Phase-driven Finite Element Model for Spatio-temporal

Tracking in Tagged Cardiac MRI

Idith Haber , Ron Kikinis , and Carl-Fredrik Westin


 Children’s Hospital, Harvard Medical School, Boston MA, USA
 idith.haber@tch.harvard.edu
Brigham and Women’s Hospital, Harvard Medical School, Boston MA, USA

Abstract. Relative motion of the left ventricle can be a useful indicator of the
degree of contraction. MRI tissue tagging produces noninvasive markers within
the muscle wall which can be uses measure motion and deformation. However,
the widespread use of tissue tagging has been limited by time-consuming image
post-processing. We describe a method for automatically calculating displace-
ments in 2D tagged images by using quadrature filters to estimating the phase.
Since the tag represents a repeating pattern, they have an intrinsic phase prop-
erty which stays the same as they move. We fit a 3D finite element model to the
phase data in from successive MR images through systole. The addition of a 3D
model allows for 1) estimating 3D motion from multiple 2D tagged images and
2) smoothing of the phase estimates which are susceptible to image artifacts and
noise. This technique is applied to data acquired from normal hearts and com-
pared with results of semi-automatic tracking.

1 Introduction
The application of MRI to the heart has allowed for better assessment of left ventricular
(LV) function [2]. Variation in signal intensities between muscle, fat, blood and air allow
for visualization of the inner and outer walls, or endocardium and epicardium. In order
to visualize motion within the muscle wall, a non-invasive tissue tagging technique, also
known as SPAMM, has been developed in the last decade [3]. With this method, radio-
frequency pulses are used to saturate parallel planes of issue prior to imaging. Images
are acquired perpendicular to these planes, resulting in dark bands in the images. The
dark bands represent actual tissue and, since the saturation lasts through systole, reveal
myocardial motion as the heart contracts. Figs. 1 and 2 show representative short and
long-axis images. In the short-axis images, 2 perpendicular sets of tagging planes can
be used to lay down a tagging grid.
The addition of the tagging grid helps in analyzing LV function. The tags provide
material markers which can be followed through the heart beat. From tag motion, twist-
ing of the heart can be revealed [14]. In addition, since the tags reveal relative motion,
strain can also be derived [24] and used as an indicator of contraction.
In this paper, we describe a method for tracking the tags in successive image frames.
by estimating the displacement at a pixel in two successive frames. Displacement is de-
rived from the local phase, a technique previously used for estimating disparity between
 (a) (b) (c)

Fig. 1. Spatio-temporal information in one short-axis location. 2D tagged MR images at (a) be-
ginning systole (b) mid-systole (c) end-systole. Dark bands represent tags which are tracked using
local phase.

two images in stereo vision and velocity measurements [10, 21]. Since the tags form a
repetitive pattern in the images, the local phase is a material property and can be tracked.
Another reason for using phase is stability: phase is independent of signal magnitude,
which makes it insensitive to illumination variation over an image scene [9]. In MR this
translates to insensitive to bias field inhomogeneities in the data. In contrast, a tracking
system based on gradient magnitude will be sensitive to such variation.
One drawback of using 2D images it that they do not capture through-plane motion.
Three-dimensional motion can be derived by combining information from tagged im-
ages taken from multiple views. We describe a finite element-based technique for 3D
motion reconstruction. This is an extension of a deformable modelling technique which
was previously applied to extracted tag data [13].
We here describe modified geometry and constraints which allows us to incorpo-
rate phase-based displacement estimates into the 3D motion reconstruction. As can be
seen in the images, the tags at end systole vary smoothly from the endocardium to the
epicardium. The 3D model ensures spatio-temporal smoothness via the finite element
stiffness. This approach allows for stabilization the phase-derived displacement esti-
mates, which are susceptible to noise and image artifacts. The technique is unique in
that it combines phase-based displacement estimates with a deformable model and it is
automatic.

