Limited Angle Tomography Reconstruction: Synthetic Reconstruction via Unsupervised Sinogram Adaptation

Abstract

Computed Tomography (CT) is commonly used in clinical procedures and limited angle tomography reconstruction has important applications in diagnostic CT, breast tomography, dental tomography, etc. However, CT images reconstructed from limited angle acquisitions suffer from severe artifacts due to incomplete sinogram data. Although existing iterative reconstruction methods improve image quality relative to filtered back projection, these methods require extensive computation and still often provide unsatisfactory images. Supervised deep learning methods have been proposed to further improve the image quality of limited angle reconstructions. However, a key limitation in supervised deep learning for this application is the lack of large-scale real sinogram-reconstruction pairs for training. Given the large number of CT images available in the wild, we can create a large number of simulated sinogram-reconstruction pairs. Thus the requirement for real paired sinogram-reconstruction data can be alleviated if simulated sinograms (e.g. monochromatic) are able to train a reconstruction network for real sinograms (e.g. polychromatic source, scattering, beam hardening). In this paper, we propose an end-to-end limited angle tomography reconstruction adversarial network (Tomo-GAN) via unsupervised sinogram adaptation without having real sinogram-reconstruction pairs. Tomo-GAN is trained by using (1) unpaired sinograms from the simulation and real domains, and (2) large-scale reconstruction images from only the simulation domain. Tomo-GAN is built based upon a cycle consistent network with similarity constrained for sinogram adaptation and a multi-scale conditional reconstruction network. Experimental results on a public dataset with a limited angle setting demonstrated a consistent improvement over previous methods while significantly reducing the reconstruction computation time.