Age estimation from thoracic CT using 3D convolutional neural networks

This study investigated the feasibility of using three-dimensional deep learning (DL) models to estimate chronological age from routine thoracic computed tomography (CT) scans and examined the contribution of different CT reconstruction windows to model performance.

This retrospective study included thoracic CT scans from 1278 adults aged 20-80 years obtained from a single tertiary center. Lung, bone, and mediastinal reconstruction windows were generated for each scan. A 3D DenseNet-based DL model was trained using either individual window inputs or a combined three-channel configuration. Five-fold cross-validation was applied during training, an independent test set was used for evaluation. Model performance was assessed using mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R²), and Pearson correlation coefficient (r). Ensemble learning, Bland-Altman analysis, and Grad-CAM was utilized to provide visual explanations of the model's focus and to identify the anatomical regions most influential in the age estimation process.

All models demonstrated good accuracy in predicting age from thoracic CT images. The bone-window model achieved the lowest MAE among single-window approaches, emphasizing the role of skeletal features. However, the three-channel model showed more consistent performance across age groups. The three-channel ensemble model achieved the best overall performance (MAE=4.89 years, R²=0.823, r = 0.915) with minimal bias. Prediction errors were lower in younger and older individuals and higher in middle-aged cohorts.

DL-based age estimation from thoracic CT images is highly accurate. Integrating multiple reconstruction windows improves performance and supports the potential clinical and forensic utility of this approach.