NeuroImage 19 (2003) 1049 –1060 www.elsevier.

com/locate/ynimg

A novel method for noninvasive detection of neuromodulatory changes

in specific neurotransmitter systems
Nathaniel M. Alpert,* Rajendra D. Badgaiyan, Elijahu Livni, and Alan J. Fischman
Division of Nuclear Medicine, Massachusetts General Hospital, Boston, MA 02114, USA
Received 2 December 2002; revised 13 February 2003; accepted 18 March 2003

Abstract
Over the last decade, it has become possible to study theories of cognition using positron emission tomography (PET) and functional magnetic
resonance imaging (fMRI). These methods yield statistical parametric maps of changes in cerebral blood flow (CBF) elicited by cognitive tasks.
A limitation of these studies is that they provide no information about the underlying neurochemistry. However, it is possible to extend the concept
of activation studies to include measurements targeting neurotransmitters and specific receptor populations. Cognitive activation increases neuronal
firing rate, increasing the endogenous neurotransmitter level. The increased neurotransmitter level can be used to alter the kinetics of specifically
bound radioligands. We describe a new approach to the design and analysis of neuromodulation experiments. This approach uses PET, a
single-scan session design, and a linear extension of the simplified reference region model (LSSRM) that accounts for changes in ligand binding
induced by cognitive tasks or drug challenge. In the LSSRM, an “activation” parameter is included that represents the presence or absence of
change in apparent dissociation rate. Activation of the neurotransmitter is detected statistically when the activation parameter is shown to violate
the null hypothesis. Simulation was used to explore the properties of the LSSRM with regard to model identifiability, effect of statistical noise,
and confounding effects of CBF-related changes. Simulation predicted that it is possible to detect and map neuromodulatory changes in
single-subject designs. A human study was conducted to confirm the predictions of simulation using 11C-raclopride and a motor planning task.
Parametric images of transport, binding potential, areas of significant dopamine release, and statistical parameters were computed. Examination
of the kinetics of activation demonstrated that maximum dopamine release occurred immediately following task initiation and then decreased with
a half-time of about 3 min. This method can be extended to explore neurotransmitter involvement in other behavioral and cognitive domains.
© 2003 Elsevier Science (USA). All rights reserved.

Introduction tion. One of the challenges for neuroimaging experiments is

to advance the methodology to the point where one can not
In recent years positron emission tomography (PET) and
only identify the regions of brain that are activated by a
functional magnetic resonance imaging (fMRI) techniques
have been used to detect changes in cerebral blood flow stimulus, but also determine the chemical systems that me-
(CBF) elicited by cognitive or behavioral tasks. These “ac- diate these responses (Coull, 1998; Granon et al., 2000;
tivation” studies have successfully identified a number of Kahkonen et al., 2001; Lawrence et al., 1998).
cortical and subcortical areas that process a variety of cog- Experiments conducted on human volunteers and labo-
nitive activities. However, these experiments are limited by ratory animals have provided reliable data suggesting that
their inability to identify cellular mechanisms associated neurotransmitters and neuromodulators significantly influ-
with cognitive changes. Since the CBF-related changes are ence a variety of cognitive functions (Jentsch et al., 2000;
nonspecific, these studies do not provide information about Kulisevsky, 2000; Oak et al., 2000; Robbins, 2000; Robbins
the nature of the underlying neurochemistry and neurotrans- et al., 1997). It has been shown that activation or inhibition
mitter systems involved in performing specific cognitive of neurotransmitter systems causes significant alteration in
tasks. Because of these limitations, activation studies are not performance of human volunteers and laboratory animals on
likely to explain neurophysiological mechanisms of cogni- a variety of tasks of spatial attention (Jackson et al., 1994;
Perry et al., 1999; Witte and Marrocco, 1997; Robbins,
* Corresponding author. Fax: ⫹1-617-726-6165. 2000), and working memory (Granon et al., 2000; Green-
E-mail address: [email protected] (N.M. Alpert). wald and Davis, 1983; Jaber et al., 1996; Kimberg et al.,

1053-8119/03/$ – see front matter © 2003 Elsevier Science (USA). All rights reserved.
doi:10.1016/S1053-8119(03)00186-1
1050 N.M. Alpert et al. / NeuroImage 19 (2003) 1049 –1060

2001). There is, however, no direct evidence that associates cally bound tracer is the same in the reference and the binding
a specific neurotransmitter system with a cognitive function. region, one can derive a series of convenient equations. By sub-
In 1995 Fisher et al. and Morris et al. suggested that it stitution for Cr , the binding equation can be rewritten as

冋 册
might be possible to use PET and specifically bound radio-
dC共t兲 K 1a dC r共t兲
ligands to detect changes in the binding kinetics of revers- ⫽ ⫹ k 2r · C r共t兲 ⫺ k 2a · C共t兲 (3)
ibly bound radioligands. The prediction was based on the dt K 1r dt
idea that neurons involved in a cognitive task increase their and directly integrated to yield
firing rate, causing an increase in the synaptic level of
endogenous neurotransmitter. This mechanism can be used
to perturb the kinetics of specifically bound receptor li-
gands, making it possible to detect the effect with PET.
冕 t
C共t兲 ⫽ RCr共t兲 ⫹ k2 Cr共u兲du ⫺ k2a C共u兲du
0
冕 t

0
(4)

