Content uploaded by Massimo Mischi
Author content
All content in this area was uploaded by Massimo Mischi
Content may be subject to copyright.
IOP PUBLISHING PHYSIOLOGICAL MEASUREMENT
Physiol. Meas. 29 (2008) 281–294 doi:10.1088/0967-3334/29/3/001
On the physical and stochastic representation of an
indicator dilution curve as a gamma variate
M Mischi
1
, J A den Boer
1
and H H M Korsten
1,2
1
Department of Electrical Engineering, Eindhoven University of Technology, Den Dolech 2,
5612 AZ Eindhoven, The Netherlands
2
Department of Anesthesiology, Catharina Hospital, Michelangelolaan 2, 5623 EJ Eindhoven,
The Netherlands
E-mail: m.mischi@tue.nl
Received 1 October 2007, accepted for publication 19 December 2007
Published 11 February 2008
Online at stacks.iop.org/PM/29/281
Abstract
The analysis of intravascular indicator dynamics is important for cardiovascular
diagnostics as well as for the assessment of tissue perfusion, aimed at the
detection of ischemic regions or cancer hypervascularization. To this end,
indicator dilution curves are measured after the intravenous injection of an
indicator bolus and fitted by parametric models for the estimation of the
hemodynamic parameters of interest. Based on heuristic reasoning, the dilution
process is often modeled by a gamma variate. In this paper, we provide both
a physical and stochastic interpretation of the gamma variate model. The
accuracy of the model is compared with the local density random walk model,
a known model based on physics principles. Dilution curves were measured
by contrast ultrasonography both in vitro and in vivo (20 patients). Blood
volume measurements were used to test the accuracy and clinical relevance of
the estimated parameters. Both models provided accurate curve fits and volume
estimates. In conclusion, the proposed interpretations of the gamma variate
model describe physics aspects of the dilution process and lead to a better
understanding of the observed parameters, increasing the value and credibility
of the model, and possibly expanding its diagnostic applications.
Keywords: indicator dilution, physiological modeling, hemodynamics,
biomedical ultrasonics, cardiovascular system
1. Introduction
Hemodynamic quantification by means of indicators injected in the bloodstream has a long
clinical history. An indicator bolus is injected upstream and its concentration-versus-time
curve, referred to as indicator dilution curve (IDC), is measured downstream. Several methods
0967-3334/08/030281+14$30.00 © 2008 Institute of Physics and Engineering in Medicine Printed in the UK 281
282 M Mischi et al
are available for IDC measurements, ranging from the classic invasive methods requiring
patient catheterization, such as dye- or thermodilution (Fegler 1954, Murray Kinsman et al
1929), to the minimally invasive contrast imaging methods, such as contrast magnetic
resonance imaging (Tofts 1997), contrast ultrasonography (Feinstein 2004) and scintigraphy
(Anger 1964).
Indicator dilution measurements have several applications in clinical diagnostics. Initially,
indicators were used for cardiovascular quantification, aiming mainly at the assessment of
blood volumes and cardiac output (Hamilton et al 1928,Zierler2000). However, recent
developments in contrast imaging have also permitted novel applications for the detection of
myocardial or brain ischemia (Donnemiller et al 1997,Weiet al 1998), as well as angiogenetic
processes related to cancer growth (Eckersley et al 2002, Ruediger et al 1999). The assessment
of blood volumes and perfusion is based on the estimation of the indicator mean transit time
(MTT). Therefore, the accurate estimation of this parameter is of major clinical importance.
Owing to the indicator recirculation in the circulatory system and the low signal-to-noise
ratio (SNR) of measured IDCs, the use of proper models for IDC fitting and extraction
of hemodynamic parameters of clinical interest is necessary. The first model that was
proposed is a monoexponential decay, representing the washout IDC of one mixing chamber
(compartment) (Hamilton et al 1928,Zierler2000). Since then, several models have been
proposed for a closer representation of the IDC. Although some models, such as the local
density random walk (LDRW) model (Sheppard and Savage 1951), provide a physical
interpretation of the dilution process, being a solution of the diffusion with drift equation,
the models that are typically implemented in clinical devices are the lognormal and the
gamma variate models (Evans 1959, Linton et al 1995, Stow and Hetzel 1954, Thompson et al
1964). These models are empirically chosen due to their resemblance with measured IDCs,
while no physical background is provided (Evans 1959, Linton et al 1995, Stow and Hetzel
1954, Thompson et al 1964).
In this paper, we show that for a simplified dilution system the probability of detecting
an indicator particle at a certain site and time can be modeled by a gamma statistics, and
the gamma variate model can be adopted as the representation of the impulse response (IR)
of this dilution system. Eventually, we present both a physical and stochastic representation
of the dilution process by a gamma variate model. The model is fitted to the IDC by the
Levenberg–Marquardt (LM) method (Marquardt 1963). A multivariate linear regression in
the logarithmic domain (see the appendix) is used for its initialization.
