Compartmental model analysis of regional TACs
This page handles the general compartmental models with arterial plasma input that are applicable to most tracers; tracer specific models are described elsewhere, as well as reference tissue input models.
Model fitting software
In TPC, Carimas is available for all researchers, and it can be used to fit one- and two-tissue compartmental models to regional data. ROIs can either be drawn and regional TACs computed inside Carimas, or regional TACs can be imported to Carimas.
Certain research groups in TPC have also acquired PMOD licenses.
Alternatively, in-house developed command-line programs (Open Source) can be used directly in Windows computers that are connected to TPC network, or can be downloaded for Windows, Linux, and macOS computers.
Nonlinear least-squares (NLLS) fitting
Programs fitk2, fitk3, fitk4, fitk5, and fitkloss can be used to fit two, three, or four compartment models to the PET TACs; fitkloss can be used to fit a three-compartment model where the last rate constant k4 or kLoss represents the efflux of labelled metabolite or radioligand directly to the venous plasma.
These programs do use the weighting information if that is included in the tissue datafile.
In many cases all of the model parameters of two-tissue model can not be reliably fitted, but some of them need to be constrained to a pre-determined value. A commonly used method is to constrain the K1/k2, representing the distribution volume of nonspecifically bound and free radioligand in the tissue, to a value that is derived from a reference region with no specific binding or uptake (k3=0). Not only does this method allow lower variances of model parameters, but it also enables us to calculate the binding potential in case of receptor studies.
As in all models applying reference tissue, also here it is assumed that K1/k2 is the same in all (brain) regions. First, two- or three-compartment model is fitted to the reference tissue curve, providing the values for K1/k2 or K1/k2 and k5/k6. These are used as constraints for these parameters when fitting the compartmental model to the regions of interest.
Constraining K1/k2 to zero with program fitk2 actually constrains k2=0, allowing the use of an extremely simple irreversible model with only two parameters, K1 and VB.
Akaike information criteria (AIC)
Various compartmental models can be constructed and used to analyze PET data. The more complicated the model is, the better is the achieved fit to the data. However, at the same time, also the variance of the fitted parameters is increased. To find the optimum model, the programs compute also Akaike information criteria values: the smaller the AIC values are, the better the model is, considering the degrees of freedom of the fit. However, the physiological interpretation of the fitted parameters is on the responsibility of the user.
General linear least squares method
Most compartmental models can be transformed into general linear least squares functions (Blomqvist, 1984), which can be solved using very fast linear methods, and are therefore suitable for computing parametric images. For regional data, program lhsol can be used to estimate the model parameters using Lawson-Hanson nonnegative least squares (NNLS) algorithm. The compartmental model can be selected with options -k1, -k2, -k3, -k4 for models excluding vascular volume fraction, or options -vk1, -vk2, -vk3, -vk4 for models including vascular volume fraction.
Note that if VA is fitted using this methods, the vascular blood TAC is assumed to be similar to the model input, i.e. metabolite corrected arterial plasma curve. This is close to truth for a few tracers only, e.g. [18F]FDG. For other studies, a fixed amount of blood background can be subtracted before the model fit using taccbv.
The fitted parameters from these programs may have non-physiological values, because there is no other constraints than non-negativity.
VT and Ki using NNLS method
In receptor binding studies distribution volume (VT) is usually the only model parameter of interest. Instead of solving separate model rate constants and calculating VT from those afterwards, more reliable estimates of VT can be obtained by solving VT directly without division (Zhou et al., 2002; Hagelberg et al., 2004). Program lhsoldv uses Lawson-Hanson non-negative least squares (NNLS) algorithm and one or two tissue compartment models (with options -1 and -2) to solve the VT without division. The noise in regional TACs does not cause bias when using this method. Two tissue compartment model is recommended, since one tissue compartment model may lead to biases with more complex tissue kinetics. By default (or with option -A), the model can be selected automatically, based on lower AIC. With option -0 the AIC weighted average (Turkheimer et al. 2002) of VT from 1- and 2-tissue compartment models is calculated.
For irreversible tracer uptake models, the influx rate constant Ki can be calculated accordingly with program lhsolki. The two tissue model is applied by default.
In-house analysis programs, including Carimas, automatically convert the sample time and radioactivity concentrations units of plasma, blood and tissue data, if the units are specified in the data files. This is not always the case, and therefore it would be safest that researcher verifies that the units are the same in all data files before proceeding to the modelling.
In TPC, the units of radioactivity concentrations in plasma and blood are by default given per volume (mL), not per mass (g). Therefore the unit of model parameter K1 is (mL plasma)*(mL tissue)-1*min-1, and the units of other rate constants k2, ... k6 are min-1. The units of the vascular blood or arterial plasma volume fractions VB and VA are mL/mL.
Steps of calculation using command-line tools
All of of the following steps can be done in Linux terminal window or MS Windows command prompt window (preferably using scripts):
- Instruction by tracer
- Analysis models for regional TACs
- Input function
- Model calculations for PET images
- Regional result files
- Processing and further analysis for regional results
Blomqvist G. On the construction of functional maps in positron emission tomography. J Cereb Blood Flow Metab. 1984; 4:629-632. doi: 10.1038/jcbfm.1984.89.
Lawson CL, Hanson RJ. Solving least squares problems. Prentice-Hall, 1974.
Turkheimer FE, Hinz R, Cunningham VJ. On the undecidability among kinetic models: from model selection to model averaging. J Cereb Blood Flow Metab. 2003; 23: 490-498. doi: 10.1097/01.WCB.0000050065.57184.BB.
Zhou Y, Brasic J, Endres CJ, Kuwabara H, Kimes A, Contoreggi C, Maini A, Ernst M, Wong DF. Binding potential image based statistical mapping for detection of dopamine release by [11C]raclopride dynamic PET. NeuroImage 2002; 16: S91.
Updated at: 2020-05-01
Created at: 2013-05-17
Written by: Vesa Oikonen