Elsevier

NeuroImage

Volume 21, Issue 2, February 2004, Pages 483-493
NeuroImage

Cluster analysis in kinetic modelling of the brain: a noninvasive alternative to arterial sampling

https://doi.org/10.1016/j.neuroimage.2003.09.058Get rights and content

Abstract

In emission tomography, quantification of brain tracer uptake, metabolism or binding requires knowledge of the cerebral input function. Traditionally, this is achieved with arterial blood sampling. We propose a noninvasive alternative via the use of a blood vessel time–activity curve (TAC) extracted directly from dynamic positron emission tomography (PET) scans by cluster analysis. Five healthy subjects were injected with the 5HT2A-receptor ligand [18F]-altanserin and blood samples were subsequently taken from the radial artery and cubital vein. Eight regions-of-interest (ROI) TACs were extracted from the PET data set. Hierarchical K-means cluster analysis was performed on the PET time series to extract a cerebral vasculature ROI. The number of clusters was varied from K = 1 to 10 for the second of the two-stage method. Determination of the correct number of clusters was performed by the ‘within-variance’ measure and by 3D visual inspection of the homogeneity of the determined clusters. The cluster-determined input curve was then used in Logan plot analysis and compared with the arterial and venous blood samples, and additionally with one of the currently used alternatives to arterial blood sampling, the Simplified Reference Tissue Model (SRTM) and Logan analysis with cerebellar TAC as an input. There was a good agreement (P < 0.05) between the values of Distribution Volume (DV) obtained from the K-means-clustered input function and those from the arterial blood samples. This work acts as a proof-of-principle that the use of cluster analysis on a PET data set could obviate the requirement for arterial cannulation when determining the input function for kinetic modelling of ligand binding, and that this may be a superior approach as compared to the other noninvasive alternatives.

Introduction

Quantification of the brain's uptake and binding of a radiolabelled tracer often involves the use of kinetic modelling of data obtained from dynamic positron emission tomography (PET) studies. This usually requires the acquisition of arterial blood samples that serve as an estimate of the cerebral input function. One alternative has been the reference tissue models (e.g. Gunn et al., 1997, Lammertsma and Hume, 1996), which avoid the requirement of any blood sampling. Their usefulness is proven by their widespread popularity, and they are advantageous in many situations. However, the tradeoff can be a possible loss in accuracy and increased bias, and the dependence on the assumption that intersubject nonspecific binding differences are negligible (Slifstein and Laruelle, 2001). Further, these methods cannot be applied in the absence of a suitable reference region, and do not work well for all radiotracers. For example, as not all ligands conform to the underlying model assumptions of the Simplified Reference Tissue Model (SRTM), it may cause an over- or underbias effect depending on whether the tracer distributes in a way such that a dynamically distinct nonspecifically bound region is present (Slifstein and Laruelle, 2001). Other alternatives include the use of standardised input curves (e.g. Kurisu et al., 2002, Ogura et al., 1999, Onishi et al., 1996, Takikawa et al., 1993, Tsuchida et al., 1999), or arterialised blood samples (e.g. Eberl et al., 1997, Huang et al., 1980). However, the former requires a reasonable population a priori to determine a robust average measure, and both approaches are limited by the requirement that interindividual differences in responses must be negligible. An alternative, but more computationally demanding, approach uses the time courses of multiple ROIs to simultaneously extract the input function and the kinetic model parameters via deconvolution techniques (e.g. Feng et al., 1997).

Another possibility, described here, is to obtain the input function from the PET data itself by extracting a cerebral vasculature time–activity curve (TAC). Apart from the obvious benefit of no arterial cannulation, there are several other advantages, including the removal of issues regarding cross-calibration (between the PET data and that obtained from the blood samples), delay, dispersion and noise issues. To extract a relevant blood-TAC from the dynamic PET data, we have used clustering to (semi)automatically extract only those voxels that belong to the brain vasculature. One of the simplest methods, K-means, was chosen for this purpose.

The idea of using a “blood-TAC” as a replacement for arterial sampling is not new. It has been around in the field of cardiac imaging for many years [e.g. Weinberg et al., 1988, for the use of the left ventricle in canines; Iida et al., 1992, who used an improved model to correct for partial volume (“spillover”) effects] as well as in other fields (e.g. Germano et al., 1992, where the abdominal aorta TAC was used in hepatic and renal studies as a replacement for the arterial input). However, its use in neuro-PET imaging is only a recent development. Wahl et al. (1999), studying [18F]6-fluoro-l-m-tyrosine brain uptake, used hand-drawn ROIs over the confluence of the venous sinuses and found it to perform better than arterialised venous blood sampling. Chen et al. (1998) used image-derived input functions in PET by summing the initial frames and hand-drawing ROIs over the carotid arteries, whereas Litton (1997) used ROIs drawn over the carotids of MR scans. Asselin et al. (2002) found that the use of the superior sagittal sinus provided several benefits for some tracers, due to its larger size and location away from other blood vessels. Asselin et al. (2001) also used clustering in the extraction of a “blood-TAC”, but only as a visual aid to the manual drawing of the vasculature ROIs. The use of clustering for automatically extracting such vasculature TACs has not been reported before. Clustering is a method of grouping voxels in a data set by the similarity of their time courses. As a result, it creates an input from functional—not spatial—data, thus creating the largest possible ROI automatically, without any spatial constraints, and obviating any possible human error of ROI definition.

