Bootstrap methods for estimating PET image noise: experimental validation and an application to evaluation of image reconstruction algorithms

Abstract

Accurate and validated methods for estimating regional PET image noise are helpful for optimizing image processing. The bootstrap is a data-based simulation method for statistical inference, which can be used to estimate the PET image noise without repeated measurements. The aim of this study was to experimentally validate bootstrap-based methods as a tool for estimating PET image noise and demonstrate its usefulness for evaluating image reconstruction algorithms. Two bootstrap-based method, the list-mode data bootstrap (LMBS) and the sinogram bootstrap (SNBS), were implemented on a clinical PET scanner. A uniform cylindrical phantom filled with (18)F solution was scanned using list-mode acquisition. A reference standard deviation (SD) map was calculated from 60 statistically independent measured list-mode data. Using one of the 60 list-mode data, 60 bootstrap replicates were generated and used to calculate bootstrap SD maps. Brain (18)F-FDG data from a healthy volunteer were also processed as an example of the bootstrap application. Three reconstruction algorithms, FBP 2D and both 2D and 3D versions of dynamic row-action maximum likelihood algorithm (DRAMA), were assessed. For all the reconstruction algorithms used, the bootstrap SD maps agreed well with the reference SD map, confirming the validity of the bootstrap methods for assessing image noise. The two bootstrap methods were equivalent with respect to the performance of image noise estimation. The bootstrap analysis of the FDG data showed the better contrast-noise relation curve for DRAMA 3D compared to DRAMA 2D and FBP 2D. The bootstrap methods provide the estimates of image noise for various reconstruction algorithms with reasonable accuracy, require only a single measurement, not repeated measures, and are, therefore, applicable for a human PET study.