217 - CarioNet: Predicting Proximal Caries Progression Using Machine Learning

Purpose: Although artificial intelligence (AI) models show promise in detecting proximal caries on bitewing radiographs, most approaches are cross-sectional and provide limited insight into lesion depth or progression (Azhari et al., 2023). Caries diagnosis is often inconsistent even at similar risk levels, contributing to variability in treatment decisions and potentially affecting tooth longevity (Henneberg et al., 2024). This study developed and evaluated CarioNet, a machine learning model designed to predict 18-month proximal caries progression using longitudinal bitewing radiographs. By shifting from detection to prediction, this approach may support more conservative, preventive care and preserve tooth structure in pediatric patients.

Methods: This retrospective study reviewed 5,862 patient charts from adolescents aged 10–17 years with permanent dentition to identify diagnostic bitewings demonstrating proximal caries in posterior teeth. Two independent models were developed. Phase I (segmentation) included 203 single-visit bitewing radiographs, while Phase II (progression) included 148 longitudinal lesion series with at least three serial recall radiographs. Lesions were staged using ICDAS classification (E1–D3) and independently reviewed by a pediatric dental resident and oral radiology resident. Reliability was assessed prior to annotation (inter-rater κ=0.65), with discrepancies resolved by consensus and adjudication. Intra-rater reliability was assessed following a three-month washout (κ=0.77). All lesions were manually annotated by the pediatric dental resident. Phase I four segmentation models, with U-Net++ model with an EfficientNet-B4 encoder pretrained on ImageNet. Phase II used logistic regression trained on lesion features, including prior ICDAS classification, lesion size from segmentation, and rate of size change through stratified resampling.

Results: Phase I performance was within the expected range for early-stage detection (F1=0.51), with U-Net++ performing best among the models tested. In Phase II, logistic regression demonstrated the best performance among seven models. The model achieved F1=0.78 and balanced accuracy=0.81, with sensitivity=0.81 (0.71, 0.90) and PPV=0.84 (0.78, 0.89). This model accurately predicted ICDAS stage for both progressing and non-progressing lesions despite class imbalance. Misclassifications occurred primarily between adjacent stages, with greatest difficulty at the E2–D1 boundary.

Conclusions: CarioNet demonstrates that proximal caries progression can be predicted from longitudinal bitewing radiographs (sensitivity=0.81(0.71, 0.90); PPV=0.84(0.78, 0.89)), supporting that prediction models may shift toward earlier, risk-based intervention and conservative management strategies in pediatric dental care.