Sub-Axial Vertebral Column Fracture CT Image Synthesis by Progressive Growing Generative Adversarial Networks (PGGANs)

Abstract

Orthopaedicians need the assistance of the Deep Learning (DL) model for easy Vertebral Column Fracture Type identification. Deep Learning models require large datasets. Due to the non-availability of large annotated data sets, the DL model needs intensive data augmentation methods. In this proposed research work, Progressive Growing Generative Adversarial Networks (PGGANs) are used to generate synthetic Vertebral Column Fracture (VCF) CT images. The synthetic CT images of VCF generated by PGGANs are high resolution, realistic yet wholly different from the real images. The PGGANs is a multi-stage generative model that generates 512 X 512 CT images that increases the accuracy of the VCF Type identification system. A total of375 vertebral column CT images were utilized for training the model, which were collected from the Spine Clinic, Orthopaedics Department, Kasturba Medical College, Manipal, Manipal Academy of Higher Education, Manipal. Among 375 images, 275 Chance fractures and 100 posterior tension band disruption fracture images were present. To analyse the effect of PGGAN augmentation on VCF type identification, lately VGG16 pre-trained model is implemented. The VGG16 model with PGGAN augmentation got an accuracy of 87.01%, which is more when compared to the model without augmentation. In conclusion, PGGAN generated VCF images are realistic and can be used for data augmentation without privacy restrictions and in VCF type identification DL models for increased performance.