Automated assessment of gingival biotype using deep learning on intraoral photographs
TL;DR
The proposed deep learning model, particularly the Vision Transformer, demonstrated high diagnostic accuracy and robust performance across gingival biotype classes, and showed improved recall and F1-scores, and better class separability in the confusion matrix, confirming its robustness and statistical reliability.
Abstract
Gingival biotype is a key factor influencing dental treatment outcomes. This study aimed to construct and validate an artificial intelligence (AI) model for objective and reproducible gingival biotype assessment based on intraoral photographs, thereby supporting clinical decision-making and personalized treatment planning. A total of 1,600 participants (aged 24 ± 2 years; 720 males and 880 females) with healthy periodontal conditions were enrolled. Gingival biotype was clinically identified using the probe transparency method and categorized as thick, medium, or thin. The dataset (640 thick, 520 medium, 440 thin) was split into a training set ( n = 1,500) and testing set ( n = 100) with proportional distribution. Weighted cross-entropy was applied to account for class imbalance. The Vision Transformer (ViT) model was trained using AdamW with an initial learning rate of 0.001 and a batch size of 8 for 10 epochs, whereas Residual Network-18 (ResNet-18) was trained using Adam with a learning rate of 1 × 10 −4 and a batch size of 32 with early stopping. Offline data augmentation (rotation ±15°, horizontal/vertical flipping, contrast/gamma adjustment, and contrast-limited adaptive histogram equalization (CLAHE)) was applied in a 1:4 ratio with fixed random seeds (123) to expand the training set from 1,500 to 7,500 images, while the test set underwent fixed preprocessing only. Model performance was evaluated using F1-score and area under the receiver operating characteristic curve (AUC), with statistical significance and 95% confidence intervals estimated via bootstrap resampling and DeLong test. The ViT model achieved superior diagnostic performance compared with ResNet-18. In the test set, the ViT model reached AUCs of 1.00 (thick), 0.98 (thin), and 0.97 (medium), significantly higher than those of ResNet-18 (0.82, 0.79, and 0.80; p < 0.001). The ViT model also showed improved recall and F1-scores, particularly for the “Thick” and “Thin” classes, and better class separability in the confusion matrix, confirming its robustness and statistical reliability. The proposed deep learning model, particularly the Vision Transformer, demonstrated high diagnostic accuracy and robust performance across gingival biotype classes. The ViT model offers predictive capability and potential to support clinical decision-making in treatment planning. This artificial intelligence (AI)-based approach enables non-invasive, objective and efficient gingival biotype assessment, facilitating early risk evaluation and personalized treatment planning. Integrated into digital diagnostic workflows, it can assist clinicians in selecting appropriate incision designs, restorative margin levels, and orthodontic force strategies according to biotype characteristics, thereby improving treatment predictability and patient outcomes.