You are on page 1of 4

EXPERIMENTAL VALIDATION OF A POWER DOPPLER PERFORMANCE

MODEL

Roger J. Zemp, Michael F. Insana
Department of Biomedical Engineering, University of California, Davis, CA, USA 95616

Abstract A quantitative performance model for
Power Doppler detection tasks is developed based on
statistical decision theory. The model gives insights
into optimizing Power Doppler system designs. Flow-
weighted and scatterer density weighted images can
be formed by changing the pulse-repetition interval.
Wideband pulses are seen to be best for clutter
limited detection of blood while high amplitude or
longer energy-rich pulses are better for noise-limited
detection.

I. INTRODUCTION

Power Doppler (PD) ultrasound has emerged in the
last several years as a sensitive way of visualizing
blood flow in vivo. PD is sensitive to changes
between successive echo signals that are not rejected
by a clutter filter. Rather than measuring blood
velocity, PD is regarded as a way of detecting moving
scatterers. PD has been used in vascular applications
to assess flow anomalies, as well as in clinical
diagnosis of organ and tumor perfusion. Rubin and
colleagues [1] were among the first to introduce the
technique, and PD capabilities are now a standard
feature on many commercial ultrasound systems.
Despite widespread implementation, there is little in
the literature about engineering aspects of PD
systems, and there has been nothing regarding
performance assessment of PD. The purpose of this
paper is to provide a framework to understand PD
system performance in a rigorous and quantitative
way. We aim to describe how blood velocity and
decorrelation, clutter, and system properties such as
the pulse design and noise all interact to affect
performance. In this way we can understand how to
optimize PD for specific clinical tasks.
We focus on performance assessment for detection
tasks. For example, detecting a region of increased or
decreased perfusion or flow. We also briefly
comment on using PD for estimation tasks. Statistical
detection theory has been used in a previous paper [2]
to describe the ideal observer (or ideal detector) for
the performance of this task. A system design strategy
aims to maximize the ideal observer performance (in
an algorithm independent way), thus providing as
much diagnostically relevant information in the raw
data as possible. The role of signal and image
processing is then to make that information
maximally apparent to human observers. This paper
will briefly explain the ideal observer model and offer
evidence for validation.
This paper also introduces a non-prewhitening
observer model that uses average power as a test
statistic. The NPW model is intended to approximate
human observer performance, and complements the
ideal observer model by suggesting how the hardware
and algorithms interact to influence performance. It is
also computationally tractable in more situations than
the ideal observer.
These models are intended to guide system,
algorithm, and experimental design for specific
clinical situations.

II. SIGNAL MODEL

We assume that the ultrasound system can take a
sequence of snapshot images of the object, and that
the object can vary over time. With L pulses, there
will be L voltage traces initiated at times t
1
, , t
L
, at a
given pulse repetition interval (PRI), the linear
system may be written as [3]
[ ] ) , ( ) ( ) ( * ) ( ) (
k k c k b k
t n t z t z h t r x x, x, x x,
x
+ + (1)
where r is the measured voltage trace due to the echo
signal and n is a signal independent noise process.
The shift-invariant function h(x) is the system pulse-
echo impulse response. The object function consists
of two components: blood and tissue clutter acoustic
impedances represented as z
b
and z
c
respectively. We
assume the blood and tissue have large numbers of
scatterers per sample volume, and hence assume a
Gaussian distributed model for the objects and echo
signals. Blood may have some translational velocity
0-7803-7922-5/03/$17.00 (c) 2003 IEEE 2003 IEEE ULTRASONICS SYMPOSIUM-861
so that (in 1D) z
b
(x,t) = z
b
(x-vt). We also consider that
the blood can decorrelate over time, so that the
autocorrelation has some temporal decay profile.
Decorrelation over time may be due to a number of
effects such as intra-pulse type motions of blood
scatterers [4].

III. IDEAL OBSERVER
In [2], a model for the performance of the optimal
Bayesian detection system was described.

Large Ensemble Size: When a large ensemble size is
used, a block-circulant approximation can be made,
which means that Fourier-approximations are
reasonable. In this case the ideal observer signal-to-
noise ratio expressed in the frequency domain is:

L
k
k n k c
k b
GNEQ
f S f S H
f S H
d SNR
1
2
2
2
) , ( ) , ( ) (
) , ( ) (
u u u
u u
u
(2)
where S
c
and S
b
are the clutter and blood object power
spectra, S
n
is the noise power spectrum, H(u) is the
(fast time) Fourier transform of the pulse. In words,
Eq. 2 tells us that the PD detectability is determined
by the ratio of blood power spectral energy (speckle
energy) over spectral energy due to electronic noise
and the clutter speckle. Coherent velocity and
decorrelation shift energy away from the clutter
region, giving improved detectability.