For validation, we compared the indicator MTT value from in vitro experimental data
using the LDRW and gamma variate model fits. When multiplied by the flow (Q), the MTT
estimation permits the volume (V) assessment as (Norwich 1977, Sheppard 1962)
V = MTT · Q.
(1)
It is important to note that the MTT adopted for the volume assessment in (1) is the transit time
of an indicator particle that travels at the velocity of the carrier flow, and this can be viewed as
the transit time of the median particle or, equivalently, the mean of the spatial distribution of
the particles (Wise 1966).
The fits of both models to IDC data measured in patients were also compared. The
IDCs were measured in the right ventricle (RV) and left ventricle (LV). The assessment
of the MTT between these two measurement sites permits interesting clinical applications,
such as the estimation of the pulmonary blood volume (PBV) (Mischi et al 2004a), which
is closely related to the cardiac function (Bindels et al 2000, Buhre et al 2001, Wiesenack
et al 2001).
On the dilution curve representation by a gamma variate 283
Figure 1. Scheme of a multisection tube model for the simplified representation of the dilution
dynamics. The distance between the indicator injection and measurement sites x
i
and x
m
is divided
into k − 1 sections (compartments) of equal length x.
2. Methodology
2.1. Gamma variate model
The gamma variate model for IDC fitting is given as
C
(t) = A
· t
α
· e
−
t
β
, (2)
where α and β are the shape and scale parameters, respectively, and the factor A
defines the
IDC amplitude (Thompson et al 1964). This model can be viewed as a formal simplification
of a gamma probability density function (PDF), which is given as
f
(t; λ, k) =
λ e
−λt
(λt)
k−1
(k)
, (3)
where λ and k are the scale and the shape factor, respectively (Lindgren 1968). The operator
(·) represents the Gamma operator, i.e., (x) =
∞
0
y
x−1
e
−y
dy with x and y ∈ R, which
is a generalization of the factorial for non-integer numbers. The gamma variate model in (2)
can be easily derived from (3)fork − 1 = α, λ = β
−1
and A
= λ
k
(k)
−1
.
In this paper we focus on the derivation of the model for integer numbers, which can then
be generalized to non-integer numbers by introducing the operator (·). Therefore, we mainly
refer to the Erlang PDF, which is given as
f
E
(t; λ, k) =
λ e
−λt
(λt)
k−1
(k − 1)!
. (4)
(k)in(3) is substituted in (4) by the factorial (k − 1)!. In fact, the gamma distribution is
a generalization of the Erlang distribution for non-integer numbers (Hillier and Lieberman
2005). The PDF in (4) represents the waiting-time probability for the occurrence of k events
and is typically employed for traffic or production control problems (Hillier and Lieberman
2005).
2.2. Multisection tube model
For dilution modeling purposes, the hemodynamic system can be represented in simplified
approximation by an infinite tube along the x-axis, as shown in figure 1. A carrier fluid flows
through the tube and transports the indicator from the injection site x
i
to the measurement
site x
m
. For simplicity and without loss of generality, we can assume x
i
= 0. The tube can
284 M Mischi et al
be viewed as a multicompartment model, whose compartments are constituted by adjacent
sections of equal length x and volume V
c
= Ax, A being the tube cross-section area.
The compartments can be enumerated as x(x)
−1
. The use of this multisection tube
model reduces the dilution process description to the representation of one-dimensional
dynamics, i.e., the indicator transport and diffusion along the tube axis, defined by the spatial
coordinate x.
2.3. Physical representation
With reference to the model in figure 1, we characterize the complete dilution system as a
cascade of equal compartments (tube sections) with a constant flow rate. Each compartment
is seen as a mixing chamber with a unidirectional indicator washout and can be represented
by its IR, given as (Chen et al 1998)
f
C
(t) =
1
τ
e
−
t
τ
, (5)
where τ is the time constant and equals the ratio QV
C
−1
between the flow through the
compartment Q and the compartment volume V
C
. As a result, the IR of the complete system
from the first to the kth section (k =x(x)
−1
) is given by the k-fold convolution of the
compartment IR, which results in the same expression as the Erlang PDF in (4).
The physical meaning of the coefficients in (4) therefore becomes clear. The scale
parameter λ represents the inverse τ
−1
of the time constant of the compartment IR, while
the parameter k represents the compartment rank number. Back to the multisection tube
representation, (4) can then be written as
f
E
(t; τ, x) =
e
−
t
τ
τ
t
τ
x
x
−1
x
x
− 1
!
