Fast detector timing resolutions with GANs

Researchers are working on improving radiation detector coincidence timing resolution to enhance time-of-flight positron emission tomography (PET) and improve image signal-to-noise ratio. Large system simulations are necessary for testing new radiation detector technologies in a scanner setting. As timing resolution strongly depends on optical transport in the detector, using detailed optical photon transport modeling in simulations might be necessary but it is time and memory-intensive, making it impractical for system-level studies. To address this, a conditional Wasserstein Generative Adversarial Network (GAN), called optiGAN, was developed to generate multidimensional optical photon distributions on the photodetector of a radiation detector without conducting individual photon tracking. The accuracy of the GAN-generated distributions has already been demonstrated by similarity values higher than 93.5% when compared to distributions obtained with traditional particle tracking simulation. This work includes the optical photon timing distribution at the photodetector face within the training dataset generated through accurate optical Monte Carlo simulations with GATE. This enables the generation of multidimensional distributions useful for studying radiation detector timing performance. The GAN-generated distributions have been used as input of an experimentally-verified SiPM model to produce detector electrical signals, allowing the estimation of the coincidence timing resolution from realistic simulation data. The optiGAN-generated time distributions show excellent agreement with the training dataset distributions, with similarity values higher than 92% for all different points where the distributions were generated, with a speed gain of up to two orders of magnitude. The SiPM signals from tracking simulations or GAN-generated distributions show very good agreement, demonstrating that GAN-generated distributions can be used to perform timing resolution studies.