Limited Ensemble Size: Often, ultrasound systems use
only 4-20 pulses and the large ensemble
approximation is inadequate. Here we consider the
case where the clutter is motionless (noise-limited
detection) and does not decorrelate. In this case
1
]
1

+
+

L H
S
S
S S H L
S H
d SNR
n
c
n c
Z
I
b
) ( ) (
) (
) ( ) (
) ( ) (
2
2
2
2
u u
u
u u
u u
u
(3)
where all temporal information has been integrated
into the factor , which describes everything about
the blood motion and decorrelation. Consider that
blood has coherent velocity v
b
, and decorrelates with
a decaying exponential trend over slow time
t
e

.
Then

'

L
m
m
a m L L L
1
) ( Re 2 ) 1 ( ) (u , (4)
where
b
t i t
e e a
v u

2
. The dot product between
u and v
b
acts to select how the system angular
spectrum H(u) will weight different flow velocities.
vanishes when there is no motion or decorrelation a
problem discussed in [5] for B-mode lesion detection.

IV. MODEL PREDICTIONS AND VALIDATION

Fig. 1 shows model predictions considering a 1D
imaging system. As coherent velocity is increased,
the blood spectral energy shifts away from the clutter
region at which point detection performance plateaus.
At twice the aliasing frequency (250 mm/s), blood
power is aliased over the wall filter and detectability
drops, as predicted in Fig. 1(a) and demonstrated
experimentally in Fig. 2.
Decorrelation actually helps rather than hinders
detection, which is in contrast to color flow
techniques for which coherence is important. When
attempting to detect a slowly decorrelating process
such as organ perfusion, detection sensitivity
increases with increasing PRF [Fig. 1(b)]. Also,
longer PRIs are less sensitive to variations in flow
rates, since statistically independent measurements
are obtained at each pulse transmission.

0 100 200 300 400 500
0
5
10
15
velocity mm/s
S
N
R
(a)
0 0.002 0.004 0.006 0.008 0.01
0
5
10
15
PRI
S
N
R
(b)


Fig. 1. Ideal observer SNR as a function of (a)
velocity and (b) PRI. In (a) detectability drops at
twice the aliasing velocity. Plot (b) had no coherent
velocity component only decorrelation.


Fig. 2. Experimental image of flow channel with
maximum velocity at twice the aliasing velocity.
2003 IEEE ULTRASONICS SYMPOSIUM-862
This suggests that different types of information can
be obtained with PD, as illustrated in Fig. 3. The
curves represent the average power across a flow
channel. A recently developed research interface on a
Siemens ANTARES ultrasound scanner allowed us to
access raw RF data and process it with an offline
MATLAB library [6]. The flow is the same for both
curves but the bottom curve uses a slower PRF. The
top curve is peaked in the center of the vessel where
the flow is fastest and we dub this a flow-weighted
image. The bottom, slow PRF curve will be termed a
scatterer-density weighted image. For detection tasks
it is desirable to use as much information about the
flow as possible hence flow-weighted images are
preferred. Other times one is concerned with
estimating microvascular density, and would prefer
that the flow-rate did not bias the observed power. In
this case a scatterer-density weighted image is
preferred. Contrast agents may be necessary.

0 1 2 3 4 5
2000
3000
4000
5000
6000
7000
8000
9000
10000
distance in channel (mm)
P
o
w
e
r

Fig. 3. Experimental PD measurements across a flow
phantom. The dashed line used a 488 Hz PRF. The
solid line used B-mode frames giving an effective
PRF of 48 Hz.

We used Monte Carlo studies to validate that the ideal
observer performance is monotonic with the observed
average Signal to Background Power Ratio (SBPR).
We simulated blood using an exponentially decaying
random process. This was superimposed on a
stationary clutter background, then blurred by a
Gaussian kernel simulating the system, followed by
addition of electronic noise. Fig. 4 shows the ideal
observer and SBPR performance with increasing
decorrelation rates. The observer monotonicity
validates the utility of the ideal observer model. The
relative inefficiency (~10%) means there is room for
improvement in our processing techniques.

0. 5 1 1. 5 2
2
4
6
8
10
12
Decorrel at ion rat e [ x PRF]
S
N
R
Id eal Ob server SNR
Mon t e Carl o SBPR

Fig. 4. Ideal Observer SNR (solid line) and SBPR
obtained from Monte Carlo simulation of a
decorrelating blood process with a large motionless
clutter component 40 dB below the blood. The clutter
to noise ratio was 26 dB.

The ideal observer tells us about the system design.
From a simplistic point of view (i.e. 1D A-scan lines),
our model predicts that wide bandwidth, high-energy
pulses give optimal detection performance. The
theory can also show that aperture broadening can
also contribute to enhanced detection.

