3D Bone Shape Reconstruction from 2D X-ray Images Using MED Generative Adversarial Network

Abstract

Three-dimensional (3D) objects are always better than two-dimensional (2D) images for visualization. Very delicate information can be detected in 3D objects compared to 2D objects. The human brain can easily infer 3D information by observing 2D objects. But retrieving 3D information from a 2D object is very challenging for machines. For this reason, 3D shape reconstruction is one of the most researched topics in computer vision, deep learning, and medical science. 3D bone shapes are needed for pre-operative surgery planning and visualization. But the conventional methods of generating high-quality 3D bone-shape from 2D images are very time-consuming. This paper proposes a framework of Generative Adversarial Network Medical Generative Adversarial Network (MED-GAN) to generate high-quality 3D bone-shape from 2D images. It generates 3D bone-shape using recent advances in convolution networks and generative adversarial networks. X-ray images are fed to the convolution network which is then converted to a D-dimensional vector by the convolution network. D-dimensional vector is fed to the GAN to reconstruct 3D bone shape.