A prediction model of pediatric bone density from plain spine radiographs using deep learning

Area of Science:

• Pediatric radiology
• Artificial intelligence in medicine
• Bone health diagnostics

Background:

• Osteoporosis in children is poorly understood, characterized by low bone mineral density (BMD) and increased fracture risk.
• Accurate assessment of pediatric BMD is crucial for early detection and management of bone diseases.

Purpose of the Study:

• To develop and validate a deep learning model for predicting pediatric BMD using standard spine radiographs.
• To classify children into low-BMD groups for targeted bone health interventions.

Main Methods:

• A two-stage deep learning model was employed, utilizing Yolov8 for vertebral body detection and ResNet-18 for BMD prediction.
• The model was trained and validated on data from 601 pediatric patients (aged 10-20 years) with available dual-energy X-ray absorptiometry and radiography.
• Performance was evaluated using metrics such as average precision, correlation coefficient, intraclass correlation coefficient, and area under the receiver operating characteristic curve.

Main Results:

• The model demonstrated high accuracy in detecting vertebral bodies (AP50=0.97) and predicting BMD (r=0.72).
• Consistent performance was observed across different vertebral segments, with good agreement (ICC=0.64).
• The model successfully classified low-BMD groups with high sensitivity (0.76) and specificity (0.87), achieving an AUC of 0.85.

Conclusions:

• The developed deep learning model shows significant promise for non-invasively predicting pediatric BMD from plain spine radiographs.
• This approach can enhance early detection of low bone density in children and streamline bone health management for high-risk populations.