Department of Radiology, Ninth People's Hospital of Suzhou, Soochow University, Suzhou, China.
Department of Dentistry and Central Lab, Ninth People's Hospital of Suzhou, Soochow University, Suzhou, China.
Clin Oral Investig. 2024 Nov 18;28(12):646. doi: 10.1007/s00784-024-06049-8.
The C-shaped mandibular second molars (MSMs) may pose an endodontic challenge. The aim of this study was to develop a convolutional neural network (CNN)-based deep learning system for the diagnosis of C-shaped MSMs on panoramic radiographs.
Panoramic radiographs and cone beam computed tomographic (CBCT) images were collected from a hospital in China and subsequently divided into two groups. In Group A, conventional panoramic images and CBCT images were derived from the same patients (n = 730 individuals), and the dataset consisted of conventional panoramic image patches of 1453 MSMs. In Group B (n = 610 individuals), the patients underwent CBCT examinations in the absence of available panoramic images; CBCT images were acquired and utilized to generate simulated panoramic images, and the dataset consisted of image patches of 1211 MSMs. Five pretrained CNN networks (ResNet-101 and - 50, DenseNet-121 and - 161, and Inception-V3) were utilized for the classification of C-shaped and non-C-shaped MSMs. Finally, the networks trained on the Group B dataset were tested on the Group A dataset. The diagnostic performance of each model was evaluated using receiver operating characteristic (ROC) curve analysis, and the CBCT images were taken as the gold standard. The results were compared with those achieved by three dental professionals.
In Group A, all five networks exhibited satisfactory diagnostic performance, with AUC values ranging from 0.875 to 0.916 and accuracies ranging from 81.8 to 86.7%. No statistical differences were detected among the five CNNs. Notably, the models trained with Group B dataset (CBCT-generated panoramic images) achieved enhanced performance as tested on Group A dataset. The AUC values reached 0.984-0.996, and the accuracies ranged between 94.5% and 98.1%. CNNs outperformed dental professionals in classification performance, and the AUC values achieved by dental specialist, novice dentist, and dental graduate student were only 0.806, 0.767 and 0.730, respectively.
CNN-based deep learning system demonstrated higher accuracy in the detection of C-shaped MSMs on panoramic radiographs compared to dental professionals. CBCT-generated panoramic images can serve as a substitute for conventional panoramic images in the training of CNN models when the quantity and quality of conventional panoramic image dataset is insufficient.
CNN-based deep learning models have demonstrated significant potential in assisting dentists with the identification of C-shaped MSMs on panoramic radiographs, which facilitating more effective, efficient and safer dental treatment.
C 形下颌第二磨牙（MSM）可能构成牙髓挑战。本研究旨在开发一种基于卷积神经网络（CNN）的深度学习系统，用于诊断全景片上的 C 形 MSM。
从中国的一家医院收集全景片和锥形束计算机断层扫描（CBCT）图像，随后将其分为两组。在组 A 中，常规全景图像和 CBCT 图像来自同一患者（n=730 人），数据集由 1453 个 MSM 的常规全景图像补丁组成。在组 B（n=610 人）中，患者在没有可用全景图像的情况下接受了 CBCT 检查；获取 CBCT 图像并用于生成模拟全景图像，数据集由 1211 个 MSM 的图像补丁组成。使用五个预训练的 CNN 网络（ResNet-101 和 -50、DenseNet-121 和 -161、Inception-V3）对 C 形和非 C 形 MSM 进行分类。最后，在组 A 数据集上测试了在组 B 数据集上训练的网络。使用接收者操作特征（ROC）曲线分析评估每个模型的诊断性能，并将 CBCT 图像作为金标准。将结果与三位牙科专业人员的结果进行比较。
在组 A 中，所有五个网络均表现出令人满意的诊断性能，AUC 值范围为 0.875 至 0.916，准确率范围为 81.8%至 86.7%。五个 CNN 之间没有检测到统计学差异。值得注意的是，使用组 B 数据集（CBCT 生成的全景图像）训练的模型在组 A 数据集上进行测试时表现出更好的性能。AUC 值达到 0.984-0.996，准确率在 94.5%至 98.1%之间。CNN 在分类性能方面优于牙科专业人员，而牙科专家、新手牙医和牙科研究生的 AUC 值仅为 0.806、0.767 和 0.730。
与牙科专业人员相比，基于 CNN 的深度学习系统在检测全景片上的 C 形 MSM 方面具有更高的准确性。当常规全景图像数据集的数量和质量不足时，CBCT 生成的全景图像可以作为训练 CNN 模型的替代方法。
基于 CNN 的深度学习模型在辅助牙医识别全景片上的 C 形 MSM 方面显示出了显著的潜力，这有助于更有效、高效和更安全的牙科治疗。
