Department of Operative and Preventive Dentistry, Charité - Universitätsmedizin Berlin, Germany.
Department of Operative and Preventive Dentistry, Charité - Universitätsmedizin Berlin, Germany; Department of Orthodontics, Dentofacial Orthopedics and Pedodontics, Charité - Universitätsmedizin Berlin, Germany.
J Dent. 2020 Jan;92:103260. doi: 10.1016/j.jdent.2019.103260. Epub 2019 Dec 9.
In this pilot study, we applied deep convolutional neural networks (CNNs) to detect caries lesions in Near-Infrared-Light Transillumination (NILT) images.
226 extracted posterior permanent human teeth (113 premolars, 113 molars) were allocated to groups of 2 + 2 teeth, and mounted in a pilot-tested diagnostic model in a dummy head. NILT images of single-tooth-segments were generated using DIAGNOcam (KaVo, Biberach). For each segment (on average 435 × 407 × 3 pixels), occlusal and/or proximal caries lesions were annotated by two experienced dentists using an in-house developed digital annotation tool. The pixel-based annotations were translated into binary class levels. We trained two state-of-the-art CNNs (Resnet18, Resnext50) and validated them via 10-fold cross validation. During the training process, we applied data augmentation (random resizing, rotations and flipping) and one-cycle-learning rate policy, setting the minimum and maximum learning rates to 10 and 10, respectively. Metrics for model performance were the area-under-the-receiver-operating-characteristics-curve (AUC), sensitivity, specificity, and positive/negative predictive values (PPV/NPV). Feature visualization was additionally applied to assess if the CNNs built on features dentists would also use.
The tooth-level prevalence of caries lesions was 41%. The two models performed similar on predicting caries on tooth segments of NILT images. The marginal better model with respect to AUC was Resnext50, where we retrained the last 9 network layers, using the Adam optimizer, a learning rate of 0.5 × 10, and a batch size of 10. The mean (95% CI) AUC was 0.74 (0.66-0.82). Sensitivity and specificity were 0.59 (0.47-0.70) and 0.76 (0.68-0.84) respectively. The resulting PPV was 0.63 (0.51-0.74), the NPV 0.73 (0.65-0.80). Visual inspection of model predictions found the model to be sensitive to areas affected by caries lesions.
A moderately deep CNN trained on a limited amount of NILT image data showed satisfying discriminatory ability to detect caries lesions.
CNNs may be useful to assist NILT-based caries detection. This could be especially relevant in non-conventional dental settings, like schools, care homes or rural outpost centers.
本初步研究旨在应用深度卷积神经网络（CNN）检测近红外光透射（NILT）图像中的龋损。
将 226 颗提取的后恒牙（113 颗前磨牙，113 颗磨牙）分为两组，每组 2 颗+2 颗，安装在模拟头中的经过初步测试的诊断模型中。使用 DIAGNOcam（KaVo，Biberach）生成单牙段的 NILT 图像。对于每个片段（平均 435×407×3 像素），由两名经验丰富的牙医使用内部开发的数字注释工具对咬合面和/或近中面的龋损进行注释。基于像素的注释被转换为二进制类别水平。我们训练了两个最先进的 CNN（Resnet18、Resnext50），并通过 10 折交叉验证进行验证。在训练过程中，我们应用了数据增强（随机调整大小、旋转和翻转）和一个周期学习率策略，将最小和最大学习率分别设置为 10 和 10。模型性能的指标为受试者工作特征曲线下的面积（AUC）、敏感性、特异性、阳性/阴性预测值（PPV/NPV）。此外，还应用了特征可视化来评估 CNN 是否基于牙医也会使用的特征。
牙齿水平的龋病患病率为 41%。这两个模型在预测 NILT 图像中牙齿龋病方面表现相似。在 AUC 方面略有优势的模型是 Resnext50，我们重新训练了最后 9 个网络层，使用 Adam 优化器，学习率为 0.5×10，批量大小为 10。平均（95%CI）AUC 为 0.74（0.66-0.82）。敏感性和特异性分别为 0.59（0.47-0.70）和 0.76（0.68-0.84）。由此产生的 PPV 为 0.63（0.51-0.74），NPV 为 0.73（0.65-0.80）。对模型预测的直观检查发现，该模型对受龋病影响的区域很敏感。
在有限数量的 NILT 图像数据上训练的中等深度 CNN 显示出令人满意的区分能力，可以检测龋病病变。
CNN 可用于辅助基于 NILT 的龋病检测。在非传统牙科环境中，如学校、养老院或农村哨所中心，这可能特别重要。
