Content uploaded by Puja Bharti
Author content
All content in this area was uploaded by Puja Bharti on May 15, 2023
Content may be subject to copyright.
Ultrasonic Imaging
1 –29
© The Author(s) 2016
Reprints and permissions:
sagepub.com/journalsPermissions.nav
DOI: 10.1177/0161734616639875
ultrasonicimaging.sagepub.com
Article
Computer-Aided Characterization
and Diagnosis of Diffuse Liver
Diseases Based on Ultrasound
Imaging: A Review
Puja Bharti1, Deepti Mittal1, and Rupa Ananthasivan2
Abstract
Diffuse liver diseases, such as hepatitis, fatty liver, and cirrhosis, are becoming a leading cause
of fatality and disability all over the world. Early detection and diagnosis of these diseases is
extremely important to save lives and improve effectiveness of treatment. Ultrasound imaging,
a noninvasive diagnostic technique, is the most commonly used modality for examining
liver abnormalities. However, the accuracy of ultrasound-based diagnosis depends highly on
expertise of radiologists. Computer-aided diagnosis systems based on ultrasound imaging assist
in fast diagnosis, provide a reliable “second opinion” for experts, and act as an effective tool
to measure response of treatment on patients undergoing clinical trials. In this review, we first
describe appearance of liver abnormalities in ultrasound images and state the practical issues
encountered in characterization of diffuse liver diseases that can be addressed by software
algorithms. We then discuss computer-aided diagnosis in general with features and classifiers
relevant to diffuse liver diseases. In later sections of this paper, we review the published studies
and describe the key findings of those studies. A concise tabular summary comparing image
database, features extraction, feature selection, and classification algorithms presented in the
published studies is also exhibited. Finally, we conclude with a summary of key findings and
directions for further improvements in the areas of accuracy and objectiveness of computer-
aided diagnosis.
Keywords
diffuse liver diseases, artificial intelligence, medical image processing, feature extraction,
classification, computer-aided characterization and diagnosis, ultrasound imaging
Introduction
Diffuse liver diseases such as hepatitis, fatty liver and cirrhosis represent a failure in hepatic
metabolic and synthesis processes.1 According to World Health Organization, viral hepatitis—a
group of infectious diseases known as hepatitis A, B, C, D, and E—affects millions of people
1Department of Electrical and Instrumentation Engineering, Thapar University, Patiala, India
2Department of Radiology, Manipal Hospital, Bangalore, India
Corresponding Author:
Puja Bharti, Department of Electrical and Instrumentation Engineering, Thapar University, Patiala 147004, India.
Email: puja.bharti@gmail.com
639875UIXXXX10.1177/0161734616639875Ultrasonic ImagingBharti et al.
research-article2016
by guest on April 20, 2016uix.sagepub.comDownloaded from
2 Ultrasonic Imaging
globally, causing acute and chronic liver disease (CLD) and killing approximately 1.4 million
people every year.2 In addition, prolonged illness is the foremost cause of liver cirrhosis and
cancer, accounting for almost 80% of all liver cancer cases.3 Rapidly rising rate of obesity and
consumption of alcohol are also contributing to health hazards such as fatty liver.4,5 These liver
diseases are of great concern as a high proportion of individuals infected with these diseases may
not experience symptoms at the early phase of the disease. Individuals affected by these diseases
usually learn about their infection when they are persistently ill.6 Hence, the early detection is
required for better and in-time treatment.7
The accurate approach to detect the severity of diffuse liver disease is biopsy,8 but biopsy is an
invasive technique and associates pain and risk of bleeding.9 The noninvasive imaging techniques
used for diffuse liver disease detection are ultrasound scan, computerized tomography (CT) scan,
and magnetic resonance imaging (MRI) scan.10 Ultrasound imaging, being cost-effective, easy to
operate, and portable, is a frequently used diagnostic technique.11,12 However, while working with
a large number of images, radiologists can commit errors due to fatigue or boredom.13 Sometimes,
there is even lack of qualitative response while comparing images of patients with a reference pat-
tern.14 In such situations, computers can play a vital role by carrying out specific and repeated tasks
to aid in diagnosis and detection of diffuse liver diseases.15-17 Apparently, many techniques, such as
artificial neural network (ANN), support vector machine (SVM), and linear discriminant analysis
(LDA), are being used to develop such systems.18-24 Furthermore, it is possible to encode knowl-
edge of many experts in a single computational procedure of these systems.25-27
In this paper, we review image processing techniques, such as feature extraction, feature
selection, and classification techniques, with a focus on ultrasound images of diffuse liver dis-
eases. We wish to provide the reader with information on (a) major issues linked to differentia-
tion of diffuse liver diseases; (b) state-of-the-art feature extraction, feature selection, and
classification techniques; and (iii) future directions in this field. This study is intended to inform
researchers in the field of ultrasound imaging of liver, diffuse liver ultrasound image processing,
and computer vision. The rest of the study is as follows: we first describe the acquisition of ultra-
sound image and ultrasound findings, and briefly describe related issues in characterization of
diffuse liver diseases. We then discuss computer-aided diagnosis (CAD) and characterization,
feature extraction and selection techniques, classifiers, and CAD assessment techniques. In a
later section of this paper, the review of published CAD studies in diffuse liver diseases is pre-
sented. In the final section, we draw conclusion and propose future directions.
Ultrasound Imaging of Diffuse Liver Diseases
Diffuse liver diseases lead to global transformation of organ, which impacts the whole liver.28
Hepatitis is defined as inflammation of liver. Acute hepatitis lasts for less than six months but it
becomes chronic if it persists longer. Chronic hepatitis may have no symptoms, or may evolve to
fibrosis and cirrhosis.29 Fatty liver is an accumulation of fat cells in the liver and is usually caused by
alcoholism, diabetes, obesity, malnutrition, chronic illness, nonalcoholic steatohepatitis, drugs, and
toxins.30-32 The fibrosis of liver is characterized by increased extracellular matrix forming hepatic
scars. Usually fibrosis is measured by Metavir score system in which fibrosis is graded on a 5-point
scale from 0 to 4 (F0: no fibrosis; F1: portal fibrosis without septa; F2: portal fibrosis with few septa;
F3: numerous septa without cirrhosis; F4: cirrhosis).33,34 Cirrhosis is the late stage of fibrosis when
the smooth liver tissue gets replaced with irregular nodules and liver becomes harder.35
Ultrasound Imaging
Ultrasound imaging technique is one of the most widely used noninvasive and real-time diagnos-
tic modality for diagnosing diffuse liver diseases.36 Ultrasound imaging uses frequencies in the
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 3
range of 2 to 18 MHz.37 The acquisition of ultrasound image is based on the principle of pulse-
echo mode. The ultrasound image is generated by ultrasound penetration into the human body.
Ultrasound waves are reflected as echoes and converted into signals by a transducer.38 The ultra-
sound image is planar, two-dimensional (2D) ultrasound. However, in recent years, three-dimen-
sional (3D) reconstruction of ultrasound images has also been developed.29,39 Accurate
information about the space, shape, size, and texture of tumor is provided by 3D/4D ultrasound.
