Related Experiment Video
Updated: Jul 14, 2026
Precision of In Vivo Quantitative Tooth Wear Measurement Using Intra-Oral Scans
Published on: July 12, 2022
Deep Learning for Assessing Caries Depth in Permanent Teeth Using Intraoral Radiographs: A Systematic Review
Laith Abu Qdais1,2, Richard Gault3, Jamie Toole1
1School of Medicine, Dentistry and Biomedical Sciences, Queen's University Belfast, Belfast, UK.
Background:
Accurate assessment of caries depth on intraoral radiographs is crucial for determining the extent of the lesion and planning appropriate treatment. Artificial intelligence (AI) models have been increasingly used for detecting and classifying carious lesions; however, the evidence focusing specifically on the classification of caries depth remains limited.
Objectives:
To systematically review studies that used AI models to classify caries depth on intraoral radiographs of permanent teeth and to assess their diagnostic performance.
Method:
PubMed, Scopus, EMBASE, MEDLINE and BASE were searched for studies published between January 2016 and September 2025. Eligible studies applied AI models to bitewing or periapical radiographs for classifying caries depth. Two reviewers independently screened titles, abstracts and full texts, with a third reviewer resolving disagreements. Risk of bias was assessed by three reviewers using the QUADAS-2 tool and the QUADAS-C extension for comparative studies. Due to substantial heterogeneity in datasets, reference standards and reported metrics, a meta-analysis was not performed and findings were synthesised narratively. The protocol was registered in PROSPERO (CRD42024512152).
Results:
Fifteen studies were included. Most used bitewing radiographs and convolutional neural-network architectures such as U-Net, ResNet, YOLO and EfficientDet. Dataset sizes ranged from 112 to 8539 radiographs. No study performed external validation and most used single-centre datasets. Reported metrics including accuracy, F1-score, AUC, mAP, recall and precision varied widely. Models generally showed higher performance for moderate-to-deep lesions than for early lesions. The risk of bias was high in the patient-selection domain for most studies, with only two rated low risk across all domains.
Conclusion:
AI models demonstrate encouraging performance for classifying caries depth on intraoral radiographs, particularly for deep caries; however, evidence is limited by small, single-centre datasets, inconsistent reference standards and high risk of bias. Future studies should employ larger multicentre datasets, standardised depth criteria and external validation before clinical translation.
Trial Registration:
ClinicalTrials.gov identifier: PROSPERO CRD42024512152.
