Predicting pediatric age from chest X-rays using deep learning: a novel approach

Area of Science:

• Radiology
• Artificial Intelligence
• Pediatrics

Background:

• Accurate pediatric age estimation is crucial for developmental assessment and forensic applications.
• Traditional methods rely on bone age assessment via wrist X-rays.
• Deep learning offers potential for utilizing other radiological modalities like chest X-rays.

Purpose of the Study:

• To evaluate the effectiveness of deep learning for pediatric age estimation using chest X-rays.
• To develop and validate a deep neural network model for this purpose.

Main Methods:

• A ResNet-based deep neural network with Coordinate Attention was developed.
• A large dataset of 128,008 pediatric chest X-rays was utilized.
• Mean Absolute Error (MAE) and Spearman correlation were key evaluation metrics.

Main Results:

• The model achieved low MAE (around 5.8-7.4 months) on internal and external validation sets.
• Spearman correlation coefficient exceeded 0.98, indicating strong age prediction accuracy.
• Heatmap analysis showed the model focused on anatomical structures like the spine and mediastinum.

Conclusions:

• Chest X-rays can be effectively used for accurate pediatric age estimation with deep learning.
• The developed model demonstrates robustness and potential as a complement to traditional bone age assessment.
• This approach may reduce radiation exposure and aid in screening abnormal growth and development.