This procedure may also be useful to evaluate the real tumor volume as well as the healthy sur-
rounding parenchyma.
Ultrasound Finding
A sequence of liver images is obtained by ultrasound imaging technique for evaluation of liver
size, contour, surface (homogeneous or coarse), echogenicity, hepatic vessel appearance, pene-
tration of ultrasound beam, and spleen size.40-45 Table 1 shows ultrasonic finding of diffuse liver
diseases, and Figure 1 shows the ultrasound images of these diseases. The high-frequency linear
(>5 Hz) array transducer is used for high-frequency gray scale imaging to assess liver surface.46-49
The low-frequency (5 MHz) convex or sector transducer is used for low-frequency gray scale
imaging, to assess liver shape, size, and parenchyma and spleen size and hepatic vessel appear-
ance or caliber.42,50-55 (Table 2)
Issues in Characterization of Diffuse Liver Diseases
When diagnosing diffuse liver diseases, ultrasound imaging is a frequently used technique in
finding useful information but it is not equally effective in determining the differences among
diffuse liver diseases or quantifying severity of these diseases. Ultrasound images are character-
ized by echogenicity that corresponds to average gray levels in image; high echogenicity means
an increased value of the average of gray levels. In case of both fatty liver and fibrosis, the
increase in echogenicity of liver tissue is almost similar, this leads to difficulty in identifying
them correctly.56 The ultrasound images of fatty liver and early cirrhosis are also similar, and it is
difficult, even for experienced radiologists, to diagnose disease type and severity.52 Diagnosing
Table 1. Causes and Ultrasound Findings of Diffuse Liver Diseases.
Disease Causes Ultrasound Findings
Hepatitis Infections from viruses, bacteria,
or parasites; medications, such as
overdose of acetaminophen
Hepatomegaly, diffuse decreased echogenicity
Fatty liver Alcohol abuse, toxins, metabolic
disorder, obesity, nutrition disorder
Fine parenchymal texture, decreased number of
vessels, hepatomegaly, increased echogenicity
Fibrosis Chronic hepatitis, alcohol abuse,
toxins, metabolic disorder,
prolonged cholestasis, hepatic
venous obstruction, immune
disorder, intestinal bypass
Normal appearance of liver, slight increase in
echogenicity, coarse echotexture
Cirrhosis Hepatitis, alcohol overconsumption,
liver toxicity, immune disorder,
Wilson’s disease, alpha 1-antitrypsin
deficiency, cystic fibrosis
Normal or slightly increased echogenicity, coarse
parenchymal texture, nodularity resulting in
surface irregularity, rounded contours, shrunken
liver, small right lobe with enlarged left and/or
caudate lobes (volume redistribution), decreased
number of vessels, regenerative nodules,
manifestations of portal hypertension
by guest on April 20, 2016uix.sagepub.comDownloaded from
4 Ultrasonic Imaging
liver fibrosis is also challenging because the visual aspects in ultrasound imaging of healthy and
fibrotic liver are much alike.57 In addition, the sensitivity and specificity of ultrasound imaging
in detecting severity of fatty liver is still in discussion.58
In CAD systems, it is observed that some of the extracted features from different diffuse liver
diseases overlap, making it difficult to achieve 100% sensitivity and specificity (Figure 2). Thus,
features to be extracted and the classifier to be used are still a topic of debate. Furthermore, the
challenge of designing a classifier with fast response time and high accuracy still exists. Finally,
there is a requirement of large database for training and validation of CAD systems.59,60
Computer-Aided Diagnostic Systems
Generally, the CAD systems for detection of diffuse liver diseases are based on four steps that are
shown in Figure 3.
1. Image Preprocessing: The key issues in ultrasound imaging are interference of speckle
and low contrast. Image preprocessing involves enhancement of image, smoothing, and
reduction of speckle without destroying significant features that are useful for
diagnosis.61-66
2. Image Segmentation: In this step, the image is partitioned into nonoverlapping regions,
and identified regions of interest (ROIs) are separated from background. This ROI is used
for feature extraction.67-71
3. Feature Extraction and Selection: In this step, features are extracted and a subset of fea-
tures is selected to build an optimal set of features that can accurately distinguish diffuse
liver diseases.72,73 The most relevant features for ultrasound images of liver are listed in
Table 3.
4. Classification: In this step, various classification techniques are applied to classify sus-
pected regions as normal or diffuse liver diseases.74-78
Some of the CAD systems do not involve image preprocessing and segmentation steps.79 In
such systems, features are obtained directly from manually selected ROIs and used for
classification.
Feature Extraction and Selection
The feature extraction and selection are key steps in CAD system and characterization of diffuse
liver diseases. The features are either combined into new features or a subset of features is
selected to build an optimal set. The optimal feature set contains relevant, effective, and discrimi-
nating features.80 The reduction of duplicate features is useful to avoid “curse of dimensionality”
problem. In classification methods, such as SVM and ANN, the dimension of feature space not
only impacts the performance of classification but also training time of the classifier. Thus,
extraction and selection of features are crucial for CAD systems.
The features of diffuse liver diseases can be divided into various categories: textural features,
acoustospectrographic features, and biomarkers.
Textural features. Texture is one of the significant characteristics of image.81 Mostly, textural
features are calculated from the entire image or ROI using the gray level values.82,83 The selected
ROI should not include artifacts such as blood vessels, costal shadows, and spikes.84 In Table 3,
distinctive and effectively proved textural features are summarized. The most frequently extracted
features are plotted in Figure 4.
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 5
FT1, gray level co-occurrence matrix (GLCM), is also known as gray level spatial depen-
dence matrix. GLCMs are built by mapping gray level co-occurrence counts or probabilities
based on the spatial relations of pixels at different angular directions. The second-order GLCMs
Figure 1. Ultrasound images of human Liver. The rectangle shows region of interest. (a) Normal liver,
(b) chronic liver, (c) fatty liver, (d) cirrhotic liver, (e) cirrhosis on which hepatocellular carcinoma (HCC)
has evolved.
by guest on April 20, 2016uix.sagepub.comDownloaded from
6 Ultrasonic Imaging
are formed by unit pixel distances and direction of displacement vectors: 0°, 45°, 90°, 135°
(Figure 5). If most of the observations in the GLCM are concentrated along the diagonals, then
the texture is coarse with respect to displacement vector. FT2 and FT3 are higher order GLCM.73
FT2, the third-order GLCM, corresponds to three pixels. The three pixels should either be col-
linear or form a right angle triangle, with the current pixel of coordinates (xi, yi) being placed in
the middle. In case of collinear, direction pairs are (0°, 180°), (90°, 270°), (45°, 225°), (135°,
315°), and in case of right angle triangle, direction pairs are (0°, 90°), (90°, 180°), (180°, 270°),
Table 2. Ultrasound Evaluation of Diffuse Liver Diseases by Use of High and Low Frequency Probes.
