Information Technology Reference
In-Depth Information
Fig. 20 Worse Initialization: Segmentation comparison under (a) no noise, noise variances
(b)
2
n
2
n
2
n
3
and (v) the proposed method. (The red and yellow colors show the contour of the ground truths and
segmented regions, respectively.)
= 0.1, (c)
= 0.25, and (d)
= 0.5. (i) Initialization. The results of (ii)
A
1, (iii)
A
2, (iv)
A
r
r
r
Slice thickness, resolution, spine column region (shape), fractures, diseases, and
spine bone edges are main factors of the classi
cation. Class A is the best data sets
which can be segmented and analyzed easily. Data sets which are classi
ed in class
C have serious problems such as diseases, fractures, weak spine edges, and low
resolution. Data sets in class B have some problems but they are better than data
sets in class C. Categorization could help to analyze the results separately.
As mentioned above, the proposed algorithm is tested on 932 CT slices/66 VBs
which are obtained from 18 (7 H and 11 L) different patients and different spine
bone regions (i.e. lumbar, thoracic, and etc.). The segmentation accuracy is given
with respect to the health condition of bone (
'
H
'
,
'
L
'
, and
'
H+L
'
), and the
classi
cation criteria (Class A, B, and C). Table
2
shows the quality measurement
results of the proposed segmentation method. The four different measurements are
given to be judged fairly. As can be interpreted from the results in the table, the
