A generative adversarial network to speed up optical Monte Carlo simulations

Abstract

Detailed simulation of optical photon transport and detection in radiation detectors is often used for crystal-based gamma detector optimization. However, the time and memory burden associated with the track-wise approach to particle transport and detection in commonly used Monte Carlo codes makes optical simulation prohibitive at a system level, where hundreds to thousands of scintillators must be modeled. Consequently, current large system simulations do not include detailed detector models to analyze the potential performance gain with new radiation detector technologies. Generative Adversarial Networks are explored as a tool to speed up the optical simulation of crystal-based detectors. These networks learn training datasets made of high-dimensional data distributions. Once trained, the resulting model can produce distributions belonging to the training data probability distribution. In this work, we present the proof of concept of using a Generative Adversarial Network to enable high-fidelity optical simulations of nuclear medicine systems, mitigating their computational complexity. The architecture of the first network version and high-fidelity training dataset is discussed. The latter is generated through accurate optical simulation with GATE/Geant4, and contains the position, direction, and energy distributions of the optical photons emitted by 511 keV gamma rays in bismuth germanate and detected on the photodetector face. We compare the GAN and simulation-generated distributions in terms of similarity using the Jensen-Shannon distance. Excellent agreement was found with similarity values higher than 93.5% for all distributions. Moreover, the GAN speeded the optical photon distribution generation by up to 2 orders of magnitude. These very promising results have the potential to drastically change the use of nuclear imaging system optical simulations by enabling high-fidelity system-level simulations in reasonable computation times. The ultimate is to integrate the GAN within GATE/Geant4 since numerous applications (large detectors, bright scintillators, Cerenkov-based timing positron emission tomography) can benefit from these improvements.

GPU optimization techniques to accelerate optiGAN—a particle simulation GAN

“…Using some effective GPU optimization techniques we were able to accelerate the training of the optiGAN (Trigila et al 2023 ) model without any loss in the model’s performance measured by the Jensen–Shannon similarity value. Figure 9 shows the comparison plot of all the techniques with different runtime regions during the training.…”

“…This data was concatenated to form a single training matrix with nine columns and 4.2 million rows, allowing us to train the Wasserstein Conditional GAN. More details can be found in Trigila et al ( 2023 ).…”

“… OptiGAN architecture (from Trigila et al ( 2023 )). It consists of a Generator (a) and a Discriminator/Critic network (b) with H = 128 hidden nodes and a ReLU activation function.…”

“…Different types of GAN models are used to produce different data-types such as texts, images, numerical, speeches (Yu et al 2017, Isola et al 2018, Xu and Veeramachaneni 2018, Donahue et al 2019. One such application of a GAN model is optiGAN, a GAN capable of generating complex multidimensional probability distributions (Trigila et al 2023). The goal of optiGAN is to mitigate the computational complexity of optical Monte Carlo simulations.…”

“…The final goal is to directly integrate the optiGAN within GATE/Geant4 to skip the optical photon tracking process and speed up the optical Monte Carlo simulations. Numerous applications (large detectors, bright scintillators, Cerenkov-based timing positron emission tomography) can benefit from these improvements (Trigila et al 2023 ).…”