【作者】 邬微微； 陈盼； 陈苏蓉； 周耕宇； 高原； 胡靖宇； 马净植；

【Author】 Weiwei Wu;Yuan Gao;Jingyu Hu;Jingzhi Ma;School of Stomatology, Tongji Medical College, Huazhong University of Science and Technology;Department of Stomatology, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology;Hubei Province Key Laboratory of Oral and Maxillofacial Development and Regeneration;Department of Cariology and Endodontics, West China Hospital of Stomatology, Sichuan University;

【机构】 华中科技大学同济医学院附属同济医院； 华中科技大学同济医学院口腔医学院； 口腔颌面发育与再生湖北省重点实验室； 四川大学华西口腔医院牙体牙髓病科；

【摘要】 【目的】利用显微计算机断层扫描技术(micro-CT)得到离体牙X线投影图,结合卷积神经网络(CNN)及迁移学习思路,提高从模拟临床的真实X线片中识别下颌第二磨牙融合根根管形态的准确率。【方法】本研究共纳入243颗融合根型下颌第二磨牙,获取micro-CT数据并生成二维X线投影图,在离体下颌骨上拍摄真实X线片模拟临床场景。根据micro-CT将所有根管分为融合型、对称型和不对称型,采用多角度投影图对数目较少的类别进行扩增。选择ResNet18、ResNet50、VGG19和EfficientNet-b5四个CNN进行根管形态识别。按照ImageNet预训练权重+X线临床模拟片,ImageNet预训练权重+X线投影图,和X线投影图权重+X线临床模拟片分组进行四个CNN的训练,并将CNN的分类结果与住院医师进行比较。【结果】ImageNet预训练权重+X线临床模拟片组的CNN在测试集上的总体准确率为65.96%,63.83%,61.70%和65.96%;ImageNet预训练权重+X线投影图组为79.25%,75.47%,73.58%和77.36%;X线投影图权重+X线临床模拟片组为72.34%,68.09%,65.96%和72.34%。住院医师对X线临床模拟片的分类结果为63.83%,57.45%和63.83%。采用ImageNet预训练权重的CNN在X线临床模拟片上对根管形态的识别能力与住院医师接近(P>0.05),结合X线投影图权重后,其识别能力有所提高(P<0.05)。【结论】在micro-CT生成的X线投影图上训练得到的权重参数有助于提高X线临床模拟片训练过程中的形态识别准确率。将X线投影图作为三维micro-CT数据和临床二维X线数据之间的桥梁,以联合不同模态影像的这一思路具有一定的可行性。更多还原

【Abstract】 Objectives: This study aims to improve the accuracy of identifying the root canal morphology of fused-rooted mandibular second molars(MSMs) from clinical radiographs by utilizing X-ray projection images obtained from micro-computed tomography(micro-CT) and integrating transfer learning with convolutional neural networks(CNNs).Methods: This study included a total of 243 fused-rooted MSMs. Micro-CT data were acquired, and two-dimensional X-ray projection images were generated. Additionally,real X-ray films of the teeth were obtained from isolated mandibles to simulate clinical scenarios. All root canals were classified into merging, symmetrical, and asymmetrical types based on the micro-CT image, and multi-angle projection methods were employed to enhance the categorization with smaller sizes. Four CNNs,ResNet18, ResNet50, VGG19, and EfficientNet-b5, were selected for root canal morphology identification. The training process was organized according to three configurations: ImageNet pretraining weights combined with X-ray clinical simulation films, ImageNet pretraining weights combined with X-ray projection images, and X-ray projection image weights combined with X-ray clinical simulation films. The classification results of the CNNs were then compared with those of endodontic residents.Results: The overall accuracies of the CNNs on the test set were 65.96%, 63.83%,61.70%, and 65.96% for the group of ImageNet pretrained weights combined with X-ray clinical simulation films; 79.25%, 75.47%, 73.58%, and 77.36% for the group of ImageNet pretrained weights combined with X-ray projection images; and 72.34%,68.09%, 65.96%, and 72.34% for the group of X-ray projection image weights combined with X-ray clinical simulation films. The classification performance of residents on X-ray clinical simulation films yielded results of 63.83%, 57.45%, and63.83%. The CNNs employing ImageNet pretrained weights demonstrated identification of root canal morphology on X-ray clinical simulation films that was comparable to that of residents(P>0.05), with performance improving when combined with X-ray projection image weights(P<0.05).Conclusions: The weighting parameters obtained from training on X-ray projection images generated by micro-CT improve the accuracy of morphology identification during the training phase with X-ray clinical simulation films. The approach of utilizing X-ray projection images as a bridge between three-dimensional micro-CT data and two-dimensional clinical X-ray data for uniting imaging from different modalities is feasible.更多还原