. (6)
In this simplified model, the flow is not completely realistic. Real flow types such as laminar
or turbulent flow could lead to imperfect mixing and deviation from (6). Here the radial
distribution of a laminar (Poiseuille) flow is reduced to the monodimensional dispersion of the
indicator along the tube axis, which is determined by the compartment size x and the time
constant τ . The presence of turbulent flow would also be reduced to the same monodimensional
interpretation.
2.4. Stochastic representation
As an alternative to the deterministic solution of a multicompartment model, a random variable
can be associated with the position of the indicator particle along the tube. In particular, a
binomial distribution can be employed for the representation of the indicator concentration
(probability density) as a result of the particle transitions (washout) from one section to the
following.
The movement (transition) of the indicator particle is represented as a result of a number
nt of trials, with n the number of trials per unit time and t the time interval. For each trial
we associate a probability p with the washout of the particle to the following section and a
probability 1 − p with the complementary case, i.e., the continued residence of the particle in
the same section. We are therefore assuming a unidirectional motion. For n np p, i.e.,
large n and small p, the Poisson distribution is a good approximation of a binomial distribution
On the dilution curve representation by a gamma variate 285
with λ = np constant (Papoulis 1991, Sheppard 1954). The probability of having k washouts
during a time t can therefore be represented by a Poisson process as
f
P
(k; t) =
e
−λt
(λt)
k
k!
. (7)
The parameter λ can be referred to as washout rate. The IR of the multisection tube dilution
system at a distance x = kx can be viewed in stochastic terms as the time PDF of having k
washouts. With T being the random variable associated with the time t and P the probability
function, this IR, h(t), can be derived as
h(t) =
∂
∂t
(1 − P [T>t|k]) =
∂
∂t
1 −
k−1
i=0
f
P
(i; t)
= f
E
(t; λ, k). (8)
Therefore, the Erlang PDF represents the dilution IR at distance kx from the injection site
when the parameter λ represents the washout rate, i.e., the expected number of transitions in
n trials.
By comparing the physical and statistical derivations of the model, we can easily conclude
that the washout rate λ = np, which corresponds to the scale parameter of the gamma
variate model, can be interpreted as the inverse of the time constant τ of the tube sections
(compartments). As a result, the MTT of the indicator from the injection to the detection site
equals kτ . By defining µ = kτ = kλ
−1
,theIRh(t) of the dilution system expressed by the
PDF in (4) and (6) can be written as
h(t) =
1
τ
·
t
τ
µ
τ
−1
·
e
−
t
τ
µ
τ
− 1
!
. (9)
The IDC at the detection site can be obtained by convolution of the IR with an input IDC. For
an input IDC modeled as a Dirac impulse of area mQ
−1
at time 0, the resulting IDC at the
measurement compartment equals
C(t) =
m
Qτ
·
t
τ
µ
τ
−1
·
e
−
t
τ
µ
τ
− 1
!
, (10)
where Q is the flow and m the injected indicator mass. According to the mass conservation
law, the integral of (10) equals mQ
−1
.AstheMTTµ represents the transit time of the median
indicator particle, which flows with the velocity of the carrier fluid, it can be multiplied by
the flow Q for the estimation of the volume between the injection and detection site. It is
interesting to note that µ is also equal to the first statistical moment of (10).
The constraint of having an integer k in the Erlang model can be again relaxed by
employing the Gamma operator (·). As a result, the factorial (µτ
−1
−1)! can be expressed
as (µτ
−1
). The integral and first statistical moment of (10) remain equal to mQ
−1
and µ,
respectively.
The derivation of the gamma variate formulation in (2) is straightforward. The MTT
is equal to (α +1)β = µ.ForA
= ((µτ
−1
))
−1
τ
−µ/τ
, the time integral of (2) from 0 to
∞ is equal to 1. As (2) represents the IR of the dilution system, the integral of the IDC,
i.e., the convolution of (2) with an input IDC, is equal to mQ
−1
, in agreement with the mass
conservation law.
This derivation of the gamma variate model results from the condition of a small value of p
in the binomial representation of the multisection tube. When this condition is not satisfied and
the probability p is sufficiently large, the binomial PDF can be approximated by a normal PDF
(Sheppard 1954, 1962). This approximation leads to a random walk type of description of the
dilution process, resulting eventually in the LDRW model. This model describes the contrast
286 M Mischi et al
Figure 2. Scheme of the in vitro setup for IDC measurements. A centrifugal pump generates a
water flow in a tube before the UCA injection site. The flow is measured by an electromagnetic
flowmeter. The injected UCA bolus is then detected by an ultrasound transducer before and after
its passage through a network of tubes that produces the volume to be estimated. Before the output
of the open fluid dynamic circuit, the static pressure is stabilized.
