Preliminary Study of X-ray Ghost Imaging Reconstruction with a Deep Neural Network Constraint

Unlike typical imaging techniques that collect photons reflecting or transmitting an object with a multi-pixel photo detector, ghost imaging (GI) is an indirect imaging technique by correlating the intensity of two beams. One of the beams that does not interact with an object provides the spatial distribution (2D information), and the other beam that actually interacts with the object provides the intensity information from the interaction (1D information). Separation of detection and imaging parts in GI could resolve the limitation of conventional imaging techniques in harsh environments such as strong turbulence. GI was perceived possible only with entangled photon pairs in quantum correlation, however; later studies have shown that classically correlated lights such as pseudo-thermal lights are also applicable. Recently, GI using classical correlation covers X-rays, and it is expected that it will bring lower exposure to radiation and higher image quality. There have been several algorithms for GI, but the nature of GI reconstruction relies on the number of random speckle patterns and measurements, making a typical ghost image blurry. GI does not have a standardized point-spread function (PSF), making the deconvolution an ill-posed problem. There are studies to cope with the problem in GI using pseudo-thermal lights with deep learning techniques, but little studies were done regarding the improvement of quality in X-ray Ghost Imaging (XGI). In this study, we attempted to resolve the quality issue in XGI using a deep learning model called GIDC (GI using Deep neural network Constraint). This model enhances the quality of pseudo-thermal light GI effectively without any training, but has not been implemented on XGI. We calculated SSIM and PSNR values to demonstrate the impact of GIDC on XGI.