The Affiliated Stomatological Hospital of Nanchang University, The Key Laboratory of Oral Biomedicine, Jiangxi Province Clinical Research Centre for Oral Diseases, Nanchang, Jiangxi Province, China.
Information Security Evaluation Section, Jiangxi Science and Technology Infrastructure Center, Nanchang, China.
Sci Rep. 2023 Oct 27;13(1):18467. doi: 10.1038/s41598-023-45757-1.
To properly repair and maintain implants, which are bone tissue implants that replace natural tooth roots, it is crucial to accurately identify their brand and specification. Deep learning has demonstrated outstanding capabilities in analysis, such as image identification and classification, by learning the inherent rules and degrees of representation of data models. The purpose of this study is to evaluate deep learning algorithms and their supporting application software for their ability to recognize and categorize three dimensional (3D) Cone Beam Computed Tomography (CBCT) images of dental implants. By using CBCT technology, the 3D imaging data of 27 implants of various sizes and brands were obtained. Following manual processing, the data were transformed into a data set that had 13,500 two-dimensional data. Nine deep learning algorithms including GoogleNet, InceptionResNetV2, InceptionV3, ResNet50, ResNet50V2, ResNet101, ResNet101V2, ResNet152 and ResNet152V2 were used to perform the data. Accuracy rates, confusion matrix, ROC curve, AUC, number of model parameters and training times were used to assess the efficacy of these algorithms. These 9 deep learning algorithms achieved training accuracy rates of 100%, 99.3%, 89.3%, 99.2%, 99.1%, 99.5%, 99.4%, 99.5%, 98.9%, test accuracy rates of 98.3%, 97.5%, 94.8%, 85.4%, 92.5%, 80.7%, 93.6%, 93.2%, 99.3%, area under the curve (AUC) values of 1.00, 1.00, 1.00, 1.00, 1.00, 1.00, 1.00, 1.00, 1.00. When used to identify implants, all nine algorithms perform satisfactorily, with ResNet152V2 achieving the highest test accuracy, classification accuracy, confusion matrix area under the curve, and receiver operating characteristic curve area under the curve area. The results showed that the ResNet152V2 has the best classification effect on identifying implants. The artificial intelligence identification system and application software based on this algorithm can efficiently and accurately identify the brands and specifications of 27 classified implants through processed 3D CBCT images in vitro, with high stability and low recognition cost.
为了正确地修复和维护种植体，这些是替代天然牙根的骨组织植入物，准确识别其品牌和规格至关重要。深度学习通过学习数据模型的内在规律和表示程度，在图像识别和分类等分析方面表现出了出色的能力。本研究的目的是评估深度学习算法及其支持的应用软件在识别和分类三种不同尺寸和品牌的牙种植体的三维（3D）锥形束 CT（CBCT）图像方面的能力。通过使用 CBCT 技术，获得了 27 个不同尺寸和品牌的种植体的 3D 成像数据。经过手动处理后，将数据转换为一个包含 13500 个二维数据的数据集。使用 9 种深度学习算法，包括 GoogleNet、InceptionResNetV2、InceptionV3、ResNet50、ResNet50V2、ResNet101、ResNet101V2、ResNet152 和 ResNet152V2，对这些数据进行处理。使用准确率、混淆矩阵、ROC 曲线、AUC、模型参数数量和训练时间来评估这些算法的效果。这 9 种深度学习算法的训练准确率为 100%、99.3%、89.3%、99.2%、99.1%、99.5%、99.4%、99.5%、98.9%，测试准确率为 98.3%、97.5%、94.8%、85.4%、92.5%、80.7%、93.6%、93.2%、99.3%，曲线下面积（AUC）值为 1.00、1.00、1.00、1.00、1.00、1.00、1.00、1.00、1.00。当用于识别种植体时，所有 9 种算法的表现都令人满意，其中 ResNet152V2 的测试准确率、分类准确率、混淆矩阵 AUC 和接收者操作特征曲线 AUC 最高。结果表明，ResNet152V2 对种植体的分类效果最好。基于该算法的人工智能识别系统和应用软件可以通过体外处理的 3D CBCT 图像，高效、准确地识别 27 个分类种植体的品牌和规格，具有高稳定性和低识别成本。
