The preliminary in vitro study and application of deep learning algorithm in cone beam computed tomography image implant recognition
Shaobo Ou-Yang,Shuqin Han,Dan Sun,Hongping Wu,Jianping Chen,Ying Cai,Dongmei Yin,Huidan Ou-Yang,Lan Liao
DOI: https://doi.org/10.1038/s41598-023-45757-1
2023-10-27
Abstract: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.