Title:Fast and memory-efficient reconstruction of sparse Poisson data in listmode with non-smooth priors with application to time-of-flight PET

Abstract:In this work, we propose and analyze a listmode (LM) version of the stochastic primal-dual hybrid gradient (SPDHG) algorithm for image reconstruction of sparse data following a Poisson distribution using subsets and non-smooth priors. Since data from state-of-the-art time of flight (TOF) PET scanners are huge but at the same time extremely sparse, listmode based algorithms, as the proposed LM-SDPHG algorithm, substantially reduces both memory requirements and computation time. Based on realistic simulations in 2D and 3D, we show that the speed of convergence of LM-SPDHG is the same as the one of the original SPDHG operating on binned data (sinograms). For a state-of-the-art TOF PET system with 400 ps TOF resolution and 25 cm axial FOV, LM SPDHG reduces the required memory from approximately 56 GB to 0.7 GB for a short dynamic frame with 1e7 prompt coincidences and to 12.4 GB for a long static acquisition with 5e8 prompt coincidences. In contrast to SPDHG, the reduced memory requirements of LM-SPDHG consequently enables a pure GPU implementation using state-of-the-art GPUs which will accelerate reconstruction times even further.