IEEE TRANSACTIONS ON MEDICAL IMAGING

Full Motion and Flow Field Recovery from Echo

Doppler Data
Muthuvel Arigovindan, Michael Suhling, Christian Jansen, Patrick Hunziker, and Michael Unser.

Abstract We present a new computational method for reconstructing a vector velocity field from scattered, pulsed-wave
ultrasound Doppler data. The main difficulty is that the Doppler
measurements are incomplete, for they do only capture the
velocity component along the beam direction. We thus propose to
combine measurements from different beam directions. However,
this is not yet sufficient to make the problem well posed because
(i) the angle between the directions is typically small; and (ii)
the data is noisy and nonuniformly sampled. We propose to
solve this reconstruction problem in the continuous domain using
regularization. The reconstruction is formulated as the minimizer
of a cost that is a weighted sum of two terms: (i) the sum of
squared difference between the Doppler data and the projected
velocities; (ii) a quadratic regularization functional that imposes
some smoothness on the velocity field. We express our solution for
this minimization problem in a B-spline basis, obtaining a sparse
system of equations that can be solved efficiently. Using synthetic
phantom data, we demonstrate the significance of tuning the
regularization according to the a priori knowledge about the
physical property of the motion. Next, we validate our method
using real phantom data for which the ground truth is known.
We then present reconstruction results obtained from clinical
data that originate from (i) blood flow in carotid bifurcation, (ii)
and cardiac wall motion.
Index Terms Pulsed wave Doppler, ultrasound Doppler, tissue
Doppler imaging, color Doppler imaging, color flow imaging,
echocardiography, non-uniform sampling, projected sampling,
velocity field reconstruction, vector field reconstruction, variational reconstruction, regularized reconstruction, shift-invariant
spaces.

I. I NTRODUCTION
Pulsed-wave ultrasound Doppler (PWD) imaging is an effective tool for the monitoring of vascular and cardiac function
[1], [2], [3]. The imaging system sends a periodic pulse train
along a set of scan lines and measures the backscattered signal.
By analyzing the Doppler frequency-shift in the received signal, the system retrieves a set of axial velocity estimates; these
are the projected components of the true 3-D velocity along
the direction of the ultrasound beam. These axial components
are resampled on a regular grid and presented in a color
coded form that is known as the color flow image. See [1]
for a comprehensive treatment of the various aspects of such
systems. In the context of clinical echocardiography, PWD
Muthuvel Arigovindan, Michael Suhling, and Michael Unser are with
Ecole Polytechnique Federale de Lausanne (EPFL), Biomedical Imaging Group, CH-1015 Lausanne, Switzerland; e-mail: mvel@ieee.org,
michael.unser@epfl.ch.
Christian Jansen, and Patrick Hunziker are with the Kantonsspital Basel, Petersgraben 5, CH-4031 Basel, Switzerland; e-mail: Christian.Jansen@unibas,ch, PHunziker@uhbs.ch.
This work was supported in part by the Swiss Heart Foundation and the
Swiss National Science Foundation under grant 200020-101821.

imaging is better known as color Doppler imaging when used

to assess blood flow, and tissue Doppler imaging, when applied
to tissue motion.
The color flow image sequence contains the instantaneous
velocity information, but it is incomplete; in particular, the
system is blind to the motion that is orthogonal to the beam.
Nevertheless, such a partial velocity field has been used for the
determination of quantitative parameters such as flow volume
[4], [5], [6]. Further, qualitative and to some extent quantitative
motion analysis obtained from this kind of data was found to
be clinically useful in several instances [7], [2], [8], [9], [10],
[3]. It can therefore be expected that the availability of full
vector field should lead to a better quantitative analysis.
An alternative to the Doppler methods is to use computational means to recover a full velocity field from a temporal
sequence of B-mode intensity images. This is typically done
using optical flow algorithms; a comparative evaluation of such
techniques in the context of echocardiography can be found
in [11]. A recent enhancement involves the use of a local
affine model that is well adapted to the motion of the heart
[12]. One can also think of improving the performance of
optical flow techniques by combining it with Doppler data to
improve the reconstruction quality of this local affine method
[13]. Even though these B-mode motion analysis techniques
hold promises for clinical practice, they have not superseded
Doppler imaging which is generally believed to be more
reliable and applicable to a wider range of situations. In that
respect, the down-sides of optical flow techniques are:
1) They do not work at all for estimating blood flow
because there is no visible structure on which to lock.
2) These methods are all based on the assumption that
a material point gives rise to the same intensity value
in successive frames. It is still not entirely clear to
which extent this assumption is valid in ultrasound
imaging because of the presence of multiple scatters that
induce speckle which is not necessarily correlated with
movement [14], [15].
3) There is some lack of temporal resolution because optical flow computations are based on frame-to-frame differences. This is in contrast with Doppler measurements
for which the equivalent averaging time (proportional to
the inverse of the pulse repetition rate) is significantly
less.
Doppler imaging would therefore come out as the method
of choice for assessing motion if it were not for its basic
limitation of only providing partial velocity measurements;
i.e., projections along the direction of the beam.
In this paper, we propose a new approach for reconstructing

IEEE TRANSACTIONS ON MEDICAL IMAGING

the true velocity field from the PWD data only. It is distinct
and complementary to the optical flow techniques in the sense
that it is also intended to work for blood flow estimation.
We formulate the task as the problem of finding a continuous
vector function that is the minimizer of a suitable criterion.
Specifically, we consider a quadratic cost functional that is
a weighted sum of a data term, and a physically inspired
smoothness functional (regularizer). The regularizer enforces
some coupling between x and y velocity components. We
demonstrate that the partial nature of the data makes this
coupling crucial.
We search for the continuous solution in a shift-invariant
space, since it allows, by choosing the step size, a trade-off between computational complexity and reconstruction accuracy.
We choose B-spline bases for the shift-invariant space, since it
gives the best trade-off, as justified in Section III.E. B-spline
solution is conveniently expressed in the form of a sparse
system of linear equations, thanks to the exact computation
of derivatives of the regularization functional [16].
Using a synthetic phantom, we demonstrate the superiority
of our coupled regularizer over the conventional regularizers.
We also validate our method using a real phantom experiment.
Finally, we present reconstruction results for blood flow in
carotid bifurcation and cardiac wall motion along with some
quantitative validation.
II. T HE R ECONSTRUCTION P ROBLEM
In this section, we give a brief description on the form of a
typical PWD data set. Then, based on a simple motion model,
we describe the indeterminacy of the Doppler data.
A. Form of the Data
A Doppler imaging system measures axial velocities at
some selected locations in the cross-sectional plane under
consideration. The selected locations {xi } are typically on a
noncartesian grid. These nonuniform measurements are scanconverted (resampled in a regular grid) and then displayed in
a color-coded form. A typical 2-D data set will contain the set
of sampling locations {xi }, the corresponding beam directions
{di }, and the Doppler measurements {mi } satisfying mi =
dTi v(xi ), where v(x) is the true velocity field. Note that
the sampling locations {xi } are nonuniform, especially when
there are multiple views.
Figure 1 gives the schematic of two commonly used
sampling geometries. The solid lines with arrows represent
the scan lines (ultrasound beam), whereas the dotted lines
represent the sampling grid along the scan lines. In the conebeam or sector-beam scheme, the beams are launched from a
single point along a series of regularly spaced angles. In the
parallel-beam acquisition, beams are launched with a fixed
angle from a set of points spaced regularly along a line. A
typical parallel-beam probe has the capability of steering the
beam angle, which means that different Doppler images can
be acquired without moving the probe (multiple acquisitions).
In this paper, we consider two types of data sets: (i) samples
of multiple acquisitions of the same cross-sectional plane
obtained from a cone-beam probe placed at different locations

(a) Cone-beam or sector-scan acquisition

Fig. 1.

(b) Parallel-beam acquisition

Scan-line schemes for 2-D acquisition

(sample locations and beam directions are transformed into

a common reference plane); (ii) multiple acquisitions from a
parallel-beam probe obtained by beam steering.
B. The deficiency of Doppler Data
Now we will demonstrate the deficiency of Doppler data
with the help of a toy example. Specifically, let us consider
a plane rotating with an angular velocity about the center
(xc , yc ); its velocity field is given by
u(x, y) = (y yc ),
v(x, y) = (x xc ),

(1)
(2)

where u and v are the x and y components of the velocity field.

