3D-2D Distance Maps Conversion Enhances Classification of Craniosynostosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Surgery

Background:

• Craniosynostosis diagnosis traditionally relies on computed tomography (CT), exposing infants to ionizing radiation.
• Photogrammetric 3D surface scans offer a radiation-free alternative but require advanced analysis methods.
• Developing effective, non-invasive diagnostic tools is crucial for early intervention in pediatric care.

Purpose of the Study:

• To develop a novel 3D surface scan to 2D distance map conversion method for craniosynostosis classification.
• To enable the application of convolutional neural networks (CNNs) for diagnosing craniosynostosis using 2D representations of 3D data.
• To evaluate the performance of this approach against traditional methods, focusing on accuracy, computational efficiency, and data augmentation potential.

Main Methods:

• A coordinate transformation, ray casting, and distance extraction process was used to create 2D distance maps from 3D surface scans.
• A CNN-based classification pipeline, specifically Resnet18, was implemented and compared with alternative classifiers.
• Investigations included low-resolution sampling, data augmentation techniques, and attribution mapping to understand classification drivers.

Main Results:

• The Resnet18 classifier achieved a high F1-score of 0.964 and an accuracy of 98.4% on a dataset of 496 patients.
• Data augmentation significantly improved classifier performance across all tested models.
• Low-resolution sampling reduced computational cost by 256-fold during ray casting while maintaining a high F1-score of 0.92.

Conclusions:

• The 2D distance map conversion method effectively enhances craniosynostosis classification using CNNs, offering benefits in data augmentation and anonymity.
• Photogrammetric 3D surface scans are a viable tool for clinical craniosynostosis diagnosis, with potential for domain transfer to CT imaging.
• This radiation-free method holds promise for reducing ionizing radiation exposure in infants and improving diagnostic workflows.