Positron emission tomography (PET) is an accurate technique for in vivo quantification of tissue function. PET tracers are molecules that are la belled with positron emitting radionuclides. The annihilation photons resulting from these decaying radionuclides can be detected accurately us ing PET and various physiological and pharmacokinetic parameters can be derived from the acquired data using tracer kinetic models. To date, various PET tracers are available for studying different tissue functions and there is an ongoing search for new tracers.
This thesis describes various methods for improving parameter estima tion during pharmacokinetic analysis, such as various filtering techniques, the use of weighting factors and non-linear regression optimisation algo rithms. Furthermore, in this thesis applicability of various pharmacokinetic models, together with simplified methods, was evaluated for tracers that play a role in the study of Parkinson's disease ([ 18 F]FP-3-CIT) and Alzheimer's disease ([ 18 F]FDDNP and [ 11 C]PIB).
Brain studies using PET are often concerned with detecting differences in neuroreceptor binding between subjects. Neuroreceptor binding is de termined using a PET tracer which shows high affinity for that neurore ceptor. First, a dynamic PET scan is made, which are a series of consecutive PET scans, in order to accurately measure the tracer uptake and clearance. Next, for each anatomical region of interest, the average time activity curve is determined (TAC). Finally, TACs are analyzed using pharmacokinetic models in order to determine the pharmacokinetic para meters. During analysis, pharmacokinetic models are fitted to the TACs using non-linear regression. Non-linear regression algorithm adjusts iter atively the pharmacokinetic parameters until a best fit is found.
Pharmacokinetic models used in PET (brain) studies describe the tracer binding in mathematical terms. Pharmacokinetic models are highly sim plified versions of reality and different simplifications are used for each tracer. The parameters of interest in these studies are the volume of dis tribution ( V T ) and the binding potential ( BP ND ). V T is the ratio of the tracer concentration in tissue relative to that in blood at equilibrium and BP ND is a measure of binding that takes into account both neuroreceptor density and affinity of the ligand for the receptor. In general, two main types of PET pharmacokinetic models can be identified: plasma input and reference tissue models. Plasma input models require accurate arte rial blood sampling during a PET scan. Arterial plasma data are then used as input function for kinetic analysis. Reference tissue models do not require arterial sampling, but rather a reference region is used as input for the regions under study. A suitable reference tissue should be similar to the region under study, but be devoid of the receptors of interest. Ref erence tissue models avoid the need for arterial sampling and, therefore, are more patient friendly and better suited for routine clinical studies. However, the accuracy of reference tissue models for a new tracer needs to be evaluated by comparison with plasma input models. In this thesis the most common pharmacokinetic models were evaluated such as: single tis sue (reversible) (1 T 2 k ), irreversible two-tissue (2 T 3 k ) and reversible two tissue (2 T 4 k ) plasma input models, together with simplified ( SRTM) and full ( FRTM) reference tissue models. BPND is the pharmacokinetic pa rameter of interest for SRTM, FRTM and (sometimes) 2 T 4 k, and V T for 1 T 2 k and (usually) 2 T 4 k . The plasma input models can also be used to es timate BPND indirectly, i.e. from the volume of distribution ratio ( DVR), according to BP ND = DVR - 1, where DVR = V target indirectly estimated BPND is comparable with BPND obtained with the reference tissue models and requires the same assumptions for the refer ence tissue.
Most pharmacokinetic models are also available in linearised versions. Linearised models may be less accurate, but are much faster and more ro bust with respect to noise than their non-linear counterparts. Linearised models are therefore ideal for rapid evaluation of kinetic parameters across the brain (i.e. at the voxel level), in particular if it is not known where abnormalities in binding can be found or when tracer binding is so hetero- geneous that abnormalities might be missed in an ROI analysis. For each application, however, accuracy of these linearised or simplified methods needs to be evaluated.
Finally, very simple semi-quantitative methods are available that can give a measure of the binding if tracer uptake has reached equilibrium. Although suffering from bias, these methods can also be useful for example in larger clinical trials or for routine clinical (e.g. diagnostic) use. In this thesis semi-quantitative methods, such as standardised uptake value ( SUV) and SUV normalized to that of the reference region ( SUV r ), were evaluated.
All studies included simulations in order to evaluate pharmacokinetic models under fully controlled conditions, such as effects of variation in rela tive flow ( R 1 = K target 1 /K reference 1 ), fractional blood volume ( V b ), binding potential ( BP ND ), and TAC noise. These simulation studies are also im portant because it is the only way to determine accuracy and precision, as the true pharmacokinetic parameters in clinical data are unknown.
