Deep convolutional neural network and image prior based super resolution for X-ray nano-tomography

K.C. Prabhat, V. DeAndrade, N. Kasturi, X. Yang
Argonne National Laboratory; The University of Chicago,
United States

Keywords: convolutional neural networks (CNN), transmission X-ray microscopy (TXM), brain cell imaging, super-resolution (SR)

Summary:

Transmission X-ray Microscopy based nano – computed tomography (CT) has taken a foothold in the domain of Materials Science. This is primarily owing to the sub-50 nm resolution and the inherent non-destructive nature of the TXM tomography. As for instance, the TXM technique has opened avenues for researchers to decipher the cellular map of nervous system at the nanoscale level. This technique is also being extensively used to deduce the 3D images of computer chips at the nm length scale. However, the TXM technique presents with its own sets of difficulties and limitations. More specifically, we are constrained by TXM’s depth of focus limitation to record images below a resolution of 20 nm. A spatial resolution of 20 nm or less is required to properly resolve features of nervous system. Likewise, images recorded at the limit of the TXM, i.e. 25 nm, suffers from motion and drift artifacts. In the case of in-situ analysis, such as imaging the components of Lithium-ion (LIBs), increasing the magnification from 50 nm to 25 nm has to be accompanied with an increase in the X-ray dose by a factor of eight. The increase in the dose makes the LIB samples exceedingly prone to irradiation. Thereby, deforming the LIB sample and skewing its microstructural and electrochemical properties. Hence, we propose the use of super resolution of x-ray images to resolve the aforementioned experimental limitations of the TXM technique. We will formulate our super resolution model making use of deep Convolutional Neural Networks (CNN) in combination with sparse image priors to promote edge preserving reconstruction. A schematic of the mapping from low resolution to high resolution is provided in fig 1. Finally, the quantitative gain of the high resolution images determined from our CNN-approach will be evaluated by the means of metrics such as the Peak signal-to-noise ratio (PSNR) and Structural Similarity Index (SSIM) for synthetic as well as experimental datasets.