Koepp et al. (1998) performed the first study testing the where:
feasibility of the idea; but their experiment involved a com-
plex task (a video game) and required two scan sessions and K 1a k 2
R⫽ ⫽
a cohort of subjects to detect an effect. A ground-breaking K 1r k 2r
study by Friston et al. (1997) proposed and demonstrated a
method for assessing time-dependent changes in ligand dis- k2
k 2a ⫽
placement by using a single-scan design and a pharmaco- 1 ⫹ BP
logical challenge. Recently, Pappata et al. (2002) have ap-
k3
plied and extended Friston’s method to study a gambling/ BP ⫽
reward task. In this study, we present a novel methodology k4
designed to increase the sensitivity of single-scan session Including the effect of activation
designs so that activation might be detected in individual
subjects. This report presents the theory, discusses its im- In this theory we consider activation to be any cognitive
plementation, evaluates elements of the theory by simula- effect or drug challenge that can render any, or all, of the model
tion, and demonstrates feasibility in a human experiment. parameters to be time dependent, i.e., the non- steady-state
situation. For this approach to be useful we have to “know” or
Theory hypothesize the time dependence of the parameters. In addition
to changes in the endogenous neurotransmitter level, changes
We derive a kinetic theory capable of describing non- in blood flow may cause time-dependent changes in the model
steady-state (i.e., time dependent) fluctuations of receptor pa- parameters (Dagher et al., 1997; Pappata et al., 2002). The
rameters. This theory rests on the reference region models non-steady-state equation is obtained by integrating Eq. (3)
(Lammertsma et al., 1996) and the simplified reference region assuming that the parameters are time dependent.
model (SRRM) of Lammertsma and Hume (1996) and Gunn et
al. (1997). Our theory extends the SRRM in two ways: first, it
allows the parameters to be time dependent and second, all
parameters enter linearly. To review briefly, the SRRM fol-
C共t兲 ⫽ 冕 t

0
R共u兲 ·
dC r
du
du ⫹ 冕 0
t
k 2共u兲 · C r共u兲du

冕
lows from the following basic equations:
t

dC共t兲 ⫺ k 2a共u兲 · C共u兲du (5)

⫽ K 1aC p共t兲 ⫺ k 2aC共t兲 (1) 0
dt
dC r共t兲 We propose a basic activation model that includes a
⫽ K 1rC p共t兲 ⫺ k 2rC r共t兲 (2) control state with parameters R, k2, and k2a. The activation
dt
state is parameterized as R ⫹ ␣ h(t), k2 ⫹ ␤ h (t), and k2a
Cp(t) denotes the plasma concentration of radioligand at time t. C ⫹ ␥ h(t), where we assume a common time dependence for
and Cr are instantaneous quantities denoting radioactivity concen- the activated state described by h(t). The basic equation is:

冕 冕
tration in a binding and reference region respectively. The refer-
t t
ence region is assumed to be devoid of specific binding. The
C共t兲 ⫽ RCr共t兲 ⫹ k2 Cr共u兲du ⫺ k2a · C共u兲du
subscript r refers to the reference region parameters and the sub-
0 0
script a refers to an “apparent” volume containing both free and

冕 冕
nonspecifically bound tracer as well as specifically bound tracer. t t
dCr
This apparent volume is assumed to be adequately described by a ⫹␣ h共u兲du ⫹ ␤ Cr共u兲h共u兲du
single compartment model, with the bound and free tracer in du
0 0
instantaneous equilibrium. In the limit of instantaneous equilib-
rium, the SRRM parameters are related to the standard receptor
model with two compartments and four parameters, K1, k2, k3, and
k4. By assuming that the volume of distribution of the nonspecifi-
冕 t
⫺ ␥ C共u兲h共u兲du
0
(6)
 N.M. Alpert et al. / NeuroImage 19 (2003) 1049 –1060 1051

Because there is a paucity of relevant experimental data, the purpose of this study we consider an exponential form
it is not possible to specify h(t) with great certainty. Based for h(t) (Endres and Carson, 1998), with activation com-
on the available data (see Fisher et al., 1995), we can mencing at time T.
assume that changes in endogenous neurotransmitter levels
occur on a time scale that is very fast compared to PET
imaging times. The available data also support the assump- h共t兲 ⫽ 冋 0, t⬍T
e ⫺␶共t⫺T兲, t ⱖ T 册 (7)
tions that the major effects of activation occur immediately
after the onset of the task and that the effects diminish over where ␶ controls the rate at which activation effects die
time. More discussion of this point is presented later. For away. The model equation

冕 冕

冤 冥
t t
RC r共t兲 ⫹ k 2 C r共u兲du ⫺ k 2a C共u兲du 0⬍t⬍T
0 0

RC r共t兲 ⫹ k 2 冕 0
t
C r共u兲du ⫺ k 2a 冕0
t
C共u兲du ⫹ . . .

冕
C共t兲 ⫽ t
(8)
⫺␶共t⫺T兲 ⫺␶共u⫺T兲
␣ 关共C r共t兲e ⫺ C r共T兲兲 ⫹ ␶ C r共u兲e du兴 ⫹ . . .
T

␤ 冕
T
t
e ⫺␶共u⫺T兲C r共u兲du ⫺ ␥ 冕T
t
e ⫺␶共u⫺T兲C共u兲du t⬎T

predicts the instantaneous concentration curve.

To derive an operational equation for this theory, we take
note of the fact that the concentrations in Eq. (8) are instan-
ref i ⫽
1
⌬i 冕
ti
ti⫹⌬i
C r共u兲du (10)

taneous quantities. Denote the measurable quantities as pet

and ref. pet is the concentration value measured by the PET where ⌬i represents the duration of the ith frame and ti
scanner in a binding region and ref is the concentration is the start time of the ith frame. Integrating Eq. (8),
measured in a region devoid of specific binding. In terms of we can closely approximate the integrals as sums, yield-
the instantaneous quantities the ith measurement is ing an operational equation suitable for numeric compu-

冕 ti⫹⌬i tation,
1
pet i ⫽ C共u兲du (9)
⌬i
ti

冤 冥
1 ⱕ i ⬍ js

R · ref i ⫹ k 2 冉冘 j⫽i
j⫽1
1
⌬ jref j ⫺ ⌬ i · ref i ⫺ k 2a
2 冊 冉冘 j⫽i
j⫽1
1
⌬ jpet j ⫺ ⌬ i · pet i
2 冊
i ⱖ js

pet i ⬇
R · ref i ⫹ k 2 冉冘
j⫽1
j⫽i 1
⌬ jref j ⫺ ⌬ i · ref i ⫺ k 2a
2 冊 冉冘 j⫽i
j⫽1
1
⌬ jpet j ⫺ ⌬ i · pet i ⫹ . . .
2 冊 (11)
␣ · 共ref i · e ⫺␶共ti⫹⌬i/ 2⫹T兲 ⫺ ref js兲 ⫹ . . .

