Quantifying dysmorphologies of the neurocranium using artificial neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Craniosynostosis involves premature fusion of cranial sutures, requiring objective methods for evaluating cranial morphology.
• Current subjective assessments of craniosynostosis lack consistency.
• This study addresses the need for quantitative, objective evaluation of craniosynostosis.

Purpose of the Study:

• To develop and validate a novel, quantitative AI-based approach for classifying craniosynostosis.
• To introduce a new metric, the Feature Prominence (FP) score, for measuring craniosynostosis severity.
• To correlate the FP score with existing clinical severity assessments.

Main Methods:

• An artificial neural network was trained on synthetic 3D head models to classify head shapes (normocephalic, trigonocephalic, scaphocephalic).
• 3D models were converted to low-dimensional shape representations using normal vector distributions for AI input, ensuring patient anonymity and invariance.
• Explainable AI methods and the novel FP score were used to analyze shape features and their clinical relevance.

Main Results:

• The AI model achieved excellent accuracy in classifying cranial shapes.
• Attention maps highlighted key regions (parietal, temporal) influencing classification.
• The FP score demonstrated a strong positive correlation with clinical severity scores for scaphocephalic (ρ=0.83) and trigonocephalic (ρ=0.64) cases.

Conclusions:

• The AI-based method provides an objective, privacy-preserving tool for quantifying cranial shape, independent of age or orientation.
• The FP score shows significant potential to aid clinical decision-making in craniosynostosis assessment.
• The study's open-source code promotes wider implementation for advancing craniofacial research and care.