July 10, 2008 / Vol. 6, No. 7 / CHINESE OPTICS LETTERS 495
Hepatic CT image retrieval based on the combination of
Gabor filters and support vector machine
Lijun Jiang (
ööö
{{{
), Yongxing Luo (
ÛÛÛ
[[[
,,,
), Jun Zhao (
ëëë
ddd
), and Tiange Zhuang (
BBB
UUU
{{{
)
Department of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai 200240
Received October 22, 2007
Content-based image retrieval has been an active area of research for more than ten years. Gabor schemes
and support vector machine (SVM) method have been proven effective in image representation and clas-
sification. In this paper, we propose a retrieval scheme based on Gabor filters and SVMs for h epatic
computed tomography (CT) images query. In our experiments, a batch of hepatic CT images containing
several types of CT findings are used for the retrieval test. Precision comparison between our scheme and
existing methods is presented.
OCIS codes: 100.2000, 100.2960, 100.6950, 100.7410.
doi: 10.3788/COL20080607.0495.
Content-based image retrieval (CBIR) has been an active
research topic for many years. The application of CBIR
techniques in medical field has also been pro posed for
the use in picture archiving and communications system
(PACS)
[1]
, case databases
[2]
, and in an even more general
sense
[3]
. In CBIR, texture is the most important infor-
mation which can be used to characterize an image
[4]
.
Up to now, several methods achieving effective feature
extraction have been proposed
[5−7]
. Among the meth-
ods, Gabor filter
[7−11]
is widely used, which is mainly
motivated by two factors
[5]
: primitives of image repre-
sentation in vision have a wavelet form similar to Ga-
bor elementary functions (EFs)
[12]
and Gabor function
represents a minimum in terms of the spread of uncer-
tainty in space and spatial frequency
[13]
. Support vector
machine (SVM) method
[14−18]
is a reliable classification
technique, which is based on statistical learning theory
and has been introduced for solving pattern recognition
problems.
Zhao et al.
[9]
proposed a scheme of hepatic computed
tomography (CT) images retrieval using Gabor features.
However, the retrieval by this scheme was only based on
the distance between the query image and the others, and
only part of the Gabor features was used, which induced
unsatisfying perfo rmance because different classes of im-
ages mixed together in feature space. In this paper, we
propose a retrieval scheme based o n the c ombination of
Gabor features and SVMs for hepatic CT images query.
All the features were used in our scheme.
For each hepatic CT image, the region of interest (ROI)
is selected ma nually and the corresponding feature vec-
tors are extracted through the Gabor transform. An im-
age is represented by a 6-feature vector defined by Porat
and Zeevi
[12]
. Then images are classified with the SVM
method. Fina lly, we execute the image retrieval based on
the classification. In the experiment, 521 hepatic CT im-
ages are used and a CBI R s ystem for hepatic CT images
is built based on these images and the method we intro-
duced above. The effectiveness of our method is verified
and the comparison between our method and that of Po-
rat a nd Zeevi is demonstrated.
According to the Gab or approach, a n image Φ(x, y)
can be represented as a linear combination of EFs
[12]
:
Φ(x, y) =
X
m
x
n
x
m
y
n
y
a
m
x
n
x
m
y
n
y
· f
m
x
n
x
m
y
n
y
(x, y), (1)
where a
m
x
n
x
m
y
n
y
is the coefficient of the order
(m
x
, n
x
, m
y
, n
y
), representing the relative weight of the
respective EF in Φ(x, y), f
m
x
n
x
m
y
n
y
is the EF of the or-
der (m
x
, n
x
, m
y
, n
y
),
f
m
x
n
x
m
y
n
y
(x, y) = g(x − m
x
D
x
, y − m
y
D
y
)
× exp(in
x
W
x
x + in
y
W
y
y), (2)
where W
x
· D
x
≤ 2π, W
y
· D
y
≤ 2π must be satisfied
and g(·, ·) is a two-dimensional (2D) normalized window
function. The function f
m
x
n
x
m
y
n
y
(x, y) is situated at the
point (x = m
x
D
x
, y = m
y
D
y
) of the Gabor lattice and
has a spatial freq uency of (ω
x
= n
x
W
x
, ω
y
= n
y
W
y
).
The constants, D
x
, D
y
and W
x
, W
y
, are the basic sam-
pling intervals along the spatial and the spatial-frequency
axes, respectively. When g(·, ·) in Eq. (2) is a Gaussian
window function, the Gabor EFs are not orthogonal and
thus the coe fficients {a
m
x
n
x
m
y
n
y
} are calculated using
an auxiliary function γ(·, ·) which is biorthog onal in a
certain sense to the window function g(·, ·):
a
m
x
n
x
m
y
n
y
=
ZZ
φ(x, y) × γ
∗
(x − m
x
D
x
, y − m
y
D
y
)
× exp(−in
x
W
x
x − in
y
W
y
y)dxdy. (3)
It should be noted that when g(·, ·) is a square window
function, the functions g(·, ·) and γ(·, ·) are identical
[19]
.
Porat and Zeevi defined six localized features to ana-
lyze the texture of an image. The six features are dom-
inant localized frequency (denoted by F ), va riance of
the dominant localized frequency (V F ), dominant ori-
entation (T ), variance of the dominant orientation (V T ),
mean of the localized intensity level (L), and variance of
1671-7694/2008/070495-04
c
2008 Chinese Op tics Letters