Biomedical Engineering Reference
In-Depth Information
3.1. Synthetic Data
Three synthetic datasets are used with the same square pattern and different
levels of Gaussian and salt pepper noise. We performed experiments using the two-
phase model and different region information to extract the square objects from the
image. The results given in Figure 4 demonstrate the stability and robustness of
our algorithms. As the image regions have different mean intensity and variance,
the results obtained using our method (Figure 4b) and the geodesic active regions
[11, 12, 13] (Figure 4c) have similar results. When the image is corrupted by high
noise and has diffuse boundary, region-based energy plays a more important role
in curve evolution. When the regions have the same mean intensity and different
variance, geodesic active regions seem to be prone to converge to local minima
as the velocity is dependent on voxel intensity and regional statistical parameters.
Opposite to this, when the proposed method applied both local mean and variance,
one can reduce the risk of converging to the local minima. Figure 4e,f shows
the robustness of the proposed method. We give experiments for region intensity
corrupted by salt-and-pepper noise and show the results in Figure 4h,i.
3.2. Real Data
In this subsection we use two 3D real data sets to validate our approach.
The first is MRI data obtained on a 1.5T GE Signa Twin Speed scanner with a
3D Spoiled Gradient-Recalled (SPGR). The imaging parameters are as follows:
TR=11
.
3 ms, TE=4
.
2 ms, and FlipAngle = 15
◦
. At first, we perform our two-
phase model to delineate brain lateral ventricles from the 3D image data. In order to
delineate lateral ventricles, a necessary preprocessing that manually crops the third
ventricle is performed to satisfy bipartitioning condition. The local neighborhood
size is 3
×
3
×
3. WM and GM will be classified to the same region during surface
evolution. Both 2D and 3D results are shown in Figure 5. The initialization surface
is a sphere centered at the centroid of the two lateral ventricles. The gray matter
and CSF nearby the cortex are discarded during preprocessing.
The second test using two-phase model are performed on DT-MRI data ac-
quired on a 1.5T Siemens Sonata with high-speed gradients. The diffusion sen-
sitizing gradients are applied along six non-collinear directions with a
b
value =
900 s/mm
2
. The imaging parameters are: TR = 8000 ms, TE = 106 ms, number
of excitations = 3, and voxel size is 0
.
9
×
0
.
9
×
4
.
0 mm
3
.
The Diffusion Tensor matrix is calculated according to the Stejskal and Tanner
equation [31]. Our algorithm is performed on the fractional anisotropy (FA) image
computed from the diffusion tensor image. In order to extract the corpus callosum
using the two-phase model, we cropped the FA image around the region of interest
so as to assure a bimodal condition. The initial surface is a sphere centered in
the image with a radius equal to 20 voxels. Figure 6 shows the evolution process
and the final result. A three-phase segmentation model is applied to brain tissue
