An Investigation of Stochastic Variance Reduction Algorithms for 3D Penalised PET Image Reconstruction

Twyman, Robert; Arridge, Simon; Kereta, Zeljko; Jin, Bangti; Brusaferri, Ludovica; Ahn, Sangtae; Stearns, Charles; Hutton, Brian; Burger, Irene A.; Kotasidis, Fotis; Thielemans, Kris

doi:10.36227/techrxiv.19497680.v1

Main Body.pdf (1.74 MB)

An Investigation of Stochastic Variance Reduction Algorithms for 3D Penalised PET Image Reconstruction

preprint

posted on 2022-04-07, 03:56 authored by Robert TwymanRobert Twyman, Simon ArridgeSimon Arridge, Zeljko KeretaZeljko Kereta, Bangti JinBangti Jin, Ludovica Brusaferri, Sangtae Ahn, Charles Stearns, Brian Hutton, Irene A. Burger, Fotis Kotasidis, Kris ThielemansKris Thielemans

Application of stochastic variance reduction algorithms to iterative PET reconstruction. We investigated the SAGA and SVRG algorithms for non-TOF PET image reconstruction. Both similated data and a patient data sets were used in the analysis. We found that the stochastic algorithms can improve convergence rate and eliminate behaviour, commonly know as limit cycle behaviour, from PET reconstruction within 5 epochs.