Assessment of caries depth on intraoral radiographs using deep learning approaches

Abstract

Accurate assessment of caries depth and the risk of pulp exposure (RoPE) on intraoral radiographs is important for diagnosis and treatment planning. However, intraoral radiographs provide a two-dimensional representation of three-dimensional tooth structures, and their interpretation may be affected by projection geometry, image quality and observer variability. This thesis aimed to develop and evaluate deep learning (DL) approaches for classifying caries depth, estimating RoPE in deep carious lesions and placing model performance in the context of human interpretation.

A systematic review was conducted to evaluate existing DL models for caries-depth assessment on intraoral radiographs. The review included 15 studies and showed generally higher model performance for moderate-to-deep lesions than for early lesions. However, most of the included studies used single-centre datasets, none performed external validation and substantial heterogeneity was identified in datasets, reference standards, analytical tasks and reported performance metrics. The risk of bias across most studies was high.

Informed by these limitations, two lesion-level DL models were then developed in this thesis. DepthNet, based on EfficientNet-B0, classified carious lesions as D1, D2 or D3 and RiskNet, based on ResNet-18, classified reference-D3 lesions as having low or high RoPE. Bitewing and periapical radiographs were obtained from Jordan University Hospital and King Abdulaziz University. For the Jordanian dataset, reference categories were assigned using intraoral radiographs together with corresponding cone-beam computed tomography (CBCT) scans. For the Saudi D3 dataset, RoPE categories were supported by intraoral radiographs, pre- and post-treatment radiographs, and documented pulp-exposure outcomes from vital pulp treatment cases. Final evaluation used patient-level internal held-out test sets. Grad-CAM was used for exploratory model interpretation. An observer survey was completed by 23 qualified dentists and 20 undergraduate dental students, who completed 13 tooth-level assessments across seven radiographs. In a separate analysis, consensus RoPE classifications made by four endodontists using intraoral radiographs alone were compared with a CBCT-supported reference assessment for 129 D3 teeth. Model–observer comparisons were descriptive. DepthNet correctly classified 182 of 251 held-out test crops, achieving 72.51% accuracy and 78.37% balanced accuracy. Recall was 92.00% for D1, 64.39% for D2 and 78.72% for D3. All errors occurred between adjacent categories, with no D1-to-D3 or D3-to-D1 errors. RiskNet achieved 72.34% accuracy, 64.88% balanced accuracy and a macro F1-score of 65.44% on 94 held-out D3 crops. Recall was 83.33% for high risk lesions and 46.43% for low risk lesions. The endodontist consensus showed sensitivity of 89.7%, specificity of 43.8%, accuracy of 78.3% and fair agreement with the CBCT-supported reference assessment (κ = 0.365). Human observer performance in caries-depth classification was variable; dentists performed better than students overall, Year 5 students performed better than Year 4 students and D2 was the most difficult depth category.

The findings provide initial evidence that DL can extract radiographic information relevant to both lesion depth and RoPE. However, the models operated on pre-localised lesion crops, were evaluated using internal rather than external test data and were less reliable for D2 and D3 low risk lesions. They are therefore not ready for independent clinical use. Larger and more balanced multicentre datasets, external validation, end-to-end model development and prospective human–AI studies are required before clinical translation.

Thesis is embargoed until 31 December 2031.

Date of Award	Dec 2026
Original language	English
Awarding Institution	Queen's University Belfast
Sponsors	Ibn Sina University for Medical Sciences
Supervisor	Ikhlas El Karim (Supervisor), Richard Gault (Supervisor) & Jamie Toole (Supervisor)