Engineering, 2013, 5, 67-72
doi:10.4236/eng.2013.55B014 Published Online May 2013 (http://www.scirp.org/journal/eng)
Improvement of Liver Segmentation by Combining High
Order Statistical Texture Features with Anatomical
Structural Features
Suhuai Luo1, Xuechen Li1, Jiaming Li2
1The University of Newcastle, Australia
2The CSIRO ICT Centre, Australia
Received 2013
ABSTRACT
Automatic segmentation of liver in medical images is challenging on the aspects of accuracy, automation and robust-
ness. A crucial stage of the liver segmentation is the selection of the image features for the segmentation. This paper
presents an accurate liver segmentation algorithm. The approach starts with a texture analysis which results in an opti-
mal set of texture features including high order statistical texture features and anatomical structural features. Then, it
creates liver distribution image by classifying the original image pixelwisely using support vector machines. Lastly, it
uses a group of morphological operations to locate the liver organ accurately in the image. The novelty of the approach
is resided in the fact that the features are so selected that both local and global texture distributions are considered,
which is important in liver organ segmentation where neighbouring tissues and organs have similar greyscale distribu-
tions. Experiment results of liver segmentation on CT images using the proposed method are presented with perform-
ance validation and discussion.
Keywords: Liver Segmentation; Texture Feature; Support Vector machine; Morphological Operation
1. Introduction
Automatic and accurate liver segmentation in medical
images such as computed tomography (CT) and magnetic
resonance imaging (MRI) is one of the most important
concentrations in medical image processing. Segmenta-
tion of liver from its surrounding organs and tissues are a
crucial yet very difficult task in building a surgical plan-
ning system for liver transplantation and resection. This
is because the boundary between the liver and its
neighbouring structures such as the heart is sometimes
barely noticeable in CT images, and the liver is nonrigid
in shape and variant in position.
Various algorithms have been proposed to deal with
liver segmentation, including live wire-based, gray level-
based, neural networks-based, model fitting-based, prob-
abilistic atlas-based, graph cut, deformable model-based,
level set-based, and machine learning-based [1-9]. Al-
though much progress has been achieved in recent years,
challenges remain on the aspects of segmentation accu-
racy, robustness and automation.
This paper presents an automatic liver segmentation by
combining high order statistical texture features with
anatomical structural features. Section two describes the
algorithm in detail, including texture analysis, liver dis-
tribution image calculation with support vector machines,
and liver organ localization with a group of morphologi-
cal operations. Section three gives the details of the liver
segmentation experiment on CT images, including the
experiment setting, performance validation and discus-
sion, and future work in the area.
2. The Approach
The proposed automatic liver segmentation in CT images
consists of three major processes as shown in Figure 1,
including texture analysis, liver distribution image cal-
culation, and liver organ localization.
The approach starts with texture analysis process,
where an optimal set of texture features including high
order statistical texture features and anatomical structural
features is calculated out of the abdominal CT images.
With these texture features as input, liver distribution
image is derived by classifying the original image pix-
elwisely using support vector machines. Since the liver
distribution image can only indicate the likely distribu-
tion of the liver organ, not the exact delineation of the
organ, liver organ localization is applied to locate accu-
rately the liver. Following subsections describe the algo-
rithm in details.