Let the measurement device be a cone beam probe located at
the origin of the coordinate system. The probe measures the
samples of the following function:


u(x, y)
vd (x, y) = [dx (x, y) dy (x, y)]
v(x, y)
Here [dx (x, y) dy (x, y)]T is the direction of the beam from the
probe, whose components, in the present geometry, are given
by
x
y
dx (x, y) = p
, dy (x, y) = p
.
2
2
2
x +y
x + y2
Then vd (x, y) reads
vd (x, y) =

xyc yxc
p
.
x2 + y 2

It is clear from the above equation that we cannot recover the

motion parameters {, xc , yc } individually from the samples
of vd (x, y), no matter how many. We can only recover the
products xc and yc . In other words, any rotating plane with
an angular velocity, 1 , and a center of rotation, (x1 , y1 ), is
a candidate solution, provided 1 x1 = xc and 1 y1 = yc .
Hence, there is no reconstruction method that can recover the
motion uniquely.
We can resolve this ambiguity if we have the value of one
of the components (u or v) at any point. In fact, the minimal
measurement set required to determine all the three parameters
is the following: (i) samples of the function vd (x, y) at two
points that do not lie on the same scan line; (ii) a sample
of one of the components. Since it is not practical to obtain a

IEEE TRANSACTIONS ON MEDICAL IMAGING

measurement of a single velocity component, the only practical

way to resolve the ambiguity of Doppler data is to combine
multiple acquisitions, which is the strategy adopted in this
work.

Then, tvd becomes

tvd = WSc.
T

Let m = [ mi ] . Putting all these elements together,

we express (5) in the following standard matrix form:

III. T HE P ROPOSED M ETHOD

JLS (v)

Let v(x) = [u(x, y) v(x, y)] be the velocity field. Our

aim is to recover v(x) from the given Doppler measurement
set {xi , di , mi }. We search for the solution in the space of
uniform B-splines; in other words, we restrict the velocity field
to be of the following form:
Ng 1 Ng 1

u(x, y) =

X X
k=0

cuk,l n (x/a k) n (y/a l)

(3)

cvk,l n (x/a k) n (y/a l)

(4)

l=0

Ng 1 Ng 1

v(x, y) =

X X
k=0

l=0

Here, n is the B-spline of degree n and a is the step size

that controls the accuracy. Ng is the grid size. The idea is to
formulate the reconstruction as the minimizer of a quadratic
functional. Consequently, the expansion coefficients ck,l are
expressed as a solution of a linear system of equation. We
study three forms of the quadratic functional.
A. Least Squares Method (LS)
The least squares B-spline solution is the minimizer of the
following quadratic cost functional:
JLS (v) =

N
X

2

(5)

Let cu = [ cuk,l ]T , cv = [ cvk,l ]T , and c =

[cTu cTv ]T . Now, the task is to express the above functional
in terms of c. Let tu = [ u(xi ) ]T and tv =
[ v(xi ) ]T . The vectors tu and tv are computed as
tu = Ss cu ,

ALS
B

= ST WT WS,
= WS.

(10)
(11)

Note that ALS is a square matrix of size 2Ng2 . Finally, the least
squares reconstructioni.e., the minimizer of (9)is given by
the solution of the following equation:
cLS = A1
LS b,

(12)

where b = BT m. In the case where ALS is not of full rank,

we consider the minimum norm solution which is given by
cLS = ALS b, where ALS is the generalized (Moore-Penrose)
inverse of ALS .
B. Regularized Least Squares Method (RLS)
The reconstruction in the regularized least squares methods
is the minimizer of an extended cost functional; it is obtained
by adding a smoothness functional to the original least squares
functional, i.e.,

A typical choice of smoothness functional, R(v), that is

frequently used for the estimation of deformation fields is
a weighted sum of membrane spline and thin-plate spline
regularizers [17], [18] applied to each component. Specifically,
we have
R(v) = 1 (D1 (u) + D1 (v)) + 2 (D2 (u) + D2 (v))
Z Z 
D1 (u)

(14)

Z Z 
D2 (u)

{Wx }ii = dxi ,

{Wy }ii = dyi ,
where dxi and dyi are the components of di . Further, let
tvd = [ dTi v(xi ) ]T .
It can be verified that
tvd = Wx tu + Wy tv = Wx Scu + Wy Scv
We now define,
(7)
(8)

u(x, y)
x

2


+

u(x, y)
y

2
dxdy,
(15)

(6)

Let Wx and Wy be the diagonal matrices defined by

[W Wy ],
 x

Ss 0
S =
.
0 Ss

(13)

where

where the sample matrix Ss is defined as

{Ss }i,Ng l+k = n (xi /a k) n (yi /a l).

(9)

where

tv = Ss cv ,

JRLS (v) = JLS (v) + R(v)

dTi v(xi ) mi

i=1

= ktvd mk = kWSc mk
= cT ALS c 2cT BT m + mT m

2

u(x, y)
+2
x2
 2
2
u(x, y)
+
dxdy,
y 2

u(x, y)
xy

2

(16)

with 1 0 and 2 0. It is important to note that this

fairly standard regularizer does not have any coupled terms,
meaning that it does not enforce any special relationship
between velocity components.
The main task now is to express D1 (.) and D2 (.) in terms of
expansion coefficients. In the appendix, we show that these can
be written down relatively simply in terms of digital filtering
operations and l2 -inner products as given below:


D1 (u) = g(k, l) cuk,l , cuk,l l (Z2 )
(17)
2


D2 (u) = l(k, l) cuk,l , cuk,l l (Z2 )
(18)
2

IEEE TRANSACTIONS ON MEDICAL IMAGING

The digital filters g and l are conveniently characterized by

their z-transforms G(z1 , z2 ) and L(z1 , z2 ). For B-splines they
take the following form:
= (z1 + 2 + z11 )B 2n1 (z1 )B 2n+1 (z2 )
+ B 2n+1 (z1 )(z2 + 2 + z21 )B 2n1 (z2 ) (19)
1 
(z1 + 2 + z11 )2 B 2n3 (z1 )B 2n+1 (z2 )
L(z1 , z2 ) =
a2
+ 2(z1 + 2 + z11 )
B 2n1 (z1 )(z2 + 2 + z21 )B 2n1 (z2 )

+ B 2n+1 (z1 )(z2 + 2 + z21 )2 B 2n3 (z2 ) (20)
P
where B n (z) = kZ n (k)z 1 denotes the z-transform of
the discrete B-spline of degree n. Equations (17) and (18) can
also be written as

velocity within an infinitesimal neighborhood at the point of

interest. We construct RE (v) as given below:
Z

G(z1 , z2 )

D1 (u) = cTu Rg cu
D2 (u) = cTu Rl cu

(21)
(22)

where Rg and Rl are the circulant matrices corresponding to

G(z1 , z2 ) and L(z1 , z2 ), respectively. This together with (9)
yields the following expression for JRLS (v):
JRLS (v) = cT ALS c 2cT BT m + mT m
T

+ 1 c Rge c + 2 c Rle c,

(23)

where


Rg
0

0
Rg

Rl
0

0
Rl

Rge =

Rle =

RE (v) = d0
Z

(27)

(Curl v(x)) dx

+ c0
Z
+ c1

k (Curl v(x))k dx.

Since the divergence gives the density change, the first term
in the equation (27) quantifies the overall compression rate,
whereas the second term gives the spatial roughness of this
compression rate. Both terms are related to the deformation
of the medium. The third term sums up the squared angular
velocity. It does not directly quantify the deformation, but will
tend to penalize rotations including rigid ones. The last term,
on the other hand, is indeed a measure of deformation, as it
captures the spatial variation of the angular velocity.
The idea behind using this regularizer is that it includes
every derivative based quantity that has a direct physical
interpretation. This allows one to incorporate some a priori
knowledge of the type of velocity field and also to specify
physically plausible solutions. Interestingly, we have verified
that any rotationally invariant functional of order lesser or
equal to two takes the form (27). The proof will be published
elsewhere.

cRLS = A1
RLS b,

(Div v(x)) dx =

where ARLS = ALS + 1 Rge + 2 Rle . In practice, we do

not compute the inverse but we solve the system using direct
or iterative methods that are efficient for sparse systems.