Study
Size of Database (No.
of Patients) Specific Technique Sensitivity Specificity
Hultcrantz et al.51 Fatty (45), Liver parenchyma 43.0 42.0
Cirr (14), echogenicity
Chronic hep (11),
Haemochromatosis
(3),
Nor (10)
Mathiesen et al.52 Fib, Liver parenchyma
echogenicity
40.0 38.6
Fatty,
Cirr
165 patients
Colli et al.49
Fibrosis, Caudate/right lobe ratio 41.0 91.0
Cirr
300 patients
Schneider et al.42 Chronic hep C (119) Spleen width 86.3 35.3
Schneider et al.42 Chronic hep C (119) Spleen length 77.5 53.0
D’Onofrio et al.44 Chronic liver disease
(105)
Caudate/right lobe ratio 32.0 99.0
D’Onofrio et al.44 Chronic liver disease
(105)
Liver parenchyma
heterogeneity
29.0 99.0
Shen et al.41 Fib (324) Diameter of portal vein 76.7 44.6
Shen et al.41 Fib (324) Diameter of spleen 60.0 78.1
Choong et al.53 Hep B (136), Liver edge 22.0 89.0
Hep C (20)
Choong et al.53 Hep B (136), Liver parenchyma 56.0 71.0
Hep C (20) echotexture
Ferral et al.45 Abnormal LFT (70) Liver surface 87.5 81.6
Colli et al.49 Fib, Liver surface 60.0 92.0
Cirr
300 patients
D’Onofrio et al.44 Chronic liver disease
(105)
Liver surface 54.0 78.0
Paggi et al.55 Increased ALT (430) Liver surface 73.0 90.0
Choong et al.53 Hep B (136), Liver surface 74.0 59.0
Hep C (20)
Vigano et al.54 Chronic Hep C (108) Liver surface 51.0 90.0
Cirr = Cirrhosis; Hep= Hepatitis; Nor = Normal; Fib = Fibrosis; ALT = Alanine aminotransferase; LFT = Liver
function test.
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 7
(0°, 270°), (45°, 135°), (135°, 225°), (225°, 315°), and (45°, 315°). These matrices are mostly
constructed for a small range of displacement vectors of d =
[]
,dx dy
, that is, 1 or 2. Similary,
for FT3, the fifth-order GLCM, the group direction of five pixels is taken as (0°, 180°, 90°, 270°)
and (45°, 225°, 135°, 315°).
FT4, co-occurrence matrix of edge orientation (EOCM), computes the local features such as
edges or edge orientations. The higher order can also be computed. Haralick features, such as
inverse difference moment (IDM), angular second moment (ASM), correlation, entropy, differ-
ence variance, contrast, sum of average, variance, sum entropy, difference entropy, and informa-
tion measures of correlation are computed from nth order GLCM and EOCM (Table 4).85
FT5, first-order statistics (FOS), are computed from the likelihood of observing a particular
pixel value at a randomly chosen location in the image. FOS depends only on individual pixel
values and not on the relation among neighboring pixel values. Average gray value, standard
deviation, uniformity, variance, skewness, energy, entropy, and kurtosis are examples of FOS.
The average/mean gray value represents the brightness or echogenicity of texture and computes
the average intensity value of pixels. The standard deviation computes average contrast. Skewness
measures the symmetry of intensity values about mean in the histogram. Entropy represents non-
uniformity or complexity of texture in image. FT6, gray level difference matrix (GLDM), is
based on the estimation of the probability density function of absolute gray level differences
Figure 2. The plot shows samples from data set containing three classes, that is, normal, chronic liver,
and cirrhosis. Note that there is considerable class overlap in the projected space that makes it difficult
to attain 100% specificity and sensitivity.
Figure 3. The architecture of CAD system for detection and classification of diffuse liver diseases.
CAD = computer-aided diagnosis.
by guest on April 20, 2016uix.sagepub.comDownloaded from
8 Ultrasonic Imaging
among pixels pairs at a given distance δ = (Δx, Δy). FT7, gray level run length matrix (GLRLM),
is defined by counting the occurrence of runs for each of the gray levels and length in the speci-
fied direction.
FT8, wavelet transform (WT), performs the decomposition of signal spectrum into numerous
details at various resolutions.86 At each level, image is decomposed into approximation coeffi-
cient, horizontal detail, vertical detail, and diagonal detail coefficients. In WT, only the approxi-
mation coefficients are analyzed at next decomposition level whereas in FT9, wavelet packet
Table 3. Textural Features.
Textural Features
FT1 Gray level co-occurrence matrix (inverse difference moment, angular second moment,
correlation, entropy, sum entropy, difference entropy, variance, difference variance,
contrast, info measures of correlation, sum average)
FT2 Third-order gray level co-occurrence matrix (correlation, autocorrelation index, variance,
energy, entropy)
FT3 Fifth-order gray level co-occurrence matrix (correlation, autocorrelation index, variance,
energy, entropy)
FT4 Third-order edge co-occurrence matrix (energy)
FT5 First-order statistics (Average gray level, standard deviation, uniformity, variance, skewness,
energy, entropy, kurtosis)
FT6 Gray level difference matrix (coarseness, texture strength, complexity, contrast, busyness)
FT7 Gray level run length matrix (short run emphasis, long run emphasis, high gray level run
emphasis, low gray level run emphasis, short run low gray level emphasis, short run high
gray level emphasis, long run low gray level emphasis, long run high gray level emphasis,
gray level nonuniformity, run length nonuniformity, run percentage)
FT8 Wavelet transform (mean, standard deviation, angular second moment, contrast, entropy
and energy, fractal)
FT9 Wavelet packet transform (mean, energy, standard deviation, mean energy)
FT10 Gabor wavelet transform
FT11 Laws’ texture (LL mean, LL stand deviation, LE stand deviation, SS skewness, RR energy, LE
energy, LS energy, LW energy)*
FT12 Fractal (Hurst fractal index, differential box count)
FT13 Mean gray level
FT14 Autocorrelation
FT15 Phase congruency (variance, contrast, covariance)
FT = Feature.
*Laws’ texture energy measures are defined from three vectors of length 3: L3 = (1, 2, 1), E3 =(-1, 0, 1) and S3=(-1,
2, -1), which indicates one-dimensional operations of center-weighted local averaging, symmetric first differencing for
edge detection and second differencing for spot detection. If these vectors are convolved with themselves or with
each other, we get five vectors of length 5: L5 = (1,4,6,4,1), S5 = (-1,0,2,0,-1), R5 = (1,-4,6,-4,1), E5 = (-1,-2,0,2,1),
and W5 = (-1,2,0,-2,1) where L5 , E5, S5, R5 and W5 are local averaging, edge, spot, ripple and wave detectors. If we
multiply the column vectors of length ‘n’ by row vectors of same length, we obtain ‘n x n’ Laws’ masks.
LL = L5’L5 = Texture energy from LL kernel.
EE = E5’E5 = Texture energy from EE kernel.
SS = S5’S5 = Texture energy from SS kernel.
RR = R5’R5 = Texture energy from RR kernel.
WW = W5’W5 = Texture energy from WW kernel.
