Modeling families of particle distributions with conditional GAN for Monte Carlo SPECT simulations

Abstract

Objective. We propose a method to model families of distributions of particles exiting a phantom with a conditional Generative Adversarial Network (condGAN) during Monte Carlo simulation of SPECT imaging devices. Approach. The proposed condGAN is trained on a low statistics dataset containing the energy, the time, the position and the direction of exiting particles. In addition, it also contains a vector of conditions composed of four dimensions: the initial energy and the position of emitted particles within the phantom (a total of 12 dimensions). The information related to the gammas absorbed within the phantom is also added in the dataset. At the end of the training process, one component of the condGAN, the generator (G), is obtained. Main results. Particles with specific energies and positions of emission within the phantom can then be generated with G to replace the tracking of particle within the phantom, allowing reduced computation time compared to conventional Monte Carlo simulation. Significance. The condGAN generator is trained only once for a given phantom but can generate particles from various activity source distributions.

The OpenGATE ecosystem for Monte Carlo simulation in medical physics

“…Proofs of concept have been published but there are still a lot of unknowns and uncertainties in those approaches that remain to be studied. There is, in particular, some ongoing work on conditional GANs that may allow to model a family of GANs, and to avoid time-consuming re-training (Saporta et al 2022).…”

Annihilation photon GAN source model for PET Monte Carlo simulation

“…In previous works (Sarrut et al 2019, 2021a, Saporta et al 2022, it has been shown that Generative Adversarial Network (GAN) can model phase-spaces such that the trained generator neural network approximates a distribution of particles, thereby being able to be used as a fast virtual source model (VSM,e ) of particles. This approach was applied for modeling photon beams from Linac head in radiation therapy treatment (Sarrut et al 2019) and for gammas exiting phantoms or patient Computerized Tomography (CT) during simulation of Single Photon Emission Computerized Tomography (SPECT) systems (Sarrut et al 2021a, Saporta et al 2022. Compared to conventional phase-space files, GAN generators are compact, a few MB instead of a few GB, and can generate particles at high speed (around 10 6 particles per second).…”

“…In previous works [1][2][3], it has been shown that Generative Adversarial Network (GAN) can model phase-spaces such that the trained generator neural network approximates a distribution of particles, thereby being able to be used as a fast virtual source model (VSM,e ) of particles. This approach was applied for modeling photon beams from Linac head in radiation therapy treatment [1] and for gammas exiting phantoms or patient CT during simulation of SPECT systems [2; 3].…”

“…2.4. Conditional GAN trainingFollowing previous works on sourceGAN (Sarrut et al 2019, 2021a, Saporta et al 2022, a GAN was developed to learn generating exiting gammas following the initial distributions. The two neural networks of the GAN, the Critic (C) and the Generator (G) are trained simultaneously and act as follows.…”