深度学习在锥形束CT下颌第二磨牙C形根管分割与分类中的应用
本研究开发深度学习模型,基于锥形束CT图像对下颌第二磨牙C形根管解剖结构进行分割与分类,并比较U-Net、残差U-Net和Xception U-Net三种架构的性能。训练验证集100例,测试集35例。Xception U-Net的Dice系数和敏感度最高,且对比度受限自适应直方图均衡化平均提升效能4.6%。
要点
- Xception U-Net在测试集上Dice系数达0.768,优于残差U-Net和U-Net。
- Xception U-Net预测C形亚类平均敏感度为0.786。
- 对比度受限自适应直方图均衡化平均提升模型整体效能4.6%。
核心结论
- 深度学习可能有助于C形根管解剖结构的检测和分类。
研究设计与样本
- 研究使用135例有限视野锥形束CT图像,其中100例用于训练和验证,35例用于测试。
- 比较U-Net、残差U-Net和Xception U-Net三种架构进行图像分割和C形解剖分类。
- 使用Dice指数评估自动标注的体素匹配精度,计算预测正确C形亚类的平均敏感度。
主要发现
- 测试集平均Dice系数:Xception U-Net为0.768 ± 0.0349,残差U-Net为0.736 ± 0.0297,U-Net为0.660 ± 0.0354。
- 三模型整体性能差异有统计学意义(方差分析,P = .000779)。
- Xception U-Net(Q = 7.23,P = .00070)和残差U-Net(Q = 5.09,P = .00951)均显著优于U-Net(事后Tukey HSD检验)。
- 平均敏感度:Xception U-Net为0.786 ± 0.0378,残差U-Net为0.746 ± 0.0391,U-Net为0.720 ± 0.0495。
- 平均阳性预测值:U-Net为77.6% ± 0.1998%,残差U-Net为78.2% ± 0.1971%,Xception U-Net为80.0% ± 0.1098%。
关键数值与推荐
- 对比度受限自适应直方图均衡化的加入使整体架构效能平均提升4.6%(P < .0001)。
临床怎么用
- 深度学习模型可辅助从锥形束CT中识别和分类下颌第二磨牙C形根管解剖,其中Xception U-Net表现最佳。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
INTRODUCTION: The identification of C-shaped root canal anatomy on radiographic images affects clinical decision making and treatment. The aims of this study were to develop a deep learning (DL) model to classify C-shaped canal anatomy in mandibular second molars from cone-beam computed tomographic (CBCT) volumes and to compare the performance of 3 different architectures. METHODS: U-Net, residual U-Net, and Xception U-Net architectures were used for image segmentation and classification of C-shaped anatomies. Model training and validation were performed on 100 of a total of 135 available limited field of view CBCT images containing mandibular molars with C-shaped anatomy. Thirty-five CBCT images were used for testing. Voxel-matching accuracy of the automated labeling of the C-shaped anatomy was assessed with the Dice index. The mean sensitivity of predicting the correct C-shape subcategory was calculated based on detection accuracy. One-way analysis of variance and post hoc Tukey honestly significant difference tests were used for statistical evaluation. RESULTS: The mean Dice coefficients were 0.768 ± 0.0349 for Xception U-Net, 0.736 ± 0.0297 for residual U-Net, and 0.660 ± 0.0354 for U-Net on the test data set. The performance of the 3 models was significantly different overall (analysis of variance, P = .000779). Both Xception U-Net (Q = 7.23, P = .00070) and residual U-Net (Q = 5.09, P = .00951) performed significantly better than U-Net (post hoc Tukey honestly significant difference test). The mean sensitivity values were 0.786 ± 0.0378 for Xception U-Net, 0.746 ± 0.0391 for residual U-Net, and 0.720 ± 0.0495 for U-Net. The mean positive predictive values were 77.6% ± 0.1998% for U-Net, 78.2% ± 0.0.1971% for residual U-Net, and 80.0% ± 0.1098% for Xception U-Net. The addition of contrast-limited adaptive histogram equalization had improved overall architecture efficacy by a mean of 4.6% (P < .0001). CONCLUSIONS: DL may aid in the detection and classification of C-shaped canal anatomy.
引用
- Sherwood AA, Sherwood AI, Setzer FC, et al. A Deep Learning Approach to Segment and Classify C-Shaped Canal Morphologies in Mandibular Second Molars Using Cone-beam Computed Tomography. J Endod. 2021. PMID: 34563507. doi:10.1016/j.joen.2021.09.009
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。