12th Conference on Industrial Computed Tomography, Fürth, Germany (iCT 2023), www.ict2023.org Deep Learning Based Scatter Estimation Markus Michen 1 , Ulf Haßler 1 , Christopher Syben 1 1 Fraunhofer IIS, Fraunhofer Institute for Integrated Circuits IIS, Division Development Center X-Ray Technology markus.michen@iis.fraunhofer.de, ulf.hassler@iis.fraunhofer.de, christopher.syben@iis.fraunhofer.de 1 Introduction X-ray computed tomography (CT) has increasingly established itself as a non-destructive testing technique in industry. Areas of application include inspection, analysis and measurement technology. In all the above fields artifacts severely degrade image quality and can complicate or even prevent subsequent algorithms operating on the reconstructed data, therefore a reliable artifact reduction method is necessary. One of the main causes of such artifacts is scattered radiation. Scattered photons hit the detector elements, superimpose with the unscattered primary radiation and therefore induce non-linearities in the imaging process. Scattering processes depend on a lot of factors, like energy, scanning geometry and the measured object. Scatter artifacts may lead to streaks, cupping and general loss of contrast in the reconstructed volume. Figure 1 illustrates the effects of scattered radiation. There is also beam hardening that often occurs in tandem with scatter. Here with increasing object thickness, the average energy of the X-rays increases, the beam gets “harder”. Beam hardening artifacts cause dark shadows and the typical cupping effect. Generally speaking, scattering and beam hardening effects can hardly be separated, since both result from non-linearities of the imaging process and manifest themselves in similar effects in the reconstructed volume. Many approaches to correct scattered radiation exist, prominent ones are Monte Carlo, kernel or machine learning (ML) based methods. In contrast to other ML-based methods, we present a Deep Learning-based scatter estimation method that optimizes the scatter distribution to yield an optimal reconstruction with respect to a quality criterion in the volume domain. The accumulated errors in the volume domain are backpropagated to projection domain and update a neural network operating on the projections. We evaluated our developed method on both simulated and real data. Additionally, we compared our results with two state-of- the-art methods, namely the iterative artifact reducation (IAR) [3, 4] and the deep scatter estimation (DSE) network [7, 8]. 2 Related Work In general scatter artifact reduction methods can be divided into physically- or algorithmically based methods. Physically, anti- scatter grids or collimators can be used to avoid scatter. Algorithmically, Monte Carlo (MC) methods, kernel based techniques or deep learning are prominent approaches to correct scatter. An iterative scheme has been developed by Kasperl et al. The iterative artifact reduction (IAR) is not a scatter correction algorithm as it only focuses on fixing beam hardening effects. However, the nonlinearities in the imaging process due to scattering and beam hardening can hardly be separated and influence each other. The algorithm consists of the following steps. First, an initial reconstruction is calculated. Then the volume is segmented and ray traced to calculate a linearization curve. This linearization curve tries to transform the measured polyenergetic values to a corresponding monoenergetic beam sum, to account for beam hardening effects. After that the corrected reconstruction is computed and can be used for the next iteration. The authors show a great improvement of image quality with respect to beam hardening and to a certain degree scatter artifacts. Their method can also replace laborious scanning of reference objects in order to derive linearization curves [3, 4]. A promising approach for scatter correction has been shown by J. Maier et al., where deep learning is used to mimic a compu- Figure 1: Scattering artifacts in the reconstructed volume and the corresponding line profile showing a cupping effect. Copyright 2022 - by the Authors. Licensed under a Creative Commons Attribution 4.0 International License. More info about this article: https://www.ndt.net/?id=27723 https://doi.org/10.58286/27723