MCDNet – A Denoising Convolutional Neural Network to Accelerate Monte Carlo Radiation Transport Simulations: A Proof of Principle With Patient Dose From X-Ray CT Imaging

Abstract

X-ray computed tomography (CT) is a popular diagnostic imaging tool that has caused public concern over potential radiation risks to the patient. Monte Carlo (MC) simulations are the most accurate methods to calculate x-ray interactions with the patient's body and voxel-wise dose distributions, but the statistical methods suffer from extremely long computing time that is required to achieve necessary statistical precision. In this paper, we propose and demonstrate the Monte Carlo Denoising Net (MCDNet), a convolutional encoder-decoder neural network, for the purpose of accelerating the MC radiation transport simulations for patient CT dosimetry. A unique set of full-body anatomically realistic adult voxel phantoms of various sizes and a GPU-based parallel MC code were used to produce adequate training and testing data for supervised learning. Gamma index passing rate (GIPR) was used to evaluate the performance of predicted dose maps. For CT scan protocols considered in this study, MCDNet is found to have the ability of predicting dose maps of $9.9\times 10 ^{\mathbf {7}}$ photons from corresponding dose maps of $1.3\times 10^{\mathbf {6}}$ photons, yielding $76\times $ speed-up in terms of photon numbers used in the MC simulations. MCDNet is the first CNN-based method to speed-up MC radiation transport simulations involving 3D and heterogeneous patient anatomies for x-ray CT. Future studies will test the feasibility of applying the deep-learning based denoising strategy to other MC radiation transport applications.

Keywords

• CT dose
• Monte Carlo simulation
• dose map
• machine learning
• deep neural network
• Gamma Index Passing Rate