To assess the validity of using a vasculature TAC as a possible replacement for the standard arterial cannulation, we compare the results of modelling the binding parameters of the 5HT2A receptor ligand [18F]-altanserin after a bolus injection using a cluster-derived estimate of the vasculature TAC extracted from the PET data set with those from arterial and venous blood sampling. A graphical model was used for analysis, and the results compared with those of a noninvasive alternative, the Simplified Reference Tissue Model.

Section snippets

Materials and methods

Five subjects (two females, age range 22–74 years) were included in the study. Subjects were excluded if they, or their first- or second-degree relatives, had a history of psychiatric or neurological illness. Exclusion criteria also included current physical illness with use of prescription or over-the-counter medications with a CNS effect and a history of alcohol or drug dependence. All subjects had normal general and neurological examinations and a normal cerebral MRI scan performed with a

Results

[18F]-altanserin remained unchanged in both whole blood and plasma for up to 30 min after sampling or processing. A paired Wilcoxon test showed that there was no significant difference in the amount of parent compound after immediate metabolite analysis or metabolite analysis 30 min after sampling of the blood sample (P > 0.05). Two hours after the [18F]-altanserin injection, approximately 40% of the radioactivity was due to the parent compound, with about 30% of the radioactivity originating

Discussion

This study shows that the results of performing kinetic modelling using the proposed alternative to blood sampling compare favourably with those obtained via arterial sampling, more so than those obtained via venous sampling (DV values, Table 1). Table 2 shows that the proposed method also does not suffer from a possible bias of BP values akin to that of the SRTM. The similarity of the cluster TAC and arterial curves (Fig. 4), particularly in the long flat tail that dominates the Logan plot

Conclusions

This work acts as a proof-of-principle that the use of cluster analysis on a PET data set could obviate the requirement for arterial cannulation when determining the input function for kinetic modelling of tracer behaviour. Apart from the obvious advantage of leaving out arterial cannulation, it has the innate benefit of requiring no cross-calibration between scanner and well counter and no (or very little) compensation for the delay and dispersion as the input curves are obtained from the same

Acknowledgements

The Danish Health Research Council, the CD2T project, the 1991 Pharmacy Foundation, the Health Insurance Fund and the Lundbeck Foundation are gratefully acknowledged for their support to this project, and The John and Birthe Meyer Foundation for the donation of the Cyclotron and PET-scanner. Karin Stahr is thanked for providing excellent technical assistance.

References (41)

  • K. Chen et al.

    The effects of measurement errors in the plasma radioactivity curve on parameter estimation in positron emission tomography

    Phys. Med. Biol.

    (1991)
  • K. Chen et al.

    Noninvasive quantification of the cerebral metabolic rate for glucose using positron emission tomography, 18F-fluoro-2-deoxyglucose, the Patlak Method, and an image-derived input function

    J. Cereb. Blood Flow Metab.

    (1998)
  • V. Dhawan et al.

    Accuracy of PET RCBF measurements: effect of time shift between blood and brain radioactivity curves

    Phys. Med. Biol.

    (1986)
  • V. Dhawan et al.

    Effect of blood curve smearing on the accuracy of parameter estimates obtained for 82Rb/PET studies of blood–brain barrier permeability

    Phys. Med. Biol.

    (1988)
  • S. Eberl et al.

    Evaluation of two population-based input functions for quantitative neurological FDG PET studies

    Eur. J. Nucl. Med.

    (1997)
  • D. Feng et al.

    A technique for extracting physiological parameters and the required input function simultaneously from PET image measurements: theory and simulation study

    IEEE Trans. Inf. Technol. Biomed.

    (1997)
  • G. Germano et al.

    Use of the abdominal aorta for arterial input function determination in hepatic and renal PET studies

    J. Nucl. Med.

    (1992)
  • X. Golay et al.

    A new correlation-based fuzzy logic clustering algorithm for fMRI

    Magn. Reson. Med.

    (1998)
  • C. Goutte et al.

    Feature-space clustering for fMRI meta-analysis

    Hum. Brain Mapp.

    (2001)
  • P. Herscovitch et al.

    Brain blood flow measured with intravenous H2(15)O: I. Theory and error analysis

    J. Nucl. Med.

    (1983)
  • Cited by (0)

    View full text