Rafif Alshenaiber, Heba Wageh Abozaed, Bandar Alzahrani, Khalid AlShehri, Fahad Alnafisah, Reem Alharbi, Asmaa Alzobidi, Aziz Alshehri, Mohammad Aldekhayyil, Qamar Hashem, Shahad Albader, Ibrahim Aljulayfi, Ziyad Alabdulsalam, Shaymaa Mudhaffer, Hunaida Tayeb
The proposed deep learning model demonstrates high accuracy and reliability in predicting tooth restorability from parallel periapical radiographs. It has the potential to serve as a supportive tool for improving consistency and efficiency in clinical decision-making.
INTRODUCTION AND AIM: Evidence-based decision-making in determining whether to preserve or extract a tooth is essential for maintaining long-term function and aesthetics. However, such decisions are often subjective and influenced by the clinician's experience. This study aimed to develop and evaluate a deep learning-based artificial intelligence model to preliminarily predict tooth restorability from radiographic data only.
MATERIALS AND METHODS: Six dental specialists developed a tooth restorability prediction score based on parallel periapical radiographs. A total of 53,035 radiographs were collected, anonymized, and screened according to predefined inclusion and exclusion criteria. Of these, 1,300 radiographs were randomly selected and independently assessed by the specialists. Intra- and inter-examiner reliability were evaluated using Cohen's kappa coefficients (P < .05). A convolutional neural network based on a modified VGG16 architecture was developed using the proposed scoring system. Data were split using 70:20:10 (train: validation: test) with 37,125 images used for training, 10,607 for validation, and 5,303 for testing.
RESULTS: Strong agreement was observed among specialists, with intra- and inter-examiner kappa values of 0.84 and 0.87, respectively. The deep learning model demonstrated high performance, achieving an accuracy of 94.74%, precision of 96.35%, recall of 92.94%, and an F1 score of 94.61%. The model also showed good discriminative ability, with an area under the receiver operating characteristic curve (AUC) of 0.888 (95% CI: 0.871-0.904).
CONCLUSION: The proposed deep learning model demonstrates high accuracy and reliability in predicting tooth restorability from parallel periapical radiographs. It has the potential to serve as a supportive tool for improving consistency and efficiency in clinical decision-making.
TRIAL REGISTRATION: This study was approved by the Research Centre of Prince Sattam bin Abdulaziz University (Approval No. SCBR-178/2023; 05/11/2023) and the Institutional Review Board of Riyadh Second Health Cluster (Approval No. FWA00018774; 22/04/2024).