Artificial Intelligence Models for the Detection and Quantification of Orthodontically Induced Root Resorption Using Cone-Beam Computed Tomography: A Systematic Review and Meta-Analysis

TL;DR

This study reviews AI models for detecting and measuring root resorption in orthodontic patients using 3D imaging, finding them highly accurate and reliable.

Contribution

The paper provides a systematic review and meta-analysis of AI models for orthodontically induced root resorption detection using CBCT.

Findings

AI models showed excellent sensitivity (0.903) and high specificity (82–98%) for detecting root resorption.

CNN-based models achieved up to 0.96 AUC, with near-perfect agreement (ICC 1.000) with manual assessments.

Linear methods were more sensitive to early resorption changes than volumetric approaches.

Abstract

Background/Objectives: Orthodontically induced root resorption (OIRR) is a well-documented but undesired consequence of orthodontic treatment. This systematic review and meta-analysis aimed to assess the diagnostic performance of artificial intelligence (AI) models applied to cone-beam computed tomography (CBCT) for detecting and quantifying OIRR while evaluating their agreement with manual reference standards and the impact of model architecture, validation design, and quantification strategy. Methods: Comprehensive searches were conducted across PubMed/MEDLINE, Scopus, Web of Science, and EMBASE up to November 2025. Studies were included if they employed AI for OIRR diagnosis using CBCT and reported relevant performance metrics. Following PRISMA guidelines, data were extracted and a random-effect meta-analysis was performed. Subgroup analyses explored the influence of model design and validation. Results: Seven studies were included. Pooled sensitivity from three eligible studies was 0.903 (95% CI: 0.818–0.989), suggesting excellent true positive rates. Specificity ranged from 82% to 98%, and area under the receiver operating characteristic curve values reached up to 0.96 across studies using EfficientNet, U-Net, and other convolutional neural network (CNN)-based architectures. The pooled intraclass correlation coefficient for agreement with manual quantification was 1.000, reflecting near-perfect concordance. Subgroup analyzes showed slightly superior performance in CNN-only models compared to hybrid approaches, and better diagnostic metrics with internal validation. Linear assessments appeared more sensitive to early apical shortening than volumetric methods. Conclusions: AI models applied to CBCT demonstrate excellent diagnostic accuracy and high concordance with expert assessments for OIRR detection. These findings support their potential integration into clinical orthodontic workflows.

Figures