Multi-Modal and Multi-View Fusion Classifier for Craniosynostosis Diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Radiology

Background:

• Craniosynostosis diagnosis is crucial for infant treatment and surgical success.
• Current methods like CT scans pose significant radiation risks to children.
• There is a need for safer, effective diagnostic tools for craniosynostosis.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for craniosynostosis detection using multi-view X-ray images.
• To assess the model's diagnostic accuracy and its potential to reduce radiation exposure compared to CT scans.
• To investigate the model's ability to localize disease-relevant regions.

Main Methods:

• A deep learning model integrating multi-view fusion (MVF) and cross-attention mechanisms was developed.
• The model processed three X-ray views (AP, lateral right, lateral left) and patient metadata (age, sex).
• Model performance was evaluated on a dataset of 882 X-ray images from 294 pediatric patients.

Main Results:

• The model achieved a high diagnostic accuracy with an AUROC of 0.975, F1-score of 0.882, sensitivity of 0.878, and specificity of 0.937.
• Grad-CAM visualizations confirmed the model's ability to identify disease-relevant areas.
• The approach demonstrated effective feature integration from multiple X-ray views and metadata.

Conclusions:

• The proposed deep learning model offers a safe, cost-effective, and accurate method for diagnosing craniosynostosis in infants.
• This AI-driven approach has the potential to significantly improve pediatric care by reducing reliance on CT scans.
• The model provides a promising alternative for early and precise detection of craniosynostosis.