Improving a Stochastic Algorithm for Regularized PET Image Reconstruction

Positron Emission Tomography (PET) image reconstruction presents challenges related to the large scale of data to be processed, which affects reconstruction speed, and the need to include regularizers to improve image quality. Among the methods proposed to overcome these challenges, the recently introduced Stochastic Primal Dual Hybrid Gradient (SPDHG) algorithm combines the ability to deal with regularizers like Total Variation and to process large datasets by random subsampling. We present two contributions regarding the step-sizes of SPDHG: i) larger step-sizes facilitated by a new formula, and ii) a numerical method to calibrate, in the context of PET reconstruction, the tradeoff between primal and dual progression, which is common to all primal-dual algorithms. We validate improvements in speed reconstruction on real PET data from the Siemens Biograph mMR.