(This figure is in colour only in the electronic version)
dilution process by a normal spatial distribution of the indicator, whose mean translates with
the velocity of the carrier flow and whose variance increases as a linear function of time. The
model formulation in the continuous domain can be given as (Bogaard et al 1984,Wise1966)
C
LDRW
(t) =
m
Qµ
e
η
ηµ
2πt
e
−
η
2
(
t
µ
+
µ
t
)
, (11)
with µ equal to the transit time between the injection and detection sites of the spatial mean of
the indicator distribution (median indicator particle) and η being an adimensional parameter
equal to the ratio between convection and diffusion in the dilution system. The parameter µ
of the LDRW model, which is also referred to as MTT, can therefore be multiplied by the
flow in order to obtain volume estimates (Wise 1966). For the LDRW model, the MTT µ is
smaller than the first statistical moment of the model, which is equal to µ(1 + η
−1
) and can be
referred to as mean residence time (MRT) (Bogaard et al 1984, Mischi et al 2004a).
The LDRW model can also be derived by solving the diffusion with drift differential
equation (Norwich 1977). This derivation clarifies the physical meaning of the parameter η,
which for a tube model equals µQ
2
(2D)
−1
A
−2
, with D and A representing the longitudinal
diffusion coefficient of the system and the tube cross-sectional area, respectively. The
longitudinal diffusion coefficient D is related to the molecular diffusion coefficient and depends
on the type of flow (Bogaard et al 1984).
2.5. Experimentation
A dedicated set of measurements was performed aimed at testing the validity of the gamma
variate hypotheses. For this purpose, the in vitro setup described in Mischi et al (2004a)
for IDC measurements by contrast ultrasonography was used (figure 2). A network of tubes
of equal diameter simulated the dilution volume. The flow was generated by a calibrated
On the dilution curve representation by a gamma variate 287
Figure 3. Example of gamma variate (dashed line) and LDRW (solid line) fits to two IDCs (gray
lines) measured in vitro by contrast ultrasonography before and after a volume network of 1080 mL
with a flow of 1 L min
−1
. The correlation coefficients of the fits are 0.996 and 0.997, respectively.
550 bio-console centrifugal pump (Medtronic, Minneapolis, MN) and measured by the
embedded electromagnetic flowmeter. Different flows, ranging from 1 to 5 L min
−1
in
five equal steps, were tested. A bolus of 5 mL of UCA SonoVue
R
(Bracco, Milan, Italy)
diluted 1:100 was injected in the system. Rapid injections, which could be approximated by
a Dirac function, were performed manually. Two measurement sites were placed before and
after the dilution volume. The adopted ultrasound scanner was a Sonos 5500 (Philips Medical
Systems, Andover, MA) equipped with an S3 probe. The scanner was set in power modulation
mode at 1.9 MHz with a low mechanical index (0.1) in order to limit the collapse of UCA
microbubbles (Chang et al 2001).
The injected UCA dose for the adopted setting of the scanner ensured an approximately
linear relationship between UCA concentration and backscattered acoustic intensity (Mischi
et al 2004a). As a result, the IDC shape was preserved by the measuring system. The
acoustic intensity was measured by Q-Lab software (Philips Medical Systems, Andover, MA)
for acoustic quantification, which permits extracting the backscattered acoustic intensity from
any region of interest without the signal distortion due to logarithmic compression, typically
adopted by ultrasound imaging systems. In general, if the compression function is known,
the video output of the ultrasound scanner is also suitable for the measurement. The adopted
sampling frequency was 25 Hz.
The aim of this experimentation was the assessment of the dilution volumes based on the
IDC MTT estimations. Different volumes were tested, each with a carefully calibrated value,
ranging in four steps from 310 to 1080 mL. In total, 40 IDCs were obtained. Figure 3 shows,
as an example, a typical observed IDC.
The experimental in vitro IDCs were comparable to real IDCs measured in patients. In
order to illustrate this, we used data from 20 IDCs measured in patients from the RV and LV at
the Department of Cardiology of the Catharina Hospital in Eindhoven (The Netherlands).
These data were obtained after consent of the local ethical committee by trans-thoracic
echocardiography. These measurements were performed using the same ultrasound scanner
and setting adopted for the in vitro experiments. The same dose of UCA that was used for
the in vitro measurements was manually injected in an antecubital vein. In order to fit the
first pass IDC and exclude recirculation curves, the IDC was fitted up to 30% of the peak
concentration along the downslope (Millard 1997). Figure 4 shows an example of RV and LV
IDCs measured in vivo from a four-chamber view. The MTT estimation from these two IDCs,
288 M Mischi et al
Figure 4. Example of gamma variate (dashed line) and LDRW (solid line) fits to two IDCs
(gray lines) measured in a patient by contrast ultrasonography in the right ventricle (RV) and the
left ventricle (LV), respectively. Both fits have a correlation coefficient higher than 0.99 before
recirculation occurs.
similar to the in vitro experiments, permits the assessment of the blood volume between the
measurement sites, i.e., the PBV (Mischi et al 2004a).
