Biomedical Engineering Reference
In-Depth Information
target images was evaluated. For DC1, landmark points selected in the starting
image were found by two independent raters in the target images and the inter-rater
variability was evaluated. The mean inter-rater variability in this case was 1.1176
mm, with a maximum of 3.2783 mm, which are similar to the respective values
of the residual deformation errors. Further, a
t
-test was performed between the
model residual errors (distances between the model-predicted landmark locations
and the corresponding locations determined by the rater, averaged over the two
raters) and the inter-rater distances. The two distributions were found to be sta-
tistically indistinguishable (
p
-value = 0.98 at a 0.05 significance level). It is also
interesting to note that the mean residual errors reported in Table 2 were found
to be highly correlated with the resolutions of the used MRI scans (
ρ
=0
.
9993
for the cross-sectional resolutions, and
ρ
=0
.
9625 for the inter-slice spacings).
These observations indicate that residual errors likely arise from the inaccuracies
in the rater's tracking of the landmark points as well as other errors that are related
to the voxel size, such as rigid registration and manual segmentation errors.
Using the optimal values of all parameters, simulated target images of the used
dataset are compared to real ones in Figure 18. Image intensities in the tumor and
a small peri-tumor region are different between the simulated and target images -
perhaps an indication of tumor infiltration or edema spread between the starting
and target images, which was not accounted for in the definition of regions
T
r
and
D
r
. Despite such signal differences inside and in the vicinity of the tumor,
the results indicate that for the optimal values of the model parameters the model
can reproduce the true deformation caused by tumor growth in the brain tissues
with sufficient accuracy and therefore can serve as a good forward model for this
deformation.
4.3.7.6. Simulations From Normal Brain Images
Model simulations with dif-
ferent values of the four model parameters
c
t
,
r
t
,
r
e
, and
P
will be used in Section
4.4.1 to generate a large number of example brain anatomies deformed by tu-
mors of different locations, sizes, and varying degrees of spread of the associated
peri-tumor edema. As argued above, in these simulations, the value of the edema
expansion strain will be fixed at
e
=0
.
35. Here, example results from two such
model simulations for the same human subject but for different model parameters
are presented. These simulations demonstrate the role of remeshing in simulating
large tumors.
A volumetric T1-weighted MR scan of a healthy elderly subject is used in
the simulations presented here. The image was segmented into brain tissue and
ventricles CSF. Figure 19 presents the results of application of the presented model
with parameter values
r
t
=5mm and
P
= 9000 Pa are presented. No peri-
tumor edema was assumed in this case, and the tumor center was chosen in an
arbitrary location in the white matter. Without the use of remeshing, the simulation
terminates before reaching the final
P
value because of severe distortion of some
tetrahedral mesh elements in vicinity of the tumor. With one-time use of adaptive