⫹␶· 冉冘
j⫽i
j⫽js
1
ref j · e ⫺␶共tj⫹⌬j/ 2⫹T兲 ⫺ ⌬ i ref ie ⫺␶共ti⫹⌬i/ 2⫹T兲 . . .
2 冊
⫺␥· 冉冘
j⫽i
j⫽js
1
⌬ j · pet j · e ⫺␶共t j ⫹ ⌬ j/ 2 ⫺ T兲 ⫺ ⌬ i · pet i · e ⫺␶共ti⫹⌬i/ 2⫺T兲
2 冊
where js is the frame in which the activation task begins.
1052 N.M. Alpert et al. / NeuroImage 19 (2003) 1049 –1060

Table 1
Simulation parameters

Reference region Binding region LSSRM

K1R k2R K1 k2 k3 k4 R BP R k2 k2a BP Notes
a
0.092 0.45 0.092 0.45 0.062 0.014 1 4.54 0.83 0.29 0.052 4.57
b
0.18 0.326 0.18 0.37 0.38 0.11 1 3.45 0.85 0.20 0.044 3.50
a,b
Adapted from aPappata et al. (2002) and bFarde et al. (1989), Subject C. The parameters listed in the Binding region section of the table were those
indicated in notes (a) and (b) and they were used to simulate positron emission tomographic (PET) curves with the standard two compartment PET receptor
model. The parameters listed under LSSRM were those estimated with the linear simplified reference region model.

Parameter estimation and hypothesis testing Simulation studies

To simplify the exposition, assume that ␶ is known; we All simulations were programmed in the MATLAB comput-
will describe how we estimated ␶ in the materials and ing environment (Release 13, The Mathworks, Natick, MA). Dif-
methods section of this report. With this assumption, the ferential equations were solved using ode23s, a solver designed
model equations are linear in the parameters. We will refer for so-called stiff equations. To mimic experimental protocols, we
to this as the linear simplified reference region model simulated an acquisition with 59 frames (10 at 30 s, 45 at 1 min,
(LSSRM). Linear least squares provides estimates of the and 4 at 5 min). Fitting parameters and their covariance matrix
model parameters and their covariance matrix. In statistical were determined analytically, using weighted linear least squares.
parlance, the effects of interest are R, k2, k2a, ␣, ␤, and ␥, The first set of simulations addressed the issue of biased
with ␥ being the parameter of greatest interest. An important estimation by comparing the LSSRM with the standard
point in this study is that while the LSSRM provides quan- four-parameter, two-compartment model of receptor bind-
titative estimates of the parameters, our interest will often be ing, as applied to 11C-raclopride. The endpoint of the sim-
satisfied if we can reject the null hypothesis ␥ ⱕ 0. Simple ulation was to determine whether the LSSRM could fit data
hypotheses are tested by using the estimated covariance generated with the standard model and whether the LSSRM
matrix to compute t statistics for the parameters of interest parameters matched the values used in the standard model.
(Searle, 1971; Friston et al., 1995; 1997). The t statistic has The instantaneous values for reference and binding region
n ⫺ p ⫹ 1 degrees of freedom where n is the number of concentrations were determined by solving the correspond-
frames and p the number of model parameters. For region of ing differential equations with parameters appropriate to
interest (ROI) analysis, the percentage points of the t dis- raclopride binding (see Table 1), using a measured input
function. The expected values for the PET curves (i.e.,
tribution can be used to test for significance. For parametric
simulated) were determined by integrating the instantaneous
imaging, the use of random field theory and the concepts of
concentration values to compute the average concentrations
SPM analysis (Friston et al., 1995, 1997) is required to set
for the simulated scan protocol. The simulated reference
appropriate significance thresholds when there are no
region curve was used as the input function with Eq. (11)
apriori localizing hypotheses. In principle, hypothesis test-
and linear least squares estimation to “fit” the simulated
ing can detect significant changes in ␣ and ␤, protecting
binding curve (␣, ␤, and ␥ were set to 0). The same calcu-
against CBF-related confounds with changes in ␥; but
lation was also performed with the original method of Gunn
whether this can be achieved in practice depends on the
et al. (1997), which solves the nonlinear SRRM equations
identifiability of the model parameters. We will return to by a combination of linear least squares and table lookup.
this point in the discussion. The second set of simulations, again with noise-free data,
studied a range of possible effects that may accompany
cognitive activation or pharmacological challenge. These
Materials and methods include the effects of increased ligand displacement and
increased blood flow. We considered possibilities such as
To evaluate the theory we performed simulation studies proportional increase in K1 and k2, so that the distribution
and experiments with human subjects. The simulations ex- volume remained constant, and separate changes in k2 and
amine basic questions about model, parameter identifiability k1. The differential equations were modified to allow for
in the face of random noise, and the confounding effects of time-dependent parameters. All of these simulations as-
changes in blood flow. The human studies attempt to dem- sumed a time-dependent parameter(s) of the form

