Age estimation through mandibular radiomic attributes in panoramic radiographs using convolutional neural networks

Artificial intelligence (AI)-based approaches to mandibular age estimation are limited by reliance on manually selected anatomical features. Thus, this study aimed to develop and validate a convolutional neural network (CNN)-based model to estimate age from panoramic radiographs through the analysis of mandibular radiomic attributes.

This was a cross-sectional study based on digital panoramic radiographs of patients aged 2 to 97 years. To develop the segmentation model, 600 radiographic images were manually annotated and used only for training. A U-Net-based CNN was implemented for the semantic segmentation task, successfully processing 7,832 radiographic images. The proposed age estimation model was built on a CNN architecture comprising three convolutional blocks, each consisting of a Conv2D layer followed by a MaxPooling2D layer, enabling progressive, hierarchical extraction of visual features.

The model achieved a mean absolute error (MAE) of 6.91 years, and a root mean square error (RMSE) of 9.41 years. The mean coefficient of determination (R²) was 0.799. In the comparison between chronological and predicted age, most observations were distributed close to the identity line. However, a slight increase in dispersion was observed in the older age groups.

The findings suggest that the proposed CNN-based model shows promising results for estimating age from panoramic radiographs. Although a modest reduction in predictive precision was noted among older individuals, overall performance remained acceptable. However, these results should be interpreted with caution, given the limited data sources and the absence of external validation.