Classifying Three-Wall Intrabony Defects from Intraoral Radiographs Using Deep Learning-Based Convolutional Neural Network Models
Issued Date
2024-01-01
Resource Type
ISSN
13057456
eISSN
13057464
Scopus ID
2-s2.0-85210282942
Journal Title
European Journal of Dentistry
Rights Holder(s)
SCOPUS
Bibliographic Citation
European Journal of Dentistry (2024)
Suggested Citation
Piroonsan K., Pimolbutr K., Tansriratanawong K. Classifying Three-Wall Intrabony Defects from Intraoral Radiographs Using Deep Learning-Based Convolutional Neural Network Models. European Journal of Dentistry (2024). doi:10.1055/s-0044-1791784 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/102262
Title
Classifying Three-Wall Intrabony Defects from Intraoral Radiographs Using Deep Learning-Based Convolutional Neural Network Models
Author(s)
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
Objective Intraoral radiographs are used in periodontal therapy to understand interdental bony health and defects. However, identifying three-wall bony defects is challenging due to their variations. Therefore, this study aimed to classify three-wall intrabony defects using deep learning-based convolutional neural network (CNN) models to distinguish between three-wall and non-Three-wall bony defects via intraoral radiographs. Materials and Methods A total of 1,369 radiographs were obtained from 556 patients who had undergone periodontal surgery. These radiographs, each featuring at least one area of intrabony defect, were categorized into 15 datasets based on the presence of three-wall or non-Three-wall intrabony defects. We then trained six CNN models-InceptionV3, InceptionResNetV2, ResNet50V2, MobileNetV3Large, EfficientNetV2B1, and VGG19-using these datasets. Model performance was assessed based on the area under curve (AUC), with an AUC value ≥ 0.7 considered acceptable. Various metrics were thoroughly examined, including accuracy, precision, recall, specificity, negative predictive value (NPV), and F1 score. Results In datasets excluding circumferential defects from bitewing radiographs, InceptionResNetV2, ResNet50V2, MobileNetV3Large, and VGG19 achieved AUC values of 0.70, 0.73, 0.77, and 0.75, respectively. Among these models, the VGG19 model exhibited the best performance, with an accuracy of 0.75, precision of 0.78, recall of 0.82, specificity of 0.67, NPV of 0.88, and an F1 score of 0.75. Conclusion The CNN models used in the study showed an AUC value of 0.7 to 0.77 for classifying three-wall intrabony defects. These values demonstrate the potential clinical application of this approach for periodontal examination, diagnosis, and treatment planning for periodontal surgery.