冋 册
onstrate the feasibility of detecting changes in receptor po tⱕT
binding due to a motor planning task. For this study we use p共t兲 ⫽ 共 p o ⫹ ␦ p兲e ⫺␶共t⫺T兲 t ⬎ T (12)
a single human study as an example to show proof of
principle; the results of the human studies have been re- Simulations were performed for perturbations beginning
ported elsewhere (Badgaiyan et al., 2003). at T ⫽ 38 min after injection. Tests were performed to
 N.M. Alpert et al. / NeuroImage 19 (2003) 1049 –1060 1053

determine the degree of bias in the parameters, the goodness Gaussian filter. Second, all frames were aligned to the frame
of fit, and model identifiability. taken at 25 min. Next, the aligned frames were summed to
The third set of simulations examined the feasibility of form a new reference volume and all frames were aligned
detecting ligand displacement in individual subjects at the again to the sum image.
voxel level. In essence, these simulations examined the A task was designed to focus on motor planning. The
effect of statistical noise on detection of effects. We simu- core element of the task was the opposition of the thumb
lated PET curves with the LSSRM and the parameters R, k2, and fingers in response to visual cues presented on a com-
and k2a shown in Table 1 for Subject C in Farde et al. puter monitor. The same visual cues, the presentation of a
(1989). The effect of increased dopamine level was set as single random digit (1, 2, 3, or 4) once every 2 s, were
described above. Simulation of noisy data was computed by visible to the subject throughout the experiment. During the
adding Gaussian random noise with variance proportional to control phase of the study, the subject was instructed to lie
Cm · e␭t/dt, where Cm was the peak value of the noise-free still without moving. In the study presented here finger
simulation and ␭ denotes the decay constant for 11C. The opposition started 38 min after raclopride injection and
proportionality constant was adjusted to yield results similar continued for the duration of the experiment. The subject
to that observed in human studies with racolpride at a spatial was instructed to oppose the right thumb and forefinger
resolution of ⬃10 mm. To gauge the level of type II error upon seeing the digit 1, their thumb and middle finger on
that might affect this type of analysis, we analyzed another seeing the digit 2, and so on. Compliance with all instruc-
ensemble of 10,000 simulations having the same noise level tions was observed by closed-circuit television.
added to a simulation with ␥ ⫽ 0. Image analysis was performed by using a combination of
ROI analysis and parametric imaging. Visual examination
of the image data was used to identify the cerebellum, a
three-dimensional ROI was defined, and a time activity
Human studies curve (TAC) was extracted to serve as the reference region
TAC. All frames were smoothed with a 10-mm FWHM
Gaussian filter. To reduce the computational load, a mask
A set of pilot studies studies were performed in 5 young indicating the “important voxels” was determined by sum-
healthy human volunteers, under the supervision of the ming all frames and using a threshold 1.3 times the mean of
Committee on Human Research at the Massachusetts Gen- all voxels. The LSSRM model with ␣ and ␤ fixed to zero
eral Hospital. The general nature of the experiments was was used to fit all voxels within the mask. Parametric
explained and subjects gave written informed consent. The images of the R, k2, k2a, BP ⫽ k2 /k2a ⫺ 1 , ␥, the
object of these initial studies was to test the predictions of covariance matrix, and the t score for ␥ were computed. A
simulations suggesting the possibility of detecting ligand presentation image of the t map for the ␥ parameter was
displacement in individual subjects during a well-defined determined by including only those voxels for which k2 ⫺
motor planning task. Because this study is methodological 1.5 · k2a ⬎ 0. In the fusion of the t map and binding
in nature and because the results of the human study will be potential, the binding potential image was filtered with a 3
of interest to a wider community, we present the results of ⫻ 3 median filter to suppress high frequency noise.
one study in this report. The analysis proceeded as a two-step process. First, data
The test subject reported here received a bolus injection were analyzed assuming h(t) was a step function (i.e., ␶ ⫽
(1 min duration) of ⬃14 mCi of 11C-raclopride, with spe- 0). Following that analysis the t map was examined for
cific activity 900 mCi/␮mol, in the left antecubittal vein. evidence of increased ligand displacement. Regions of in-
The subject was imaged in the supine position with head creased displacement were further analyzed by ROI analy-
restrained by a custom-molded thermoplastic mask, using sis. TACs were extracted and fit with a model that allowed
an ECAT EXACT HR⫹ PET scanner operating in the three- ␶ to vary. Following estimation of ␶, the parametric maps
dimensional mode. The ECAT HR⫹ records 63 slices si- were recomputed.
multaneously, with an intrinsic resolution of about 4-mm
FWHM in all spatial dimensions (Adam et al., 1997). The
acquisition protocol recorded frames over a 90-min period. Statistical analysis
Images were reconstructed as 128 ⫻ 128 ⫻ 63 element
volumes using an OSEM algorithm, with corrections for The LSSRM model uses weighted linear least squares
photon attenuation, random coincidences, scatter, and dead analysis for parameter estimation. The theory of least
time. To minimize the residual effects of head movements, squares analysis also provides an estimate of the covariance
image registration was performed to align each frame to a matrix of the parameters [see Friston et al. (1997) for rele-
common orientation (Alpert et al., 1996). To effect the vant details]. Accordingly, parameter values are presented
image registration, we used the following multistep proce- with estimates of their standard error (SE). The hypothesis
dure: First, all frames were smoothed with a 10-mm FWHM ␥ ⱕ 0 is tested by computing t ⫽ ␥/SE(␥).
1054 N.M. Alpert et al. / NeuroImage 19 (2003) 1049 –1060

Fig. 1. Simulation of raclopride kinetics. In all cases the solid line is the control condition (see Table 1 for details) and all effects begin 38 min post injection
with ␶ ⫽ 0.22 1/min. (A) 33% increase in endogenous dopamine. (B) 20% increase in both K1 and k2. (C) 20% increase in k2 with K1 unchanged. (D) 20%
increase in K1.

