Machine Learning-Based Modeling of the Anode Heel Effect in X-ray Beam Monte Carlo Simulations

Hussein Harb, Didier Benoit, Axel Rannou, Chi-Hieu Pham, Valentin Tissot, Bahaa Nasr, Julien Bert · 2025

Read Machine Learning-Based Modeling of the Anode Heel Effect in X-ray Beam Monte Carlo Simulations by Hussein Harb, Didier Benoit, Axel Rannou, Chi-Hieu Pham, Valentin Tissot, Bahaa Nasr, Julien Bert (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-19155.txt.

Introduction

Hussein Harb and co-authors use machine learning to fix a flaw in simulated X-ray beams: the anode heel effect, which makes beam intensity uneven along one axis. Monte Carlo simulators assume a symmetric beam, so their dose maps drift from clinical reality. The team trained regression models on measured intensity weights across tube voltages, then built the best model into two simulation toolkits.

Gradient boosting won, predicting intensities within five percent at every energy. A calibration protocol needed six detector positions per energy level, cutting measurement effort by about two thirds while adding under two percent error. In simulations the learned model reproduced clinical beam profiles and beat symmetric assumptions. The authors offer the method as generalizable across machines. For dosimetry and radiation protection, realistic beams mean safer, better-checked treatment plans. The work turns a tedious calibration into a short measurement session.

Catalog record

CREATORHussein Harb, Didier Benoit, Axel Rannou, Chi-Hieu Pham, Valentin Tissot, Bahaa Nasr, Julien Bert
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-19155

