Proton path tracking in proton computed tomography using machine learning.

Ramezani, Zohreh; Riazi, Zafar; Vosoughi, Naser; Alibeigi, Ehsan

doi:10.24200/nst.2025.1719.2070

Proton path tracking in proton computed tomography using machine learning.

Articles in Press

Document Type : Research Paper

Authors

zohreh ramezani ¹

zafar riazi ²

Naser vosoughi ³

Ehsan Alibeigi ⁴

¹ faculty of energy,sharif university of technology

² دکتر

³ Department of Energy Engineering,. Sharif university of technology

⁴ PhD, Radiation Application Research School Nuclear Science and Technology Research Institute, Tehran&ndash; Iran

https://doi.org/10.24200/nst.2025.1719.2070

Abstract

One of the main problems in proton computed tomography is the reduction of the spatial resolution of the image due to the multiple coulomb scattering of protons while passing through the material. To deal with this problem, it is necessary to model the most likely path of protons (MLP). In this regard, artificial neural networks, as one of the innovative solutions in the field of machine learning, have helped to improve the accuracy of estimating the most likely path of protons. In this study, an artificial neural network with adaptive moment estimation optimization algorithm was designed to estimate the most likely proton path (MLP). The training of the network was done using the data obtained from the simulation of the proton computed tomography system in Geant4, so that 60% of this data was allocated for training, 20% of the data for validation, and 20% for the model test. These data contained information such as the position of entry and exit of protons, energy deposit, exit angle of protons and 10 points forming the path for each proton. The image matrix was modified with the help of the most likely path estimated by the neural network (MLP) and the cubic spline path (CSP), and the image was reconstructed using the Filtered Back Projection algorithm. The results of this study showed that the MLP method was able to achieve a spatial resolution of 5 line pairs/cm, while this value was equal to 3 line pairs/cm for the CSP method.

Keywords

Proton computed tomography

Geant4

Most likely path

Neural network

Main Subjects