TY - JOUR

T1 - Simulating X-ray beam energy and detector signal processing of an industrial CT using implicit neural representations

AU - Blum, Edwin

AU - Burmeister, Moritz

AU - Stamer, Florian

AU - Lanza, Gisela

N1 - Conference code: 14

PY - 2025/2/19

Y1 - 2025/2/19

N2 - Simulating computed tomography (CT) systems offers numerous advantages, including the optimization of scan parameters, training of specialist personnel, and quantification of measurement uncertainties. Current simulation approaches, often referred to as virtual computed tomography (vCT), predominantly rely on analytical models. However, these models require extensive system-specific tuning to produce realistic synthetic measurements, creating a significant barrier to broader adoption and efficiency. To address this challenge, this work explores the potential of implicit neural representation (INR) as an alternative to the analytical models used in vCT. INRs excel at representing complex, high-dimensional data in a continuous and differentiable manner, making them a promising substitute for traditional analytical models. As a first building block, we propose a two-stage approach for simulating the X-ray beam energy and detector signal processing in industrial CT systems. This method is trained and evaluated using real-world data. Results demonstrate that the proposed INR-based architecture can accurately generate synthetic projections for parameter configurations within the training dataset. However, its poor performance on the test dataset highlights limitations in generalization beyond the training data. Potential methods to address these shortcomings are discussed.This study underscores the potential of INRs as a flexible framework for simulating complex CT systems. By capturing subtle system-specific characteristics and reducing dependence on explicitly defined parameterizations, INRs could pave the way formore versatile and efficient simulation models.

AB - Simulating computed tomography (CT) systems offers numerous advantages, including the optimization of scan parameters, training of specialist personnel, and quantification of measurement uncertainties. Current simulation approaches, often referred to as virtual computed tomography (vCT), predominantly rely on analytical models. However, these models require extensive system-specific tuning to produce realistic synthetic measurements, creating a significant barrier to broader adoption and efficiency. To address this challenge, this work explores the potential of implicit neural representation (INR) as an alternative to the analytical models used in vCT. INRs excel at representing complex, high-dimensional data in a continuous and differentiable manner, making them a promising substitute for traditional analytical models. As a first building block, we propose a two-stage approach for simulating the X-ray beam energy and detector signal processing in industrial CT systems. This method is trained and evaluated using real-world data. Results demonstrate that the proposed INR-based architecture can accurately generate synthetic projections for parameter configurations within the training dataset. However, its poor performance on the test dataset highlights limitations in generalization beyond the training data. Potential methods to address these shortcomings are discussed.This study underscores the potential of INRs as a flexible framework for simulating complex CT systems. By capturing subtle system-specific characteristics and reducing dependence on explicitly defined parameterizations, INRs could pave the way formore versatile and efficient simulation models.

KW - Engineering

KW - Machine Learning

KW - Simulation

KW - vCT

KW - Implicit Neural Representations

U2 - 10.58286/30752

DO - 10.58286/30752

M3 - Conference article in journal

VL - 30

JO - E-Journal of nondestructive testing

JF - E-Journal of nondestructive testing

SN - 1435-4934

IS - 2

T2 - 14th Conference on Industrial Computed Tomography - iCT 2025

Y2 - 4 February 2025 through 7 February 2025

ER -