nbul 2004
length.
nl
thickness -
ches.
ranch
ength
S
5-9
107-19
20—358
407
70 pixel
ty of
vessel, a
Is on the
earching
ors and
ind it is
branch-
'e of the
'essel to
; again,
tion.
t of the
tery. In
osterior
ertebral
by the
ing our
ed and
'cted as
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Figure 8. Effectiveness of the connection search: Left: not
applied, Right: applied.
2.6 Evaluation of the Results
It is a difficult problem how to evaluate the results of
segmentation. We can evaluate only subjectively the results by
observing 3-D images of the extracted region. Moreover, as the
shapes and location of blood vessels are not stable like brain or
other organs, it is quite difficult to judge whether the results of
detection are success or failure by observing 3-D images.
Consequently, the objective evaluation method, in which the
result is measured and shown by numeric value, is necessary for
blood vessel segmentation.
Properly segmented 3-D reference data for the ordinal organs
are obtainable by the manual segmentation on each slice data.
In this case the numeric evaluation is possible by measuring the
difference between the result of the segmentation and 3-D
reference data. But in the case of blood vessels, manual
segmentation is impossible because it is quite difficult to
recognize blood vessels on MRA slice images correctly.
Here we propose a new evaluation method which compares two
projection images. One is a group of 2-D vessel region images,
which are manually extracted from MIP and used as reference
data. Hereinafter we call it “segmented MIP” for short. It is
easy to distinguish vessel region on MIP and 2-D manual
segmentation can be performed without fail. The result of 3-D
segmentation is also projected to the same directional plane as
the MIP data. We call it projected result and it is used for the
projection image for evaluation with segmented MIP.
Faise
Binalized negative
MIP image
(true region) ae )
False
Projection positive
ofextracted
region (er
region
Figure 9. Evaluation method of segmentation.
The evaluation value is obtained by counting the number of the
pixels on the image given by subtraction operation of two
projection images mentioned above. More precisely, as shown
in Figure 9, the shortage regions are acquired by subtracting
projected result from segmented MIP. The excess regions are
acquired by subtracting segmented MIP from projected result.
Thus we can get false negative and false positive error index
value, respectively.
In this evaluation process it may often happen that the
thicknesses of vessels are different on two images and it causes
erroneous edges of the vessels appear on the differential images.
So, the vessels on the image going subtraction process are
thinned previous to the subtraction so that the erroneous edges
do not appear.
An example of differential images is shown in Figure 10. The
black lines in the left image indicate the shortage region, and
those in the right indicate the excess region. Each image is
obtained along x, y and z axis, and the sum of pixels on these
three images is used as index values of the extraction error.
ao
f AH 2 CA AT xf D
e i ] ;
UN M CAE
MEE [Fv] Hae | : ix
‘ vi S TU f STE TR X
V S ness S à
= Xe E t ' Jui À
X y : i 5 4 ; :
ips
CE % 1 d a. x. Y M
C. S |
^ / uh id,
4 * Hf T { 2 ^x
1 Pu } ue
N Ia e À P
NA ] Y T
Figure 10. Differential image for evaluation.
left: pixels of shortage right: pixels of excess.
3. DISCUSSION -
3.1 MRA data
We evaluate our segmentation method using 5 head MRA data
obtained from 3 volunteers. Data-I is acquired from 1.5T MRI
and the others are acquired from 3T MRI. Both MRI machines
are manufactured by Siemens. Specifications of each data set
are shown in Table 2. The data depth is 16 bits / voxel and 9-10
bits are effectively used. To reduce variation of intensity range
between the data sets, the intensity values are normalized to
8bits / voxel.
Data | Person Resolution Slices
| A 512x512 168
II B 384 x 512 72
III C 384 x 512 72
IV B 288 x 384 128
V C 288 x 384 128
Table 2. Specifications of MRA data sets
3.2 Experiment
We performed vessel region segmentation for each data set
using the following methods.