2 Related Work
Several researchers have approached the problem of tracking tags in 2D images. The
methods include optical flow [11, 7], template matching [8, 12] and least-squares min-
imization [12], Fourier transformation methods [25] and active contours [24]. Another
group have developed a method which calculates strain directly from phase images [18].
Both model-based [17, 5, 19, 16, 22, 20] and non-model based [15, 6] approaches
have been applied to reconstructing the 3D motion from multiple views. Some tech-
 (a) (b) (c)

Fig. 2. Long-axis images at (a) beginning systole (b) mid-systole (c) end-systole

niques are applied directly to the images [17, 22, 20], while the rest operate on tag data
extracted from the 2D images using one of the methods described above. Model-based
approaches to 3D motion reconstruction include: fitting a 3D displacement field com-
posed of an analytic series [17] or to predefined functions derived from known LV
motion [5, 19] to data; using spline curve and surface reconstructions of tag data as in-
put to a linear optimization procedure [16]; fitting of a cubic B-spline model [20] and
nonlinear least-squares fitting of a finite element model [23].

3 Methods
3.1 Image Pre-processing
In an ideal sinusoidal signal, the phase is the argument of the analytical signal. In a more
complicated signal, the local phase is the argument of the output of convolving it with
complex quadrature filters. Since for a sinusoidal pattern the phase uniquely defines
a spatial position modulo the wavelength, this phase can be used for estimating local
displacements, assuming displacements smaller than the wavelength of the pattern.
The spherically separable quadrature filters are defined as:

    (1)
 the vector valued frequency variable,   . The directional function,
 ! , isisoriented
where
along the direction of the tags. The radial frequency function

is a Gaussian function on a logarithmic scale, called a lognormal function:
 "#%$ ')(+& * , (.- / (02143517698 (2)
<;
ter frequency of the filter.
=) : is the relative bandwidth in octaves and
is the radial frequency function. is the cen-
defines the frequency characteristics of the quadrature
filters.
 >@?
The tag spacing, is known at the time of imaging or can be measured from the
@CDFE<5G
image. For each location and tag direction, two successive images are considered. For
HB 
=! JI A B A B
each pair
.
, the images are first convolved with the quadrature filter:

The phase difference KL

is calculated as:

KL NMPO Q. H I H
R (3)

Also, the magnitude of H
I H
provides a measure for the certainty of the estimate. The
phase disparity is used to estimate displacement, : S
S  %G KT
U L @> ? (4)

3.2 FEM Deformable Model

We incorporate smoothing by fitting a solid finite element model to the displacement

data. The model is initiallly a parllelipipid block, with elements aligned with the tag
planes. During the fitting, elements for which the phase estimate is uncertain are elimi-
nated.

Dynamics The energy formulation for the dynamics of a deformable solid dictates
that for small displacements, the energy from external forces is absorbed by inertial,
viscous, and the elastic energy of the solid itself [4]:
VWVXVZY[ ]\)^_ VXVZ` a\Pbc VdVXV EU+G%e7fhgie\)^
(5)

where the external and internal energy terms are on the and and right hand sides, re-
Y[ `  e g
spectively. The displacement and strains fields are, and , respectively. is the stress-
strain matrix, contains body forces, contains surface tractions, is the density of

j
the material, and is the damping parameter.
For each element, the displacement field and its derivatives are approximated as
Nkc=e 9lm9no @f
p (6)
p
where are the displacements at the nodes of the mesh, and are the interpolation,
k
local element coordinate system
=e @l.9no
or shape functions. These functions have as dependent variables the non-stationary,
. The shape functions can be seen as weighting
e @l.@n k
the value (i.e. displacement) global coordinate of a node according to where a point lies
within the system of the element. Since are only dependent only position the
derivatives are simple to calculate. We use eight-noded parallelepiped elements, whose
linear shape functions are given in [4].
The discretized strain field is a function of discretized displacements:
eorqDst (7)
 q
where is derived by 1) converting strain to a sum of partial derivative of displace-
ments; 2) using the inverse of the Jacobian to relate derivatives in the global coordinate
system to those in the local element coordinate system, and 3) using the derivative of
the shape functions to relate these local derivatives to nodal displacements, .
s
Eqs. 5 and 7 can be combined to obtain the following:
uwxvp _zy vp _|{}p}r~€ (8)
where the vector the vector
~ contains the image-derived forces. The inertial term is
neglected and the viscous damping (with a coefficient of one) is used in order to allow
the model to come to rest when it reaches the position most consistent with the data.
{}p
Stiffness can also be viewed as an internal resistance which results in a force equal to
. Using these considerations, the equations of motion is written for each node, , as
C
vp B r
 ~ =B  B2‚%ƒ=„€…9‚<†7‡ _ˆ~ B= „€‰Šƒ=„€…9‚<†7‡ (9)
p ~
where B is the 3D nodal displacement, B B‹‚Œƒ=„€…5‚+†7‡
Ž {}pm B ~
and B= „€‰Šƒ=„€…5‚+†7‡
is an ex-
ternal, image-derived force. It is important to note that these forces are only used to
deform our model and are not meant to replicate the actual forces or material properties
of cardiac tissue.
For each iteration, the external forces are calculated, the internal element forces are
calculated and distributed to the nodes, and finally the nodal equations are solved using
adaptive Euler integration.