Results holding the distribution volume constant has negligible ef-

fect; whereas, 20% increases in k2 alone may be difficult to
Basic simulations separate from displacement due to increasing levels of en-
dogenous dopamine. Increasing K1 results in an increased
In all cases the LSSRM made excellent fits to the sim- PET concentration. Fitting the LSSRM model to these data
ulated noise-free data. With graphical display, only very yielded good fits to the simulation curves and the differ-
small differences between data and fit could be seen. How- ences between simulation and fit were difficult to visualize
ever, as seen in Table 1, the parameter values for the with graphical display. All cases could be fit with ␣ ⫽ 0 and
LSSRM were biased and, as expected from the structure ␤ ⫽ 0; in the case of increased K1, the estimates of ␥ were
of the model, the covariance matrix of the parameter esti- negative. When two or more of the “activation” parameters
mates showed high correlation between k2 and k2a. Despite were included, the estimates were unreliable. Examination
this bias, values of binding potential determined from BP of the covariance matrix of the parameter estimates showed
⫽ k2 /k2a ⫺ 1 were close to the “true value.” The same very high correlation among two or more activation param-
simulation performed with the original nonlinear method, eters. We will discuss this point more fully later.
described by Gunn et al. (1997), gave very similar results,
with the small differences likely due to the difference in
Simulation 3, effect of statistical noise
fitting method employed.
Fig. 2 shows an example of the LSSRM fit to noisy data.
Simulation The parameters of this fit were close to the input values, R
⫽ 0.854 ⫾ 0.020, k2 ⫽ 0.195 ⫾ 0.004, k2a ⫽ 0.043 ⫾
Fig. 1 shows simulations that predict the kinetics of 0.001, and ␥ ⫽ 0.020 ⫾ 0.004. For an ensemble of 10,000
raclopride under the following four conditions: (1) displace- simulations with ␥ ⫽ 0, the mean t score for estimation of
ment resulting from increased levels of endogenous dopa- ␥ ⫽ 0 was 3.7. Examination of the distribution of t scores
mine; (2) perturbation resulting from increases in K1 and k2 obtained for the ensemble showed that t was greater than 3
(holding the distribution volume constant); (3) displacement in 93% of the simulations. When the same simulation was
resulting from increase of k2; and (4) perturbation resulting run with ␥ ⫽ 0, we found the mean t score for estimation of
from increase in K1. These data illustrate the magnitude of ␥ ⫽ 0 was ⫺0.07, with 7% of t scores greater than 2 and
the effects. Increasing transport and clearance 20% while 0.7% greater than 3.
 N.M. Alpert et al. / NeuroImage 19 (2003) 1049 –1060 1055

Fig. 2. Effect of noise on raclopride kinetics. The solid line is the control condition (see Table 1 for details) and all effects begin 38 min post injection with
␶ ⫽ 0.22 1/min. Simulation of 33% increase in endogenous dopamine plus noise (circles). Fit to the noisy data (dashed line). See text for details.

Human study raclopride by the motor planning task. Bilateral activation

was detected in the putamen and the caudate nucleus. Fig. 3
Analysis of measurements analyzed with the LSSRM shows the TAC extracted from an activated region and the
and ␶ ⫽ 0 showed clear evidence for displacement of fit obtained after varying ␶. Immediately following the ini-

Fig. 3. Time activity curve (TAC) and fit to data from a human study. The circles symbols indicate the positron emission tomographic (PET) concentration
in an “activated” putamen voxel. The vertical line at t ⫽ 38 min denotes the start time of the activation task. The ⫻ symbols indicate PET concentrations
from the reference region TAC. The solid line is the linear simplified reference region model (LSSRM) fit with ␶ ⫽ 0.221/min.
1056 N.M. Alpert et al. / NeuroImage 19 (2003) 1049 –1060

Fig. 4. Parametric images in transverse section. (A) Image of R. (B) Image of k2. (C) Image of k2a. (D) Image of binding potential. (E) Image of ␥. (F) Image
of SE(␥).

tiation of the task, a change in the putamen TAC (circle and C as k2/k2a ⫺ 1. Panel E shows the image of ␥; whereas,
symbols) suggestive of ligand displacement can be ob- panel F shows the standard error of ␥ as computed voxel-
served; whereas, no changes is seen in the reference TAC. by-voxel from the estimate of the covariance matrix. Panel
The LSSRM parameter estimates were R ⫽ 1.154 ⫾ 0.0451, E presents both positive (yellow 3 red) and negative esti-
k2 ⫽ 0.242 ⫾ 0.010, k2a ⫽ 0.0653 ⫾ 0.0035, and ␥ ⫽ mates of ␥ (blue 3 yellow). Examination of panel F shows
0.0283 ⫾ 0.0054. This leads to an estimate of BP ⫽ 2.704 high values of the standard error in ␥ everywhere except in
and an observed t score of 5.3, with df ⫽ 56, rejects the null the striatum. Accordingly, computation of t scores will
hypothesis ␥ ⱕ 0. The parametric maps were recomputed typically yield smaller values outside the striatum.
with ␶ ⫽ 0.22 min⫺1. Fig. 5 illustrates the fusion of the t score (t ⬎ 3, df ⫽ 56)
Fig. 4 presents an example of the parametric imaging computed voxel-by-voxel as the ratio ␥/SE(␥). Visual in-
results. There are six panels, all from the same slice. The spection shows that there is statistically significant increase
basic parameters of the SRRM are illustrated in the top row. in ligand displacement rate bilaterally in the putamen and
The binding potential (panel D) is computed from panel B right caudate nucleus.
 N.M. Alpert et al. / NeuroImage 19 (2003) 1049 –1060 1057

Fig. 5. Displacement of ligand by cognitive activation: The two panels show transverse sections through the putamen and caudate nucleus. The sections start
near the dorsal aspect of the putamen, top of left panel, and progress in 2.4-mm steps to the basal aspect, bottom of right panel. At each level on the right
is shown the parametric image of binding potential; on the left is a fusion of the t map (t ⬎ 3) for ligand displacement with the binding potential map.