LE = L5’E5 = (LE+EL)/2 = Average texture energy from LE and EL kernels.
LW = L5’W5 = (LW+WL)/2 = Average texture energy from LW and WL kernels.
Similarly, LS, LR, EL, ES, ER, EW, RL, RE, RW, WL, WE, WS and WR can be obtained.
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 9
transform (WPT), each approximation and details coefficients are further broken down till the
required level of decomposition.87 Then from these coefficients, statistical features, such as
mean, standard deviation, ASM, contrast, entropy, and energy features, are calculated. FT10,
Gabor wavelet, removes redundant information and is optimal in minimizing the joint 2D uncer-
tainty in space and frequency.88 It provides a more flexible decomposition of the entire frequency
band. Ahmadian et al.88 have shown that nonseparable Gabor WT outperformed dyadic WT for
texture classification.
FT11, Laws’ textural features, is computed by using local masks to detect local averaging,
edge detection, spot detection, wave detection, and ripple detection in texture. It measures the
amount of variation within a fixed-size window. These features are calculated via the special
1D filters of length 3, 5, 7, and 9. FT12, fractal, provides a measure of the complexity of gray
level structure in a ROI, having the property of self-similarity at different scales. It is good to
measure the degree of roughness of liver image surfaces. FT14, autocorrelation, is a basic and
traditional textural feature that evaluates linear spatial relationship between primitives. It
measures coarseness as well as the amount of regularity present in ROI. FT15, phase congru-
ency, detects a broad range of features, and it is invariant to local change in illumination and
contrast.
Acoustospectrographic features. Ultrasound images are obtained by sending an ultrasound pulse
into tissue via an ultrasound transducer (probe). The sound reflects (echoes) from parts of the
tissue; these echoes are recorded and displayed as a B-mode scan image to the operator.27 The use
of radio frequency (RF) signals facilitates access to the echoes before going through the nonlin-
ear process of envelope detection and filtering that results in a B-mode image.89 The processing
of RF signals is used for estimating acoustospectrographic tissue characteristics. Backscattered
ultrasound signals include the information of the internal scattering structure and absorption
properties of tissue through which they propagate.90 In case of diffuse liver diseases, tissue scat-
ter properties get changed, which in turn provide potential useful information for disease detec-
tion and diagnosis. RF signals give an alternative to gray scale image analysis.91 Theoretically,
the analysis of RF signals provides a more precise and reproducible method for measurement of
Figure 4. The frequently extracted features for the detection of diffuse liver diseases. FOS = first-order
statistics; GLCM = gray level co-occurrence matrix; GLRLM = gray level run length matrix. Source of
data is Table 6.
by guest on April 20, 2016uix.sagepub.comDownloaded from
10 Ultrasonic Imaging
tissue properties because signals are not subject to machine dependent processing, subsampling,
interpolation, quantization, and operator-dependent settings. Gaitini et al.92 calculated mean
backscatter coefficient and mean attenuation for detecting fatty infiltration. Wan et al.93 observed
that, if probability density function of echoes does not fit the Rayleigh model, the most likely
outcome is abnormal liver (containing fibrosis). The disadvantage of acoustospectrographic fea-
tures is that the background is quite complex, and estimation of parameters is very complicated.
Joint textural and acoustospectrographic features. The combination of both textural and acous-
tospectrographic features provides remarkable results in the field of liver characterization.94 The
feature vector is constructed by concatenating image textural and acoustospectrographic features
with the intent to combine the appearance of the tissues with information obtained from their
spectral content. The key benefit of texture and RF signal based analysis will be providing com-
plementary data for more accurate appearance of tissue composition. Thus, a rich set of features
is available for selection and classification.
Biomarkers. Biomarkers are physical, functional, or biochemical indicators that can reflect the
severity of disease.95-101 The knowledge of biomarkers clubbed with textural features can aid
experts in decision making.55,102,103 Biomarkers used in CAD systems for diffuse liver diseases
are summarized in Table 5. Biomarkers have the potential to reflect the status of liver.104,105 How-
ever, they do not correlate with the degree of liver damage.
Dimensionality reduction. In the presence of so many features, finding an optimal feature set with
low dimensionality is crucial.106 The dimensionality reduction can be achieved by discarding
noisy or irrelevant features or combining relevant features. The feature extraction targets trans-
formation of coordinate system to improve a determined goal, whereas feature selection targets
reduction of dimensionality without changing the coordinate system of data.
Feature extraction. Feature extraction reduces the amount of resources required to describe a large
set of data. The most well-known feature extraction technique is principal components analysis
(PCA). PCA reduces the dimension of feature space by projecting the original data along the
directions of greater variance.107 The directions of maximum variance of original data set are
known as principal components. These principal components are orthogonal to each other and
are of reduced dimensions. The less varying and overlapping data are eliminated, as they do not
present meaningful information for classification. The independent component analysis, factor
analysis, and discriminant analysis are few other feature extraction techniques that can be used to
reduce feature dimensions.
Figure 5. The main directions for the displacement vectors of superior order GLCM. GLCM = gray
level co-occurrence matrix.
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 11
Table 4. Haralick Features.
Features Description
Inverse difference moment
i
N
j
N
gg
ij
sij
==
∑∑
()
11
1
1+
(, )
2
−
Angular second moment
ij
{(,)}2
si j
Correlation
iij sij
==
∑∑
()()
−
11
NN
xy
xy
gg
j,
µµ
σσ
Where
µµ
σσ
xy
xy
and,, are the mean and standard
deviation of
ss
xand y
Entropy −
()
()
==
∑∑
i
N
j
N
gg
sijsij
11
,log(
,)
Sum entropy −
()
∑
i
N
xy xy
g
si
si
=2
2
++
log( ())
Difference entropy −
=
∑
i
N
xy xy
g
si
si
0
1
()log( ())
−
−−
Contrast
k
N
xy
ks k
=
∑
0
g−12-
((
))
Variance
i
N
j
N
gg
is
ij
==
∑∑
()
11
2
(),−
µ
Difference variance variance of s
xy
−
Sum of average
i
N
x+y
g
is i
=
∑
2
2
()
Information measures of
correlation [H1 and H2] H1 =
HXYHXY1
max(HX,HY)
−
H2 = [1- exp (2 )]
1
2
−−HXY2 HXY
()
where, S(i,j) is the probability of having two pixels with gray value i and gray value j a fixed distance apart.
