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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Surgery

Background:

• Craniosynostosis, premature fusion of infant skull bones, can cause disfigurement and impede brain growth.
• Current diagnosis often relies on computed tomography (CT), involving radiation exposure and potential delays.
• Early detection is crucial for less invasive surgical interventions and improved patient outcomes.

Purpose of the Study:

• To evaluate the efficacy of state-of-the-art AI architectures for detecting craniosynostosis using standard 2D clinical photographs.
• To compare the performance of EfficientNet-B7, ResNet-50, and ResNet-152 models in binary and multi-class craniosynostosis classification.
• To establish a non-invasive, radiation-free diagnostic tool for craniosynostosis.

Main Methods:

• Trained and tested EfficientNet-B7, ResNet-50, and ResNet-152 models on 688 clinical 2D photographs.
• Utilized five random seeds for robust evaluation of binary and multi-class classification tasks.
• Focused on standard photographs, avoiding specialized imaging like X-ray, CT, or photogrammetry.

Main Results:

• All models achieved high accuracy in binary classification, with EfficientNet-B7 reaching 99.7% accuracy and perfect sensitivity.
• EfficientNet-B7 demonstrated superior performance in multi-class classification (96.6% accuracy), outperforming ResNet-50.
• AI models successfully classified craniosynostosis subtypes using only standard 2D images.

Conclusions:

• AI models trained on 2D photographs provide accurate, cost-effective, and non-invasive craniosynostosis detection and classification.
• This approach supports timely intervention, potentially expanding access to care in diverse healthcare settings.
• This study pioneers high-accuracy craniosynostosis classification using readily available clinical photographs.