TomoGAN: unsupervised learning-based reconstruction of tomography

Abstract

Tomography enables three-dimensional imaging of object of interest, ranging from nano-materials to biological cells and human organs. Although tomographic imaging provides invaluable information in many fields including medicine, material science and biology, in many cases the quality of the image is degraded due to insufficient measurements. The missing cone problem in optical diffraction tomography (ODT) limited angle X-ray computed tomography (CT), and sparse-view reconstruction in electron dispersive X-ray (EDX) tomography are representative obstacles that demarcate the resolution of the object being imaged. In this paper, a universal method to solve these problems through unsupervised deep learning, dubbed TomoGAN, is presented. TomoGAN, which is generally applicable to 3D tomography, works by enhancing the resolution of projection measurements by learning the optimal transport map. The algorithm presented in this paper has great advantages in that it does not require high resolution ground truth data for training. The applications that are covered in this paper is as follows: ODT, limited angle CT, and extreme sparse-view EDX tomography, and extreme sparse-view dual energy CT.