Ng = number of distinct gray levels in the quantized image.
s(i,j) = (i,j)/R ; R =
i
N
j
N
gg
S(i, j);
==
∑∑
11
sx(i) =
j
Ng
s(i, j);
=
∑
1
sy(j) =
i
Ng
s(i, j);
=
∑
1
sx+y (k) =
i
N
j
N
i+ j,k
gg
s(i, j),
==
∑∑
11
δ
k = 2,3,…,2Ng
sx-y (k) =
i
N
j
N
|i j|,k
gg
s(i, j),
==
∑∑
11
δ−
k = 0,1,…,Ng -1
where
δ
m,n= 1 when m = n; 0 when m n;
HX = −
() ()
∑
i
xx
si silog( ) = entropy of sx
HY = −
() ()
∑
j
yy
sj
sj
log( ) = entropy of sy
HXY = −
()
()
==
∑∑
i
N
j
N
gg
si,j i, j
11
log(
s)
HXY1 = −
() ()
∑∑
ij
xy
sijsis j(, )log{ }
HXY2 = −
() () () ()
∑∑
ij
yxy
is jsis jslog{ }
x
by guest on April 20, 2016uix.sagepub.comDownloaded from
12 Ultrasonic Imaging
Feature selection. Feature selection is the process of choosing optimal or suboptimal subset from
original feature space while maintaining the distinguishing capability among classes.80 The indi-
vidually poor features, which have potential to become distinct features when combined, are also
considered. Generally, these optimal subsets improve classifier performance and reduce time and
cost of obtaining features. The feature selection algorithm that is based on correlation, outputs a
Table 5. Biomarkers.
Biomarkers
Age
Gender
Duration of infection
HCV genotype
HCV RNA level
Serum total protein
Serum albumin (%)
Serum albumin (g/dL)
1 globulin (%)
1 globulin (g/dL)
2 globulin (%)
2 globulin (g/dL)
globulin (%)
globulin (g/dL)
globulin (%)
globulin (g/dL)
Albumin/globulin ratio
Total cholesterol
AST
ALT
LDH
ALP
-GTP
LAP
CHE
Total bilirubin
AST/ALT ratio
Platelet count
Serum iron
HPT
Prothrombin time (%)
Prothrombin time (s)
APTT (%)
Ferritin
TTT
ZTT
Hyaluronic acid
AST = aspartate aminotransferase; ALT = alanine aminotransferase; LDH = lactate dehydrogenase; ALP = alkaline
phosphatase; -GTP = gamma glutamyl transpeptidase; LAP = leucine aminopeptidase; CHE = cholinesterase; HPT
= hepaplastin test; APTT = activated partial thromboplastin time; TTT = thymol turbidity test; ZTT = zinc sulfate
turbidity test; HCV = hepatitis C virus.
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 13
set of features with high correlation with respect to a class. Genetic algorithm (GA), which helps
in feature selection, is a stochastic algorithm that mimics natural evolution. The basic iterative
model of GA is an evaluation–selection–reproduction loop.83 Chi-square feature selection evalu-
ates each attribute with respect to class by providing ranks to attributes regardless of their cor-
relation with each other. For both feature selection and feature extraction, overfitting should be
avoided. Sequential forward selection and sequential backward selection are also common tech-
niques for feature selection. Infogain attribute evaluation in conjunction with ranker method and
gain ratio attribute evaluation in combination with ranker method are also used for feature selec-
tion. Exhaustive search technique selects the most successful feature subset; however, it is a slow
feature selection approach.73
Classifiers
After the extraction and selection of features, the observation values are provided as input to clas-
sifier to classify liver images into normal and abnormal or into different abnormalities of diffuse
liver diseases.108,109 Majority of the publications focused on classifying normal and cirrhotic liver
or normal and fatty liver; some of the articles focused on classifying severity levels of fatty liver,
chronic hepatitis, and cirrhosis; and only few of them focused on progression of chronic liver
stages: normal, inflammation, fibrosis, cirrhosis, and liver failure/hepatocellular carcinoma
(HCC). In Table 6, classifiers used in diffuse liver disease classification are presented. Figure 6
shows the most commonly used classifiers in the published research.
ANN. ANN is a computational model that imitates the way the human brain deals with informa-
tion.135 ANN is self-learning and self-organizing, and has strong fault tolerance ability. ANN is a
network of highly interconnected neurons operating in parallel. The neurons are organized into
three layers: input, hidden, and output. The values of input layer are multiplied by weight and
passed on to hidden layer. Several hidden layers can exist in one neural network. In hidden layers,
neurons combine the weighted inputs according to activation function and threshold value, and
use them to determine the output. Detection and classification neural networks, such as back
propagation neural network (BPNN), probabilistic neural network (PNN), and self-organizing
map (SOM), are often used in the field of diffuse liver diseases.136
BPNN. BPNN is a feed forward ANN. BPNN is trained to produce desired output in response to
a set of training inputs. The initial weights are randomly set. After iteration, weights are adjusted
by using back propagation so that mean square error gets reduced. Kalyan et al.131 used two-layer
feed forward NN, trained by using back propagation algorithm. Lee et al.118 used two-layer struc-
ture, 15-neuron hidden layer with hyperbolic tangent transfer function, and three-neuron output
layer with linear transfer function, and used Levenberg–Marquardt algorithm as learning rule to
speedup convergence.
PNN. PNN is a multilayer feed forward ANN with supervised learning process. It uses kernel-
based approximation to form an estimate of the probability density functions of classes for clas-
sification. The main advantages of PNN are fast training process and guarantee to converge to an
optimal classifier. In case the size of representative training set increases, training samples can be
added or removed without extensive retraining. Huang et al.137 used 100 ultrasound images to
train network. The recognition rate of fatty and normal liver was above 85% and 82%.
SOM. SOM explores multidimensional data set to find out patterns or structures hidden within
data and is a totally unsupervised method. It assesses input patterns, organizes them to learn the
by guest on April 20, 2016uix.sagepub.comDownloaded from
14
Table 6. Summary of Diffuse Liver Diseases of Ultrasound Images Presented in Published Papers.
