Computational Method for Harmonization of Brain PET Images

Felix Carbonell, Ph.D.¹, Alex P. Zijdenbos, Ph.D.¹, Evan Hempel ², Mihály Hajós, Ph.D.², Barry J. Bedell, M.D., Ph.D.^1,3

¹ Biospective Inc., Montréal, QC, Canada
² Cognito Therapeutics, Inc., Cambridge, MA, USA
³ Research Institute of the McGill University Health Centre, Montréal, QC, Canada

The most common approach for estimating the spatial resolution of brain positron emission tomography (PET) images in multi-center studies typically uses Hoffman phantom data as a surrogate. Specifically, the phantom-based matching resolution approach assumes that scanned phantom PET images are well approximated by a ground truth, noise-free digital phantom convolved with a Gaussian kernel of unknown size. The size of the kernel is then estimated by an exhaustive search on the amount of blurring needed to match the smoothed digital phantom to a particular scanned phantom image.

Unfortunately, Hoffman phantom images may not always be readily available, and phantom-based approaches may yield sub-optimal results. To overcome these limitations, we propose a new, computational approach, called SPITFIRE™, that allows estimation of spatial resolution directly from the PET image itself. We generalized the so-called logarithmic intensity plots method to the 3D case to perform a spatial resolution estimation in both axial and in-plane directions of the PET images.

Our approach was applied to two different cohorts. The first cohort consisted of [18F]florbetapir amyloid PET images and matching phantoms coming from a Phase II clinical trial and includes different scanner models and/or orientation and grid reconstructions. The second cohort included β-amyloid, tau, and FDG PET images from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) study.

We obtained in-plane and axial resolution estimators that vary between 3.5 mm and 8.5 mm for both PET and matching phantom images. In both cases, we obtained small across-subject variability in groups of images sharing the same PET scanner model and reconstruction parameters. For human PET images, we also obtained a strong cross-tracer and longitudinal consistency in the spatial resolution estimators.

Our novel approach does not only eliminate the need for surrogate brain phantom data, but also provides a general framework that can be applied to a wide range of tracers and other image modalities, such as SPECT.