Related Experiment Video
Updated: Apr 11, 2026
Digital Hybrid Model Preparation for Virtual Planning of Reconstructive Dentoalveolar Surgical Procedures
Published on: August 5, 2021
Deep learning-based segmentation of caries, implants, fixed prosthesis, and restorations on bitewing radiographs: A
Amisha Parekh1, Rohan Jagtap2, Yalamanchili Samata3
1Department of Biomedical Materials Science, School of Dentistry, University of Mississippi Medical Center, Jackson, MS, USA.
Abstract:
ObjectiveThe present retrospective diagnostic accuracy study aimed to evaluate the performance of an AI-based system for automated detection of teeth, caries, implants, restorations, and fixed prostheses on bitewing radiographs.MethodsA total of 407 bitewing radiographs from 315 adult patients were analyzed using an AI system developed by VELMENI Inc. and compared with reference annotations on individual tooth-level provided by two oral and maxillofacial radiologists. Every tooth was encoded for the absence (0) or presence (1) of radiographic findings: caries, restorations, fixed prosthesis, and implants. Cohen's kappa (κ) with 95% bootstrap confidence intervals was used for assessment of inter-rater reliability. The AI system's diagnostic accuracy was evaluated for sensitivity and specificity using human consensus reference standards.ResultsThe annotated dataset consisted of 2,829 tooth-level observations. The two radiologists showed substantial to near-perfect agreement for prosthesis (κ = 0.925) and restorations (κ = 0.872) detection, moderate for caries detection (κ = 0.804), and lowest for implants detection (κ = 0.726). The AI system showed substantial agreement with the human observers for restorations (κ = 0.812-0.871) and prosthesis detection (κ = 0.882-0.940), moderate for caries (κ = 0.454-0.508), and the highest agreement for implant detection (κ = 0.763-0.974). Post-filtration for human consensus, the AI system showed high sensitivity for implant (1.000), prosthesis (0.984), restorations (0.974), and caries (0.972). The AI system showed high specificity for implant (1.000) and prosthesis (0.984), and restorations (0.936) detection, but slightly lower specificity for caries detection (0.842).ConclusionsThe AI system demonstrated diagnostic performance comparable to that of oral and maxillofacial radiologists for detecting multiple dental findings on bitewing radiographs, including restorations, prosthesis, and implants, and slightly lower for caries detection. These findings support the potential role of the AI system as a clinical adjunct to improve efficiency and consistency in routine dental imaging interpretation.