Paper Name
Diseases (No. of
Patients) Features Extracted
Feature
Selection Classifiers Remarks
Pavlopoulos
et al.110
Nor,
Fatty,
Cirr
FOS, GLDM, GLRLM, GLCM, fractal
dimension
Fuzzy NN Accuracy (%): 82.67
Horng et al.111 Nor (30),
Cirr (30),
Hep 30)
TFCM, GLCM, statistical feature matrix,
fractal dimension, textural spectrum
Maximum likelihood Accuracy (%):86.7
Lee et al. 112 Nor (23),
Cirr (46),
Hepatoma (44)
WT, fractal dimension, energy, magnitude,
GLCM, Gabor filter
Bayes Accuracy (%):
Nor & Abnor: 96.7
Cirr & Hepatoma:
93.6
Lee et al.77 Nor (50),
Cirr (50),
Hepatoma (50)
M-band WT, fractal dimension Bayes, k-NN, PNN, BPNN,
Modified PNN
Accuracy (%):
Statistical classifier: 90
ANN: 92
Gaitini et al.92 Nor,
Fatty
65 patients
GLCM, attenuation/backscattering (slope,
offset, mean)
Logistic regression AUC:
Abnormal & Nor:
0.86
Pure fatty & Nor: 1
Pure fattu & mixed:
0.92
Cao et al.113 Nor (18),
Fib (18)
Fractal dimension; EOCM (entropy) FLD, SVM Accuracy (%):
FLD: 85.2
SVM: 84.4
Cao et al.114 Nor (30),
Fib (30),
Cirr (20)
2DPhaseCongruency (variance, contrast,
covariance, sum variance)
SVM Accuracy (%):
Nor: 96.7
Fib: 86.6
Cirr: 95
Horng et al.115 Nor (13),
Chronic Hep (20),
Cirr (7)
GLCM, TFCM, CM-score Correlation Radial function network Accuracy (%): 92.5
(continued)
by guest on April 20, 2016uix.sagepub.comDownloaded from
15
Paper Name
Diseases (No. of
Patients) Features Extracted
Feature
Selection Classifiers Remarks
Lu et al.116 Nor (64),
Cirr (30)
Mean, standard deviation, variance, fractal
dimension
PCA Kernel density estimation Bootstrap prediction
error: 5.29%
Sriraam et al.117 Nor,
Fatty,
Malignant
Gabor WT, gray level mean, variance BPNN Accuracy (%): 94
Lee et al.118 Nor,
Cirr,
Hepatoma
M-band WT, fractal dimension, proposed
algorithm
Bayes, k-NN, BPNN, PNN Accuracy (%): 93.23
Vicas et al.13 Chronic Hep C
(125)
FOS, GLCM, multiresolution fractal
dimension, Gabor filter, Laws’ energy,
EOCM, phase congruency-based
edge detection, TFCM, GLDM, FPS,
morphological fractal dimension,
differential box counting
Logistic regression Accuracy (%):
Nor & Cirr: 69.5
Mitrea et al.119 Cirr with HCC,
HCC
Mean gray levels, 2nd, 3rd, 5th order
GLCM, 2nd, 3rd order EOCM, WT
(entropy), Laws’ filters
CFS with
genetic
search
SVM, ANN, AdaBoost + SVM,
AdaBoost + ANN
Accuracy (%):
SVM: 77.94
Mitrea et al.73 Cirr with HCC,
HCC
(130)
3rd, 5th, 7th order GLCM, edge-based
statistics, Laws’ filters, WT (entropy)
CFS with
genetic
search
RF, SVM, Bagging + RF,
Bagging + SVM, AdaBoost +
RF,
AdaBoost +SVM,
Stacking: MLP + RF + SVM
Accuracy:
Bagging + RF: 81.78%
Vicas et al.120 Chronic Hep C
(125)
FOS, GLCM, EOCM, fractal dimension,
Gabor filter, Laws’ energy, phase
congruency-based edge detection, TFCM
Correlation Logistic regression Accuracy > 97.9%
Wu et al.121 Nor (90),
Cirr (176),
HCC (166)
GLCM, WT (fractal dimension, energy,
energy deviation), Gabor WT (fractal
dimensions, energy, energy deviation),
genetic algorithm
Genetic
algorithm
k-NN, fuzzy k-NN, PNN, SVM Accuracy (%):
fused features: 96.62
(continued)
Table 6. (continued)
by guest on April 20, 2016uix.sagepub.comDownloaded from
16
Paper Name
Diseases (No. of
Patients) Features Extracted
Feature
Selection Classifiers Remarks
Moldovanu
et al.122
Nor (10),
Fatty (10)
Fractal analyses, Euler number, RF5(the
proportion of iso-segments)
Correlation ANN Accuracy (%): 94
Sensitivity (%): 96
Specificity (%): 92
Singh et al.123 Nor (15),
Fatty (15)
Laws’ energy, statistical feature matrix
(coarseness, periodicity, roughness),
Fourier power spectrum, fractal
dimension, proposed matrix, GLCM
FLD Accuracy (%): 92
Sensitivity (%): 100
Andrade et al.124 Nor,
Fatty
FOS, GLCM, GLRLM, fractal dimensions,
Laws’ texture energy
Stepwise
regression
ANN, SVM, k-NN Accuracy (%):
ANN: 76.92,
SVM: 79.77
k-NN: 74.05
Minhas et al.125 Nor (39),
Fatty (30),
Heterogeneous (19)
WPT (median, standard deviation,
interquartile range)
SVM Overall accuracy (%):
95
Icer et al.126 Nor (45),
Fatty (95)
Mean gray level, gray relational grades,
AST & ALT liver enzymes
AUC:
Nor & GradeI: 0.975
GradeII & III: 0.958
GradeII & III: 0.949
Virmani et al.87 Nor (15),
Cirr (16),
HCC (25)
WPT (mean, standard deviation, energy) GA-SVM SVM Accuracy (%): 88.8
Virmani et al.127 Nor (15),
Cirr (16)
WT, WPT, Gabor WT Exhaustive
search
SVM Accuracy (%): 98.33
Sensitivity (%): 100
Ribeiro et al.94 Nor,
Chronic hep (30),
Compensated Cirr
(34),
Decompensated
Cirr (36)
GLCM, WT, autoregressive model (least
squares algorithm), laboratory and
clinical features
Bayes, SVM, k-NN Accuracy (%):
Nor:98.67,
Chronic hepatitis:
87.45
Cirr: 95.71
Table 6. (continued)
(continued)
by guest on April 20, 2016uix.sagepub.comDownloaded from
17
Paper Name
Diseases (No. of
Patients) Features Extracted
Feature
Selection Classifiers Remarks
Wu et al.128 Nor (23),
Cirr (46),
Hepatoma (44)
WT (fractal, energy), Gabor (fractal,
energy), GLCM
PSO, GA k-NN, PNN, fuzzy k-NN, SVM Accuracy (%) of SVM
with:
PSO: 96.11
GA: 96.25
Singh et al.129 Nor (80),
Fatty (100)
GLCM, GLDM, FOS, FPS, statistical
feature matrix, Laws’ filter, fractal
dimension
Correlation Proposed Accuracy (%): 95
Lee et al.130 Nor,
Cirr,
Hepatoma
M-band WT, Gabor (energy, energy
deviation, fractal)
Minimum distance classifier,
fuzzy k-NN, PNN, SVM
Accuracy (%): 95.69
Kalyan et al.131 Nor (30),
Fatty (10),
Cirr (10),
Hepatomegaly (10)
Intensity histogram, GLCM, GLRLM,
Invariant moments
Random
search,
genetic
search
BPNN Accuracy of mixed
features (%): 92.5
Santos et al.132 Nor (68),
Fatty (52)
FOS, GLCM, GLRLM, Gabor filter, Laws’
filter, fractal dimension, lacunarity,
hepatorenal coefficient, attenuation
Stepwise
regression
ANN, SVM, k-NN, Bayes,
Decision Tree
Accuracy (%) with
classifiers fusion: 79
Owjimer et al.133 Nor (39),
Fatty (30),
Heterogeneous (19)
WPT (median, standard deviation,
interquartile range)
SVM, k-NN Accuracy (%) with
SVM: 97.9
Kim et al.134 Nor (10),
Mild fatty (7),
Moderate fatty (8)
Hepatorenal index difference SOM
Nor = normal; Cirr = cirrhosis; Fib = fibrosis ; Hep = hepatitis; HCC = heptocellular carcinoma; FOS = first-order statistics; GLDM = gray level difference matrix; GLRLM =
gray level run length matrix; GLCM = gray level co-occurrence matrix; TFCM = texture feature co-occurrence matrix; EOCM = edge orientation co-occurrence matrix;WPT
= wavelet packet transform; CM = computer morphometry, FPS = fourier power spectrum; AST = aspartate aminotransferase; ALT = alanine aminotransferase ;NN = nearest
neighbor; WT = wavelet transform; BPNN = back propagation neural network; ANN = artificial neural network; PNN = probabilistic neural network, FLD = fisher linear
discriminant;;;; SVM = support vector machine; PCA = principal components analysis; RF = random forest;; PSO = particle swarm optimization; SOM = self-organizing map; MLP =
multilayer perceptron ; CFS = correlation based feature selection; GA = genetic algorithm; AUC = area under the curve.