V. NONPREWHITENING OBSERVER

We now model an observer that does not prewhiten
the data as does the ideal observer. We take the
average power in a region of interest as a statistic.
The test statistic can be histogrammed over a large
number of target present and target absent images.
The better the separation of the target present and
target absent distributions the better the detection
performance. A threshold can be set on the
distributions to make a decision. A Receiver
Operating Characteristic curve is a plot of the true
positive probability/fraction (TPF) for a fixed false
positive probability level (FPF), and is the gold
standard for quantifying decision tasks. In some cases
the squared peak separation normalized by the
average variance [SNR

2
] is a useful measure of
performance.
We consider 2-D correlation regions of the 3-D
RF data. Correlation regions can be defined as the
spatial extent of data correlations. These regions can
be broader than speckle spots because successive A-
scan lines may demonstrate coherent motion a
source of correlation. With the correlation region
2003 IEEE ULTRASONICS SYMPOSIUM-863

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPF
T
P
F

long pulse
short pulse with same amplitude
short pulse with equal energy
concept in hand, we can write the SNR

2
of the test
average power as:
[ ]
( )
2
2 2
2
2
2 /
o
c
I
SNR
S
A
SNR
+

+
+


, (5)
where A is the target area and S
c
is the area of the
correlation region. SNR
o
is the signal-to-noise of a PD
pixel, and may be expressed as:
[ ] { }
[ ] [ ]
2 2
2
2
+
+
+

K K
K K
tr tr
tr
SNR
o
(6)
where the signal present covariance is the covariance
of post wall-filtered slow-time echo data r
o
at a single
spatial location. This can also be decomposed into a
spectral representation by a discrete Karhunen-Loeve
transformation (i.e. eigendecomposition)
[ ] + +
+ + n c b
H t
o o
r r K . (7)
Thus
[ ] { }
( ) [ ] ( ) [ ] { } 2 /
2 2
2
2
n c n c b
b
o
tr tr
tr
SNR
+ + +

+
. (8)
When considering decorrelation processes, the
spectra are approximately independent of the pulse
shape, only pulse energy. Noise limited detection is
when the noise is the dominant factor hindering the
task. Clutter limited detection occurs when the post-
wall filtered clutter spectrum is the primary factor
confounding detection, and when the noise spectrum
is relatively insignificant.
Long or at least energy-rich pulses are
advantageous in noise limited situations. Coded
excitation techniques are one way of loading more
energy into a pulse. Matched filtering or
deconvolution [7] can be used to recover spatial
resolution. For clutter limited situations, short pulses
(broadband) are seen to be better (Fig. 5). This is
because broadband pulses give more independent
correlation regions to make a decision. Thus
shrinking the pulse duration shrinks the variance of
the average power (and thus bigger A/S
c
), while
having little effect on SNR
o
.

VI. SUMMARY AND CONCLUSIONS

Changing the PRI can lead to different types of PD
information: flow/perfusion-rate weighted images and
scatterer density-weighted images. PD levels are seen
to drop at twice the aliasing velocity. Short pulses
(wideband) pulses are preferred in clutter-limited
situations, while narrowband or coded excitation
pulses are preferred in noise-limited cases.


Fig. 5. ROC curve from Monte Carlo simulations
aimed to detect slow decorrelating flow in the
presence of decorrelating clutter.

VII. ACKNOWLEGEMENTS

Thanks to Jerome Mai and Siemens Medical
Solutions for work on the ANTARES URI and OPT.
We acknowledge funding from NIH R01 CA 82497.

VIII. REFERENCES

[1] J.M. Rubin, R.O. Bude, P.L. Carson, R.L. Bree, R.S.
Alder, Power Doppler US: A Potentially Useful
Alternative to Mean Frequency-based Color Doppler US,
Radiology, March 1994.
[2] R.J. Zemp, C.K. Abbey, M.F. Insana, Ideal Observer
Model for Detection of Blood Perfusion and Flow Using
Ultrasound Information Processing in Medical Imaging
2003, Ambleside, UK.
[3] R. J. Zemp, C.K. Abbey, M.F. Insana, Linear System
Models in Ultrasound: Application to Signal Statistics,
IEEE Trans. Ultrason. Ferroelect, Freq. Contr. Vol. 50, No.
6, 642-654, June 2003.
[4] V.L. Newhouse, D. Censor, T. Voutz, Ultrasound
Doppler Probing of Flows Transverse with respect to Beam
Axis, IEEE Trans. Biomed. Eng. Vol. 34, pg779-89, 1987
[5] R. J. Zemp, C.K. Abbey, M.F. Insana, Generalized
NEQ for Assessment of Ultrasound Image Quality, SPIE,
Physics of Medical Imaging, 2003.
[6] http://www.bme.ucdavis.edu/URI/
[7] Bruno Haider, Peter A. Lewin and Kai E. Thomenius,
"Pulse Elongation and Deconvolution Filtering for Medical
Ultrasonic Imaging", IEEE on UFFC, pp 98-113, vol 45,
No 1, 1998
2003 IEEE ULTRASONICS SYMPOSIUM-864