Information Technology Reference
In-Depth Information
Fig. 23 Average
segmentation accuracy of the
VB segmentation method on
12 ESP CT images. After
each VB is detected, the size
is reduced to 120
×
120 that
the segmentation
measurements are calculated
Table 1 Classi
cation of clinical data sets used in the experiments: There are totally 18 data sets
in the data base
Class A
Class B
Class C
Slice thickness
Usually
<2.5 mm
Usually
2.5 mm,
but may be <2.5 mm
≥
Usually
3.00 mm,
but can be smaller
if disease exists
≥
Resolution
High
Usually low
Lower
Shape
Straight
Straight/Curvy
Usually curvy but it
can be straight
Bone degeneration or
osteophyte
No
May have disease
May have disease
Fracture
No
No serious fracture
May have serious
fractures
Edge
Strong
Strong/Weak
Usually weak
Note
Optimum
data
This class has some
problems
This class has very
serious problems
Class A, B, and C have 7, 5, and 6 data sets, respectively
Table 2 Segmentation results of each data class based on different measures
'
H
'
'
L
'
'
H+L
'
Class A
Class B
Class C
Accuracy, %
98.2
97.9
97.6
99.0
98.7
98.0
Precision, %
91.1
86.6
88.6
90.9
89.9
84.4
Jaccard coefficient, %
87.6
83.0
85.0
87.7
86.9
80.3
Dice
'
s coef
cient, %
93.1
90.4
91.5
93.8
92.9
89.0
The measurements are based on data sets with 120
120
Z
size where
Z
represents number of
slices
Jaccard coef
cient gives the lowest quality score respect to the others. Also, the
accuracy gives the highest quality score respect to the other measurements. By
using this information, the proposed segmentation reaches the scores of
'
Jaccard
