Deep focus approach for accurate bone age estimation from lateral cephalogram.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Pediatric Endocrinology

Background:

• Bone age assessment is crucial for evaluating children's growth and development.
• Deep learning techniques show significant potential for improving bone age estimation accuracy.
• Current methods often rely on hand-wrist radiographs, prompting exploration of alternative imaging modalities.

Purpose of the Study:

• To develop and validate a deep learning approach for accurate bone age estimation.
• To utilize cervical vertebrae on lateral cephalograms for bone age assessment.
• To employ image segmentation techniques for precise feature extraction from cephalometric images.

Main Methods:

• A dataset of 900 participants (aged 4-18 years) with simultaneous lateral cephalograms and hand-wrist radiographs was used.
• Cervical vertebrae segmentation was performed using the DeepLabv3+ architecture.
• Bone age estimation was conducted via transfer learning with an Inception-ResNet-v2 regression model on segmented images.

Main Results:

• The cervical vertebrae segmentation model achieved high performance metrics (accuracy: 0.956, IoU: 0.913, F1: 0.895).
• The bone age estimation model demonstrated excellent accuracy with a mean absolute error of 0.300 years and R-squared of 0.983.
• Gradient-weighted class activation mapping visualized key regions used for prediction, enhancing model interpretability.

Conclusions:

• The proposed deep learning method accurately estimates bone age using cervical vertebrae from lateral cephalograms.
• This approach offers a reliable and potentially more convenient alternative for bone age assessment in growing children.
• The findings support the clinical utility of AI-driven analysis of cephalograms for pediatric growth evaluation.