Generation of Multi-Vertebrae Using 3D Deep Learning Models

In the field of medical image processing, the application of 3D deep learning technology has received increasing attention. This study explores a composite 3D deep learning architecture for generating images of single vertebrae and multiple vertebrae. First, we build on Generative Adversarial Nets (GAN) and expand it into a 3D GAN architecture. By training the generator and discriminator, we use it to generate 3D images of a single vertebra. However, 3D single-vertebrae limits practical usability. We hope to further explore another challenge in vertebrae image generation, namely generating 3D images of multiple vertebrae. The biggest problem with multiple vertebrae is the continuity between vertebrae. To this end, we created a composite architecture called 3D U-Net+GAN, which combines the advantages of U-Net and GAN. This architecture can not only generate high-quality multi-vertebral images, but also maintain the continuity and consistency of the images. We use two publicly available datasets, xVertSeg.v1 and VerSe2020, to conduct experiments to demonstrate the superiority of the 3D U-Net + GAN architecture in generating multi-vertebral images. The two 3D deep learning architectures proposed in this study provide effective solutions for single-segment and multi-segment vertebral image generation, which are of great significance for improving medical image processing technology.