Multi energy computed tomography imaging optimization

Abstract

X-Ray Computed Tomography (CT) is a powerful imaging method because it can map spatial X-ray attenuation properties of an object. Lab-based X-ray sources generate photons of varying energies (the source spectrum), while laboratory-based X-ray detectors generally cannot resolve X-ray energies. These detectors measure the total amount of energy. This method is imperfect for the quantitative characterization of materials, as both material density and material composition will change overall X-ray transmission. Therefore, for material decomposition, an accurate model of the source spectrum and the detector’s spectral response is needed in order to estimate absorption spectra when using non-energy resolving detectors. We develop computational methods that allow us to use a standard lab-based X-ray CT system to resolve the full range of possible materials using a multi-energy X-ray source and an energy integrating detector. We suggested achieving this by using a multi-energy scan approach coupled with the optimization of a more accurate X-ray absorption model. To overcome the ill-conditioning, we implement and study deep learning methods to constrain the inverse problem. We generated objects that have two chemical elements using NIST data at energies between 20keV and 200keV. To learn a low dimensional model of absorption spectra, we used different auto-encoder models. We tested the method on a simplified 10-pixel problem with spectra quantised into 20 energy levels and generated one deep learning model to estimate material densities from the recovered absorption spectra. Our deep learning model was able to reliably predict one material of the two elements in the generated object.

More information

Identifiers

ORCID for Thomas Blumensath: orcid.org/0000-0002-7489-265X

Catalogue record

Export record