S. H. LUO ET AL.
68
Abdominal CT images
Texture feature analysis
Liver distribution image calculation
Liver region localization
Segmented liver organ
Figure 1. The diagram of the proposed automatic liver seg-
mentation.
2.1. Texture Feature Analysis
In abdominal CT images, liver shows similar grey scales
and textures to its neighbouring structures such as heart
and stomach. Therefore, three important considerations
are taken in developing the proposed algorithm. First,
liver segmentation based on greyscale parameters alone
is not sufficient; second, high order texture parameters
can better deal with liver segmentation; lastly, optimal
liver segmentation can be achieved when global ana-
tomical structural features are used.
Since homogeneity and consistency characterize liver
segmentation where multiple slices and different patients
are dealt with, texture features are considered. Textures
are complex visual patterns that are composed of entities
or subpatterns that have characteristic brightness, colour,
slope, size, etc. [10]. It can be regarded as a similarity
grouping in an image. Methods of texture analysis can be
broadly classified into four categories, including: struc-
tural approach, which represents texture by defined
primitives; statistical approach, which represents texture
by non-deterministic properties that govern the distribu-
tions and relationships between grey levels of an image;
model-based approach, which uses fractal and stochastic
models to interpret image texture; and transform ap-
proach, which represents image in a space where texture
is well characterised. The statistical approach and the
transform approach are investigated and adopted in the
proposed segmentation method, mainly based on the
facts that the statistical approach has the advantage of
representing texture inexplicitly and the transform ap-
proach has the advantage of representing texture at vari-
ous scales.
2.1.1. Grey Level Co-occurrence Matrix and Haralick
Texture Descript ors
In characterising the distribution and relationship of pix-
els, i.e. texture, in a grey scale image, the joint probabil-
ity distribution of pairs of pixels is used. It is defined as
the co-occurrence matrix. Its normalised form is noted as
Cij(d, θ) [11].
In an image Im of size H by W pixels and with G in-
tensity levels, for every pixel centered on a neighborhood
I(x,y) of size N by N, Cij(d, θ) is defined as the total
numbers of times that, within the N by N neighborhood:
I(x1,y1) = i and I(x1+dcosθ, y1+dsinθ) = j (1)
where x1 = 0, 1, …, N-1, is the row number in the
neighborhood ; y1= 0, 1, …, N-1, is the column number
in the neighborhood; i = 0,1,…,G-1, is the row number in
the co-occurrence matrix ; j = 0,1,…,G-1, is the column
number in the co-occurrence matrix; d = 1,…,N-1, is the
displacement distance along θ; θ = 0°, 45°, 90°, 135°, is
the angle between the pair.
Figure 2 illustrates the calculation of Cij(d, θ) at a
pixel in an abdominal CT image. The co-occurrence ma-
trix is square and has a size G, which is the intensity lev-
el in original image. Note that in Figure 2 (b), for the
sake of easy description, the maximal intensity value is
supposed to be 5, much less than that of original liver CT
image. For a normalised CT image, the maximal inten-
sity value is usually 255. In the figure, the curved line
connecting (b) and (c) shows how C00(1, 0) is calculated:
for d = 1 and θ = 0, within the 5 by 5 neighborhood, the
total number of times that a pair of 0-value pixels ap-
pears is one, so C00(1, 0) = 1.
1 1 0 0 0 2
0 1 0 0 0 1
1 2 0 0 0 0
0 0 2 0 0 0
1 0 1 0 1 1
0 2 0 2 0 1
0 5 3 2 1
4 2 1 5 1
0 0 5 5 1
4 4 0 1 1
4 5 3 2 0
(a)The original CT image Im, with size
1024x1024 and 256 intensity levels
(b)I(x,y), a neighborhood at a
position inside (a).
(c)Corresponding co-occurrence matrix C
ij
(d, θ),
with d=1 and θ= 0.
Figure 2. The calculation of Cij(d, θ) at a position in an ab-
dominal CT images.

S. H. LUO ET AL. 69
The co-occurrence matrix basically keeps track of all
the pixel-pair counts. It is also called spatial dependence
matrix. Since representing an image with its co-occur-
rence matrix will result in much more data (e.g., for an
image of size 1024 by 1024 pixels and with 256 intensity
levels, the co-occurrence matrix will be of size 256 by
256 by 1024 by 1024), a set of features with much less
size yet reflecting the co-occurrence characters was pro-
posed, known as Haralick texture descriptors [11]. The
nine Haralick texture descriptors can be defined and cal-
culated as below.
Entropy: measures the randomness of gray-level dis-
tribution:
lj
G
i
G
j
ij CC log

 (2)
Energy: measures the occurrence of repeated pairs:

