1995 IEEE-EMBC and CMBEC Theme 2: Imaging
3D Motion Analysis
of MR
Imaging
Using Optical Flow Method
Siaw-Hwa Huang,
Shu-Tzu
Wang, Jyh-Horng Chen
.
Electrical Engineering, National Taiwan University, Taipei, Taiwan, R0.C.
PurposeOptical flow is the velocity disbn’bution of each pixel in an image. In this paper, we extend 2D optical flow method to 3D optical flow analysis and also improve its performance around
boundary. Dynamic Analysis of heart and knee is under rotation and contraction.
investigation. Result Display To
Introduction distribution, we overlay
between the variation of the gray level gradient and the velocity of each pixel in the whole image[ 11.
E%$+Eyv+Et=o
,
while the gray level distriiution of the static background is: The motion we simulate includes,The optical flow constraint equation
each pixel.
Cine
Study Short axisview
of left ventricle is imaged using gated Cine sequence with GE 1.5 Signa MR imager. TheROI
is 64x64. In the spatial domain, we use diffusion filter[3] to get In 1993, Amartur and Vessderivative method t
median filter is and therefore ne s time in computation.
But
it usesa
larger toTrajectory Display
3x3 neighborhood and fit its gray scale with a polynomial and obtain the first, second order derivative
h m
this
polynomial. This method improves a lot in the calculated optical flow at boundary region and, moreover, get more accurate gradient values in general.constraint equation as follows,
-
2 0 Confraction We compare
the
performance between previous method(Fig.2a) and our method(Fig.2b) to calculate the gradient. This is a 2D contracting heart simulation, where weer boundary regions. To extend 2D optical flow algorithm to 3D, we derive the 3D
;t
E ,:;
E ,:]
E , x[:I
v = -[t]
E , E, E , E ,
Here, subscript means partial derivative,
u,v,w denotes the
velocity in the x,y,z direction respectively.Material & Methods the
We use Sun sparc workstation and MATLAB to implement
m is simulated as the 3 0 Motion The phantom contracts a pixel toward the left ventricle myocardium(Fig. 1). The phantom’s resolution is
32x32 and the gray level distribution of this phantom is set to be,
center in the xy plane and translates a pixel in the z axis at the same time. Fig.3a shows the velocity in the xy plane and Fig.3b
Icx w
16 16 16
E(x,y,z) = 32+ 4(1 +cos(-). oos(~)*cos(-)) shows the velocity in the z
axis
direction
. n e results matcheswell with the true motion field.
Fig. 3 (a)
Fig.3 (b)
Fig4 (c)
Fig.4 (d)
or Mostvelo '
spatial domain or time
axis,
can also get more accurate gradient and obtain better results.Conclusion
In this study, we improve the optical flow methods to have better algorithm in gradient calculations and extend 2D velocity field to 3D motion study. Results shows better performance at boundary and g e n d continuity of velocity field. To evaluate the clinical
data
results, we are comparing it with the MRW&g study.
Joint analysis is also und m
185-203
2.A new approach to Stu
of Cine M R Images S
1993 3 . N o h
Medical Imaging, Vol. 11