Discussion ally more complex than it appears on the surface. The

analysis depends on quantitative comparison of the bind-
There are a number of ways to design and analyze a ing potential, but measurement of binding potential in the
cognitive neuromodulation study. Koepp et al. (1998) activated state assumes that the subject is in steady state
measured binding potential in two separate studies, a during activation. In fact, such measurements take 40 –90
control and an activation state. This design has the virtue min and the effects attributed to adaptation, learning, and
of simplicity of analysis because standard techniques habituation make the steady-state assumption question-
such as the SRRM can be used to measure binding able. Furthermore, the need for two separate measure-
potential in the two states. However, this design is actu- ments and quantitative comparison of the binding poten-
1058 N.M. Alpert et al. / NeuroImage 19 (2003) 1049 –1060

tial introduces a source of experimental error that may or task strategies to make more complex alterations in the
reduce the sensitivity of this design. kinetics and hence to improve identifiability. With ␶ ⫽ 0.22
The design employed by Pappata et al. (2002) is a clear min⫺1 the duration of the perturbation is limited and the
advance since it accepts that effects may be transient and covariance structure of model parameters reduces the model
uses alternating blocked stimuli (i.e., different tasks) in a identifiability, a finding that can be understood on an intu-
single-scan design. Their analysis follows the approach of itive basis. That is, when the perturbing effects are short-
Friston et al. (1997). The PET curve is modeled as a lived, there are likely to be an infinite number of parameter
weighted sum of curve shapes describing the null hypothe- sets that can fit the data about as well as one another.
sis (i.e., no activation), shapes describing the effects of Simulations with noisy data, similar in amplitude to that
cognitive ligand displacement, and CBF-related changes. observed in experiment, show that a high detection thresh-
The curve shape for the null hypothesis of no activation is old will be necessary for single-subject designs to control
based on previous measurements in other subjects; whereas, the false positive rate because with noisy data and low
curve shapes modeling cognitive ligand displacement or thresholds one expects a higher type II error. A more con-
CBF-related changes come from simulation. Hypothesis servative approach to control of type I and type II error
testing is effected within the general linear statistical model would be to use cohorts of subjects, combining the LSSRM
by testing whether the weight of a particular curve shape is with the methodology described by Aston et al. (2000) or
significantly greater than zero. However, the curve shapes possibly with an SPM-like analysis that accounts for be-
for the control state are not specific to the subjects analyzed; tween-subject variability.
and while the simulated curves are likely to be similar to the Dagher et al. (1997) and Pappata et al. (2002) have
effects they model, they are not derived from real data. On discussed the ambiguities that can arise from CBF-related
this basis, one expects the model predictions to resemble the effects. As noted in these studies, the available data are not
measured data but not to fit it in detail. No data were definitive regarding how serious these effects are in prac-
presented regarding the ability of this model to fit measured tice. We know from activation studies with 15O-H2O (e.g.,
data and no data were presented regarding the covariance of Hurtig et al., 1994) and fMRI that sensory-motor tasks
the model parameters. The use of canonical curve shapes is produce large flow changes. Changes from cognitive acti-
guaranteed to inflate the sum of squared differences be- vation tasks are usually much smaller, typically less than
tween the model prediction and the data, and since the error 15% (Andreasen et al., 1996). Jueptner et al. (1997) re-
variance is directly proportional to the sum of squares, the ported striatial activation in motor learning tasks with CBF
variance of parameter estimates will increase as well. Ac- increases in the range of 2– 4%. Simulations of the activa-
cordingly, it is likely that the approach of Friston and tion state demonstrated that the LSSRM fits the data well
colleagues (1997) is not optimal in a statistical sense. even in situations where CBF-related effects were on the
This study proposes an analysis, the LSSRM, that is order of 20%. We distinguish the following three cases: (1)
based on a kinetic model of the processes involved. The ligand transport rate increases with constant distribution
effects, or lack thereof, are estimated by fitting the model to volume, (2) plasma to tissue transport (K1) increases more
data from individual subjects. Non-steady-state effects are than clearance rate, and (3) clearance rate increases more
included by making the parameters time dependent. The than transport. In the first case, our simulations as well as
simulations reported here were informed and focused by those of Dahger et al. (1997) and Pappata et al. (2002) show
analyzing data from human studies. We used results from a that the changes in the PET binding curve are negligible. In
human study to set ␶, the rate of decline in activation effect, the second case, transient increase in the PET binding curve
in our simulations. We found that ␶ ⬇ 0.2 min⫺1, suggest- was demonstrated by simulation. When analyzed with the
ing, for the task studied, that the effect is nearly abolished LSRRM (fixing ␣ ⫽ 0 and ␤ ⫽ 0 and letting ␥ vary), we
by 10 min after the activation is initiated. A series of were able to fit the data with negative values of ␥. We
simulation studies, using the standard receptor model as interpret that to mean that increased transport, without li-
input, illustrate the basic properties of the method. We gand displacement, can be distinguished from the effect of
found using simulations of the “no activation” state with interest. In the third case, the binding curve is similar to that
noise-free data that the LSSRM fit the data very well. There obtained with ligand displacement and the LSRRM is un-
was significant bias in the parameter estimates; but the bias able to distinguish CBF-related effects. We do not know if,
in the binding potential was very small. Due to the structure or to what degree, case 3 exists in practice; however, it does
of the model, there is high correlation between k2 and k2a. not seem likely that CBF-related changes will be a serious
The bias in the parameter estimates is a characteristic of the problem in ligand displacement studies using cognitive ac-
simplified reference region model and arises from the fact tivation.
that the assumption of instantaneous equilibrium is not truly The results of our simulations should be evaluated in the
satisfied. light of the assumptions we made. First, we assumed that
In our examination of the LSRRM, we found that model changes in model parameters occurred immediately after
identifiability is limited to one time-dependent parameter. activation was initiated. Second, we not only assumed that
This finding is due to the simple shape of the PET curves. the activation effect diminished with time but also that the
It may be possible to employ different experimental designs decrease in effect was exponential. Therefore, these simu-
 N.M. Alpert et al. / NeuroImage 19 (2003) 1049 –1060 1059

