深度学习从X线图像识别融合根下颌第二磨牙根管形态
本研究纳入271颗融合根下颌第二磨牙,以显微CT三维根管形态为金标准,利用X线图像训练VGG19、ResNet18、ResNet50和EfficientNet-b5识别根管形态(融合型、对称型、不对称型)。ResNet18结合多角度投影增强表现最佳,总体准确率79.25%,且卷积神经网络分类表现优于牙髓科住院医师(平均准确率60.38%,P < .05)。
要点
- ResNet18结合多角度投影增强识别融合根下颌第二磨牙根管形态总体准确率79.25%。
- 卷积神经网络分类表现优于牙髓科住院医师(60.38%,P < .05)。
- 多角度投影增强在多数卷积神经网络中优于传统增强方法。
核心结论
- 卷积神经网络在识别融合根下颌第二磨牙三维根管形态方面比牙髓科住院医师更有效。
- 卷积神经网络有能力利用二维图像辅助三维诊断。
研究设计与样本
- 研究纳入271颗融合根下颌第二磨牙,获取显微CT重建图像和二维X线投影图像。
- 以显微CT图像确定三维根管形态金标准,分为融合型、对称型和不对称型。
- 采用Python包Augmentor的传统增强和多角度投影方法扩增X线图像数据集。
- 使用预训练的VGG19、ResNet18、ResNet50和EfficientNet-b5在X线图像上进行根管形态识别。
- 将卷积神经网络分类结果与牙髓科住院医师的分类结果进行比较。
主要发现
- 多角度投影增强方法在除EfficientNet-b5外的所有卷积神经网络中均优于传统增强方法。
- ResNet18结合多角度投影方法优于所有其他组合,总体准确率为79.25%。
- 在具体分类中,融合型、对称型和不对称型的准确率分别为81.13%、86.79%和90.57%。
- 卷积神经网络分类表现优于牙髓科住院医师,住院医师平均准确率仅为60.38%(P < .05)。
关键数值与推荐
- ResNet18结合多角度投影方法总体准确率79.25%。
- 融合型准确率81.13%,对称型准确率86.79%,不对称型准确率90.57%。
- 牙髓科住院医师平均准确率60.38%(P < .05)。
临床怎么用
- 卷积神经网络可基于二维X线图像辅助识别融合根下颌第二磨牙的三维根管形态。
- ResNet18结合多角度投影增强可作为该任务的优选模型组合。
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: Understanding the intricate anatomical morphology of fused-rooted mandibular second molars (MSMs) is essential for root canal treatment. The present study utilized a deep learning approach to identify the three-dimensional root canal morphology of MSMs from two-dimensional X-ray images. METHODS: A total of 271 fused-rooted MSMs were included in the study. Micro-computed tomography reconstruction images and two-dimensional X-ray projection images were obtained. The ground truth of three-dimensional root canal morphology was determined through micro-computed tomography images, which were classified into merging, symmetrical, and asymmetrical types. To amplify the X-ray image dataset, traditional augmentation techniques from the python package Augmentor and a multiangle projection method were employed. Identification of root canal morphology was conducted using the pretrained VGG19, ResNet18, ResNet50, and EfficientNet-b5 on X-ray images. The classification results from convolutional neural networks (CNNs) were then compared with those performed by endodontic residents. RESULTS: The multiangle projection augmentation method outperformed the traditional approach in all CNNs except for EfficientNet-b5. ResNet18 combined with the multiangle projection method outperformed all other combinations, with an overall accuracy of 79.25%. In specific classifications, accuracies of 81.13%, 86.79%, and 90.57% were achieved for merging, symmetrical, and asymmetrical types, respectively. Notably, CNNs surpassed endodontic residents in classification performance; the average accuracy for endodontic residents was only 60.38% (P < .05). CONCLUSIONS: CNNs were more effective than endodontic residents in identifying the three-dimensional root canal morphology of MSMs. The result indicates that CNNs possess the capacity to employ two-dimensional images effectively in aiding three-dimensional diagnoses.
引用
- Wu W, Chen S, Chen P, et al. Identification of Root Canal Morphology in Fused-rooted Mandibular Second Molars From X-ray Images Based on Deep Learning. J Endod. 2024. PMID: 38821263. doi:10.1016/j.joen.2024.05.014
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。