Related Experiment Video
Updated: Jul 3, 2026
Reliability of Artificial Intelligence-Based Cone Beam Computed Tomography Integration with Digital Dental Images
Published on: February 23, 2024
Deep learning-based dental age classification from panoramic radiographs: a multi-granularity, multi-model
Esin Akol Görgün1, Ezgi Eroğlu Çakmakoğlu2, Savaş Sağmak3
1Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Adıyaman University, Altinsehir Neighborhood, 3005 Street No:13, Adiyaman, 02040, Turkey. esinakol@adiyaman.edu.tr.
Objective:
To evaluate the performance of a hybrid deep learning architecture, MaxViT-T, for fine-grained dental age estimation (DAE) and coarse-grained age-group classification using panoramic radiographs, and to interpret the model's decision-making process through Explainable AI (XAI).
Materials And Methods:
A dataset of n = 2,761 panoramic radiographs from patients aged 6-18 years was used. Two classification tasks were performed: a 13-class fine-grained age estimation and a 3-group coarse-grained age classification (age ranges 6-9, 10-13, and 14-18 years, approximating the early mixed, late mixed, and permanent dentition periods). The hybrid MaxViT-T model was compared against ten other architectures, including CNNs and Transformers, using 5-fold stratified cross-validation. Model interpretability was assessed using Grad-CAM and SHAP.
Results:
MaxViT-T achieved the highest performance in both tasks. For the 13-class age estimation, it reached an accuracy of 0.4980 with a Mean Absolute Error (MAE) of 0.71 years. Tolerance-based accuracy increased to 84.9% and 96.1% within 1 and 2 years, respectively. For the 3-group age classification, the model achieved 91.13% accuracy and an AUC of 0.9669. XAI analyses confirmed that the model's attention shifted from deciduous teeth and erupting permanent teeth in younger ages to third molar development in older ages, predominantly focusing on the mandibular region.
Conclusion:
The hybrid MaxViT-T architecture achieved the highest accuracy for dental age estimation among the evaluated models, although its advantage over the leading CNN and Transformer architectures was not statistically significant. Its decision behavior aligned with established odontogenic developmental landmarks, supporting the biological plausibility of the learned features.
