Convolutional neural network-based classification of craniosynostosis and suture lines from multi-view cranial X-rays

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Surgery

Background:

• Craniosynostosis (CSO) requires early diagnosis for effective infant treatment.
• Computed tomography (CT) provides detailed imaging but involves significant radiation exposure risks for children.
• There is a need for safer, reliable diagnostic methods for CSO.

Purpose of the Study:

• To develop and evaluate a deep learning model for diagnosing CSO and classifying cranial suture lines using 2D X-rays.
• To minimize radiation exposure associated with traditional diagnostic imaging techniques.
• To enhance the accuracy and efficiency of CSO diagnosis in clinical settings.

Main Methods:

• A deep learning approach utilizing Convolutional Neural Networks (CNNs) was developed for CSO and suture-line classification.
• The model incorporated preprocessing steps including X-ray marker removal, head-pose standardization, and skull cropping.
• Data from 1,047 normal and 277 CSO cases (2006-2023) were used to train and validate the CNN models.

Main Results:

• The CNN models achieved high diagnostic performance for CSO detection, with F1-scores exceeding 0.96 and sensitivity/specificity above 0.9.
• Preprocessing strategies further improved accuracy, yielding the highest F1-scores, precision, and specificity.
• The suture-line classification model accurately distinguished five suture lines with accuracy greater than 0.9.

Conclusions:

• The proposed deep learning model offers a reliable and low-radiation alternative for diagnosing craniosynostosis from 2D X-rays.
• The model demonstrates significant potential for clinical application, improving diagnostic accuracy and efficiency.
• This AI-driven approach can streamline CSO diagnosis, reducing diagnostic time and labor in clinical practice.