The data from the in vitro as well as the in vivo experiments were used to obtain model fits
for both the gamma variate and the LDWR model. Before fitting, the IDC data were filtered
by a linear phase filter (50 taps) with a cut-off frequency at 2 Hz. The fitting was obtained
in two steps. The first step was a multivariate linear regression in the logarithmic domain,
as proposed in Mischi et al (2004b) for the LDRW fitting. This method includes an iterative
search for the ‘injection time’ t = 0 that results in the minimum mean squared error of the IDC
fit. The fitting parameters obtained in this step were used for the initialization of the second
step: a LM nonlinear iterative fitting. Examples of fitted curves are given in figures 3 (in vitro
data) and 4 (in vivo data).
The MTT of the indicator was derived as the parameter µ of the IDC model fits. The
value of the dilution volume was estimated as the product between the flow and the observed
MTT difference (MTT) derived from the IDC fits as
MTT = µ
2
− µ
1
, (12)
where µ
1
and µ
2
are the µ parameters of the first and the second IDC model fit, respectively.
Even though the injection time of the two IDCs was the same, usually their estimates did not
coincide, and the injection time estimate for the second IDC was delayed. In fact, the injection
time estimates that provide the best IDC fits do not always correspond to the real injection
time (Bogaard et al 1984), resulting typically in delayed estimates (Wise 1966). This time
difference was therefore added to the MTT in (12). The values of µ
1
and µ
2
were estimated
by both the gamma variate and the LDRW model fits. In vitro, the flow was measured by
the electromagnetic flowmeter embedded in the system. In vivo, ultrasound Doppler time
integration across the aorta was used for the cardiac output estimation (Huntsman et al 1983).
On the dilution curve representation by a gamma variate 289
Figure 5. In vitro volume estimates by the LDRW (squares) and the gamma variate (triangles)
models. The solid lines indicate the real volumes. The measurements are reported for flows
ranging between 1 and 5 L min
−1
.
3. Results
The fitting of all in vitro IDCs provided excellent results, with average correlation coefficients
R = 0.981 (standard deviation equal to 0.03) and R = 0.988 (standard deviation equal to 0.02)
for the LDRW and the gamma variate model, respectively. The improvement of the correlation
coefficient R between the initial linear regression and the subsequent LM fitting was 0.034
and 0.003 for the LDRW and the gamma variate model, respectively. In order to evaluate
whether the linear regression initialization was necessary, the sensitivity of the LM fitting to
the parameter initialization was tested. A variation of 5% to the initial values estimated by
the linear regression was therefore applied. Due to the low SNR of the measured IDCs, this
resulted in a decrease of the fit accuracy to an average correlation coefficient of 0.936 and
0.981 for the LDRW and the gamma variate fits, respectively. As a result, we concluded that,
especially for the LDRW model, an accurate parameter initialization was required.
The ranges of the observed parameters in the in vitro experiments were α ∈ [0.9, 20] and
β ∈ [0.05 s, 5.8 s] for the gamma variate model, and η ∈ [1.76, 179.5] and µ ∈ [0.53s,19.2s]
for the LDRW model.
The in vitro dilution volume estimates for all flows and volumes are reported in figure
5. The correlation between the estimated and the real volumes resulted in determination
coefficients R
2
= 0.998 and R
2
= 0.999 for LDRW and the gamma variate model, respectively.
The average standard deviation of each volume estimate over the five different flows was equal
to 2.0% and 1.0% of the calibrated (real) volume for the LDRW and the gamma variate model,
respectively. Both models underestimated the largest volume. However, for larger volumes
the gamma variate model provided larger volume estimates than the LDRW model. As a
result, the average underestimation of the largest volume was more limited for the gamma
variate than for the LDRW model (3.4% against 7.4%).
The average values of the model parameters estimated in patients were α = 1.3 ± 1.0,
β = 3.3 ± 2.4 s, η = 2.9 ± 2.5 and µ = 5.2 ± 3.1 s, for the 20 RV IDCs, and α = 2.1 ± 1.1,
β = 5.2 ± 3.0 s, η = 5.0 ± 2.6 and µ = 13.5 ± 5.2 s for the 20 LV IDCs. The average
290 M Mischi et al
Figure 6. Bland–Altman plot of the PBV estimates in patients by the two models. The difference
between the gamma variate and the LDRW volume estimates (Y-axis) is reported against the
average value of the two estimates (X-axis). The mean difference (black line) and the standard
deviation (dashed lines) are also shown.
values of α and β increased (p<0.02) from the RV to the LV IDC. The increase of α is a
clear consequence of the addition of a compartment (dilution volume) such as the lungs. The
increase of β can be interpreted as a result of an increase, for the same flow (cardiac output),
of the average time constant τ = β, due to the large volume of the added compartment.
