Noise and bias characteristics of standardized uptake value (SUV) derived with point spread function (PSF) image reconstruction: should PSF be used for PET tumor uptake quantification ?

Martin Lodge, Ronald Boellaard

Research output: Contribution to journalArticleAcademicpeer-review

Abstract

614Objectives: PET image reconstruction with PSF modelling improves spatial resolution and may lead to better performance for clinical visual assessment tasks. Unfortunately SUV quantification from PSF images is generally not consistent with conventional (non-PSF) reconstruction. Tumor SUVmax is typically overestimated and may have higher variability, meaning that PSF images are not directly compatible with existing guidelines for tumor response assessment (e.g. QIBA FDG PET/CT Profile, https://www.rsna.org/qiba/). Using phantom images with and without PSF we assessed aspects of noise and bias for two different SUV implementations: SUVmax and SUVpeak.Methods: The NEMA image quality phantom was prepared with an initial 18F background concentration of 5.3 kBq/mL and a 4-to-1 sphere-to-background ratio. Five sequential scans were performed using a Biograph mCT, each for 3 min / bed position. Images were reconstructed using a conventional non-PSF protocol (OSEM+TOF, 2i, 21s, 5 mm Gaussian) and a high resolution PSF protocol (OSEM+TOF+PSF, 2i, 21s, no post smoothing). SUVmax and SUVpeak were determined for each sphere using commercial software (Syngo Via) that implements the peak region as a fixed size 1 mL spherical volume positioned so as to maximize the enclosed average. Recovery coefficients (RCs) were calculated as SUVT / (4×SUVB) where SUVT was either SUVmax or SUVpeak and SUVB was the mean SUV in the background. For each reconstruction protocol, the 5 image replicates were used to calculate the coefficient of variation (CoV) for both SUVmax and SUVpeak as a function of sphere size.Results: RCs for SUVmax in conjunction with PSF were higher than the upper limit recommended by EARL (Boellaard et al, EJNMMI 2015) for all spheres. The 37 mm sphere (expected to be minimally influenced by partial volume) had an RC of 1.21 ± 0.03. Edge or Gibbs artifacts are a problematic feature of PSF, particularly when using SUVmax, and led to an even greater bias for the smaller 17 mm sphere, RC of 1.32 ± 0.08. When SUVpeak was used in conjunction with PSF, recovery was consistent with the EARL range for all spheres. The greater volume averaging of SUVpeak suppressed the effect of PSF Gibbs artifacts, resulting in RCs of 1.06 ± 0.02 and 0.92 ± 0.02 for the 37 and 17 mm spheres respectively. CoV for SUVmax was poorer with PSF compared to non-PSF, ranging from 2.3 - 11.6 % (PSF) and 1.7 - 9.1 % (non-PSF) for the largest to the smallest spheres. For SUVpeak & PSF, CoV ranged from 2.1 - 5.3 %, indicating that test-retest repeatability in clinical imaging is likely to be at least as good for PSF & SUVpeak as for the conventional approach involving non-PSF & SUVmax.Conclusion: This study indicates that SUVpeak may be preferable to SUVmax when image reconstruction includes PSF modelling. SUVpeak improves the quantitative characteristics of PSF images, reducing positive bias and decreasing the variability of tumor SUVs. Multiple implementations of PSF are recognized and further testing is ongoing but these data suggest that SUVpeak could enable improved tumor quantification for high resolution PSF images. Research Support: This work was funded by a grant from the National Institutes of Health (HHSN268201500021C).
Original languageEnglish
Number of pages1
JournalJournal of Nuclear Medicine
Volume58
Publication statusPublished - 2017

Cite this