G
i
G
j
ij
C2 (3)
Contrast: measures the local contrast:
2
()
GG
lj
ij
ijC
 (4)
Sum Average: measures the average of the gray-level:
1(
2
GG
lj lj
ij
iC jC
 )
(5)
Variance: measures the variation of gray-level distri-
bution:
2
1(()() )
2
GG
rlj clj
ij
iCjC  
 2
(6)
Correlation: measures a correlation of pixel pairs:
22
()( )
GG rc
ij rc
ij 



 ij
C
(7)
Maximum Probability (MP): gives the most predomi-
nant pixel pair:
,
,
GG
ij
ij
M
axC (8)
Inverse Difference Moment (IDM): measures the
smoothness:
2
1
GG ij
ij
C
ij
 (9)
Cluster Tendency: measures the grouping of pixels
that have similar gray-level values:
2
(
GG
rc
ij
ij  
 )
ij
C (10)
where Cij is the normalised co-occurrence matrix with
displacement distance d and angle θ; r

, c

, 2
r

, and
2
c

are the means and variance of row and column in
Cij(d,θ).
2.1.2. Wavelet Coef fi ci ents
Wavelet coefficients are the output of wavelet transform
(WT) [12] which is the decomposition of a signal into a
set of basis functions consisting of contractions, expan-
sions and translations of a mother wavelet

.
The wavelet transform of a signal f(x) is defined as
*
1
,
(,),()( )
us s
tu
Wf usff tdt
s





 (11)
where the mother wavelet

is a zero average function,
centered around zero with a finite energy. The family of
vectors is obtained by translations and dilatations of the
mother wavelet:
,
1
()( )
us
tu
t
s
s


 (12)
In image processing applications, the wavelet trans-
form is usually computed with dyadic wavelet transform
which is implemented by filter banks. The filtering is
done along both row and column with pairs of lowpass
filter and highpass filter [12]. Figure 3 illustrates the
process of deriving wavelet coefficients for an image
using the dyadic wavelet transform. Figure 3 (a) gives a
one-scale wavelet decomposition result which has four
blocks of components: LL is the downsampling of the
lowpass filtering along both row and column, LH is the
downsampling of the lowpass filtering along row and
highpass filtering along column, HL is the downsampling
of the highpass filtering along row and lowpass filtering
along column, and HH is the downsampling of the high-
pass filtering along both row and column. Such filtering
or decomposition can be done further on LL, resulting a
two-scale wavelet decomposition of an image as shown
in Figure 3 (b). Note that the number of total wavelet
coefficients equals to the number of the pixels in the im-
age, no matter being a one-scale decomposition or two-
scale decomposition. In general, there will be 4 + 3* (S -
1) blocks for an S-scale decomposition.
LL LH
HL HH HL HH
LH
LL
LL
LL
LH
LL
HL
LL
HH
(a) one-scale (b) two-scale
Figure 3. The process of deriving wavelet coefficients for an
image using dyadic wavelet transform. Where L is a low-
pass filter, H is a high-pass filter.

S. H. LUO ET AL.
70
Comparing to other transforms such as Fourier [13]
and Gabor [14], Wavelet transform has two advantages
in segmentation application. One is that it can represent
textures at the most suitable scale by varying the spatial
resolution. The other is that wavelets best suit for texture
analysis in a specific application can be chosen because
of a wide range of choices for the wavelet function.
2.1.3. Combin i ng Hi gh Order Statistical T exture
Features with An at omi c al St ructural Features
As discussed before, the grey level co-occurrence matrix
and related Haralick texture descriptors are second-order
statistical texture features. They have the advantages of
describing the statistic relationships among neighboring
pixels. However, in practical segmentation applications,
they are confined by two factors – small local range cov-
erage and huge computation load. The small local range
coverage is the fact that the co-occurrence matrix is cal-
culated within a neighboring of N by N pixels. N is usu-
ally a single digit value, e.g., 5, considering the computa-
tion load. To consider the statistical texture representa-
tions for an image, the computation load is huge. For
example, for an image of size 1024 by 1024 pixels and
with 256 intensity levels, if three kinds of pixel-pairs are
considered (i.e., the displacement distances d = 1, 2, 3),
and only one direction is considered (i.e., angle between
the pair θ =0), 3 × 9 × 1024 × 1024 Haralick texture
descriptors will be calculated, with each calculation be-
ing propotional to the task of deriving the co-occurrence
matrix of size 256 by 256.
Wavelet coefficients can compensate Haralick de-
scriptors in specifying texture, by providing features to
describe anatomical structure at a large scope with vari-
ous resolutions. For example, for a WT of 3 scales and
filter length 9, a coefficient in the lowest resolution block
can represent the texture of 8 × 9 pixels, which will well
cover the important anatomical structure around liver.
Therefore, to fully take the advantages of high order
statistical texture features and anatomical structural fea-
tures, both Haralick texture descriptors and WT coeffi-
cients are used as the inputs to liver distribution image
calculation stage.
2.2. Liver Distribution Image Calculation
The liver distribution image of an abdominal CT image is
a binary image. In the distribution image, the values of
the pixels are one if the pixels have the most possibility
of being liver, whereas the values of the other pixels are
zero. Support vector machines (SVMs) [15] are imple-
mented as classifier to pixelwisely derive the distribution
image. SVMs are a set of discriminative classifiers which
are defined by an optimal separating hyperplane. View-
ing input data as two sets of vectors in an n-dimensional
space, the hyperplane will maximize the margin between
the two data sets.
The SVMs classifiers are built in a training process. In
the process, assume the training set is {(xi,yi), i =1,2,…l},
where xi is the input with xi ∈ Rn, yi is the output with yi ∈
R R={-1,+1}, and l is the number of input samples. Then
an optimal hyperplane in canonical form must satisfy the
following constraints:
() 0xb