Table 6. (continued)
by guest on April 20, 2016uix.sagepub.comDownloaded from
18 Ultrasonic Imaging
similarities among the input sets, and clusters them accordingly. The disadvantage of SOM is that
the number of parameters increases exponentially with the dimensions of input space, and it
becomes difficult to decide the number of classes. Kim et al.134 applied SOM algorithm to clas-
sify normal and fatty liver images.
SVM. SVM is both a linear and nonlinear data classifier and supervised learning process. The
kernel function is used for nonlinear mapping to transform original training data into higher
dimensions. In SVM, training can be a slow process but accuracy is high owing to its ability to
model complex nonlinear decision boundaries.
Virmani et al.87 built a hybrid classifier by combining evolutionary algorithm and SVM. GA
was used to output the most significant attributes. The performance superiority of GA-SVM was
statistically significant in comparison to SVM. Ding et al.138 proposed fuzzy SVM in which a
membership function was added in training set to reduce the impact of noise and outliers. The
high membership was awarded to support vectors and low membership was awarded to nonsup-
port vectors such as noise and outliers.
k-nearest neighbor (k-NN). This is a simple and well-known supervised classifier. It assumes that
observations, which are close, are likely to have same classification. The performance of k-NN
largely depends on the size and the quality of training set. Lee et al.77 found the performance of
k-NN classifier decreases if the feature set available is limited to the required feature set. Riberio
et al.94 observed the same for classifying images into normal, chronic hepatitis, and cirrhosis.
k-NN outperformed Bayes when optimal feature set was used.
Bayes. Bayes is a simple and fast converging classifier. It classifies observations to class label that
has highest posterior probability among class labels. Bayes could also be used to classify normal,
fatty, cirrhosis, and hepatitis liver ultrasound images.139,140 Lee et al.112 proposed hierarchical Bayes.
At first stage, normal and abnormal liver images were classified. At second stage, abnormal cases
were further classified as cirrhosis or hepatoma. The overall accuracy by hierarchical Bayes classi-
fier was above 93.06% and proved to be better than conventional Bayes classifier.
Kernel fisher discriminant. KFD, derived from LDA, is a binary classifier that uses Gaussian ker-
nel or polynomial or tangent hyperbolic kernel, to separate classes.141 KFD was used by Salca
et al.142 to distinguish severity level of fibrosis in liver.
Other classifier. The hybrid classification techniques improve performance in recognition pro-
cess, speed, and accuracy. Hybrid classification also increases robustness, stability, and
Figure 6. The frequently used classifiers for the classification of diffuse liver diseases. NN = nearest
neighbor; SVM = support vector machine.
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 19
generalization capacity of classification. ANN, fuzzy logic, and GA could be combined to form
hybrid classifier to develop CAD.20 AdaBoost provides boosting process to increase recognition
accuracy by reconsidering the incorrectly classified observations. Adaboost when combined with
SVM and MLP provided improved results.73
Assessment of CAD Systems
We have presented and reviewed different approaches for semi-automated detection of diffuse
liver diseases. It is observed that the potential of CAD system depends on the number of true-
positive and false-positive markers. Receiver-operator characteristic curve is a graphical plot that
illustrates the performance of a binary classifier system, it shows the trade-off between the frac-
tion of true-positive and false-positive markers as a function of a threshold on classifier output.59
The automatic classification efficiency of CAD system can also be evaluated by using the correct
classification rate (CCR), sensitivity, and specificity rates.143
True positive (TP) = a patient with diffuse liver disease is diagnosed with same diffuse
liver disease.
False positive (FP) = normal or some diffuse liver disease patient is diagnosed with some
other diffuse liver disease.
True negative (TN) = a normal person is correctly diagnosed as normal.
False negative (FN) = a patient with some diffuse liver disease is falsely diagnosed as
normal.
CCR =TP TN
TP TN FP FN
+
+++
Specificity = TN
TN FP+
Sensitivity = TP
TP FN+
Positive predictive value =TP
TP FP+
Negative predictive value =TN
TN FN+
To be constructive, performance of CAD system should be close to that of expert radiologist,
that is, at the specificity and sensitivity level of expert.
In most of the published research, to evaluate CAD systems justly and precisely, same
database is used for both training and testing. Most of the researchers have used individual
databases for their research and analysis. The databases differ from each other in several
aspects such as size, number of images of each disease category, and above all they are taken
from different machines and by different experts. Even if the same evaluation criteria are
used, it is still tough to make the evaluation just and precise. Large databases that represent
clinical practice and are cautiously interpreted provide basis for valid comparative research.
Therefore, a publically available reference database would give an immense boost to CAD
system research. Collection of large annotated database can attract more research groups into
this field to develop CAD systems. The size and performance of some CAD systems are listed
in Table 6.
by guest on April 20, 2016uix.sagepub.comDownloaded from
20 Ultrasonic Imaging
Review of Published CAD Studies in Diffuse Liver Diseases
Pavlopoulos et al.110 developed a technique to classify normal, fatty, and cirrhosis liver images.
FOS, GLDS, GLRLM, GLCM, and fractal dimension based features were extracted and used in
five-layered feed forward neural network, and 82.67% accuracy was obtained. Horng et al.111
proposed a texture analysis technique, called texture feature co-occurrence matrix to categorize
normal, hepatitis, and cirrhotic liver. An accuracy of 86.7% was achieved by using Maximum
likelihood classifier. Gaitini et al.92 extracted a combination of image texture and attenuation/
backscattering-based features to classify normal and fatty liver ultrasound images.
Lee et al.112 proposed feature extraction algorithm based on spatial-frequency decomposition
and fractal geometry. Accuracies of 96.7% in distinction of normal and abnormal liver and 93.6%
in cirrhosis and hepatoma were obtained by using Bayes classifier. Authors further investigated77
the accuracy of statistic classifiers i.e. Bayes, k-NN, and ANNs i.e. BPNN, PNN and modified
PNN for classification of ultrasonic liver images with increased data set. The accuracy obtained
with statistic classifier was at most 90.7% and with ANN was at least 92%. Lee et al.118 proposed
a method for computing fractal dimension, and with the proposed method, an accuracy of 93.23%
was obtained. Lee et al.130 proposed an ensemble-based data fusion method for characterizing
normal, cirrhosis, and hepatoma. The algorithm selected adequate classifier with high recogni-
tion rate and diversity. An accuracy of 95.67% was achieved.
