Ye Ye, Fang Shuobo, Lu Huitong, Liu Mingqian, Wu Xueying
Department of Prosthodontics, Shanghai Stomatological Hospital and School of Stomatology, Fudan University. Shanghai 200001, China. E-mail:
Shanghai Kou Qiang Yi Xue. 2025 Apr;34(2):119-125.
To train the U-net of convolutional neural network to establish a method for detecting and segmenting the mandibular canal and its bifurcation, and validate its accuracy based on the ground truth labeled by experts.
A total of 290 CBCT scans were collected from Shanghai Stomatological Hospital from Jan. 2022 to Dec. 2022, which were divided into training set of 200 scans and test set of 90 scans. Model training included two steps. In the first step, bilateral mandibular canals and its bifurcation of 50 CBCT scans were labeled in 3D Slicer image computing platform by investigators. Three dimensional U-net segmentation model were trained initially with data enhancement. A morphological post-processing method was applied to the predicted results. In the second step, pseudo label method was employed to help annotating the mandibular canal and corresponding bifurcations on remaining 150 CBCTs, which would be included in training set after revision. Three dimensional U-net model was trained based on these 200 data. During test phase, totally 90 scans were labeled by two doctors and U-net model respectively. Consistency check was conducted to evaluate the labels between two doctors. Dice similarity coefficient and Hausdorff distance were calculated to evaluate the labels between doctors and the model. The detection rate of bifurcation was calculated. SPSS 20.0 software package was used for data analysis.
In 90 CBCT test set, the Kappa value between two dentists' annotations was 0.667. The average Dice and Hausdorff distance between predictions and labels of doctors were (0.739±0.068) and (0.988±1.14) mm. In bifurcation detection, the detection rate was 91.30% on scans with clear bifurcations.
The dentification and segmentation U-net model of mandibular canal on dental CBCT can be reliable and practical for its high segmentation precision and predicting speed.
训练卷积神经网络的U-net，建立下颌管及其分支的检测与分割方法，并基于专家标注的真值验证其准确性。
收集2022年1月至2022年12月上海口腔医院的290例CBCT扫描数据，分为200例扫描的训练集和90例扫描的测试集。模型训练包括两个步骤。第一步，研究人员在3D Slicer图像计算平台上对50例CBCT扫描的双侧下颌管及其分支进行标注。首先对三维U-net分割模型进行数据增强训练。对预测结果应用形态学后处理方法。第二步，采用伪标签法辅助对其余150例CBCT上的下颌管及相应分支进行标注，修订后纳入训练集。基于这200个数据训练三维U-net模型。在测试阶段，分别由两名医生和U-net模型对90例扫描进行标注。进行一致性检查以评估两名医生之间的标注。计算Dice相似系数和豪斯多夫距离以评估医生与模型之间的标注。计算分支的检出率。采用SPSS 20.0软件包进行数据分析。
在90例CBCT测试集中，两名牙医标注之间的Kappa值为0.667。预测与医生标注之间的平均Dice和豪斯多夫距离分别为(0.739±0.068)和(0.988±1.14)mm。在分支检测中，在分支清晰的扫描上检出率为91.30%。
牙科CBCT下颌管识别与分割U-net模型分割精度高、预测速度快，可靠实用。