 (13)
where b
∈ R,

is a normal vector, and ()
x

is an
inner product induced feature map that maps the input
space into a high dimension linear space.
SVMs convert the task of finding the optimal hyper-
plane into a task of quadratic programming problem as:
2
1
2
1
min( )
l
i
i
C




subject to
()1, {1
ii ii
yxby ,1}



 (14)
Applying Lagrange multipliers, the optimal quadratic
programming problem can be solved as the following
dual optimal problem:
1
2
111
max{( ,)}
j
ll
iijiji
iij
yyK xx 


 j
i
subject to
1
0, and 0
l
ii
i
Cy 



 (15)
where i

is support value, the xi corresponding to
0iC


 is support vector (SV), and the xi corre-
sponding to 0iC


 is normal support vector
(NSV).
1[(
NSV
ij
ijji
N
xNSVx SV
byyKx 

 ,)]
j
x

 (16)
where NNSV is the number of NSV, (, )
ij
K
xx is kernel
function. Typical kernel functions are linear, polynomial,
radial basis function, and sigmoid.
The training process will derive i

, b, and (, )
ij
K
xx .
Then the SVM as a classifier can classify any input data
x with the following classify function:
1
(){( ,)}
l
ii i
i
f
xsign yKxxb 



 (17)
2.3. Liver Region Localization
The liver distribution image derived with SVMs is a bi-
nary image. It can indicate most of the liver correctly.
Figure 4 illustrates one such example, where the left is
an original abdominal CT image and the right is the out-
put of SVMs on the image. In the image, liver is at the
top-left corner, indicated with the white curve.

S. H. LUO ET AL. 71
Figure 4. An example of liver distribution image left: an
original image; right: output of SVMs.
Observing the distribution image, two issues need be
tackled. One is that the classification is not perfect, re-
sulting misclassified pixels both within the liver and out-
side the liver. The other is that the shape and spatial in-
formation is not considered, making the classification
sensitive to the noise produced by the misclassified pix-
els. Therefore, a set of binary morphological operations
[16] is specifically designed to get the delineation of the
liver out of the distribution image.
The morphological operation starts with dilate and
erode on the distribution image to get connected regions.
Morphologic operations are described by the shape and
size of the structural element used. Considering both the
anatomical structural knowledge of the abdomen and the
CT image resolutions, a square structural element with a
diameter of 6 pixels is chosen. The second stage of the
morphological operation is to further purify the outcome
of the first stage. This includes: to retain the largest ob-
ject; to remove the pixels that are liver yet are misclassi-
fied as non-liver using hole filling operation; and to de-
lete the spurs and smooth the contour along edges with
erode and dilate operation.
3. Experiments and Discussions
The proposed automatic liver segmentation algorithm
was applied to human abdominal CT images obtained
from [17]. All the images were enhanced with contrast
agent and scanned in the central venous phase on a vari-
ety of scanners ranging from 4 to 16 and 64 detector
rows. All the data were acquired in transversal direction.
The pixel spacing varied between 0.55 and 0.80 mm, the
inter-slice distance varied from 1 to 3 mm. In the ex-
periments, eight images from one subject were chosen as
training set to train the SVM classifier, and testing set
were the images from another subject.
Segmentation performance validation was done by
comparing the automatic segmentation results with the
benchmark provided by the data supplier. Three metrics
are designed to evaluate the algorithm as below.
 False positive volume fraction (FPVF)
FPVF is defined as the amount of the pixels that are
falsely classified as the liver by the proposed method, as
a fraction of the total amount of pixels that are identified
as the liver in the benchmark. It can be expressed as:
SVM man
man
LL
FPVF L