+
+
Z

C. Vector Regularized Least Squares (VRLS)

(Curl v(x)) dx =

We now introduce the concept of a vector regularization that

relies on an extended smoothness functional that couples the
velocity components. The criterion to minimize is
(24)

k (Div v(x))k dx =

where RE (v) is our extended smoothness functional. The

functional, RE (v), is based on the divergence and curl of
the velocity field, which are defined as
Div v(x) = x u(x, y) + y v(x, y),
Curl v(x) = y u(x, y) + x v(x, y).

k (Div v(x))k dx

+ d1

As shown in the appendix, the various terms in (27) can be

written as

Thus the minimizer of (23) can be formally expressed as

JV RLS (v) = JLS (v) + RE (v),

(Div v(x)) dx

(25)
(26)

The divergence of the velocity field quantifies the rate change

of the density of the medium at a given point. The curl of a
velocity field, on the other hand, is equal to twice the angular

+
+
Z

k (Curl v(x))k dx =
+

r11 (k, l) cuk,l , cuk,l


r12 (k, l) cvk,l , cvk,l


2 r13 (k, l) cuk,l , cvk,l ,
(28)


u
u
r11 (k, l) ck,l , ck,l


r12 (k, l) cvk,l , cvk,l


2 r13 (k, l) cuk,l , cvk,l
(29)


u
u
r21 (k, l) ck,l , ck,l


r22 (k, l) cvk,l , cvk,l


2 r23 (k, l) cuk,l , cvk,l
(30)


r21 (k, l) cuk,l , cuk,l


r22 (k, l) cvk,l , cvk,l


2 r23 (k, l) cuk,l , cvk,l .
(31)

IEEE TRANSACTIONS ON MEDICAL IMAGING

where the underlying digital filters are defined as follows:

R11 (z1 , z2 )

R12 (z1 , z2 )

R13 (z1 , z2 )

R21 (z1 , z2 )

=
+

R22 (z1 , z2 )

=
+

R23 (z1 , z2 )

=
+

(z1 + 2 + z11 )
B 2n1 (z1 )B 2n+1 (z2 )
B 2n+1 (z1 )
(z2 + 2 + z21 )B 2n1 (z2 )
B 02n+1 (z1 )B 02n+1 (z21 )
1 
(z1 + 2 + z11 )2
a2
B 2n3 (z1 )B 2n+1 (z2 )
(z1 + 2 + z11 )B 2n1 (z1 )

(z2 + 2 + z21 )B 2n1 (z2 )
1  2n+1
B
(z1 )
a2
(z2 + 2 + z21 )2 B 2n3 (z2 )
(z1 + 2 + z11 )B 2n1 (z1 )

(z2 + 2 + z21 )B 2n1 (z2 )
1 
(z1 + 2 + z11 )
a2
D2n2 (z1 )B 02n+1 (z21 )
D2n+1 (z1 )

(z2 + 2 + z21 )B 02n2 (z21 )

(32)

Similarly, using (35), (36), and (20), we get

(33)
(34)

L(z1 , z2 ) = R21 (z1 , z2 ) + R22 (z1 , z2 ).

These two relations in turn imply
Rd0 + Rc0 = Rge
Rd1 + Rc1 = Rle

(35)

Thus we conclude that RE (v) = R(v) if d0 = c0 = 1 and

d1 = c1 = 2 . In fact, it can be shown that this relation is
valid for any vector function and not just for those that are
included in our B-spline reconstruction space.
Finally, the solution for the VRLS methodi.e., the minimizer of (38)is given by

(36)

cV RLS = A1
V RLS b,
where
AV RLS = ALS +d0 Rd0 +c0 Rc0 +d1 Rd1 +c1 Rc1 (43)

(37)

P
n
n
1
Here Dn (z) =
. In
kZ ( (x + 1/2) (x 1/2)) z
matrix form, Equations (28), (29), (30), and (31) read
Z
2
(Div v(x)) dx = cTu R11 cu + cTv R12 cv
+2cTu R13 cv
Z

Note that the functional, RE (v), includes R(v) as a special

case. To show this, we first observe from (32), (33) and (19)
that
G(z1 , z2 ) = R11 (z1 , z2 ) + R12 (z1 , z2 ).

(Curl v(x)) dx = cTu R12 cu + cTv R11 cv

Note that this kind of regularized reconstruction method

that involves the projected sampling model has been proposed
for velocity field reconstruction from tagged MRI data [19]
. However, the approach of Suter et al. differs from the
proposed method by two respects: (i) it does not involve
any physics based regularization; (ii) the method involves
the computation of globally supported functions which is
computationally expensive, whereas the proposed method is
tailored for computational efficiency as will be demonstrated
in the subsequent sections.

2cTu R13 cv
Z

D. Structure of the Linear System of Equations

k (Div v(x))k dx = cTu R21 cu + cTv R22 cv

+2cTu R23 cv

k (Curl v(x))k dx = cTu R22 cu + cTv R21 cv

2cTu R23 cv

The cost functional, JV RLS (v), now becomes

JV RLS (v)

= cT ALS c 2cT BT m + mT m
+d0 cT Rd0 c + c0 cT Rc0 c
+d1 cT Rd1 c + c1 cT Rd1 c,

(38)

Let us now have a closer look at the structure of the matrices

involved and make a few comments on the numerical methods
for solving the linear system of equations. We only consider
the linear system that correspond to VRLS method since the
other methods are some special cases of the VRLS method. It
can be shown that the matrix given in (43) can be express as


Au Auv
AV RLS =
Auv Av

 

Du Duv
Ru Ruv
=
+
(44)
Duv Dv
Ruv Rv
where

where


Rd0
Rc0
Rd1
Rc1

R11 R13
=
R13 R12


R12 R13
=
R13 R11


R21 R23
=
R23 R22


R22 R23
=
R23 R21

d2xi n (xi /a j) n (yi /a k)

{Dv }Ng k+j,Ng m+l

n (xi /a l) n (yi /a m),

(45)
X
2 n
n
=
dyi (xi /a j) (yi /a k)

{Duv }Ng k+j,Ng m+l

n (xi /a l) n (yi /a m),

(46)
X
n
n
=
dxi dyi (xi /a j) (yi /a k)

(39)
(40)

{Du }Ng k+j,Ng m+l

(41)
(42)

n (xi /a l) n (yi /a m),

(47)

IEEE TRANSACTIONS ON MEDICAL IMAGING

and Ru , Ruv and Ruv are the circulant matrices corresponding to some filters Ru (z1 , z2 ), Ruv (z1 , z2 ) and Ruv (z1 , z2 )
respectively. The filters are given by
Ru (z1 , z2 )

= d0 R11 (z1 , z2 ) + c0 R12 (z1 , z2 )

+d1 R21 (z1 , z2 ) + c1 R22 (z1 , z2 )
= d0 R12 (z1 , z2 ) + c0 R11 (z1 , z2 )
+d1 R22 (z1 , z2 ) + c1 R21 (z1 , z2 )
= (d0 c0 )R13 (z1 , z2 )
+(d1 c1 )R23 (z1 , z2 )

Rv (z1 , z2 )
Ruv (z1 , z2 )

