A novel AI system for preliminary triage support in single-tooth edentulous spaces using intraoral images.

This study aimed to demonstrate the integration of deep learning (DL) and machine learning (ML) using only occlusal photographs to provide preliminary triage support between orthodontic and prosthodontic treatments in cases requiring alternatives to implant-based interventions. Occlusal photographs (n = 2,962) were collected under routine conditions using smartphones and digital cameras. Two groups of dental specialists independently annotated the dataset. A YOLOv8m model was used for the localization of single-tooth edentulous spaces (S-TES), which were defined as localized areas within the dental arch where a single tooth is absent while adjacent teeth remain present, as well as their mesial and distal adjacent teeth. ResNet-50/ResNet-101 and VGG-16/VGG-19 models were employed to classify the clinical conditions and anatomical categories of teeth adjacent to S-TES. A deterministic function converted the mesiodistal width of the S-TES from pixels to millimeters using mean central incisor widths. Logistic Regression (LR) and XGBoost (XGB) models were used to predict the preliminary triage outputs. Among the classifiers for the clinical tooth condition classification task, VGG-19 showed the highest macro F1 score (0.928) and ResNet-101 yielded the highest macro AUC (0.961) and weighted kappa (0.927), with the narrowest 95% confidence intervals (95% CI) for both metrics. For the anatomical categorization task, ResNet-101 showed the highest F1 score, recall, and precision. For the triage support models, LR showed the highest AUC (0.896) with an expected calibration error (ECE) of 0.041, and XGB demonstrated the highest accuracy (0.869), sensitivity (0.874), and specificity (0.862) with an ECE of 0.024. The proof-of-concept demonstrates that the integration of DL and ML can achieve acceptable performance for preliminary triage support between orthodontic and prosthodontic treatment from occlusal images, within the acknowledged limitations. Further research is needed to improve model generalizability. However, the findings indicate a promising direction for future work. While in the long term, such technology has the potential to support preliminary triage between treatment options and become a clinically applicable tool, our current findings show that external validation and improved generalizability of the proposed AI framework are necessary before such clinical applications can be realized.