Chapter 2 describes the evaluation of the effects of pharmacokinetic analysis using incorrect weighting factors, the performance of optimisa tion algorithms commonly used in PET (i.e. interior-reflective Newton methods), and a newly developed simulated annealing (SA) based me thod. Only reversible plasma input models (i.e. 1 T 2 k and 2 T 4 k ) were investigated and data were taken from [ 15 O ]H 2 O, [ 11 C ]Flumazenil and (R) -[ 11 C ]PK11195 studies. SA is a method that automatically produces appropriate new starting parameters for repeated optimisation. There fore, in contrast to the commonly used interior-reflective Newton method, SA was able to produce accurate results without the need for selecting appropriate starting values for (kinetic) parameters. The Newton method yielded biased results, unless it was modified to restart over a range of initial parameter estimates. For patient studies, where large variability can be expected, both SA and the extended Newton method provided accurate results. Small to intermediate mismatches between variance in data and weighting factors used did not significantly affect the outcome of the fits. Therefore approximately correct weighting models are required for good accuracy. It was concluded that selection of specific optimisation algorithms and weighting factors can have a large effect on the accuracy and precision of PET pharmacokinetic analyses.
In chapter 3, improvements by wavelets based denoising of (R) -[ 11 C ]PK11195 TACs are discussed. Wavelets allow for filtering frequency components at selected time intervals and could therefore be ideal for filtering TACs, because TAC noise levels are time dependent. In simulations, when using optimised settings, all wavelet filters tested reduced noise without biasing TACs. Furthermore, for both clinical and simulated data, plasma Logan V T values increased after filtering. This increase in plasma Logan V T sug gests a reduction of noise-induced bias by wavelet based denoising, as was seen during simulations. However, after filtering, no improvements were seen for reference Logan DVR outcomes. Wavelet denoising of TACs for (R) -[ 11 C ]PK11195 PET studies might therefore be especially useful when parametric Logan based V T is the parameter of interest.
Chapter 4 describes a study on the quantification of [ 18 F ]FP- 3 -CIT, a tracer of the dopamine transporter used for human PET studies. Simula tion studies showed poor fits (Akaike criterion) for plasma input models at typical noise levels ( COV ~ 2.5%) and scan durations (< 90 min). These poor fits are due to the relatively slow kinetics of [ 18 F ]FP- 3 -CIT, which approaches irreversible kinetics for short scan times. However, reference tissue models provided more reliable fits, which were nearly independent of noise and scan duration. Similar results were obtained in clinical studies. SRTM provided best discrimination between patients and controls. When differentiating between patients and controls, SUV r performed almost equally well as SRTM, although contrast between striatum and background was lower. Therefore, SRTM is the method of choice for quantitative [ 18 F ]FP- 3 -CIT studies. SUV r, however, might be an alter native for larger clinical trials.
Chapter 5 describes a study on the quantification of [ 18 F ]FDDNP, a tracer of amyloid deposition. Blood data showed rapid metabolism of [ 18 F ]FDDNP, with a large number of polar metabolites being formed. Recently, it has been demonstrated that the latter metabolites may enter the brain and, therefore, they should be accounted for. To do so, evaluation of analytical methods included a modified 2 T 4 k plasma input model with an additional compartment for metabolites (2 T 1 M ). In clinical stu dies, based on the Akaike criterion, the 2 T 1 M model was preferred over the standard 2 T 4 k model. SRTM showed better correlation with 2 T 1 M then with 2 T 4 k . Furthermore, in simulations, SRTM showed relatively constant bias with best precision, even when it was assumed that metabo lites could enter the brain. It was concluded that SRTM is the method of choice for quantitative analysis of [ 18 F ]FDDNP studies, even if it is unclear whether labelled metabolites enter the brain.
In chapter 6 various reference tissue based parametric methods for improving quantification of [ 11 C ]PIB studies are evaluated. The follow ing parametric methods were evaluated: receptor parametric mapping (basis function implementation of the simplified reference tissue model with and without fixed k 0 2 ), reference Logan, and several multi-linear ref erence tissue methods (again with and without fixed k 0 2 ). In addition a semi-quantitative method, SUV r, was evaluated. For clinical studies, most parametric methods showed comparable performance, with poorest results for SUV r . Best performance was obtained for receptor paramet ric mapping ( RPM 2) and one of the multi-linear reference tissue mod els ( MRTM 2), both with fixed (reference tissue clearance constant) k 0 2 . BP ND was least affected by noise and generated images showed better contrast than with other methods. In addition, RPM 2 and MRTM 2 pro vided the most accurate and precise BPND estimates. Therefore, RPM 2 and MRTM 2 are the methods of choice for parametric analysis of clinical [ 11 C ]PIB studies.
Finally, in chapter 7 the evaluation of several parametric methods for improving quantification of [ 18 F ]FDDNP studies is described. This study was similar in design to the [ 11 C] PIB study of the previous chapter. In clinical studies, again best performance was obtained using RPM 2 and MRTM 2. Both methods showed good correlation with SRTM, BP ND was least affected by noise and parametric images showed good contrast. Similar results were found in the simulations. Therefore, RPM 2 and MRTM 2 also are the methods of choice for parametric analysis of clinical [ 18 F ]FDDNP studies.