lations could be taken as a point of departure for experi- Aston, J.A., Gunn, R.N., Worsley, K.J., Ma, Y., Evans, A.C., Dagher, A.,
mental designs that might seek to increase sensitivity by 2000. A statistical method for the analysis of positron emission tomog-
raphy neuroreceptor ligand data. Neuroimage 12, 245–256.
prolonging the effects or otherwise blunting habituation or
Badgaiyan, R.D., Fischman, A.J., Alpert, N.M., 2003. Striatal dopamine
overlearning of a cognitive or behavioral task. release during unrewarded motor task in human volunteers. NeuroRe-
A number of observations can be made about the port 14(8), in press.
LSSRM approach. First, LSSRM offers a simple computa- Coull, J.T., 1998. Neural correlates of attention and arousal: insights from
tional scheme. The data indicate it can make high quality electrophysiology, functional neuroimaging and psychopharmacology.
parametric maps of the parameters of the SRRM of Gunn et Prog. Neurobiol. 55, 343–361.
Dagher, A., Gunn, R.N., Lockwood, G., Cunningham, V.J., Grasby, P.M.,
al. (1997). Second, we found strong evidence of ligand
Brooks, D.J., 1998. Measuring neurotransmitter release with Positron
displacement. Using an ROI targeted to a single voxel in the Emission Tomography: methodological issues, in: Carson, R., Daube-
activated region we extracted a TAC whose shape was very Witherspoon, M., Herscovitch, P. (Eds.), Quantitative Functional Brain
similar to the predictions of the simulations. Third, these Imaging with Positron Emission Tomography. Academic Press, San
data provide the first direct evidence for the short duration Diego, pp. 449 – 454.
of neuromodulatory effects in human subjects performing a Endres, C.J., Carson, R.E., 1998. Assessment of dynamic neurotransmitter
cognitive task. Finally, further research is needed to evalu- changes with bolus or infusion delivery of neuroreceptor ligands.
J. Cereb. Blood Flow Metab. 18, 1196 –210.
ate the possibility of extending the method to different
Farde, L., Eriksson, L., Blomquist, G., Halldin, C., 1989. Kinetic analysis
dopaminergic receptors and to other neurotransmitters. of central [11C]Raclopride binding D2dopamine receptors studied by
PET—a comparison to the equilibrium analysis. J. Cereb. Blood Flow
Metab. 9, 696 –708.
Conclusions Fisher, R.E., Morris, E.D., Alpert, N.M., Fischman, A.J., 1995. In vivo
imaging of neuromodulatory synaptic transmission using PET: a re-
view of relevant neurophysiology. Hum. Brain Mapp. 3, 24 –34.
This study describes a novel approach to the design and
Friston, K.J., Holmes, A.P., Worsley, K.J., Poline, J.B., Frith, C.D., Frack-
analysis of experiments whose goal is to detect changes in owiak, R.S.J., 1995. Statistical parametric maps in functional imaging:
radioligand binding elicited by cognitive/behavioral tasks or a general approach. Hum. Brain Mapp. 2, 189 –210.
pharmacological challenge. We have presented a new the- Friston, K.J., Malizia, A.L., Wilson, S., Cunningham, V.J., Jones, T., Nutt,
oretical approach, the LSSRM, that is model driven and D.J., 1997. Analysis of dynamic radioligand displacement or “activa-
facilitates computation of parametric maps. The properties tion” studies. J. Cereb. Blood Flow Metab. 17, 80 –93.
Granon, S., Passetti, F., Thomas, K.L., Dalley, J.W., Everitt, B.J., Robbins,
of the LSSRM were studied by simulation, showing that the
T.W., 2000. Enhanced and impaired attentional performance after in-
theory and analytic approach can accommodate the effects fusion of D1 dopaminergic receptor agents into rat prefrontal cortex.
of CBF-related changes as well as ligand displacement. J. Neurosci. 20, 1208 –1215.
Simulations of the noise properties of the LSSRM suggested Greenwald, B.S., Davis, K.L., 1983. Experimental pharmacology of Alz-
that under realistic circumstances it is possible to detect and heimer disease. Adv. Neurol. 38, 87–102.
map neuromodulation. Mapping of increased dopamine re- Gunn, R.N., Lammertsma, A.A., Hume, S.P., Cunningham, V.J., 1997.
Parametric imaging of ligand-receptor binding in PET using a simpli-
lease was demonstrated in a human study.
fied reference region model. Neuroimage 6, 279 –287.
Hurtig, R.R., Hichwa, R.D., O’leary, D.S., Boles-Ponto, L.L., Narayana,
S., Watkins, L., Andreasen, N.C., 1994. The effects of timing and
Acknowledgments duration of cognitive activation in 15O water PET studies. J. Cereb.
Blood Flow Metab. 14, 423– 440.
The authors wish to thank the staff of the PET Imaging Jaber, M., Robinson, S.W., Missale, C., Caron, M.G., 1996. Dopamine
receptors and brain function. Neuropharmacology 35, 1503–1519.
and Cyclotron Laboratories for their technical assistance.
Jackson, S.R., Marrocco, R., Posner, M.I., 1994. Networks of anatomical
We also thank Evan Morris for valuable suggestions. areas controlling visuospatial attention: Models of neurodynamics and
R.D.B. acknowledges the support of grant T32CA09362. behavior. Neural Networks 7, 925–944.
Jentsch, J.D., Roth, R.H., Taylor, J.R., 2000. Role for dopamine in the
behavioral functions of the prefrontal corticostriatal system: implica-
References tions for mental disorders and psychotropic drug action. Prog. Brain
Res. 126, 433– 453.
Jueptner, M., Frith, C.D., Brooks, D.J., Frackowiak, R.S., Passingham,
Adam, L.E., Zaers, J., Ostertag, H., Trojan, H., Bellemann, M.E., Brix, G.,
1997. Performance evaluation of the whole-body PET scanner ECAT R.E., 1997. Anatomy of motor learning. II. Subcortical structures and
EXACT HR⫹ following the IEC standard. IEEE Trans. Nucl. Sci. 44, learning by trial and error. J. Neurophysiol. 77, 1325–1337.
1172–1179. Kahkonen, S., Ahveninen, J., Jaaskelainen, I.P., Kaakkola, S., Naatanen,
Alpert, N., Badgaiyan, R., Fischman, A., 2002. Detection of neuromodu- R., Huttunen, J., Pekkonen, E., 2001. Effects of haloperidol on selective
latory changes in specific neurotransmitter systems: experimental de- attention: a combined whole-head MEG and high-resolution EEG
sign and strategy. Neuroimage 16, Part 2, Abst. S53. study. Neuropsychopharmacology 25, 498 –504.
Alpert, N.M., Berdichevsky, D., Levin, Z., Morris, E.D., Fischman, A.J., Kimberg, D.Y., Aguirre, G.K., Lease, J., D’Esposito, M., 2001. Cortical
1996. Improved methods for image registration. Neuroimage 3, 10 –18. effects of bromocriptine, a D-2 dopamine receptor agonist, in human
Andreasen, N.C., Arndt, S.C., Cizadlo, T., O’Leary, D.S., Watkins, G.L., subjects, revealed by fMRI. Hum. Brain Mapp. 12, 246 –257.
Boles-Ponto, L.L., Hichwa, R.D., 1996. Sample size and statistical Koepp, M.J., Gunn, R.N., Lawrence, A.D., Cunningham, V.J., Dagher, A.,
power in [15O]H2O studies of human cognition. J. Cereb. Blood Flow Jones, T., Brooks, D.J., Bench, C.J., Grasby, P.M., 1998. Evidence for
Metab. 16, 804 – 816. striatal dopamine release during a video game. Nature 393, 266 –268.
1060 N.M. Alpert et al. / NeuroImage 19 (2003) 1049 –1060

