Information Technology Reference
In-Depth Information
Fig. 12 An axial cross-section of a CT spine image I. a Edge points
f
p
m
g
, extracted by the 3D
Canny edge detector. b Image intensity gradient vectors
, extracted by the 3D Sobel
gradient operator (the number of gradient vectors was considerably reduced for visualization
purposes). c An illustration of the determination of a pair of opposite edge points
f
g
ð
p
m
Þg
p
m
Þ
ð
p
m
;
in the
direction of gradient vector
þ
g
ð
p
m
Þ
. d The resulting accumulator A
I
with superimposed edge
points
extracted edges in the 3D image, while their magnitude is proportional to the
strength of the extracted edges in the 3D image. For a given 3D image, the edge
points
of anatomical structures (Fig.
12
a) can
be extracted by the 3D Canny edge detector (using e.g. high threshold t
high
that
captures 50 % of image pixels and low threshold t
low
¼
f
p
m
¼
ð
x
m
;
y
m
;
z
m
Þ;
m
¼
1
;
2
; ...;
M
g
4t
high
for hysteresis
thresholding), while the corresponding image intensity gradient vectors
0
:
f
g
ð
p
m
Þ
¼
½
(Fig.
12
b) can be extracted by the 3D
Sobel gradient operator (using e.g. kernel size of 3
g
x
ð
x
m
Þ;
g
y
ð
y
m
Þ;
g
z
ð
z
m
Þ;
m
¼
1
;
2
; ...;
M
g
3mm
3
) and normalized
3
so that
1. Before computing the edges and gradient
vectors, the images can be smoothed with a 3D Gaussian
8
g
ð
p
m
Þ)
0
k
g
ð
p
m
Þ
k
filter (with e.g. standard
deviation of
r
¼
5mm).
For each edge point p
m
, a search for the opposite edge point p
m
is therefore
performed in the direction of the normalized gradient vector g
1
:
ð
p
m
Þ
, however,