Dataset

This study received ethical approval from the Inha University Institutional Review Board and the Inha University Data Review Board (IRB No. 2024–08-034; DRB No. DRB-2024–10-001). All panoramic radiographs were acquired at the Dentistry Centre of Inha University Hospital utilizing a dental panoramic X-ray machine (ORTHOPHOS XG 5 & CEPH, Dentsply Sirona, Bensheim, Germany). A total of 607 radiographic images were collected between 2021 and 2023. Images from patients exhibiting primary or mixed dentition, as well as those affected by significant image distortion, were excluded from the dataset to ensure consistency and reliability in the subsequent model training process. All patient data were fully anonymized prior to analysis in accordance with institutional data protection protocols and international ethical standards, thereby safeguarding the confidentiality of personal information throughout the study.

Annotation and image processing

All images were annotated utilizing the Computer Vision Annotation Tool (CVAT), a platform extensively employed in medical image analysis for precise manual labeling. The annotation process was conducted independently by two dentists possessing over five years of clinical experience in diagnosing periodontitis. Any discrepancies between their annotations were reviewed and reconciled through consensus with a senior dentist who provided the final verification. This multi-layered validation process ensured a high degree of consistency and minimized subjective variability within the dataset.

The annotation protocol was specifically designed to support a two-stage deep learning architecture based on YOLOv11. In the first stage, each tooth in the panoramic radiograph was segmented and classified into one of four anatomical categories: incisor, canine, premolar, and molar. This segmentation facilitated the model’s capability to isolate and analyze each tooth individually, providing a standardized input for subsequent landmark detection. In the second stage, the model was trained to detect six anatomical keypoints per tooth (m1–d3), corresponding to the mesial and distal CEJ, alveolar crest levels, and root apex. For single-rooted teeth, the mesial and distal apex (m3 and d3) were annotated as a singular point due to their anatomical overlap. These six key anatomical landmarks, summarized in Table 1, were selected based on their clinical significance for measuring radiographic bone loss.

Table 1 Description of six anatomical key points used for keypoint detection

Following tooth segmentation and classification, each tooth was cropped and annotated with the corresponding key landmarks. Given that certain keypoints (particularly m1, m3, d1, and d3) were frequently located near the image borders, a 30-pixel white margin was incorporated around the cropped images to mitigate the risk of misinterpretation during both training and inference. This preprocessing step enhanced the stability and accuracy of keypoint localization by ensuring that no landmark was truncated or positioned too close to the image edge, which could otherwise diminish detection precision.

Periodontitis staging standard

The staging of periodontitis was determined based on RBL measurements, calculated by selecting the higher bone loss percentage between the mesial and distal aspects of each tooth. This methodology adhered to the approach outlined by Jiang et al. (2022) and complied with the 2017 Consensus Classification jointly issued by the American Academy of Periodontology and the European Federation of Periodontology. Unerupted or abnormally positioned third molars were excluded from analysis due to their inconsistent bone morphology and limited clinical relevance to generalized periodontitis staging.

Teeth exhibiting alveolar bone loss of less than 2 mm from the CEJ were not clinically diagnosed as having bone resorption. Nonetheless, given the technical challenges associated with accurately measuring 2 mm on panoramic radiographs, this study did not differentiate between teeth without bone loss and those classified as Stage I. The staging criteria and clinical features utilized in this study are summarized in Table 2. This standardization ensured that the model’s staging predictions were clinically meaningful and directly aligned with widely accepted diagnostic guidelines.

Table 2 Staging criteria for periodontitis based on radiographic bone loss (RBL) and associated clinical featuresModel training procedure

All experiments were conducted on Google Colab utilizing an NVIDIA L4 GPU. Transfer learning was employed to enhance model convergence and improve performance, particularly given the relatively limited dataset size. Training proceeded for a maximum of 200 epochs, commencing with an initial learning rate of 0.001. An early stopping strategy, characterized by a patience parameter of 30 epochs, was implemented to mitigate overfitting; this ensured that training was terminated once performance on the validation set ceased to exhibit improvement. The weights of the best-performing model were automatically preserved for subsequent evaluation.