The above matrices are of size Ng2 Ng2 . From Equations

(45), (46), and (47), it can be observed that the sample
matrices Du , Dv , and Duv are sparse. In particular, assuming
that each grid interval has a sample point (worst case sparsity)
it can be verified that an element {}Ng k+j,Ng m+l is zero for
max(|l j|, |m k|) > n (finite support of B-splines). Hence
the matrices are block-band diagonal of width 2n+1 and each
of the constituent block of size Ng Ng is also band diagonal
of width 2n + 1. Moreover, these matrices are symmetric and
each constituent block is symmetric as well. Next, from the
expression (32), (33), (34), (35), (36), and (37), the filters are
symmetric and of length n + 1. This implies the matrices Ru ,
Rv , and Ruv have the same structure as that of the sample
matrices (with the additional property of being circulant). As
a result, the matrices Au , Av , and Auv also have the same
sparsity and symmetry.
Hence, the overall matrix AV RLS is sparse and symmetric.
Since AV RLS is also typically non-singular as in any regularized linear problem, it is positive definite as well.
It can be shown that the RHS of the linear system of
equations is given by


bu
b=
,
(48)
bv
=

E. Computational Complexity
The computational task essentially consists of two parts: (i)
constructing the linear system of equations, (ii) and solving
the system. The first is much less demanding than the second.
The first part is essentially the computation of the matrices
Du , Dv , and Duv using (45), (46), and (47) and then adding
the circulent matrices. The latter is negligible whereas the
former is proportional to N (n + 1) where N is the number of
samples points.
The second part, which is the computation of the solution,
has a complexity that is proportional to the number of non V RLS . Because of the
redundant and non-zero elements in A
symmetry and the sparsity properties identified before, the
number of non-zero elements is equal to (Ng (n + 1))2 . Hence
the
of solving the linear system is proportional to
i
h complexity
(n+1)
a

dxi n (xi /a j) n (yi /a k)mi ,

. On the other hand, since the approximation error

for B-spline is in the O(an+1 ), the reconstruction accuracy is
proportional to an+1 . As a consequence, a becomes a trade-off
factor that allows a compromise between the computational
complexity and the reconstruction accuracy. In this regard,
B-splines give the best trade-off since they have the highest
approximation order for a given support size [20] (support size
determines the computational complexity).

dyi n (xi /a j) n (yi /a k)mi .

IV. E XPERIMENTS

where
{bu }Ng k+j

The main advantage of this modified linear system is that the

retains the sparse structure of its constituent matrices
matrix A
Au , Av , and Auv with 2 2 matrices playing the role of the
scalars elements in the latter ones. In this configuration, the
non-zero elements are all concentrated near the diagonal. This
allows us to conclude that the solution can be obtained by
solving a sparse, well-conditioned, diagonally-dominant, and
positive-definite linear system of equations. Hence we have
at our disposal a rich set of computationally efficient iterative
methods to achieve the required reconstruction. This is the
primary advantage of the proposed scheme over an analytical
method such as the one presented in [19].

{bv }Ng k+j

We now consider an important point on the ordering of

coefficients {cuk,l } and {cvk,l } to get the coefficient vector c. Until now, we used the concatenated ordering c =
[ cuk,l cvk,l ]T to facilitate the derivations. However,
there is an alternative ordering that is physically more meaningful:
c = [ cuk,l cvk,l ]T .
The corresponding linear system of equations is given by

= b,
Ac

A. Tuning the parameters of the algorithm

where

 

1 0
0
Au
+ Av
0 0
0



0 1
+ Auv
1 0
 
 
1
0

b = bu
+ bv
0
1

To demonstrate the feasibility of the proposed formulation

and to compare different methods, we first study the reconstruction of synthetic data. We then move to real phantom data,
where the motion is controlled and known a priori. Finally,
we present reconstruction results obtained from two types of
clinical data for the assessment of : (i) blood flow in carotid
bifurcation, and (ii) cardiac wall motion. We use a Gaussian
noise model in our experiments, since the noise in Doppler
systems is usually found to be Gaussian [1].

0
1



(49)
(50)

We tune the regularization parameters by a two step procedure. First, the parameter search space is constrained by
incorporating some a priori knowledge on the type motion
if available. Next, by performing multiple reconstructions,
the parameters are updated by an iterative search within the
constrained search space such that the difference between the
data term of the cost functional, JLS , and the noise variance

IEEE TRANSACTIONS ON MEDICAL IMAGING

v1 (x) and v2 (x), where

s1

vj (x)

s2
Fig. 2.

Schematic for synthetic phantom experiment

is minimum. In other words, the parameters are obtained as

the minimizer of the the following cost:


JLS (vr ) 2 
,
(51)
C=
2
where vr (x) is the reconstructed velocity field, and 2 is
the noise variance. The idea behind this strategy is that, if
the reconstruction function coincides with the underlying true
velocity field, then the error in the data term is entirely due
to noise implying that JLS (vr ) 2 .
Since JLS (vr ) is an increasing function with respect to each
regularization parameter, we can achieve a faster convergence
if we keep the parameters proportional to 2 .
B. Validating the reconstructed function
While it is easy to validate the reconstructed field in a test
case where the ground truth is known, it is less straightforward
to do so in a clinical scenario. Here, we propose a crossvalidation strategy for the quantitative assessment of our
method. The idea is to use only part of the available data for
field reconstruction by the above mentioned method (including
the iterative search for optimal parameter determination). We
then evaluate (51) using the data that was left out from the
reconstruction process. If the cost (51) is sufficiently close
to the minimum attained during the iterative reconstruction,
then we are able to conclude that the reconstructed field is
consistent with the full data set, suggesting that it is a good
approximation of the underlying true velocity field.
C. Synthetic Rotating Phantom
Here we consider the vector field scenario described in
Section II-B, which corresponds to a planar rotation with a
constant angular velocity. We demonstrated earlier that the
single probe measurement set is ambiguous and does not
permit the recovery of the motion parameters. Hence we
consider measurements from two probe locations, since it is
the only practical way to resolve the ambiguity. We study the
performance of all three method on such a measurement set.
The required Doppler data is simulated for two sector scan
probes. Let s1 = [x1 y1 ]T and s2 = [x2 y2 ]T be the locations,
and let be the angle subtended by the probe locations with
respect to the center of rotation (figure 2). The idealized
phantom data is now obtained by sampling two functions

= vj (x, y),
= dTj (x)v(x),

=

[dxj (x, y) dyj (x, y)]

u( x, y)
v( x, y)


.

Here, u(x, y) = (y yc ), v(x, y) = (x xc ), and

x xj
dxj (x, y) = p
,
(x xj )2 + (y yj )2
y yj
dyj (x, y) = p
.
(x xj )2 + (y yj )2
We are also adding noise to make the problem more realistic.
Each of the functions v1 (x) and v2 (x) are uniformly sampled
with respect to the polar coordinate systems with origins s1 ,
and s2 , respectively. However, note that the samples are nonuniform with respect to our cartesian reconstruction system.
Let {xji , i = 1, . . . , N } be the set of sample locations for the
probe sj . The input for the algorithm is given by the following
list of triplets:
{xji , dji , mji , j = 1, 2; i = 1, . . . , N },
where dji = dj (xji ), and mji = dTj (xji )v(xji ) + nji
with nji being some i.i.d Gaussian measurement noise with
variance 2 . Note that the richness of the data set is controlled
by the magnitude of , and that the reconstruction becomes
easier as tends to 90o .
Now we consider the choice of regularization parameters.
For a rigid rotation, the divergence is zero. The gradient
of curl is zero as well except at the boundaries. Hence the
appropriate setting will be {d0 , d1 = 0, c0 =
0, c1 }. However, since the gradient of curl is nonzero at the boundaries, the choice c1 will lead to a
sizable error therein. A good compromise can be achieved by
choosing {d0 , d1 = 0, c0 = 0, c1 = k 2 }, where k
is an adequately chosen positive real number. The constant k is
chosen iteratively to match the variance as described before.
For the RLS method, we need to set 1 = 0, since a rigid
rotation contains the first order spatial variation. Hence the
appropriate setting is {1 = 0, 2 = k 2 }.
Before presenting our results, we define the following
quantities:

!
E dTji v(xji )2
Input SNR = 10 log
2
#!
"
2
kv(k)k
Recons. SNR = 10 log E
2
kvr (k) v(k)k
where v(x) is the true velocity field. Note that the input SNR
is computed over the available sample locations, whereas the
reconstruction SNR is computed over the reconstruction grid.
Figure 3 gives reconstruction SNR for VRLS, RLS, and
LS as a function of , for input SNR equal to 44.5 dB. It
can be observed that the reconstruction error decreases with
increasing , as one expects. Also, one can clearly see the
superiority of VRLS. Figure 4 gives the reconstruction SNR
for all three methods as a function of the input SNR with

IEEE TRANSACTIONS ON MEDICAL IMAGING

50

In this experiment too, the input data are the measurements

simulated for two sector scan probes (Figure 2) with some
i.i.d Gaussian noise.
Since the velocity field has a non-zero divergence and a nonzero curl, we have to choose d0 = c0 = 0. The remaining
parameters d1 and c1 are chosen iteratively. Interestingly,
we found that the resulting parameters satisfy

Reconstruction SNR (dB)

40

30
VRLS, Input SNR = 44.5 dB
RLS, Input SNR = 44.5 dB
LS, Input SNR = 44.5 dB

20

d1 = k 2 2 ,
10

c1 = k 2 2 ,

10

Fig. 3.