As expected, also the parameter µ of the LDRW model increased significantly (p<10
−6
)
between the two IDCs. The average correlation coefficients of the fits in patients were R =
0.967 and R = 0.963 (standard deviations equal to 0.037 and 0.041) for the LDRW and the
gamma variate fits, respectively.
Based on the measured IDCs and the two models, PBV estimates in patients were also
derived. The comparison of the results provided by the two models is reported in the Bland–
Altman plot in figure 6 (Bland and Altman 1986). The mean difference between the gamma
and the LDRW PBV estimates was 63.1 mL, showing an increase for higher volumes. This
result confirms the results obtained by the in vitro measurements. The standard deviation
of the volume difference was 69.7 mL, i.e., 11.1% of the mean volume, which was equal to
623.1 mL and was estimated as the mean provided by the two models.
4. Discussion
A new formalism for the physical and stochastic representation of indicator dilution dynamics
by a gamma process is introduced. The model parameters are simultaneously interpreted as
characterizing unidirectional indicator washout from tube sections or waiting times between
successes, e.g., transitions of the indicator particles. These interlaced interpretations lead to a
representation of the indicator dilution systems in terms of indicator IR or PDF, respectively.
The dilution system identification reduces to the estimation of the time constant τ of the
compartment exponential washout curve and the number of compartments k representing the
dilution system, or, equivalently, to the estimate of the washout rate (success rate) λ and
the mean waiting time (the MTT µ) to have a number of washouts (successes) equal to k.
On the dilution curve representation by a gamma variate 291
While the use of the gamma variate model in previous publications was mainly based on
its heuristic resemblance with an IDC (Thompson et al 1964), this paper presents a structured
derivation of the model based on both physical and stochastic principles. Since the parameters
of the LDRW model were already known to have a physical meaning, being the LDRW
model derived as a solution of the diffusion equation, this brought us in a position that
allows a comparison of the applicability of these physical principles to real dilution systems.
Compared to the complexity of real hemodynamic systems, the proposed model derivation is
based on extremely simple assumptions. However, even the use of the gamma variate model,
a simple model with only three parameters, can already provide an accurate description of the
human pulmonary hemodynamic system, permitting the estimation of clinical parameters of
diagnostic relevance, such as for instance the PBV.
A dedicated model fitting based on a linear regression in the logarithmic domain was used
to initialize a subsequent nonlinear iterative regression. This is a critical issue as variations
of 5% of the initial values already resulted in a clear deterioration of the fit accuracy. In our
in vitro data, although the results of the linear regression were already accurate, the nonlinear
regression always led to increased correlations. However, the improvement was in most cases
not significant. In vivo, probably due to the lower SNR of the measured IDCs, a systematic
improvement was no longer recognizable.
In vitro and in vivo data were collected during controlled and careful measurements. The
quality of the IDC fits (R > 0.9 for all in vivo as well as in vitro data) suggests that both models
provide parameters that could be of clinical use. The shapes of the in vitro IDCs resembled
those observed in vivo and the distribution of the parameters estimated in vivo was amply
covered by the parameter intervals observed in vitro. Therefore, the in vitro parameters are
suitable for further interpretation aimed at clinical applications.
Obviously, the predictive value of our in vitro results for the hemodynamics in humans
is limited, due to the complexity of physiological dilution systems. Conclusions pertaining to
that system will require further validation by means of in vivo measurements. The presented
measurements in patients, however, already suggest a meaningful relationship between the
dilution system morphology and the estimated parameters. An example is the increase of
the number of compartments α and the time constant β when a large compartment, such as
the lungs, is included in the measured dilution system. In general, differently from other
authors (Thompson et al 1964), we did not find any contradiction between the estimated
values of the model parameters, both in vitro and in vivo, and the hypotheses from which
the models are derived. In particular, values of β that are larger than 1 do not represent a
contradiction according to the presented derivations.
The two analyzed models appear to be equally suited as a means for IDC fitting. A
comparison between the volume measurements based on the gamma variate and the LDRW
model showed a slightly higher correlation for the gamma variate model estimates. This
is a first confirmation of the gamma variate hypothesis, which, differently from the LDRW
model, is derived under the assumption of unidirectional movement of the indicator particles.
As a result, the transit time of the median indicator particle τ k = µ is equal to the mean
of the transit time distribution, i.e., the first statistical moment of the model. This differs
from the LDRW model, where the MTT µ is smaller than the first statistical moment of the
model, i.e., the MRT. Although further investigation is required, this might explain the volume
underestimation provided by the LDRW model for increasing volumes, confirmed by both our
in vitro and in vivo measurements.