where Lman denotes the total amount of pixels that are
identified as the liver by benchmark. LSVM denotes the
total amount of the pixels that are classified by the pro-
posed method as the liver. |LSVM- Lman | is the set differ-
ence between LSVM and Lman.
 False negative volume fraction (FNVF)
FNVF is defined as the amount of the pixels that are
falsely classified by the proposed method as non-liver, as
a fraction of the total amount of pixels that are identified
as the non-liver in the benchmark. It can be expressed as:
man SVM
man
NL NL
FNVF NL


where NLman denotes the total amount of pixels that are
identified as non-liver in the benchmark. NL SVM denotes
the total amount of the pixels that are classified by the
proposed method as non-liver. |NLman-NLSVM| is the set
difference between NLman and NLSVM.
 True positive volume fraction (TPVF)
TPVF is defined as the amount of the pixels that are
classified as liver by both the proposed method and in the
benchmark, as a fraction of the total amount of pixels
that are identified as the liver in the benchmark. It can be
expressed as:
p
roposed man
man
LL
TPVF L


where Lproposed denotes the total amount of the pixels that
are classified as the liver by the proposed method.
The procedure of the experiments are so designed that
the performance comparison is done between the method
using high order statistical texture features only and the
method using both high order statistical texture features
and anatomical structural features. Two experiments had
been done. In experiment 1, nine Haralick texture de-
scriptors (as defined in equations 2 to 10) were used to
derive the liver distribution image. Where d = 2, θ = 0,
and the intensity was normalized to 256 levels. In ex-
periment 2, in addition to the nine Haralick texture de-
scriptors, Wavelet coefficients were used. Where scale
number S = 3. In both the experiments, the parameters
for SVMs are the same, including using a polynomial
kernel function.
Table 1 shows the performance comparison of the two
experiments. From the table, it can be seen that when
both the high order statistical texture features and ana-
tomical structural features are used, the total segmenta-
tion performance is apparently improved than high order