10

20

30

40

50
(degress)

60

70

80

90

Reconstruction error for synthetic phantom experiment.

50

40

VRLS
RLS
LS

30

Output SNR (dB)

20

10

10

20

30

40

Fig. 4.

10

15

20
25
Input SNR (dB)

30

35

40

45

Reconstruction error for synthetic phantom experiment.

= 30o . Note that VRLS performs the best, and that the
reconstruction error for the conventional LS approach is far
worse (off by more than 30 dB). All the reconstructions were
done on a 16 16 grid with a = 4 and with the number of
samples equal to the number of grid points.
Note that when there is no input noise, the reconstruction
error is dominated by the minimum amount of smoothing
that is required to make the linear system of equations well
conditioned.
D. Synthetic Phantom with Non-Rigid Motion
In this experiment, we consider a more complex model
for the synthetic phantom. Specifically, we adopt a non-rigid
motion model given by




(kxk)
y (kxk)
v(x) = x
+

(52)

y (kxk)
x (kxk)
2

1 r /2
where (r) = 2
e
, 0, and 0. The first term is
the curl-free component and the second is the divergence-free
component. In continuum mechanics, they are identified as the
irrotational and solenoidal components of the velocity field.

where k is the parameters that is tuned iteratively. In other

words, the ratio dc1 is equal to the ratio of the energies

1
of divergence-free and curl-free parts of the original velocity
field.
The chosen range for the reconstruction grid is [3, 3]
(essential support of the Gaussian) and the average sampling
density is 28 samples per unit area. Figure 5 compares the
reconstruction result for VRLS obtained for different values
of the step size a with = 30o , = 0.25, and = 0.75.
The input SNR is 10 dB. The reconstruction SNR is nearly
constant when a is sufficiently small and falls off drastically
when it becomes too large. For practical purpose, the choice
a = 2 offers a good compromise in terms of quality and
computational cost (SNR = 14.4 dB).
We also performed reconstruction with two modified probe
setting for a = 2: (i) three views with the third view in the
middle of the first two views (angle between the outer views
is 30o and angle between adjacent views is 15o ); (ii) two
views with angle 45o . The first modification yields an SNR
of 15.2 dB, whereas second yields 25.5 dB. Qualitatively, we
can say that the improvement in the first case is due to a
better noise removal, whereas the improvement in the second
is due to an increase in the amount of information, which is
more significant than that of the first. We observed that, for
the second, when the angle is increased beyond 90o , the SNR
starts decreasing. With some simplifying assumptions on the
probe geometries, the reason for this behavior can be explained
analytically for any type of velocity field. We are skipping this
proof due to space limitations.
E. Real Phantom Experiment
The real phantom that we constructed for this experiment
is a cylindrical tissue-mimicking object (gelatin) of radius 7.5
cms immersed in a water container. The object rotates with
a constant angular velocity of 0.5 rad/s. Doppler data were
acquired using a sector scan probe, and reconstruction was performed from views differing by 10 degrees. The regularization
parameters were tuned in the same way as that of the synthetic
phantom experiment. Figure 6 shows one frame of the B-mode
intensity image with the super-imposed reconstructed motion
field. Since we know the true velocity field, we can compute
the ideal measurements at the sample locations. Hence we
can compute the input noise variance, which was found to be
14.27 dB. The reconstruction SNR for the VRLS method is
14.67 dB. Average angular reconstruction error is 3 degrees.
The average radial velocity in the reconstructed velocity field
is about 3.5 % of the peak velocity in the true velocity field.

IEEE TRANSACTIONS ON MEDICAL IMAGING

16

two cycles. Figure 7 gives one of the extracted frames for

each view angle. Note that the top of the color scale represents
the peak negative axial velocity and the bottom represents the
peak positive velocity. Also note that the color map has been
inverted for the first view when compared to the other two
views. The noise variance of the Doppler data was estimated
by comparing frames corresponding to different cycles. The
noise variance was estimated to be 9% of the signal energy.

14

Reconstruction SNR (dB)

12

10

4
Step size a

Fig. 5. VRLS Reconstruction error for synthetic phantom with non-rigid

motion. Input SNR = 10 dB.

(a) = 70o

(b) = 90o

(c) = 110o
Fig. 7.

Fig. 6.

Reconstructed motion field for the rotating real phantom.

F. Blood Flow in Carotid Bifurcation

Blood flow patterns in the carotid bifurcation have been
reported to have significant influence on the development of
artherosclerosis [21] which is a leading cause of heart attack
and stroke. In this experiment, we attempted to reconstruct
the complete blood-flow distribution in the carotid bifurcation
from multiple-view Doppler data acquired using a parallelbeam probe setup with steering capability. We performed
three acquisitions with beam angles = 70o , 90o , 110o . Note
that we have acquisitions obtained from three views without
moving the probe.
Imaging was performed using Acuson Sequoia Echo 512
machine with Siemens 6L3 probe. An ultrasound frequency
of 3.5 MHz was used. The penetration depth is 30 mm. Color
calibration was done for the axial velocity ranging from - 0.17
m/s to 0.17 m/s.
We extracted the frames corresponding to the R-peak over

Selected color Doppler images of carotid bifurcation.

We used 75% of the total samples that were selected

randomly (data set A) for reconstruction and the remaining
samples (data set B) for validation. Since we do not have
any a priori knowledge about the flow pattern, all the four
regularization parameters were tuned iteratively. The minimum
attained (equation (51)) was in the order of 104 . Figure 8(a)
gives a visualization of the reconstructed flow field, whereas
the figure 8(b) gives the flow profiles across the vessels before
and after the bifurcation. We observe that, in the flow profiles
after bifurcation, the maximum velocity is shifted towards the
interior of the bifurcation, whereas in the flow profile before
bifurcation, the maximum velocity stays approximately in the
center of the vessel. This agrees with model-based predictions
[22], and constitutes a significant clinical finding. Figures 9(a)
and 9(b) visualizes the reconstructed velocity components in
a color coded form.
Next, we cross validated the reconstructed flow field using
the data set B. Specifically, we evaluated the cost (51) for
the reconstructed function. The cost was close to the attained
minimum during the iterative search; i.e., in the order of 104 .
This indicates that the chosen values for the regularization
parameters are matched to the data, and that the reconstructed
velocity must be close to the underlying true flow field.
G. Myocardial Motion Recovery
The data for this experiment are from a healthy individual
and were acquired using a sector scan probe from two echocar-

10

IEEE TRANSACTIONS ON MEDICAL IMAGING

(a) Reconstructed flow field

(b) Velocity profile across different parts of the vessels

Fig. 8.

Reconstructed results for the carotid bifurcation

reconstruct the motion field in one image plane, one therefore

has to register the two data sets and superimpose them onto
a common reference system. To this end, we computed the
required transformation by identifying a pair of landmarks
(base of the mitral valves) in both the sequences.
The noise variance estimated using periodicity in this case is
22% of the signal energy. In this experiment too, we used 75%
of the samples for reconstruction and the rest for validation.
Parameter tuning was performed as in the previous experiment.
The attained minimum of the cost (51) was in the order of
103 in this case and the cost for the validation data set was
also in the same order.
Note that the minimum attained in this experiment is less
than that of the previous experiment. This is probably due
to the fact that there are two additional sources of error:
(i) misalignment of cross-sectional planes, and (ii) error in
registering two images. The effect of the later can be reduced
if one uses more landmarks.
By looking at the B-mode images, we manually generated
regions of interest (binary masks) specifying the cardiac wall
that is visible in the given images. Figure 10 shows the
reconstructed motion fields displayed over the regions of
interest for the frame corresponding to the leading edge of
the T-wave of the associated ECG. Note that, even though the
reconstructed motion fields in the apical and parasternal planes
are related by a transformation, the displayed motions fields
do not look so, since the regions of interest are not related by
a transformation.
The reconstructed velocity field was also validated in a
qualitative manner by comparison of the motion directions determined from grayscale images by several echocardiographic
experts. In this way, we could confirm that the results found
by our algorithm are generally in good agreement with the
expert readings.
H. Discussion

0.25 m/s

0.25 m/s

0.25 m/s

0.25 m/s

(a) Horizontal component (left to right (b) Vertical component (downwards is

is positive)
positive)
Fig. 9.