The results of the gamma variate model in terms of accuracy of the IDC fits and MTT
estimations are comparable to those obtained by the LDRW model, although the models are
based on different physical assumptions. We may therefore conclude that the unidirectional
292 M Mischi et al
movement condition is a good representation of the real dilution system. In the physiological
system, the presence of valves in the heart as well as in the venous circulation might in fact
support the hypothesis of unidirectional motion.
5. Conclusions
This paper provides for the first time a combined physical and stochastic background to the
derivation of the gamma variate model for the representation of indicator dilution dynamics.
The theoretical derivations are supported by experimental data, involving a comparison with
the LDRW model.
As a result of this study, a meaningful interpretation of the model parameters can be
achieved, possibly adding diagnostic potential to the clinical use of the gamma variate model
for indicator dilution curve analysis. An immediate application, as shown in this paper, can
be the measurement of the PBV based on the MTT analysis of the RV and LV IDCs. This and
other possible applications require future work for the analysis of the parameters in vivo.
Appendix
The fitting of a model f(t; θ), which is a function of the parameter vector θ ,toanIDCC(t),
is performed by finding the vector
θ that minimizes the mean square error ε(θ) between the
real data and the model as
ε(θ) =
∞
0
(C(t) − f(t; θ))
2
dt. (A.1)
For simplicity and without loss of generality, we assume the injection time (beginning of the
IDC) equal to 0. If the model f(t; θ) can be represented as a linear combination of functions
x
i
(t) with i = 1, 2, ..., n,
f(t; θ) = θ
0
+ θ
1
x
1
(t) + θ
2
x
2
(t) + ···+ θ
n
x
n
(t), (A.2)
then the vector
θ that minimizes the mean square error (A.1) can be derived in the discrete
domain as (Krzanowski 1998)
θ = ([X]
t
[X])
−1
[X]
t
C =
⎡
⎢
⎢
⎢
⎣
θ
1
θ
2
.
.
.
θ
n
⎤
⎥
⎥
⎥
⎦
, (A.3)
where C is the vector containing the IDC samples and [X] is a matrix whose columns are filled
by the samples of the model functions x
n
(t), i.e.,
[X] =
⎛
⎜
⎝
1 x
11
... x
1n
1
.
.
.
.
.
.
.
.
.
1 x
n1
··· x
nn
⎞
⎟
⎠
. (A.4)
Although the gamma variate is a nonlinear function, a transformation in the logarithmic domain
ln(C
(t)) allows the use of a linear model as given in (A.2) for its representation. Therefore, a
linear regression can be applied for the model fitting, and the estimate for the vector θ can be
On the dilution curve representation by a gamma variate 293
directly derived by (A.3). In fact, the gamma variate model is defined by a three-dimensional
parameter vector:
θ =
⎡
⎢
⎢
⎢
⎣
ln(k)
−
1
β
α
⎤
⎥
⎥
⎥
⎦
, (A.5)
with x
1
(t) and x
2
(t) in (A.2) equal to t and ln(t), respectively.
References
Anger H O 1964 Scintillation camera with multichannel collimators J. Nucl. Med. 5 515–31
Bindels A J G H, van der Hoeven J G, Graafland A D, de Koning J and Meinders A E 2000 Relationships between
volume and pressure measurements and stroke volume in critically ill patients Crit. Care 4 193–9
Bland J M and Altman D G 1986 Statistical methods for assessing agreement between two methods of clinical
measurement Lancet i 307–10
Bogaard J M, Smith S J, Versprille A, Wise M E and Hagemeijer F 1984 Physiological interpretation of skewness of
indicator-dilution curves; theoretical considerations and practical application Basic Res. Cardiol. 79 479–93
Buhre W, Kazamaier S, Sonntag H and Weyland A 2001 Changes in cardiac output and intrathoracic blood volume:
a mathematical coupling of data? Acta Anaesthesiol. Scand. 45 863–7
Chang P P, Wen-Shiang C, Mourad P D, Poliachik S L and Crum L A 2001 Thresholds for inertial cavitation in albunex
suspensions under pulsed ultrasound conditions IEEE Trans. Ultrason. Ferroelectr. Freq. Control 48 161–70
Chen X, Schwarz K Q, Phillips D, Steinmetz S D and Schlief R 1998 A mathematical model for the assessment of
hemodynamic parameters using quantitative contrast echocardiography IEEE Trans. Biomed. Eng. 45 754–65
Donnemiller E, Heilmann J, Wenning G K, Berger W, Decristoforo C, Moncayo R, Poewe W and Ransmayr G
1997 Brain perfusion scintigraphy with
99m
Tc-HMPAO or
99m
Tc-ECD and
123
I-β-CIT single-photon emission
tomography in dementia of the Alzheimer-type and diffuse Lewy body disease Eur. J. Nucl. Med. Mol. Imaging
24 320–5
Eckersley R J, Sedelaar J P, Blomley M J, Wijkstra H, de Souza N M, Cosgrove D O and de la Rosette J J 2002
Quantitative microbubble enhanced transrectal ultrasound as a tool for monitoring hormonal treatment of prostate
carcinoma Prostate Cancer Prostatic Dis. 51 256–67
Evans R L 1959 Two comments on the estimation of blood flow and central volume from dye-dilution curves J. Appl.
