Generative-Adversarial-Network-Based Data Augmentation for the Classification of Craniosynostosis

Abstract Craniosynostosis is a congenital disease characterized by the premature closure of one or multiple sutures infant’s skull. For diagnosis, 3D photogrammetric scans are radiation-free alternative to computed tomography. However, data only sparsely available and role augmentation for classification craniosynostosis has not yet been analyzed. In this work, we use 2D distance map representation infants’ heads with convolutional-neural-network-based classifier employ generative adversarial network (GAN) augmentation. We simulate two scarcity scenarios 15% 10% training test influence different degrees added synthetic balancing underrepresented classes. used total accuracy F1-score as metric evaluate final classifiers. data, GAN-augmented dataset showed an increased up 0.1 3 %. both metrics decreased. present deep convolutional GAN capable creating craniosynostosis. Using moderate amount using slightly better performance, but had little effect overall. The simulated scenario may have limited model’s ability learn underlying distribution.