Color coded reconstructed velocity components

diographic standard views: (i) apical long axis view, and (ii)
parasternal long axis view.
Imaging was performed using Philips SONOS 5500 system
with S4 ultraband cardiac transducer. An ultrasound frequency
of 1.8 MHz was used. The penetration depth is 150 mm. Color
calibration was done for the axial velocity ranging from - 12
m/s to 12 m/s.
Along with the Doppler images, the data set contains Bmode intensity images that reveal the anatomical structure of
the cross sectional plane. The acquisition was done by an
experienced echocardiographer who adjusted the probe such
that both acquisitions are from the same cross-sectional plane.
Hence, the premise of this data set is that the image planes
are related by a rigid transformation (translation+rotation). To

The above experimental examples demonstrate the effectiveness of the proposed method. We observe a striking difference
between the performance of the vector regularizer and the
more traditional thin-plate spline regularizer. This is due to
the fact that the vector regularizer introduces a coupling
between the x and y components whereas the thin-plate spline
regularizerwhich works well when the sample are complete
(both components) treats the components independently.
The incompleteness of the Doppler data makes the coupling
crucial, and hence the improvement of coupled regularizer over
the uncoupled one is significant. Another point that makes the
proposed vector regularizer attractive is that it allows one to
incorporate some a priori knowledge on the motion. This term
can be specified to penalize pure (or elastic) deformation only;
in other words, the null space of the regularization operator
can be chosen to include all forms of rigid motion such as
translation and rotation.
Our reconstruction method is formulated in the continuous
domain. Our choice of spline reconstruction space is computationally attractive; it allows an optimal trade-off between the
complexity and accuracy, since B-splines have the maximum

IEEE TRANSACTIONS ON MEDICAL IMAGING

(a) Apical long axis view.

11

algorithm developed in [24].

It should be noted that, as far as we know, our method is
the first attempt to recover vector velocity field from Doppler
data only. Our method opens up new possibilities in clinical
diagnosis. In particular, we can now recover full velocity field
from blood flow, which has not been possible before. Further,
for tissue motion, our method yields more accurate result
compared to intensity-based approaches due to the reasons
explained in the introduction. We believe that integrating
our algorithm in imaging machines that have capability of
obtaining multiple view acquisitions will be a promising
direction, since this will eliminate the inaccuracies due to
manual alignment.
Another strength of the proposed method is that it can
integrate multiple acquisitions efficiently. This is due fact that
(i) the method is formulated for arbitrarily spaced samples,
and (ii) number of sample points does not significantly affect
the computational complexity.
Finally, it should be noted that the proposed formulation is
perfectly extendable to 3D. The only practical difficulty there
is to acquire synchronized multiple view data. Fortunately, the
prospects are getting better with the recent generation of 3D
probes.
V. C ONCLUSION

(b) Parasternal long axis view.

Fig. 10.

Reconstructed motion field from two view cardiac Doppler data.

approximation order for a given support size [20] (O(an+1 )

convergence in the L2 sense). Further, the exact solution is
expressed simply in the form of a sparse linear system of
equation. The proposed computational framework is general
and flexible. The user can easily fine-tune the regularization
for his particular application. Practically, changing the regularization functional only amounts to adjusting some weights
or eventually adding a convolution matrix. In fact, one can
use a set of templates of such matrices to get any quadratic
functional as a weighted sum.
All the reconstructions were done with the grid size in the
range 32 50. The corresponding size of the linear system
is in the range 2048 5000. Thanks to the right choice of
the basis functions, the matrices are sparse and hence it takes
only 3 9 seconds to solve by direct methods using matlab
on a Apple G5 system. We also implemented a multi-grid
algorithm [23] for the case d0 = c0 , and d1 = c1 .
We verified that the multigrid solver will speed up by at
least one order of magnitude. The same strategy should also
work for the more general case but will require a significant
programming effort as well as a re-engineering of the structure
of the current program, which is an adaptation of the scalar

We demonstrated the feasibility of recovering true velocity

field from pulsed wave Doppler data. We adopted a minimization approach, where the cost to be minimized is a weighted
sum of data error and a smoothness functional. We proposed
a derivative-based smoothness functional that allows one to
incorporate an a priori knowledge about the motion and set
up to penalize the deformation. We proposed a continuous
solution for this minimization problem, and showed how the
solution could be obtained by solving a system of linear equations. We demonstrated the effectiveness of the method using
experimental examples. We also demonstrated the importance
of tuning the smoothness functional, and showed how the data
indeterminacy can be overcome by a suitable choice of vector
regularization.
A PPENDIX
Our goal is to find expressions for R(v) and RE (v) in
terms of the B-spline expansion coefficients. Each constitutive
term is a bilinear functional involving the velocity components.
They are of the following form:
Z


B(f, g, p1 , q1 , p2 , q2 ) =
xp1 yq1 f (x, y)
(53)


p2 q2
x y g(x, y) dxdy.
where f and g are expressed as
X f
f (x, y) =
ck,l n (x/a k) n (y/a l)

(54)

k,l

g(x, y) =

cgk,l n (x/a m) n (y/a n)

(55)

m,n

Here, f and g are the symbolic replacements for the velocity

components, u and v; they either refer to the same component

12

IEEE TRANSACTIONS ON MEDICAL IMAGING

(square terms) or different components (cross terms). This

representation facilitates deriving a generic computational formula that will be useful for both square terms and cross terms.
To derive the formula, we first define
dp
{ n (x/a)} .
dxp
Substituting (54) and (55) in (53), we get
XX f
B(f, g, p1 , q1 , p2 , q2 ) =
ck,l cgm,n
p,a (x) =

(56)

Let an (x) = n (x/a). Then,

Consequently, the Fourier expression for p,a (x) defined in

the equation (56) is given by

p
(j) an ()
p
(j) an (a)

p1 ,a (x ka)p2 ,a (x ma)dx

(j) a

q1 ,a (y la)q2 ,a (y na)dy.

p,a ()

l,n k,m

Z
Z
This in turn yields
B(f, g, p1 , q1 , p2 , q2 ) =

XX

cfk,l cgm,n
(57)

l,n k,m

p1 ,p2 (x) = p1 ,a (x) p2 ,a (x).

(58)

(59)
(60)

Substituting we get,
=

XX

cgk,l cfm,n

p1 ,p2 ()

m,n

p1 ,q1 ,p2 ,q2 (m k, n l),

D
E
=
p1 ,q1 ,p2 ,q2 (k, l) cfk,l , cgk,l
In other words, the bilinear form B(f, g, p1 , q1 , p2 , q2 ) that
involves an integral over R2 is re-expressed as a bilinear form
that involves a discrete inner product over l2 . Hence (61) gives
a basic building block to represent any vector regularization
functional for B-splines by discrete convolutions. The discrete
filter p1 ,q1 ,p2 ,q2 (k, l) is completely determined by the quadruple of derivative orders {p1 , q1 , p2 , q2 }, the degree of the Bspline, and the scale a (equations (56), (58), (59), (60)). Let
p1 ,p2 (z) and Mp1 ,q1 ,p2 ,q2 (z1 , z2 ) be z-transforms of p1 ,p2 (k)
and p1 ,q1 ,p2 ,q2 (k, l) respectively. Then
(62)

The next step now is to find explicit expression for

p1 ,p2 (z). To this end, we first write the Fourier expressions
for B-spline [16]:
(ej/2 ej/2 )n+1
.
F [ n (x)] = n () =
(j)n+1

