Phantom measurement predicts the impact of image reconstruction on Florbetapir SUVR quantification

Abstract

Objectives: PET quantification of Αß is a critical component of research studies and clinical trials into Alzheimer’s disease (Sperling et al, Sci Transl Med 2014). Due to the large numbers of subjects in these studies, PET imaging data is typically multi-site and multi-vendor, giving rise to potentially significant heterogeneity in data acquisition. To date, there has been limited investigation into the role that image reconstruction settings play in the quantification of Αß. Our objectives were therefore to describe SUVr variation caused by reconstruction parameter choices, and further to determine the utility of resolution estimates from a Gallium-68 phantom (Lodge et al, J Nuc Med 2018) in predicting SUVr variation. The specific aims of the study were as follows: 1. To demonstrate the effect of image reconstruction parameters on Αß Florbetapir SUVR quantification; 2. To characterise the relationship between Ga68 phantom-derived image resolution measurements and Aβ Florbetapir SUVr quantification; 3. To investigate the stability of Ga68 phantom-derived image resolution measurements across scanner manufacturers.

Methods: Αß Florbetapir scans from 50 Amyloid negative subjects collected on a Siemens Biograph 128 PET scanner were retrospectively analysed using an offline computing cluster. Each subject’s raw data was reconstructed using 72 different reconstruction settings, including three algorithms (OSEM, with PSF, with TOF) and a clinically relevant range of iterations, subsets and Gaussian smoothing filter parameters, for a total of 3600 reconstructions. The mean brain cortical SUVr for each reconstruction was calculated using CapAIBL (Bougeat et al, NeuroImage 2018). Phantom data was collected from a standard Gallium-68 barrel phantom rotated 90 degrees on the same scanner. To measure image resolution for each reconstruction algorithm, a cumulative Gaussian distribution function was fitted to a line profile perpendicular to the barrel edge, and the Gaussian distribution fit using MATLAB to estimate the full-width-half-maximum (FWHM). The square of the FWHM (proportional to variance) was correlated to the mean SUVr values. The barrel phantom study was repeated using a second 68-Ga barrel phantom on a GE Discovery 960 PET scanner, and reconstructed using a subset of eight of the 72 parameter combinations.

Results: As demonstrated in Fig.1A, in which the 72 reconstruction settings are ordered according to decreasing SUVr value, reconstruction settings lead to more than 10% change in SUVr values in the Amyloid negative group. The mean SUVr values for each reconstruction setting showed a strong linear relationship with the square of the Ga68 phantom FWHM (R2=0.985) (Fig.1B). Further, the FWHM of the two scanners (each using a different Ga68 phantom) were found to be strongly correlated (R2=0.993) for matched reconstruction parameter settings (Fig.1C).

Conclusions: Choices of reconstruction algorithm and parameter settings were shown to impact Αß Florbetapir SUVr quantification in an Amyloid negative control group. We have demonstrated, however, a predictive mapping between SUVr values and Ga68 phantom resolution for matched reconstruction settings, that held for two scanner types. This offers the potential to harmonise across multi-site, multi-vendor scanners for matched SUVr quantification in Αß Florbetapir PET imaging studies, using phantom resolution measurement.