Physiol. 14 457
Fegler G 1954 Measurement of cardiac output in anaesthetised animals by a thermodilution method Exp. Physiol.
39 153–64
Feinstein S B 2004 The powerful microbubble: from bench to bedside, from intravascular tracer to therapeutic delivery
system, and beyond Am.J.Physiol.287 H450–7
Hamilton W F, Moore J W, Kinsman J M and Spurling R G 1928 Simultaneous determination of the pulmonary and
systemic circulation times in man and of a figure related to cardiac output Am. J. Physiol. 84 338–4
Hillier F S and Lieberman G J 2005 Introduction to Operations Research (London: McGraw-Hill)
Huntsman L L, Stewart D K, Barnes S R, Franklin S B, Colocousis J S and Hessel E A 1983 Noninvasive Doppler
determination of cardiac output in man. Clinical validation Circulation 67 593–602
Krzanowski W J 1998 An Introduction to Statistical Modelling (London: Edward Arnold)
Lindgren B W 1968 Statistical Theory (New York: Macmillan)
Linton R A F, Linton N W F and Band D M 1995 A new method of analysing indicator dilution curves Cardiovasc.
Res. 30 930–8
Marquardt D 1963 An algorithm for least-squares estimation of nonlinear parameters SIAM J. Appl. Math. 11 431–41
Millard R K 1997 Indicator-dilution dispersion models and cardiac output computing methods Am.J.Physiol.Heart
Circ. Physiol. 272 H2004–12
Mischi M, Kalker A A C M and Korsten H H M 2004a Contrast echocardiography for pulmonary blood volume
quantification IEEE Trans. Ultrason. Ferroelectr. Freq. Control 51 1137–47
Mischi M, Kalker A A C M and Korsten H H M 2004b Videodensitometric methods for cardiac output measurements
Eurasip J. Appl. Signal Process. 2003 479–89
Murray Kinsman J, Moore J W and Hamilton W F 1929 Studies on the circulation: I. Injection method: physical and
mathematical considerations Am.J.Physiol.89 322–30
Norwich K H 1977 Molecular Dynamics in Biosystems (Oxford: Pergamon)
294 M Mischi et al
Papoulis A 1991 Probability, Random Variables, and Stochastic Processes (Singapore: McGraw-Hill)
Ruediger E P, Knopp M V, Hoffmann U, Milker Zabel S and Brix G 1999 Multicompartment analysis of gadolinium
chelate kinetics: blood-tissue exchange in mammary tumors by dynamic MR imaging J. Magn. Reson. Imaging
10 233–41
Sheppard C W 1954 Mathematical considerations on indicator dilution techniques Minn. Med. 93–104
Sheppard C W 1962 Basic Principles of Tracer Methods: Introduction to Mathematical Tracer Kinetics (New York:
Wiley)
Sheppard C W and Savage L J 1951 The random walk problem in relation to the physiology of circulatory mixing
Phys. Rev. 83 489–90
Stow R W and Hetzel P S 1954 An empirical formula for indicator-dilution curves as obtained in human beings
J. Appl. Physiol. 7 161–7
Thompson H K, Starmer C F, Whalen R E and Mcintosh H D 1964 Indicator transit time considered as a gamma
variate Circ. Res. 14 502–15
Tofts P S 1997 Modelling tracer kinetics in dynamic Gd-DPTA MR imaging J. Magn. Reson. Imaging 7 91–101
Wei K, Jayaweera A R, Firoozan S, Linka A, Skyba D M and Kaul S 1998 Quantification of myocardial blood flow
with ultrasound-induced destruction of microbubbles administrated as a constant venous infusion Circulation
97 473–83
Wiesenack C, Prasser C, Keyl C and Rodig G 2001 Assessment of intrathoracic blood volume as an indicator
of cardiac preload: single transpulmonary thermodilution technique versus assessment of pressure preload
parameters derived from a pulmonary artery catheter J. Cardiothorac. Vasc. Anesth. 15 584–8
Wise M E 1966 Tracer dilution curves in cardiology and random walk and lognormal distributions Acta Physiol.
Pharmacol. Neerl. 14 175–204
Zierler K 2000 Indicator dilution methods for measuring blood flow, volume, and other properties of biological
systems: a brief history and memoir Ann. Biomed. Eng. 28 836–48
