Noise Aware System Generative Model (NASGM) of PET: a deep learning-based model for PET image simulation with quantitative assessment

Positron emission tomography (PET) simulation is widely applied in the development and validation of image processing algorithms as well as the optimization of imaging protocols because of its reproducibility and ability to provide ground truth. However, the simulation of PET images is generally time and computationally intensive. Furthermore, the non-differentiable nature of stochastic forward projection and iterative image reconstruction algorithms hinders our access to the underlying object space. To address these issues, we develop a deep learning-based PET simulation method, the noise-aware system generative model (NASGM). Specifically, we introduce a novel dual-domain discriminator to build a conditional generative adversarial network that processes activity and attenuation maps as inputs and generates simulated PET images by learning various noise characteristics of simulated PET images for different acquisition times. To prepare the dataset, a publicly available PET/CT clinical dataset is utilized as activity and attenuation maps, and FAST-PET (fast analytical simulator of tracer-PET), an analytical PET simulation tool we developed, is applied to simulate PET images. Validations in terms of recovery coefficient and normalized mean absolute error show that the NASGM-generated images have the highest quantitative accuracy and noise levels most closely matching those of the traditionally simulated images compared to other networks. In conclusion, the proposed NASGM functions as an efficient PET simulator that models the entire PET imaging and reconstruction system using a differentiable neural network and bridges the gap between the object and the image domain.