Artificial Intelligence-based multi-stage pipeline for detecting caries and restorations in pediatric primary molars from intraoral photographs

Dental caries remain a prevalent concern, particularly in children, and accurate diagnosis prior to orthodontic treatment is essential to prevent further demineralization and maintain oral health. In this study, we developed a multi-stage deep learning pipeline for automated detection and classification of posterior primary molars (D and E) using intraoral photographs. A YOLOv8 model was trained to localize primary first molars (D) and primary second molars (E), followed by SAM-based segmentation and ResNet50 classification into five categories: Healthy, Stained, Decayed, Stainless Steel Crown (SSC), and Amalgam Restoration. The dataset comprised 725 occlusal intraoral photographs from 725 children aged 6–12 years. For tooth detection, photographs were divided into training, validation, and test sets at a 70%/15%/15% ratio (508/109/108 images; 303 annotated teeth in the test set). For the primary tooth-level classification analysis, 1,946 tooth samples were divided at a 70%/20%/10% ratio into training, validation, and test sets (1,362/389/195 teeth). A separate patient/source-photograph-level sensitivity analysis used a 70%/15%/15% split of the 725 source photographs. On the held-out YOLOv8 test set of 108 photographs containing 303 annotated teeth, the detector achieved a precision of 0.963, recall of 0.972, and mAP@0.5 of 0.983. In the patient/source-photograph-level sensitivity analysis, an independently retrained ResNet50 classifier was evaluated on 290 held-out tooth samples and achieved an accuracy of 85.9%, macro-averaged precision of 85.6%, recall of 85.7%, F1-score of 85.6%, and Cohen’s κ of 0.822. The original tooth-level classification analysis, which did not restrict teeth from the same source photograph to a single subset, yielded an accuracy of 87.7% on 195 test teeth. Grad-CAM was used as a qualitative visualization of image regions contributing to model predictions. An integrated AI-based pipeline may be capable of providing accurate and interpretable multi-class dental health classification from intraoral images. While the observed performance is encouraging, additional validation on larger and more diverse multi-center datasets will be necessary to better assess generalizability and robustness, particularly in scenarios involving confounding restorative materials. Overall, this approach suggests the potential of AI-assisted diagnostics to contribute to the standardization of dental assessment, support standardized photographic assessment, and improve workflow efficiency in pediatric dentistry.