Constraints During the image pre-processing, the 3D field of displacements for 3

different directions has been derived from both short and long-axis images. Prior to
fitting between a pair of image frames, a target displacement, , is assigned to each SB
node by interpolating from the displacement field data. During each iteration, the forces
on the nodes are set to be equal to the target displacement subtracted by the nodal
displacement from previous iterations:
~ B B‹‚Œƒ=„€…5‚+†7‡  SB’‘”“•B (10)
The fitting continues until all forces are minimized, or the nodal displacement is
approximately equal to the target displacement.

4 Results

The local phase was calculated and the difference was used to calculate the the displace-
ment. Each row of Fig. 3 shows the phase pairs and resulting displacement images.
In order to test the phase-driven deformable modeling technique, we applied it to
simulated data. Given a cylindrical geometry and displacement field, we generated syn-

– R —+˜]˜
thetic image data. The geometry and motion was similar to that found in the heart. The

– R ™EŠ˜]˜
difference between reconstructed and prescribed tag locations was rms while
the difference in nodal deformations was , which are less than the pixel reso-
lution.
 time 1 time 2 Displacement

Fig. 3. Pairs of phase images and displacement derived from them. Locations are same as those
shown above (1st 2 rows = short-axis , 3rd row = long-axis).

Image data from normal patients was also analyzed using the method. Fig. 4 shows
the the intersection of the reconstructed model with the original image plane locations,

– R›šŒœ<˜]˜
along with the image. When these results were compared with those found using a semi-
automatic tracking technique [1] an average discrepancy of was calculated
between the two data sets.

5 Discussion

We describe a 3D tracking method in which phase-derived displacements are used as

input into a deformable model. The model allows for the regularization of the phase
estimates and integration of motion information from multiple image planes. This tech-
nique can easily be extended to 3D tagged images, where the model would serve as a
regularizing force.
 a b

Fig. 4. Motion reconstruction results overlayed on original images. Representative a) short-axis

and b) long-axis images .

