Uncertainty analysis of MR-PET image registration for precision neuro-PET imaging

Pawel J Markiewicz,Julian C Matthews,John Ashburner,David M Cash,David L Thomas,Enrico De Vita,Anna Barnes,M Jorge Cardoso,Marc Modat,Richard Brown,Kris Thielemans,Casper Da Costa-Luis,Isadora Lopes Alves,Juan Domingo Gispert,Mark E Schmidt,Paul Marsden,Alexander Hammers,Sebastien Ourselin,Frederik Barkhof

doi:10.1016/j.neuroimage.2021.117821

Abstract

Accurate regional brain quantitative PET measurements, particularly when using partial volume correction, rely on robust image registration between PET and MR images. We argue here that the precision, and hence the uncertainty, of MR-PET image registration is mainly driven by the registration implementation and the quality of PET images due to their lower resolution and higher noise compared to the structural MR images. We propose a dedicated uncertainty analysis for quantifying the precision of MR-PET registration, centred around the bootstrap resampling of PET list-mode events to generate multiple PET image realisations with different noise (count) levels. The effects of PET image reconstruction parameters, such as the use of attenuation and scatter corrections and different number of iterations, on the precision and accuracy of MR-PET registration were investigated. In addition, the performance of four software packages with their default settings for rigid inter-modality image registration were considered: NiftyReg, Vinci, FSL and SPM. Four distinct PET image distributions made of two early time frames (similar to cortical FDG) and two late frames using two amyloid PET dynamic acquisitions of one amyloid positive and one amyloid negative participants were investigated.For the investigated four PET frames, the biggest impact on the uncertainty was observed between registration software packages (up to 10-fold difference in precision) followed by the reconstruction parameters. On average, the lowest uncertainty for different PET frames and brain regions was observed with SPM and two iterations of fully quantitative image reconstruction. The observed uncertainty for the varying PET count-level (from 5% to 60%) was slightly lower than for the reconstruction parameters. We also observed that the registration uncertainty in quantitative PET analysis depends on amyloid status of the considered PET frames, with increased uncertainty (up to three times) when using post-reconstruction partial volume correction. This analysis is applicable for PET data obtained from either PET/MR or PET/CT scanners.

Highlights

1 joint senior authorship.patient, and reconstruction-specific factors Frey et al (2012)
The MR-positron emission tomography (PET) registration uncertainty is presented across nine ROIs in Fig. 6, using the metric of standard deviation (SD) of the distributions of the Dice coefficient for all reconstruction setups, the four registration software packages and the four PET frames
The observed difference in precision between software implementations was up to 10-fold, which was observed for the temporal lobe of the early frame of negative amyloid scan reconstructed with quantitative reconstruction (QNT) ordered subsets expectation maximization (OSEM) and two iterations

Summary

Introduction

1 joint senior authorship.patient-, and reconstruction-specific factors Frey et al (2012). The leading physical aspects limiting the precision of PET are the radiotracer together with image noise and limited spatial resolution ( referred to as partial volume effects—PVEs) Barrett et al (1994); Erlandsson et al (2012); Hooker and Carson (2019). The spatial distribution of various neuro PET radiotracers—combined with the PET noise and limited resolution—can negatively affect the precision of the MR-to-PET (MRPET) rigid-body registration and further limit the precision of PET quantification it is important to determine the uncertainty of MR-PET registration and how this propagates to the accuracy and robustness of the quantitative parameters. The effect of MR-PET registration uncertainty is illustrated, where the same brain with two noisy PET image realisations is considered, A and B. Despite the considerable overlap of the cortical regions shown in blue, there is a significant number of voxels that do not overlap (shown in white and red for realisations A and B, respectively)

Methods

Results

Conclusion