An efficient method for determining the mass attenuation coefficient of contrast agent-water mixtures and analysis of influencing factors in medical imaging systems.
Authors
Affiliations (3)
Affiliations (3)
- School of Nuclear Science and Technology, University of South China, Hengyang, 421001, China.
- School of Nuclear Science and Technology, University of South China, Hengyang, 421001, China. Electronic address: [email protected].
- Hepatobiliary Surgery Department, The First Affiliated Hospital of University of South China, Hengyang, 421001, China.
Abstract
The mass attenuation coefficient (MAC) plays a key parameter in computed tomography (CT) imaging and the development of novel contrast agents. In practical applications, the MAC depends not only on photon energy but also on complex, interrelated factors such as material composition, contrast agent concentration, and mixture density, exhibiting pronounced nonlinear characteristics. Existing theoretical databases and pointwise simulation methods are often inadequate for rapid, high-resolution predictions across broad parameter spaces. In this work, we propose a new, high-throughput MAC prediction framework that integrates high-resolution Geant4 Monte Carlo simulations with a random forest machine learning regression model. The X-ray attenuation properties of contrast agent-water mixtures were systematically modeled under varying photon energies, molecular weights, concentrations, electron densities, and bulk densities, producing a comprehensive and high-resolution simulation dataset within clinically relevant energy and concentration ranges. Using five key physical parameters as input, the random forest model was trained with normalized preprocessing, grid search hyperparameter optimization, and K-fold cross-validation, yielding good generalization and predictive performance. The proposed method demonstrated strong agreement with Monte Carlo simulations for both training and test sets, as indicated by R<sup>2</sup>, MAE, and RMSE metrics, while improving computational efficiency for large-sample, multivariable cases. Compared with conventional approaches, this study establishes-for the first time-a high-resolution simulation dataset encompassing the major parameter ranges relevant to CT imaging, and realizes a stable, scalable MAC prediction framework based on machine learning. The developed framework enables automated, high-throughput MAC estimation across vast, multi-dimensional parameter spaces, facilitating real-time adaptation to diverse clinical or engineering scenarios.