1
(eja/2 eja/2 )p1
ap1 +p2
(eja/2 eja/2 )p2
 ja/2

eja/2 )n+1p1
2 (e
a
(ja)n+1p1

 ja/2
(e
eja/2 )n+1p2
(ja)n+1p2

1
ap1 +p2 1

(eja/2 eja/2 )p1

(eja/2 eja/2 )p2


 ja/2
(e
eja/2 )2n+2p1 p2
a
(ja)2n+2p1 p2
1
=
(eja/2 eja/2 )p1
ap1 +p2 1
(eja/2 eja/2 )p2 a2n+1p1 p2 (a)
1
=
(eja/2 eja/2 )p1
p
+p
1
2 1
a
(eja/2 eja/2 )p2 a2n+1p1 p2 ()
1
=
(p1 ,p2 ) (eja )
ap1 +p2 1
a2n+1p1 p2 ()
(64)

(61)

Mp1 ,q1 ,p2 ,q2 (z1 , z2 ) = p1 ,p2 (z1 )q1 ,q2 (z1 )

p ,a ()
=
p1 ,a ()
2

Now let us consider the case when p1 + p2 is even. We rewrite

the above equation as

l,n k,m

p1 ,q1 ,p2 ,q2 (m k, n l),

X g X
=
ck,l
cfm,n
k,l

(63)

Substituting (63) yields

Now define

B(f, g, p1 , q1 , p2 , q2 )

(eja/2 eja/2 )n+1

(ja)n+1

1 ja/2
(e
eja/2 )p
p
a


(eja/2 eja/2 )n+1p
a
(ja)n+1p

p1 ,p2 ()

p1 ,p2 ()

p1 ,p2 (k) = p1 ,p2 (ka),

p1 ,q1 ,p2 ,q2 (k, l) = p1 ,p2 (k)q1 ,q2 (l).

Now, p1 ,p2 (x) defined in the equation (58) is given by

p1 ,p2 ((m k)a)q1 ,q2 ((n l)a),

where p1 ,p2 (x) is defined by the following convolution:

eja/2 )n+1
(ja)n+1

ja/2

(e
F [an (x)] = an () = an (a) = a

where
(p1 ,p2 ) (eja ) = (eja/2 eja/2 )p1 (eja/2 eja/2 )p2 .
Since p1 + p2 is even (p1 ,p2 ) (eja ) is a polynomial in eja
(or even polynomial in eja/2 ). Let
X (p ,p )
(p1 ,p2 ) (eja ) =
l 1 2 ejla .
(65)
l

IEEE TRANSACTIONS ON MEDICAL IMAGING

13

Substituting in (64) and applying an inverse Fourier transformation, we get

X (p ,p )
1
l 1 2 a2n+1p1 p2 (x la)
p1 ,p2 (x) =
p
+p
1
a 1 2
l
X (p ,p )
1
=
l 1 2 2n+1p1 p2 (x/a l)
ap1 +p2 1
l

This yields
p1 ,p2 (k)

= p1 ,p2 (ka)
X (p ,p )
1
l 1 2 2n+1p1 p2 (k l)
=
p
+p
1
a 1 2
l

(p ,p )
1 2
ap1 +p2 1 k

2n+1p1 p2

D (z)
(k)

p1 ,p2 (z 1 ) = (1)p1 +p2 p1 ,p2 (z)

Now, let p1 + p2 be odd. We rewrite (64) as follows:

1
(p1 1,p2 ) (eja )
ap1 +p2 1
(eja/2 eja/2 )a2n+1p1 p2 ()

Substituting the above equation in (66) and applying inverse

Fourier transformation, we get
X (p 1,p )
1
2
p1 ,p2 (x) =
l 1
p
+p
1
1
2
a
l

a2n+1p1 p2 (x la + a/2)


a2n+1p1 p2 (x la a/2)
X (p 1,p )
1
2
l 1
p
+p
1
1
2
a
l

2n+1p1 p2 (x/a l + 1/2)

2n+1p1 p2 (x l 1/2)

p2 ,p1 (z) = p1 ,p2 (z 1 )

(69)

This these two relations together yield

p2 ,p1 (z) = (1)p1 +p2 p1 ,p2 (z)

(70)

The next step is to use (67) to get explicit expressions for

each of the terms in R(v) and RE (v). We first consider D1 (u)
. Using (53), (15) can be written as
D1 (u) = B(u, u, 1, 0, 1, 0) + B(u, u, 0, 1, 0, 1).
Substituting (61) in the above equation yields


D1 (u) = 1,0,1,0 (k, l) cuk,l , cuk,l


+ 0,1,0,1 (k, l) cuk,l , cuk,l


= (1,0,1,0 (k, l) + 0,1,0,1 (k, l)) cuk,l , cuk,l
(71)
The constituent filters in z-domain read
M1,0,1,0 (z1 , z2 ) = 1,1 (z1 )0,0 (z2 )

This yields
p1 ,p2 (k)

(68)

On the other hand, (58) and (59) imply

(66)

Note that (p1 1,p2 ) (eja ) is now a polynomial in eja , and

let
X (p 1,p )
2
(p1 1,p2 ) (eja ) =
l 1
ejla .

( n (k + 1/2) n (k 1/2))z k .

The filter, p1 ,p2 (z), have some symmetry properties. In

(67), all the terms in the right hand side are symmetric except
02n+1p1 p2 (z), which is anti-symmetric. Hence we get

= (p1 ,p2 ) (z)B 2n+1p1 p2 (z)

p1 ,p2 () =

k
X
k

Hence in z-domain we get,

p1 ,p2 (z)

In summary, the filter p1 ,p2 (z) is given by

1
p1 +p2 1 (p ,p )
1 2 (z)B 2n+1p1 p2 (z),

a
if p1 + p2 is even, and


p1 ,p2 (z) =
1 p1 +p2 1 (p1 1,p2 )

(z)D2n+1p1 p2 (z),

a
otherwise,
(67)
where
p1 
p2

z 1/2 z 1/2
,
(p1 ,p2 ) (z) =
z 1/2 z 1/2
X
B n (z) =
n (k)z k ,

= p1 ,p2 (ka)
X (p 1,p )
1
2
=
l 1
ap1 +p2 1
(k l + 1/2)


2n+1p1 p2 (k l 1/2)
1
(p 1,p2 )
1
d2n+1p1 p2 (k),
ap1 +p2 1 k

l
2n+1p1 p2

where dn (x) = n (x + 1/2) n (x 1/2). In z-domain it

reads
p1 ,p2 (z)

= (p1 1,p2 ) (z)D2n+1p1 p2 (z),

where
Dn (z) =

X
k

dn (k)z k .

M0,1,0,1 (z1 , z2 ) = 0,0 (z1 )1,1 (z2 )

Using (67), the filters 1,1 (z) and 0,0 (z) are expressed as
0,0 (z) = B 2n+1 (z)
1,1 (z) = (z + 2 + z 1 )B 2n1 (z)
Substituting these expressions in (71) yields (17). In a similar
way, we can establish (18).
R Next, we2 intend prove (28). We first expand
(Div v(x)) dx using (25) and then substitute (53)
for each resulting term. We get
Z
2
(Div v(x)) dx = B(u, u, 1, 0, 1, 0) + B(v, v, 0, 1, 0, 1)
+ 2B(u, v, 1, 0, 0, 1)

14

IEEE TRANSACTIONS ON MEDICAL IMAGING

Using (61), this yields

Z


2
(Div v(x)) dx = r11 (k, l) cuk,l , cuk,l


+ r12 (k, l) cvk,l , cvk,l


+ r13 (k, l) cuk,l , cvk,l

Applying the formula (67) in each of the terms in equations

(72), (73), and (74), we finally get
1 
(z1 + 2 + z11 )2 B 2n3 (z1 )B 2n+1 (z2 )
R21 (z1 , z2 ) =
a2
+(z1 + 2 + z11 )B 2n1 (z1 )

(z2 + 2 + z21 )B 2n1 (z2 )
1  2n+1
B
(z1 )(z2 + 2 + z21 )2 B 2n3 (z2 )
R22 (z1 , z2 ) =
a2
+(z1 + 2 + z11 )B 2n1 (z1 )

(z2 + 2 + z21 )B 2n1 (z2 )
1 
(z1 + 2 + z11 )D2n2 (z1 )D2n+1 (z21 )
R23 (z1 , z2 ) =
a2

+D2n+1 (z1 )(z2 + 2 + z21 )D2n2 (z21 )

where
r11 (k, l) = 1,0,1,0 (k, l)
r12 (k, l) = 0,1,0,1 (k, l)
r13 (k, l) = 1,0,0,1 (k, l)
Using (62) and (67) results

This proves (30). (31) is established in a similar way.

