Title: Validation of Blood Flow Simulations in Intracranial Aneurysms
1Validation of Blood Flow Simulations in
Intracranial Aneurysms
Final-Project Presentation (Registration)?
Brown University
2Objective
- Finished
- Generate 3d patient-specific mesh from Dicom
files. - Simulate concentration field inside with the
mesh. - Now
- Fit the 3d results with 2d dye-injection image by
2d-3d image registration technique.
3Registration
- For every iteration of the registration algorithm
a 3D rigid-body geometric transform is applied to
the CT volume to produce a change in the 3D
position of the arteries. - The 3D volume is then reduced to a 2D digitally
reconstructed radiograph (DRR) by summing the
voxel values of the transformed CT volume in the
z direction.
z
Compare with
y
rotate translate
x
project
2d fluoroscopy frame
3d object
DRR
4Registration
- Assume pixel values of the filtered DRR are
denoted by Ii and pixel values of the fluoroscopy
frame are denoted by Ri, by minimizing the
objective function - where
- and is the histogram bin which includes
Ri.
. I_i
. R_i
- NOTE
- I didn't filter the data, because in our case
not only the shape should match, the density on
each pixel should also match. - Since the data size is huge (536by536by536), I
took the R_i instead of its average.
5Registration
- To optimize the objective function S(m), with
Taylor expansion for the update vector p, - we can get an approximation for p as
- where m(Tx,Ty,Tz,Rx,Ry,Rz) contains the
information for translation (T) and rotation (R). - In the implementation, I use the matlab
optimization function - x
fminunc(fun,x0)? - Instead of optimizing all six parameters at
one time, I optimize S with respect to rotation R
first, then to translation T, and repeat this
process for five times.
6Simple Tests
- Translation only
- Rotation only
-
3D DATA size 656565 when 32ltx,y,zlt40 density1
2D DATA size 3535 when 17ltx,y,zlt25 density1
FITTING RESULT T(15, 15, 0.111) s2.58e-13
Initial T(17, 17, 1)?
2D DATA size 3535 when 17ltx,y,zlt25 density1,
rotate pi/3
3D DATA size 656565 when 32ltx,y,zlt40 density1
FITTING RESULT R(1.047, -5e-5, -5e-6) s2.54e-7
Initial T(17, 17, 1)?
7Simple Tests
FITTING RESULT T(14.5, 15.3, 0.082) R(6.95,
-2.91, 0.45)? s3.56
2D DATA size 3535 created by 3D DATA rotating
with R(pi/3,pi/4,0)?
Initial T(15, 15, 1/9)? R(0, 0, 0)?
3D DATA size 656565 when 32ltx,y,zlt40 density1
FITTING RESULT T(22.5, 17, 0.083) R(0.52,
0.79, 0)? s6.79e-3
Initial T(15, 15, 1/9)? R(pi/3, pi/4, 0)?
8Comparison for arterial data qualitative
T0.22 (sec)?
T0.72 (sec)?
T0.22 (sec)?
T0.72 (sec)?
T1.22 (sec)?
T1.72 (sec)?
T1.22 (sec)?
T1.72 (sec)?
9Quantitative comparison Prepare Data
- 2D data
- Considering the geometric differences near the
aneurysm part, we cut upstream areas in 2d
angiograms for comparison.
- 3D data
- Invert plt concentration field data into
536by536by536 matlab 3d matrix.
For easier comparison, change the 2d and 3d data
to black background, that is, the values for
background pixels are zero.
10Quantitative comparison Coarse to fine
- Coarse
- Condense both the 2d and 3d data into 1/16 of
their original sizes and apply the fitting
algorithm, get optimal parameters T_small and
R_small.
- Fine
- Now apply the algorithm to data with original
size, with initial values for T and R as - T16T_small
- RR_small
- Because of the lack of time, we use data with
¼ of the original size as our fine results.
11Quantitative comparison Results
T0.22 (sec)?
T0.72 (sec)?
T1.22 (sec)?
2D data
Fitted 3D data
Relative error I-R
5.61
5.80
4.05
12Conclusions
- Conclusion
- For rotation or translation only, the fitting
algorithm gives satisfying results for different
initial values. However, to fit with both
rotation and translation effects, a good guess
for initial values is important for reasonable
results. - The concentration field calculated from simulated
velocity field matches well with the angiograms
from dye injection (relative error I-R around
5).
13References
- Juan R. Cebral, Alessandro Radaelli, Alejandro
Frangi, and Christopher M. Putman, Qualitative
Comparison of Intra-aneurysmal Flow Structures
Determined from Conventional and Virtual
Angiograms, Medical Imaging 2007 Physiology,
Function, and Structure from Medical Images. - Matthew D. Ford, Gordan R. Stuhne, Hristo N.
Nikolov, Damiaan F. Habets, Stephen P. Lownie,
David W. Holdsworth, and David A. Steinman,
Virtual Angiography for Visualization and
Validation of Computational Models of Aneurysm
Hemodynamics, IEEE Transactions on Medical
Imaging, Vol. 25, No. 12, 2005. - M. Pickering, A. Muhit, J. Scarvell, and P.
Smith, A new multimodal similarity measure for
fast gradient-based 2D-3D image registration, in
Proc. IEEE Int. Conf. on Engineering in Medicine
and Biology (EMBC), Minneapolis, USA, 2009, pp.
5821-5824.
Thank you!