Investigation of the Effects of Image Signal-to-Noise Ratio on TSPO PET Quantification of Neuroinflammation

Abstract

Neuroinflammation may be imaged using positron emission tomography (PET) and the tracer [11C]-PK11195. Accurate and precise quantification of 18 kilodalton Translocator Protein (TSPO) binding parameters in the brain has proven difficult with this tracer, due to an unfavourable combination of low target concentration in tissue, low brain uptake of the tracer and relatively high non-specific binding, all of which leads to higher levels of relative image noise. To address these limitations, research into new radioligands for the TSPO, with higher brain uptake and lower non-specific binding relative to [11C]-PK11195, is being conducted world-wide. However, factors other than radioligand properties are known to influence signal-to-noise ratio in quantitative PET studies, including the scanner sensitivity, image reconstruction algorithms and data analysis methodology. The aim of this thesis was to investigate and validate computational tools for predicting image noise in dynamic TSPO PET studies, and to employ those tools to investigate the factors that affect image SNR and reliability of TSPO quantification in the human brain. The feasibility of performing multiple (n≥40) independent Monte Carlo simulations for each dynamic [11C]-PK11195 frame- with realistic modelling of the radioactivity source, attenuation and PET tomograph geometries- was investigated. A Beowulf-type high performance computer cluster, constructed from commodity components, was found to be well suited to this task. Timing tests on a single desktop computer system indicated that a computer cluster capable of simulating an hour-long dynamic [11C]-PK11195 PET scan, with 40 independent repeats, and with a total simulation time of less than 6 weeks, could be constructed for less than 10,000 Australian dollars. A computer cluster containing 44 computing cores was therefore assembled, and a peak simulation rate of 2.84x105 photon pairs per second was achieved using the GEANT4 Application for Tomographic Emission (GATE) Monte Carlo simulation software. A simulated PET tomograph was developed in GATE that closely modelled the performance characteristics of several real-world clinical PET systems in terms of spatial resolution, sensitivity, scatter fraction and counting rate performance. The simulated PET system was validated using adaptations of the National Electrical Manufacturers Association (NEMA) quality assurance procedures within GATE. Image noise in dynamic TSPO PET scans was estimated by performing n=40 independent Monte Carlo simulations of an hour-long [11C]-PK11195 scan, and of an hour- long dynamic scan for a hypothetical TSPO ligand with double the brain activity concentration of [11C]-PK11195. From these data an analytical noise model was developed that allowed image noise to be predicted for any combination of brain tissue activity concentration and scan duration. The noise model was validated for the purpose of determining the precision of kinetic parameter estimates for TSPO PET. An investigation was made into the effects of activity concentration in tissue, radionuclide half-life, injected dose and compartmental model complexity on the reproducibility of kinetic parameters. Injecting 555 MBq of carbon-11 labelled TSPO tracer produced similar binding parameter precision to 185 MBq of fluorine-18, and a moderate (20%) reduction in precision was observed for the reduced carbon-11 dose of 370 MBq. Results indicated that a factor of 2 increase in frame count level (relative to [11C]-PK11195, and due for example to higher ligand uptake, injected dose or absolute scanner sensitivity) is required to obtain reliable binding parameter estimates for small regions of interest when fitting a two-tissue compartment, four-parameter compartmental model. However, compartmental model complexity had a similarly large effect, with the reduction of model complexity from the two-tissue compartment, four-parameter to a one-tissue compartment, two-parameter model producing a 78% reduction in coefficient of variation of the binding parameter estimates at each tissue activity level and region size studied. In summary, this thesis describes the development and validation of Monte Carlo methods for estimating image noise in dynamic TSPO PET scans, and analytical methods for predicting relative image noise for a wide range of tissue activity concentration and acquisition durations. The findings of this research suggest that a broader consideration of the kinetic properties of novel TSPO radioligands, with a view to selection of ligands that are potentially amenable to analysis with a simple one-tissue compartment model, is at least as important as efforts directed towards reducing image noise, such as higher brain uptake, in the search for the next generation of TSPO PET tracers.