R11 (z1 , z2 )

=
=
R12 (z1 , z2 ) =
=
R13 (z1 , z2 ) =
=
=

11 (z1 )00 (z2 )

(z1 + 2 + z11 )B 2n1 (z1 )B 2n+1 (z2 )
00 (z1 )11 (z2 )
B 2n+1 (z1 )(z2 + 2 + z21 )B 2n1 (z2 )
10 (z1 )01 (z2 )
10 (z1 )10 (z21 ) (using (69))
B 02n+1 (z1 )B 02n+1 (z21 )

which proves (28). We obtain (29) in a similar way.

Next, to establish (30), we first expand it as follow:
Z
2
k (Div v(x))k dx = B(u, u, 2, 0, 2, 0)
+B(v, v, 0, 2, 0, 2)
+B(u, u, 1, 1, 1, 1)
+B(v, v, 1, 1, 1, 1)
+B(u, v, 2, 0, 1, 1)
+B(u, v, 1, 1, 0, 2)
Using (61) yields
Z


2
k (Div v(x))k dx = r21 (k, l) cuk,l , cuk,l


+ r22 (k, l) cvk,l , cvk,l


+ r23 (k, l) cuk,l , cvk,l
where
R21 (z1 , z2 )

=
=
R22 (z1 , z2 ) =
=
R23 (z1 , z2 ) =
=

M2,0,2,0 (z1 , z2 ) + M1,1,1,1 (z1 , z2 )

22 (z1 )00 (z2 ) + 11 (z1 )11 (z2 ) (72)
M0,2,0,2 (z1 , z2 ) + M1,1,1,1 (z1 , z2 )
00 (z1 )22 (z2 ) + 11 (z1 )11 (z2 ) (73)
M2,0,1,1 (z1 , z2 ) + M1,1,0,2 (z1 , z2 )
21 (z1 )01 (z2 ) + 10 (z1 )12 (z2 )

Now, using (70),

(12 (z1 ))(10 (z2 ))
R23 (z1 , z2 ) becomes

we get 21 (z1 )01 (z2 )

=
=
12 (z1 )10 (z2 ). Hence,

R23 (z1 , z2 ) = 12 (z1 )10 (z2 ) + 10 (z1 )12 (z2 )

(74)

R EFERENCES
[1] D. Evans and W. McDicken, Doppler Ultrasound: Physics, Instrumentation and Signal Processing. Wiley, 2000.
[2] G. Sutherland, M. Steward, K. Grounstroem, C. Moran, A. Fleming,
F. Guell-Peris, R. Riemersma, L. Fenn, K. Fox, and W. Mc Dicken,
Color doppler myocardial imaging: a new technique for assessment
of myocardial function, J. Am. Soc. Echocardiogr., vol. 7, no. 5, pp.
441458, Sept.Oct. 1994.
[3] P. Hunziker, M. Picard, N. Jander, M. Scherrer-Crosbie, M. Pfisterer, and
P. Buser, Regional wall motion assessment in stress echocardiography
by tissue Doppler bulls-eyes. J. Am. Soc. Echocardiography, vol. 12,
no. 3, pp. 196202, Mar. 1999.
[4] H. Tsujino et al., Combination of pulsed-wave doppler and real-time
three-dimensional color doppler echocardiography for quantifying the
stroke volume in the left ventricular outflow tract, Ultrasound in
Medicine and biology, vol. 30, no. 11, pp. 14411446, 2004.
[5] C. Holland et al., Volumetric flow estimation in vivo and in vitro
using pulsed-doppler ultrasound, Ultrasound in Medicine and biology,
vol. 22, no. 5, pp. 591603, 1996.
[6] C. DeGroot et al., Evaluation of 3-D color doppler ultrasound for
the measurement of proximal isovelocity surface area, Ultrasound in
Medicine and biology, vol. 26, no. 6, pp. 989999, 2000.
[7] W. McDicken, G. Sutherland, C. Moran, and L. Gordon, Colour doppler
velocity imaging of the myocardium, Ultrasound Med. Biol., vol. 18,
no. 6-7, pp. 651654, 1992.
[8] K. Miyatake, M. Yamagishi, N. Tanaka, M. Uematsu, N. Yamazaki,
Y. Mine, A. Sano, and M. Hirama, New method of evaluating left
ventricular wall motion by color-coded tissue Doppler imaging: in vitro
and in vivo studies, J. Am. Coll. Cardiol., vol. 25, no. 3, pp. 717724,
Mar. 1995.
[9] J. Gorcsan, V. Gulati, W. Mandarino, and W. Katz, Color-coded
measures of myocardial velocity throughout the cardiac cycle by tissue
Doppler imaging to quantify regional left ventricular function, Am.
Heart J., vol. 131, no. 6, pp. 12031213, June 1996.
[10] L. Galiuto, G. Ignone, and A. DeMaria, Contraction and relaxation
velocities of the normal left ventricle using pulsed-wave tissue Doppler
echocardiography, Am. J. Cardiology, vol. 81, no. 5, pp. 609614, Mar.
1998.
[11] P. Baraldi, A. Sarti, C. Lamberti, A. Prandini, and F. Sgallari, Evaluation of differential optical flow techniques on synthesized echo images,
IEEE Trans. Biomed. Eng., vol. 43, no. 3, pp. 259272, Mar. 1996.
[12] M. Suhling, M. Arigovindan, C. Jansen, P. Hunziker, and M. Unser,
Myocardial motion analysis from B-mode echocardiograms, IEEE
Trans. Image Processing, 2004, in press.
[13] M. Suhling, M. Arigovindan, C. Jansen, P. Hunziker, and M. Unser,
Bimodal myocardial motion analysis from B-mode and tissue Doppler
ultrasound, in Proceedings of the Second 2004 IEEE International
Symposium on Biomedical Imaging: From Nano to Macro (ISBI04),
Arlington VA, USA, April 15-18, 2004, pp. 308311.
[14] J. Meunier and M. Betrand, Echocardiographic image mean gray
levels changes with tissue dynamics, IEEE Transactions on Biomedical
Engineering, vol. 42, no. 4, pp. 403410, April 1995.
[15] , Ultrasonic texture motion analysis: Theory and simulation, IEEE
Transactions on Medical Imaging, vol. 14, pp. 293300, Jun 1995.

IEEE TRANSACTIONS ON MEDICAL IMAGING

[16] M. Unser, Splines: A perfect fit for signal and image processing, IEEE
Signal Processing Magazine, vol. 16, No. 2, pp. 2238, Nov. 1999.
[17] J. Duchon, Splines minimizing rotation-invariant semi-norms in sobolev
spaces, Constructive Theory of Functions of Several Variables, pp. 85
100, 1977.
[18] F. L. Bookstein, Morphometric Tools for Landmark Data. Cambridge
University Press, 1991.
[19] D. Suter and F. Chen, Left ventricular motion reconstruction based on
elastic vector splines, IEEE Transactions on Medical Imaging, vol. 19,
pp. 295305, 2000.
[20] T. Blu, P. Thevenaz, and M. Unser, MOMS: Maximal-order interpolation of minimal support, IEEE Transactions on Image Processing,
vol. 10, pp. 10691080, 2001.
[21] D. Ku, D. Giddens, C. Zarins, and S. Glagov, Pulsatile flow and
artherosclerosis in the human carotid bifurcationpositive correlation
between plaque location and oscillating shear stress, Artherosclerosis,
pp. 293302, 1985.
[22] D. Ku, Blood flow in arteries, Annual reviews of fluid mechanics, pp.
399434, 1991.
[23] W. Hakbusch, Iterative Solutions of Large Sparse Systems of Equations.
Springer-Verlag, 1994.
[24] M. Arigovindan, M. Suhling, P. Hunziker, and M. Unser, Variational
image reconstruction from arbitrarily spaced samples: A fast multiresolution spline solution, IEEE Transactions on Image Processing, vol. 14,
no. 4, pp. 450460, April 2005.

15