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Machine Learning-Based Modeling of the Anode Heel Effect in X-ray Beam Monte Carlo Simulations Hussein Harb, Didier Benoit, Axel Rannou, Chi-Hieu Pham, Valentin Tissot, Bahaa Nasr, Julien Bert Abstract This study enhances Monte Carlo simulation accuracy in X-ray imaging by developing an AI-driven model for the anode heel effect, achieving improved beam intensity distribution and dosimetric precision. Through dynamic adjustments to beam weights on the anode and cathode sides of the X-ray tube, our machine learning model effectively replicates the asymmetry characteristic of clinical X-ray beams. Experimental results reveal dose rate increases of up to 9.6% on the cathode side and reductions of up to 12.5% on the anode side, for energy levels between 50 and 120 kVp. These experimentally optimized beam weights were integrated into the OpenGATE and GGEMS Monte Carlo toolkits, significantly advancing dosimetric simulation accuracy and the image quality which closely resembles the clinical imaging. Validation with fluence and dose actors demonstrated that the AI-based model closely mirrors clinical beam behavior, providing substantial improvements in dose consistency and accuracy over conventional X-ray models. This approach provides a robust framework for improving X-ray dosimetry, with potential applications in dose optimization, imaging quality enhancement, and radiation safety in both clinical and research settings. Keywords: Monte Carlo Simulation, Artificial Intelligence, Anode Heel Effect, Dosimetry, Medical Imaging, Medical Physics. 1 Introduction X-ray imaging plays a pivotal role in both diagnostic and interventional medicine. Due to its ionizing nature, optimizing imaging systems, protocols, and reconstruction methods is crucial to maintaining high image quality while minimizing patient exposure. In this context, Monte Carlo simulation remains an indispensable tool in medical physics, offering unmatched accuracy in modeling photon-matter interactions and dose deposition [ 1 , 2 ] . This approach allows the simulation of the entire X-ray tube system [ 3 ] , enabling the generation of highly realistic photon spectra and spatial distributions. Nevertheless, Monte Carlo simulations are computationally intensive. A common approach to reduce computational costs is to divide the simulation into two phases: (1) a simulation of the X-ray tube alone, generating a phase space file, and (2) a patient-dependent phase where this file is reused, eliminating the need for repeated tube simulations. However, for simulations that require a large number of particles and extensive statistics, the phase space file can reach hundreds of gigabytes, making the simulation impractical. Generally, a virtual source is preferred. This refers to an analytical mathematical model that represents the distribution and correlation of particles in the phase space [ 4 ] . Such model requires only few parameters saving stored memory and can rapidly produce particle distributions on demand. The main drawback is that such models are often an approximation and only represents the predominant characteristics of the source. For example, the anode heel effect of the X-ray tube is rarely considered. This affect the beam intensity asymmetry arising from the angled geometry of the X-ray tube anode. This effect causes X-rays directed toward the cathode side to exhibit higher intensities, while those toward the anode side are attenuated due to self-absorption within the anode [ 5 , 6 ] . This asymmetry impacts both dose distribution and image quality, particularly in wide-field imaging applications [ 7 ] . Neglecting the anode heel effect in MC simulations can lead to significant discrepancies between simulated and real measurements [ 8 ] . Recent studies have explored various approaches to model the anode heel effect, including empirical adjustments and analytical formulations [ 9 ] . However, these methods often lack flexibility or fail to account for energy-dependent variations in beam intensity. Advances in artificial intelligence (AI) present a promising solution for capturing complex, non-linear phenomena like the anode heel effect. Machine learning models, trained on experimental data, can generalize across a wide range of energy levels and beam geometries, offering a scalable alternative to traditional methods [ 10 ] . In this study, we propose an AI-driven approach to accurately model the anode heel effect and integrate it into MC simulations. By combining experimental measurements with a Gradient Boosting Regression model, we predict beam intensity distributions across energy levels ranging from 50 to 120 kVp. These predictions are seamlessly incorporated into the OpenGATE [ 2 ] and GGEMS [ 11 ] MC toolkits, enabling precise simulation in both research and clinical settings. Our approach not only addresses a critical gap in X-ray beam modeling but also lays the groundwork for future advancements in radiation safety and imaging optimization. 2 Materials and Methods In this study, we modeled a virtual source that replicate the characteristics and behavior of a real X-ray tube. This tube consists of a cathode that emits electrons through thermionic emission and a tungsten anode where these electrons are accelerated under high voltage and converted into X-ray photons via bremsstrahlung and characteristic interactions [ 12 , 13 ] . Tube voltages ranging from 50 to 120 kVp were employed to simulate various clinical imaging scenarios. The anode’s angled geometry plays a critical role in producing the non-uniform intensity distribution of the beam, contributing to the anode heel effect. 2.1 Standard X-ray tube virtual source model A virtual source model consists of mathematically modeling the emission of photons from the tube. In Monte Carlo simulations, for each randomly emitted particle, the parameters of the photon have to be determined. Each photon \gamma(E,\mathbf{p},\mathbf{d}) , is defined by its energy E , its position \mathbf{p}=\{p_{x},p_{y},p_{z}\}^{T} , and its momentum (direction) \mathbf{d}=\{d_{x},d_{y},d_{z}\}^{T} . The position \mathbf{p} is defined by the emission area on the anode, which corresponds to a rectangular surface measured in millimeters (mm). For each new particle, this position is determined randomly and uniformly over this surface, \mathbf{p}\sim\mathcal{U}(S) , where \mathcal{U} represents the uniform distribution, and S denotes the surface. The particle’s momentum is defined by the Euler angles (\alpha:altitude,\beta:azimuthal) of the particle emission. In the standard model particles are emitted following an angular uniform distribution. In this case the random sampling should not be done over the altitude angle \alpha , but rather over its cosine \cos(\alpha) . This can be expressed as: \frac{d\Omega}{d\cos\alpha}=1 (1) where d\Omega is the differential solid angle and \cos(\alpha) is the cosine of the polar angle \alpha . Therefore, the momentum direction \mathbf{d} is sampled randomly with a probability distribution proportional to \mathcal{U}(\cos\alpha) . The azimuthal angle \beta is also sampled randomly and uniformly between 0 and 2\pi : \beta\sim\mathcal{U}(0,2\pi) . However, the angle of the beam’s aperture \mathcal{A} must also be taken into account. The angle \alpha is then obtained as follows: \alpha=\cos^{-1}\left(1+\xi\cdot\left[\cos(\mathcal{A})-1\right]\right) (2) where \xi is a uniform random number between [0,1) . Based on the two angles \alpha and \beta , the direction of the particle \mathbf{d} can be defined as follows: \mathbf{d}=\left[\begin{array}[]{c}\cos\beta\cdot\sin\alpha\\ \sin\beta\cdot\sin\alpha\\ \cos\alpha\end{array}\right] (3) Most of the time, the tube is also oriented around the patient, then a transformation including rotation and translation must be applied to the particle: \left\{\begin{array}[]{l}\mathbf{p^{\prime}}=\mathbf{T}_{xyz}\mathbf{p}\\ \mathbf{d^{\prime}}=\mathbf{R}_{\psi\theta\phi}\mathbf{d}\end{array}\right. (4) where \mathbf{p^{\prime}} is the final particle emitting position, \mathbf{d^{\prime}} is the final particle momentum, \mathbf{R} and \mathbf{T} are the rotation and translation matrices of the X-ray tube respectively. Collimation can also be applied, whether to achieve a fan beam in diagnostic CT or to reduce the field of view in fluoroscopy. Two strategies are

Read the complete text