Biomedical Engineering Reference
In-Depth Information
Table 1. Comparison of our Method with
the Phantom Ground Truth
WM GM Overall
pn=3% rf=0%
0.90
0.76
0.83
pn=9%, rf=0%
0.88
0.74
0.81
pn=3%, rf=40%
0.90
0.73
0.82
pn=9%, rf=40%
0.88
0.74
0.81
tion. Moreover, our algorithm performs robustly in the presence of severe thermal
noise and RF inhomogeneities.
5.3.2. Validation on real MR brain, data
To further demonstrate the performance of our algorithm under various con-
ditions, we used real MR data as test samples and performed a comparison with
manually outlined surfaces. We utilized the MR data provided by the ISBR of
center from morphometric analysis done at the Massachusetts General Hospital
(20 normal MR brain datasets), along with manual segmentation performed on
positionally normalized scans by trained investigators from the ISBR.
The data we obtained and utilized here are 20 normal T1-weighted brains in
8-bit format. The coronal three dimensional T1-weighted spoiled gradient echo
MRI scans were performed on two different imaging systems. Ten FLASH scans
on four males and six females were performed on 1.5-tesla Siemens magnetic MR
systems with the following parameters: TR = 40 ms, TE = 8 ms, flip angle =
50 degrees, field of view = 30 cm, slice thickness = contiguous 3.0 mm, matrix
= 256, and averages = 1. Ten 3D-CAPRY scans on six males and four females
were performed on a 1.5-tesla General Electric Signa MR System (Milwaukee,
WI), with the following parameters: TR=50ms, TE=9ms, flip angle = 50
degress, field of view = 24 cm, slice thickness = contiguous 3.0 mm, matrix =
256, and averages = 1. The 8-bit version of the 20 normal T1-weighted brains
were reduced from 16-bit data by thresholding intensities above or below 99.9%
of the total number of different intensities, and scaling the remaining range to
0 ··· 255 if and only the range did not already fit within 0 ··· 255. All the datasets
were positionally normalized by imposing a Talairah coordinate on each 3D MR
scan. The repositioned scans were then resliced into normalized 3.0-mm coronal,
1.0-mm axial, and 1.0-mm sagittal scans for subsequent analysis.
The experimental results with our algorithm are depicted in Figure 17 and
quantitatively evaluated in Table 2. Fortunately, we can note that the exactly
same dataset was extensively used in other published papers, and the results were
gathered by Rajapakse and Kruggel [102]. Therefore, it is easy to compare our
 
Search WWH ::




Custom Search