To retain anatomical details, the segmentation model was trained on images resized to 1280 × 1280 pixels, while the keypoint detection model utilized a resolution of 640 × 640 pixels to achieve a balance between accuracy and computational efficiency. The entire dataset was randomly partitioned into training, validation, and test sets in an 8:1:1 ratio. The training set was designated for model fitting and parameter optimization, the validation set was employed to fine-tune the model and alleviate overfitting, and the test set provided an unbiased assessment of model performance on previously unseen data.

Data augmentation techniques were implemented to address class imbalance and enhance the model’s generalizability. These techniques included brightness adjustment, random rotation of up to ± 10°, multiscale resizing, mosaic augmentation, and copy-paste augmentation. Such methodologies introduced controlled variability into the training data, thereby enabling the model to learn robust features and maintain accuracy across diverse imaging conditions. The overall workflow consisted of two primary stages. Initially, the segmentation model isolated and classified individual teeth from panoramic radiographs. Subsequently, four keypoint detection models were applied to the segmented regions to identify anatomical landmarks necessary for calculating radiographic bone loss and staging periodontitis. The performance of the final optimized model was subsequently compared with clinical assessments conducted by experienced dentists on an independent test set (Fig. 1).

Fig. 1The alternative text for this image may have been generated using AI.

Workflow of tooth segmentation and anatomical keypoint detection using YOLOv11

Evaluation metrics

The performance of the segmentation and keypoint detection models was evaluated using multiple standard metrics to provide a comprehensive assessment of predictive accuracy and robustness. For segmentation, the training process was monitored through loss curves, where a consistently decreasing loss over epochs indicated successful learning and optimization. A confusion matrix was generated to visualize classification performance across various tooth types, facilitating the identification of potential misclassification patterns. Precision was calculated to ascertain the proportion of correct positive predictions, while recall quantified the model’s ability to identify true positive cases. The F1-score, defined as the harmonic mean of precision and recall, served as a balanced measure of the model’s classification accuracy.

The degree of spatial overlap between predicted and ground-truth tooth boundaries was quantified using Intersection over Union (IoU), also referred to as the Jaccard index. Additionally, mean average precision (mAP) was computed to summarize the model’s segmentation and localization performance across all tooth classes. Evaluating mAP at multiple IoU thresholds allowed for the assessment of model reliability under both lenient and stringent conditions, thereby providing a more robust indication of performance stability.

For keypoint detection, analogous evaluation metrics were applied, supplemented by Object Keypoint Similarity (OKS), which measures the spatial correspondence between predicted and ground-truth keypoints. OKS functions similarly to IoU but focuses on landmark localization rather than segmentation boundaries. High OKS values indicated accurate localization of clinically relevant anatomical points, which is essential for reliable measurement of radiographic bone loss.

Comparative evaluation with dentists

To establish a reference standard and assess clinical validity, three dentists with over five years of clinical experience at the Dentistry Centre, Inha University Hospital, independently conducted periodontitis staging and grading on the test dataset. These clinicians were blinded to the model’s predictions to mitigate bias and ensure an objective comparison between human and AI-based assessments.

The agreement between the model’s outputs and the evaluations provided by the dentists was assessed using two complementary statistical methods. The Pearson correlation coefficient (PCC) was calculated to determine the strength and direction of the linear relationship between the model’s predictions and the reference standard. Additionally, the intraclass correlation coefficient (ICC) was employed to evaluate the consistency and agreement between raters, specifically between the dentists and the model. All statistical analyses were conducted using IBM SPSS Statistics (Version 26.0, Armonk, NY, USA) and Python 3.9.13. A statistical significance threshold of p < 0.05 was established, and 95% confidence intervals were reported where applicable. These methodological approaches ensured that the model’s diagnostic performance was not only technically robust but also clinically comparable to expert human judgment.