Towards Automated Craniosynostosis Diagnosis Using EfficientNet-Based Artificial Intelligence Models: A Two-Class and Multi-Class Approach.

Craniosynostosis is a congenital condition characterized by premature fusion of a newborn's cranial bones, impeding brain growth and causing disfiguration. Early diagnosis allows less invasive surgery and better outcomes, but diagnosis is often delayed and current practice often relies on computed tomography (CT), exposing infants to radiation. Routine head photography offers a safer alternative, and advances in artificial intelligence (AI) provide an opportunity to detect this condition earlier. We evaluated three state-of-the-art architectures: EfficientNet-B7, ResNet-50, and ResNet-152, trained and tested across five random seeds for binary and multi-class craniosynostosis classification using 688 real-world clinical photographs. Unlike prior studies relying on X-ray, CT, or photogrammetry, our work uses standard 2D photographs, addressing limitations of small datasets, conventional machine learning approaches, and dependence on specialized imaging equipment. In binary classification, all models achieved near-perfect performance; EfficientNet-B7 attained the highest mean accuracy (99.7%) with perfect sensitivity. In multi-class classification, EfficientNet-B7 achieved the highest mean accuracy (96.6%), significantly outperforming ResNet-50 (p < 0.005) while performing comparably to ResNet-152. These findings demonstrate that AI models trained on standard 2D photographs can deliver accurate, cost-effective, and non-invasive craniosynostosis detection and subtype classification, supporting timely intervention and expanding access to care in both specialized and resource-limited settings. To our knowledge, this is the first study to achieve high-accuracy binary and multi-class craniosynostosis classification using real-world 2D clinical photographs rather than radiation-based or photogrammetry-derived imaging.