Automated Sagittal Craniosynostosis Classification from CT Images Using Transfer Learning.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Pediatric neurosurgery

Background:

• Sagittal craniosynostosis (CSO) is a condition where the sagittal suture fuses prematurely in children.
• Current surgical treatments for CSO rely on physician classification of subtypes from CT images.
• Previous methods using hand-crafted features had limitations in accurately representing complex image features.

Purpose of the Study:

• To develop a deep learning-based method for improved classification of sagittal craniosynostosis (CSO) subtypes.
• To enhance feature extraction efficiency and classification accuracy over traditional methods.
• To reduce subjectivity in surgical technique selection by providing objective diagnostic insights.

Main Methods:

• 3D skulls were segmented from CT slices using a Hounsfield Unit (HU) threshold.
• Hemispherical projection mapped 3D skulls to 2D binary images (512x512 resolution).
• Deep convolutional neural networks were trained using augmented data and transfer learning.

Main Results:

• Deep learning models achieved over 90% prediction accuracy, surpassing previous methods (72%).
• The senior surgeon's classification yielded the highest performing model (75% accuracy on unseen data).
• Significant inter-observer variability (54%) was noted among surgeons' manual classifications.

Conclusions:

• Deep learning methods are superior to hand-crafted features for sagittal CSO classification.
• Model performance depends on data quality; deep learning can reduce inter-observer variability.
• AI models show potential to approximate physician diagnostic performance and guide clinical treatment decisions.