Kulisevsky, J., 2000. Role of dopamine in learning and memory: implica- Pappata, S., Dehaene, S., Poline, J.B., Gregoire, M.C., Jobert, A., Delforge,
tions for the treatment of cognitive dysfunction in patients with Par- J., Frouin, V., Bottlaender, M., Dolle, F., Di Giamberardino, L., Syrota,
kinson’s disease. Drugs Aging 16, 365–379. A., 2002. In vivo detection of striatal dopamine release during reward:
Lammertsma, A.A., Bench, C.J., Hume, S.P., Osman, S., Gunn, K., a PET study with [11C]raclopride and a single dynamic scan approach.
Brooks, D.J., Frackowiak, R.S., 1996. Comparison of methods for Neuroimage 16, 1015–1027.
analysis of clinical [11C]raclopride studies. J. Cereb. Blood Flow Perry, E., Walker, M., Grace, J., Perry, R., 1999. Acetylcholine in mind: a
Metab. 16, 42–52. neurotransmitter correlate of consciousness? Trends Neurosci. 22, 273–
Lammertsma, A.A., Hume, S.P., 1996. Simplified reference tissue model 280.
for PET receptor studies. Neuroimage 4, 153–158. Robbins, T.W., 2000. Chemical neuromodulation of frontal-executive
Lawrence, A.D., Weeks, R.A., Brooks, D.J., Andrews, T.C., Watkins, functions in humans and other animals. Exp. Brain Res. 133, 130 –
L.H., Harding, A.E., Robbins, T.W., Sahakian, B.J., 1998. The rela- 138.
tionship between striatal dopamine receptor binding and cognitive Robbins, T.W., McAlonan, G., Muir, J.L., Everitt, B.J., 1997. Cognitive
performance in Huntington’s disease. Brain 121, 1343–1355. enhancers in theory and practice: studies of the cholinergic hypothesis
Morris, E.D., Fisher, R.E., Alpert, N.M., Rauch, S.L., Fischman, A.J., of cognitive deficits in Alzheimer’s disease. Behav. Brain Res. 83,
1995. In vivo imaging of neuromodulation using positron emission 15–23.
tomography: optimal ligand characteristics and task length for detec- Searle, S.R., 1971. Linear Models. John Wiley & Sons, New York.
tion of activation. Hum. Brain Mapp. 3, 35–55. Witte, E.A., Marrocco, R.T., 1997. Alteration of brain noradrenergic ac-
Oak, J.N., Oldenhof, J., Van Tol, H.H., 2000. The dopamine D(4) receptor: tivity in rhesus monkeys affects the alerting component of covert
one decade of research. Eur.J. Pharmacol. 405, 303–327. orienting. Psychopharmacology (Berl.) 132, 315–323.