Thanks to the short wavelength of X-rays and their ability to penetrate deep into matter, they are uniquely suited for nanoscale 3D imaging of materials ranging in size from micrometers to centimeters. Full-field X-ray microscopes enable rapid imaging of a material’s nanostructure and are more suitable for in situ and operando transmission imaging. Scanning allows X-ray fluorescence mapping of elemental distributions in samples, as well as structural imaging of extended areas beyond lens limits when using a coherent beam of illumination. While initial X-ray microscopes targeted 2D imaging, the journey from 2D to 3D imaging started with conventional tomography, as developed for medical imaging. Eventually, the method was adopted in electron and X-ray microscopy. Today, the X-ray tomography systems offered at synchrotron facilities around the world have grown into a vital imaging tool, offering sub-100 nanometer isotropic 3D resolution for materials and biological research.
However, while X-ray microscopy is instrumental to many scientific studies, there are unrealized opportunities for advancing imaging by using modern computational imaging methods. The two fundamental challenges in imaging are the uncertainties related to the image formation process, and the limited amount of observational data; both set the limit for imaging capability. The common approach to address these challenges has been using computationally costly iterative reconstruction methods that allow one to incorporate prior models about the solution. Instead of this costly approach, we propose an alternative approach by using generative models (i.e., a deep generative adversarial network, called TomoGAN).
In this setting, the deep network could remove noise and artifacts from an initial reconstruction obtained by fast but less accurate methods such as filtered-backprojection (FBP), thereby avoiding computationally costly iterative optimization techniques. Experiments on both synthetic and experimental data showed that TomoGAN ran faster and delivered better image quality than the costly iterative reconstruction method. Thanks to the high generality of TomoGAN, it has also been used in other scenarios such as realtime streaming CT reconstruction as well as learned priors for robust Ptycho-Tomography reconstruction.
The source code of TomoGAN is publicly available at: https://github.com/lzhengchun/TomoGAN.
The tutorials are also available at https://github.com/AIScienceTutorial/Denoising, for those who wish to learn uses of deep learning for denoising of reconstructed X-ray images.