5887.
[18]
ZhangJ, LiC, SongQ, et al. Automatic 3D tooth segmentation using convolutional neural networks in harmonic parameter space[J]. Graphical Models, 2020, 109: 101071.
[19]
QiuL, YeC, ChenP, et al. Darch: dental arch prior-assisted 3D tooth instance segmentation with weak annotations[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans: IEEE, 2022: 20752-20761.
[20]
CuiZ, LiC, ChenN, et al. TSegNet: an efficient and accurate tooth segmentation network on 3D dental model[J]. Medical Image Analysis, 2021, 69: 101949.
[21]
Ben-HamadouA, SmaouiO, RekikA, et al. 3DTeethSeg’22: 3D teeth scan segmentation and labeling challenge[J]. arXiv Preprint, arXiv: 2305.18277, 2023.
JinZhixin, SunHaohang, LiLin, et al. Motion object recognition in complex scenes based on multi-attention mechanism fusion[J]. Science Technology and Engineering, 2025, 25(29): 12570-12578.
ZhaiYongjie, ZhuHaoshuo, LiJiawei, et al. VDD-YOLO: method for vehicle deformation and damage detection[J]. Science Technology and Engineering, 2025, 25(29): 12560-12569.
[28]
DosovitskiyA. An image is worth 16×16 words: Transformers for image recognition at scale[J]. arXiv Preprint, arXiv: 2010.11929, 2020.
ChengT, SongL, GeY, et al. YOLO-world: real-time open-vocabulary object detection[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 16901-16911.
[31]
ZhaoY, LvW, XuS, et al. DETRs beat YOLOs on real-time object detection[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 16965-16974.
[32]
ZhangH, LiF, LiuS, et al. DINO: DETR with improved denoising anchor boxes for end-to-end object detection[J]. arXiv Preprint, arXiv: 2203.03605, 2022.
[33]
MuzahidA A M, WanW, SohelF, et al. CurveNet: curvature-based multitask learning deep networks for 3D object recognition[C]// IEEE/CAA Journal of Automatica Sinica. New York: 2020: 1177-1187.
[34]
QianG C, LiY, PengH, et al. PointNeXt: revisiting PointNet++ with improved training and scaling strategies[C]// Advances in Neural Information Processing Systems. New Orleans: Curran Associates, Inc., 2022: 35.