S. H. LUO ET AL.
72
Table 1. Performance metrics (%) of the two experiments.
FPVF FNVF TPVF
Experiment 1 14.7 6.3 93.8
Experiment 2 11.1 5.1 97.3
statistical texture features only are used. The perform-
ance improvement is across all the metrics, with about
four percent improvement on TPVF.
4. Conclusions
This paper presents an accurate liver segmentation algo-
rithm. The main focus of the discussion is how to im-
prove segmentation performance by selecting most suit-
able image features. There are three major steps in the
proposed method, including texture analysis which re-
sults in a suitable set of texture features, calculation of
liver distribution image using support vector machines,
and accurate liver organ localization using a group of
morphological operations to locate the liver organ. The
novelty of the approach is resided in the fact that the
features are so selected that both local and global texture
distributions are considered. Out of detailed methodology
description and segmentation experiments, it has shown
that the proposed method can accurately segment liver in
CT image, achieving as high as 97.3% on true positive
volume fraction.
REFERENCES
[1] A. M. Mharib, A. R. Ramli, S. Mashohor and R. B. Mah-
mood, “Survey on liver CT image segmentation methods”,
Artificial Intelligence Review, Vol. 37, No. 2, 2012, pp.
83-95. doi:10.1007/s10462-011-9220-3
[2] H. Bourquain, et al., “Hepavision2—A Software Assis-
tant for Preoperative Planning in Living Related Liver
Transplantation and Oncologic Liver Surgery,” Computer
Assisted Radiology&Surgery, 2002, pp. 341-346.
[3] H. P. Meinzer, M. Thorn and C. Cardenas, “Computer-
ized Planning of Liver Surgery: An Overview,” Com-
puters and Graphics, Vol. 26, No. 4, 2002, pp. 569-576.
[4] P. Campadelli, E. Casiraghi and A. Esposito, “Liver Seg-
mentation from Computed Tomography Scans: A Survey
and a New Algorithm,” Artificial Intelligence in Medicine,
Vol. 45, No. 2-3, 2009, pp. 185-196.
doi:10.1016/j.artmed.2008.07.020
[5] S. Luo, Q. Hu, X. He, J. Li, S. J. Jin, S. Chalup and M.
Park, “Automatic Liver Parenchyma Segmentation from
Abdominal CT Images Using Support Vector Machines,”
2009 IEEE/CME Int. Conf on Complex Medical Engi-
neering, April 9-11, 2009, Tempe, USA, paper 10071.
[6] V. Pamulapati, A. Venkatesan, B. J. Wood and M. G.
Linguraru, “Liver Segmental Anatomy and Analysis from
Vessel and Tumor Segmentation via Optimized Graph
Cuts,” MICCAI'11 Proceedings of the Third international
conference on Abdominal Imaging: Computational and
Clinical Applications, 2011, pp. 189-197.
[7] T. Dima and J. Domingo, “A Local Level Set Method for
Liver Segmentation in Functional MR Imaging,” IEEE
Nuclear Science Symposium Conference Record, 2011, pp.
3158-3161.
[8] J. Lu, L. Shi, M. Deng, S. C. H. Yu and P. A. Heng, “An
Interactive Approach to Liver Segmentation in CT Based
on Deformable Model Integrated with Attractor Force,”
Machine Learning and Cybernetics (ICMLC) 2011, pp.
1660-1665.
[9] Y. Zhao, Y. Zan, X. Wang and G. Li, “Fuzzy C-means
Clustering-Based Multilayer Perceptron Neural Network
for Liver CT Images Automatic Segmentation,” Control
and Decision Conference (CCDC) 2010, pp. 3423-3427.
[10] A. Materka and M. Strzelecki, “Texture Analysis Meth-
ods - a Review,” Technical Report, Technical University
of Lodz, Institute of Electronics, 1998.
[11] R. M. Haralick, “Statistical and Structural Approaches to
Texture,” Proceedings of the IEEE, Vol. 67, No. 5, 1979,
786-804. doi:10.1109/PROC.1979.11328
[12] S. Mallat, “Multifrequency Channel Decomposition of
Images and Wavelet Models”, IEEE Trans. Acoustic,
Speech and Signal Processing, Vol. 37, No. 12, 1989, pp.
2091-2110.
[13] A. Rosenfeld and J. Weszka, “Picture Recognition in
Digital Pattern Recognition,”K. Fu (Ed.), Springer-Verlag,
1980, pp. 135-166.
doi:10.1007/978-3-642-67740-3_5
[14] J. Daugman, “Uncertainty Relation for Resolution in
Space, Spatial Frequency and Orientation Optimised by
Two-Dimensional Visual Cortical Filters,” Journal of the
Optical Society of America, Vol. 2, 1985, pp. 1160-1169.
doi;10.1364/JOSAA.2.001160
[15] N. Cristianini and J. Shawe-Taylor, “An Introduction to
Support Vector Machines and Other Kernel-based Learn-
ing Methods,” Cambridge University Press, ISBN
0521780195, 2000. doi:10.1017/CBO9780511801389
[16] J. Serra, “Image Analysis and Mathematical Morphol-
ogy,” Theoretical Advances, New York: Academic, Vol.
2, 1998.
[17] T. Heimann, M. Styner and B. van Ginneken, “3D Seg-
mentation in the Clinic: A Grand Challenge,” MICCAI
2007, the 10th Intel Conf. on Medical Image Computing
and Computer Assisted Intervention, 29 Oct. to 2 Nov.
2007, Brisbane, Australia, pp. 7-15