背景:这项研究旨在评估使用深度卷积神经网络(DCNN)算法来检测临床特征并预测术前根尖周X线片上牙髓治疗的三年结果。
方法:准备了一个单根前磨牙数据库,该数据库接受了牙髓治疗或由牙髓医师进行的再治疗,具有三年的结果(n=598)。我们构建了一个具有自我注意层的17层DCNN(PRESSAN-17),训练了模型,已验证,并测试1)检测七个临床特征,即,全覆盖恢复(FCR),近端牙齿(PRX)的存在,日冕缺损(COD),根休息(RRS),运河能见度(CAV),先前的根填充(PRF),和根尖周射线可透性(PAR),和2)通过分析术前根尖X线片作为输入来预测三年的牙髓预后。在预测测试中,对没有自我注意层的常规DCNN(RESNET-18)进行了比较测试。主要评估接收器工作特征(ROC)曲线(AUC)下的准确性和面积,以进行性能比较。使用梯度加权类别激活图(Grad-CAM)可视化加权热图。
结果:PRESSAN-17检测到FCR(AUC=0.975),PRX(0.866),COD(0.672),RRS(0.989),PRF(0.879)和PAR(0.690)显著,与无信息率相比(p<0.05)。比较两个模型的5倍验证的平均准确性,PRESSAN-17(67.0%)与RESNET-18(63.4%,p<0.05)。此外,PRESSAN-17的平均ROC下面积为0.638,与无信息率相比有显著差异.Grad-CAM证明PRESSAN-17正确识别了临床特征。
结论:深度卷积神经网络可能有助于牙髓治疗结果的预测。
BACKGROUND: This study aimed to evaluate the use of deep convolutional neural network (DCNN) algorithms to detect clinical features and predict the three-year outcome of endodontic treatment on preoperative periapical radiographs.
METHODS: A database of single-root premolars that received endodontic treatment or retreatment by endodontists with presence of three-year outcome was prepared (n = 598). We constructed a 17-layered DCNN with a self-attention layer (Periapical Radiograph Explanatory System with Self-Attention Network [PRESSAN-17]), and the model was trained, validated, and tested to 1) detect 7 clinical features, that is, full coverage restoration, presence of proximal teeth, coronal defect, root rest, canal visibility, previous root filling, and periapical radiolucency and 2) predict the three-year endodontic prognosis by analyzing preoperative periapical radiographs as an input. During the prognostication test, a conventional DCNN without a self-attention layer (residual neural network [RESNET]-18) was tested for comparison. Accuracy and area under the receiver-operating-characteristic curve were mainly evaluated for performance comparison. Gradient-weighted class activation mapping was used to visualize weighted heatmaps.
RESULTS: PRESSAN-17 detected full coverage restoration (area under the receiver-operating-characteristic curve = 0.975), presence of proximal teeth (0.866), coronal defect (0.672), root rest (0.989), previous root filling (0.879), and periapical radiolucency (0.690) significantly, compared to the no-information rate (P < .05). Comparing the mean accuracy of 5-fold validation of 2 models, PRESSAN-17 (67.0%) showed a significant difference to RESNET-18 (63.4%, P < .05). Also, the area under average receiver-operating-characteristic of PRESSAN-17 was 0.638, which was significantly different compared to the no-information rate. Gradient-weighted class activation mapping demonstrated that PRESSAN-17 correctly identified clinical features.
CONCLUSIONS: Deep convolutional neural networks can detect several clinical features in periapical radiographs accurately. Based on our findings, well-developed artificial intelligence can support clinical decisions related to endodontic treatments in dentists.