Conditional GAN for Monte Carlo SPECT simulation

International audienceIn a previous work [1], a method was proposed to model by a Generative Adversarial Network (GAN) [2] the distribution of particles exiting a patient or a phantom during Monte Carlo simulation of SPECT imaging devices. This approach allows reduced computation time (few seconds) compared to conventional Monte Carlo simulation (few minutes), as there is no need to track again the particles within the phantom. The file containing the GAN parameters is smaller (few MB) than the phase space file (few GB). However, this approach requires training a new GAN each time a parameter is modified. In this work, we extend the architecture developed in [1] such that the GAN has to be trained only once for a family of activity distributions[1] D. Sarrut, A. Etxebeste, N. Krah, and J. Létang, “Modeling Complex Particles Phase Space with Gan for Monte Carlo SpectSimulations: A Proof of Concept,”Physics in Medicine & Biology, vol. 66,p. 055014, Mar. 2021[2] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville and Y. Bengio. Generative adversarialnets. In Advances in Neural Information Processing Systems, volume 2, pages 2672–2680, 2014