Using the bootstrap method to evaluate image noise for investigation of axial collimation in hybrid PET

Abstract

The purpose of this work was to investigate the use of the bootstrap method to evaluate noise in maximum-likelihood, ordered subsets list-mode iterative reconstructions from data simulated on a hybrid PET system. In addition, image noise from reconstructions acquired with various levels of axial collimation was investigated. A warm cylinder (20-cm diameter and 70-cm long) was simulated using the GEANT 3.21 Monte-Carlo simulation code for different collimator designs: open-frame, parallel axial (PAR), and hybrid parallel fan-beam (HPF) axial collimators. Ten noisy sets of list-mode data were simulated for each collimator. In addition, events from the first noisy list-mode data set of each category (open-frame, PAR axial and HPF axial collimators) were chosen at random with replacement to produce ten new sets of the same size. List-mode data were reconstructed using maximum-likelihood, ordered-subsets list-mode iterative reconstruction using 2 iterations and 5 subsets. The total counts within volumes of interest (VOI) were computed, and the variance of this estimate over noise realizations was evaluated. The F-test was used to test statistical significance between variances of VOI estimates from noisy and bootstrap samples. The F-test results indicate that there is no significant difference in the variance of simulations and bootstrap data when using the open-frame, the PAR axial and the HPF axial collimators. The test showed that the precision of estimates vary with the axial spatial location within the object and the axial collimator. In addition, noise variation within list-mode reconstructed images appears to be related to the 3D sensitivity map.