基于深度学习对CBCT图像采用多视角投影的牙齿分类  

Classification of teeth in CBCT images using deep learning with multi-view projection

作　　者：刘沐然 谭敏慧 张煜[1] LIU Muran;TAN Minhui;ZHANG Yu(School of Biomedical Engineering,Southern Medical University,Guangzhou 510515,China;School of Biomedical Engineering,ShanghaiTech University,Shanghai 201210,China)

机构地区：[1]南方医科大学生物医学工程学院,广东广州510515 [2]上海科技大学生物医学工程学院,上海201210

出　　处：《中国医学物理学杂志》2025年第3期313-319,共7页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(82472056);广东省自然科学基金(2024A1515012004)。

摘　　要：目的:针对当前基于锥形束计算机断层扫描(CBCT)图像的牙齿分类方法过于依赖精确分割,缺少对牙齿形态和位置信息的综合利用,提出一种基于多视角投影和Transformer架构的牙齿分类方法,可对全年龄层的CBCT图像中的牙齿(包括儿童病例)进行准确的52分类。方法:通过引入多视角投影,结合Transformer架构,融合语义分割和实例分割,由粗至细进行牙齿分类任务,增强对牙齿空间位置信息的利用。采用国际牙科联盟(FDI)两位数牙位标记法对CBCT图像中的牙齿进行分类,并对多视角融合效果进行评估。结果:改进后的方法能够准确区分恒牙与乳牙,同时有效地进行牙齿编号,牙齿层面的分类准确率达到0.982。结论:基于多视角投影与Transformer架构的牙齿分类方法实现对牙齿类别与位置信息的有效融合,提高牙齿分类的精度,为个性化治疗方案的制定提供更为精确的基础。Objective To address the issue that current methods for classifying teeth in cone beam computed tomography(CBCT) images overly rely on precise segmentation and lack utilization of tooth morphology and positional information,a tooth classification method based on multi-view projection and Transformer architecture is proposed for accurately classifying teeth from CBCT images across all age groups,including pediatric cases,into 52 categories.Methods The coarseto-fine tooth classification task was accomplished after enhancing the utilization of spatial positional information of the teeth by incorporating multi-view projection,integrating Transformer architecture,and combining semantic segmentation with instance segmentation.The two-digit notation system of the Federation Dentaire Internationale was adopted for classifying the teeth in CBCT images,and evaluating the effectiveness of multi-view fusion.Results The improved method effectively classified and numbered both permanent and deciduous teeth,achieving a tooth-level classification accuracy of 0.982.Conclusion The tooth classification method based on multi-view projection and Transformer architecture successfully integrates tooth category and positional information,and improves the accuracies of tooth classification and numbering,providing a more precise foundation for the formulation of personalized treatment schemes.

关 键 词：深度学习 锥形束计算机断层扫描 牙齿识别 多视角投影 

分 类 号：R318[医药卫生—生物医学工程]