References
1. L. Axel, D. Bloomgarden, C.N. Chang, D. Kraitchman, and A.A. Young. SPAMMVU: a
program for the analysis of dynamic tagged MRI. In Proceedings of the Soc. of Magnetic
Resonance in Medicine 12th Annual Meeting, page 724, 1993.
2. L. Axel, R.C. Concalves, and D. Bloomgarden. Regional heart wall motion: Two–
dimensional analysis and functional imaging with MR imaging. Radiology, 283:745–50,
1992.
3. L. Axel and L. Dougherty. Heart wall motion: Improved method of spatial modulation of
magnetization for MR imaging. Radiology, 172(2):349–50, Aug 1989.
4. R.D. Cook, D.S. Malkus, and M.E. Plesha. Concepts and Applications of Finite Element
Analysis. John Wiley and Sons, New York, third edition, 1989.
5. J. Declerck, N. Ayache, and E.R. McVeigh. Use of 4D planispheric transformation for the
tracking and the analysis of LV motion with tagged MR images. In Proceedings of the SPIE
Medical Imaging, San Diego, CA, February 1999.
6. T.S. Denney, Jr and J.L. Prince. Reconstruction of 3D left ventricular motion from planar
tagged cardiac MR images: an estimation theoretic approach. IEEE Trans on Med Imaging,
4(4):625–635, December 1995.
7. L. Dougherty, J.C. Asmuth, A.S. Blom, L. Axel, and R. Kumar. Validation of an optical flow
method for tag displacement estimation[letter]. IEEE Transactions on Medical Imaging,
18(4):359–63, April 1999.
8. D. J. Fisher, S.K. Tadikonda, J.C. Ehrhardt, W.T.C. Yuh, and S.M. Collins. Sub-pixel de-
tections of CINE-SPAMM magnetic resonance tags. Book of Abstract: Society of Magnetic
Resonance in Medicine, page 2511, 1992.
9. D.J. Fleet and A.D. Jepson. Stability of phase information. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 15(12):1253–1268, 1993.
10. D.J. Fleet, A.D. Jepson, and M. Jenkin. Phase-based disparity measurement. CVGIP: Image
Understanding, 53(2):198–210, 1991.
11. S.N. Gupta and J.L. Prince. On variable brightness optical flow for tagged mri. In IPMI,
Dordrecht, 1995.
12. M. A. Gutttman, J.L. Prince, and E.R. McVeigh. Tag and contour detection in tagged MRI
images of the left ventricle. IEEE Trans on Med Imaging, 13(1):74–88, March 1994.
13. I. Haber, D.N. Metaxas, and L. Axel. Three-dimensional motion reconstruction and analysis
of the right ventricle using tagged MRI. Med Image Analysis J, 4(4):335–355, 2000.
14. S.S. Henson, S.K. Song, J.S. Pastorek, J.J. Ackerman, and C.H. Lorenz. Left ventricular
torsion is equal in mice and humans. Am J Physiol, 278:H1117–1123, 2000.
15. W. Kerwin and J. Prince. Cardiac material markers from tagged MR images. Medical Image
Analysis, 2(4):339–53, 1998.
16. M.J. Moulton, L.L. Cresswell, S.W. Downing, R.L. Actis, B.A. Szabo, M.W. Vannier, and
M.K. Pasque. Spline surface interpolation for calculating 3-D ventricular strains from MRI
tissue tagging. Am J Physiol, 39:H281–H297, 1996.
17. W.G. O’Dell, C.C. Moore, W.C. Hunter, E.A. Zerhouni, and E.R. McVeigh. 3-dimensional
myocardial deformations: calculation with displacement field fitting to tagged MR images.
Radiology, 195(3):829–35, Jun 1995.
18. N. F. Osman, E.R. McVeigh, and J.L. Prince. Imaging heart motion using harmonic phase
mri. IEEE Trans on Med Imaging, 19(3):186–201, March 2000.
19. J. Park. Model–based shape and motion analysis: Left ventricle of a heart. PhD thesis,
University of Pennsylvania, Philadelphia, 1996.
20. P. Radeva, A. Amini, and J. Huang. Deformable B-solids and implicit snakes for localization
and tracking of MRI-SPAMM data. Computer Vision and Image Understanding, 66(2):163–
178, May 1997.
21. T. Sanger. Stereo disparity computation using gabor filters. Biol Cybern, 59:405–418, 1988.
22. A.A. Young. Model tags: Direct 3D tracking of heart wall motion from tagged MR images.
In Medical Image Computing and Computer-Assisted Intervention - MICCAI, pages 92–101,
Cambridge, MA, 1998.
23. A.A. Young and L. Axel. Three-dimensional motion and deformation of the heart wall:
estimation with spatial modulation of magnetization - a model-based approach. Radiology,
185:241–47, 1992.
24. A.A. Young, D. L. Kraitchman, L. Dougherty, and L. Axel. Tracking and finite element
analysis of stripe deformation in magnetic resonance tagging. IEEE Trans on Med Imaging,
14(3):413–21, September 1995.
25. S. Zhang, M.A. Douglas, L. Yaroslavsky, R.M. Summers, V. Dilsizian, L. Fananpazir, and
S.L. Bacharach. Fourier based algorithm for tracking SPAMM tags in gated magnetic reso-
nance cardiac images. Medical Physics, 23(8):1359–69, August 1996.