Self-Absorption Correction in X-Ray Fluorescence-Computed Tomography With Deep Convolutional Neural Network

Abstract

Data collected from the X-ray fluorescence-computed tomography (XFCT) is frequently reconstructed with algorithms proposed for X-ray transmission tomography. As these algorithms do not model the self-absorption effect inherent to XFCT, their capacity on accurately reconstructing the elemental distribution is limited. Although algorithms specialized for XFCT reconstruction have been developed, the majority of them impose strict requirements on the samples and the acquisition setup. To relax these prerequisites, a deep convolutional neural network is proposed to correct the self-absorption effect in the sinogram domain. Through quantitative evaluation, we conclude that the well-trained neural network can correct fluorescence sinograms affected by the self-absorption effect. Furthermore, we demonstrate that such corrections enable conventional algorithms to reconstruct the elemental distribution with high fidelity. As the only input required by the proposed neural network is the fluorescence sinogram, it is fully automatic and is applicable to different scan setups and samples.