Cao et al.113 extracted fractal dimension and entropies of texture edge co-occurrence matrix to
classify fibrosis and normal liver images. By using these features in Fisher linear discriminant
and SVM, they obtained accuracies of 85.2% and 84.4%, respectively. Cao et al.114 extracted
liver features using 2D phase congruency to distinguish among normal, fibrosis, and cirrhosis of
liver. Accuracies of 96.27% for normal liver, 86.6% for fibrosis, and 95% for cirrhosis were
achieved.
Horng et al.115 developed ultrasonic scoring system for assessing CLD by analyzing the char-
acteristics of liver echotexture. An accuracy of 92.5% was attained, and the false-negative frac-
tion was 4.17%. Lu et al.116 developed a CAD system by deriving features from ultrasound
images of liver and accompanying spleen. The smallest bootstrap prediction error of 5.29% was
obtained by combining dimension reduction, kernel density estimator (KDE), textural features,
and simultaneous comparisons of echo textures for liver versus accompanied spleen. Sriraam
et al.117 extracted mean gray level and variance after applying Gabor WT and used them in
BPNN. An accuracy of 96.8% was obtained for classifying normal, fatty, and malignant liver
images.
Vicas et al.13 concluded that the texture analysis methods have better discrimination power
compared with human experts. Authors further evaluated the performance of human experts and
texture analysis systems in predicting chronic hepatitis and concluded that the performance of
software depends highly upon the human expert who establishes the ROI.120
Mitrea et al.144 analyzed superior order GLCM and EOCM for classifying HCC and cirrhotic
liver parenchyma on which HCC had evolved. Authors further evaluated third-, fifth-, and sev-
enth-order GLCM with combination of Bagging and Random forest classifiers. They reported a
recognition rate of 81.78% for HCC and cirrhosis on which HCC had evolved.119
Wu et al.121 proposed two-stage feature fusion method to classify ultrasonic images of liver
tissue in three classes: normal, cirrhosis, and hepatitis. GLCM, multiresolution fractal feature,
and multiresolution energy feature were extracted, and the resulting fused feature set was used in
SVM. As a result 96.25 ± 1.91 accuracy was obtained. Wu et al.128 presented an evolution-based
hierarchical feature fusion method to classify hepatoma, cirrhotic, and normal liver images.
Moldovanu et al.122 extracted fractal dimension, Euler number, and RF5 textural features from
normal and fatty liver images. The extracted features were used in ANN classifier. A classification
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 21
rate of 94%, sensitivity of 96%, and specificity of 92% were obtained as a result. Andrade et al.124
extracted 325 features, that is, 10 using FOS, 44 using GLRLM, 198 using GLCM, 70 using
Laws’ texture energy measure (TEM), and three using fractal approach and applied stepwise
regression feature selection technique to obtain an optimal subset of seven features. It was
observed that accuracy improved by 12.8%, 14.38%, and 0.23% for ANN, SVM, and k-NN,
respectively.
Singh et al.123 developed a matrix inspired from visual criteria that radiologists use to classify
normal and fatty liver. An accuracy of 92% and a sensitivity of 100% were reported. Singh
et al.129 used GLCM, GLDS, FOS, FPS, statistical feature matrix (SFM), Laws’, and fractal fea-
tures to differentiate between normal and fatty liver. Pearson correlation coefficient was applied
to a set of 35 features, and seven features were selected. Furthermore, the selected seven features
were fused and used with proposed classifier. A classification accuracy of 95% and a specificity
of 100% were reported as a result.
Minhas et al.125 presented a technique based on multiscale capability of WPT to detect fatty
liver and heterogeneous liver by exploring features, such as echogenicity, granularity, and homo-
geneity of ultrasound liver images. A classification accuracy of 95% was achieved by multiclass
linear SVM. Authors suggested extending the work by calculating features based upon the differ-
ences in echogenicity of liver from that of spleen and renal cortex. Içer et al.126 evaluated the
correlation between enzymes (AST & ALT) and gray relational grades for fatty liver having dif-
ferent grades. Kalyan et al.131 developed a technique to differentiate among fatty, cirrhotic, and
hepatomegaly liver. GLCM, GLRLM, and intensity histogram were extracted and used in MLP.
As a result an accuracy of greater than 90% was obtained. Owjimer et al.133 decomposed normal,
fatty, and heterogeneous liver ultrasound images by WPT. SVM outperformed k-NN, attaining an
overall accuracy of 97.9% with a sensitivity of 100%, 100%, and 95.1% for heterogeneous, fatty,
and normal, respectively.
Virmani et al.87 developed a CAD system for characterizing normal, cirrhosis, and HCC
evolved on cirrhotic liver, based on multiresolution texture descriptors. GA-SVM was used for
feature selection, and an accuracy of 88.8% was obtained. In a similar study, the same group of
investigators obtained an accuracy of 98.33% by using a combination of WT, WPT, and Gabor
WT. 127 Riberio et al.94 built a hierarchical method, by integrating ultrasound images, laboratory
tests, and clinical records, to categorize chronic liver patients. Accuracies of 98.67% for normal
liver, 87.45% for chronic hepatitis, and 95.71% for cirrhosis were obtained.
Santos et al.132 presented three approaches to classify normal and fatty liver. First, textural
features were analyzed, and use of classifiers fusion provided accuracy of 79%. Second, heptore-
nal coefficient was calculated, followed by a statistical analysis. A sensitivity of 90% and speci-
ficity of 88% were obtained. Third, acoustic attenuation coefficient was calculated. Author also
observed that hepatorenal coefficient was not influenced by ultrasound machine parameter. Kim
et al.134 also extracted heptorenal index (HRI) difference between liver and kidney to differentiate
multilevel hepatic steatosis. SOM was used to form clusters, and analysis of these clusters gave
statistical delineation of intensity distribution in terms of HRI difference among different levels
of fatty liver. Table 6 presents the summary of these studies.
To summarize, liver abnormalities can be detected and differentiated using ultrasound imag-
ing by analyzing size, contour, echogenicity, structure, and penetration of ultrasound beam, that
is, posterior beam attenuation. The roughness of liver surface can be measured by fractal dimen-
sions. As speckle is present in high-frequency components of ultrasound image, 2D WPT, which
is considered as richer space-frequency multiresolution analysis scheme, may offer appropriate
texture descriptor. The multiresolution fractal analysis can add more information about the signal
heterogeneity. The mean of gray levels and GLCM (correlation) can indicate differences in gran-
ularity between HCC and cirrhotic liver.
by guest on April 20, 2016uix.sagepub.comDownloaded from
22 Ultrasonic Imaging
Future Directions and Conclusion
Future Directions
Currently, in the field of diffuse liver diseases, most of the work focuses on categorization of normal
and cirrhotic liver, and normal and fatty liver. This can be further extended to progression of diffuse
liver disease and severity of stage. Mostly, only liver images are used to develop CAD systems.
Practically, experts consider other information such as biomarkers and spleen and kidney appearance
along with liver images to draw conclusions. Hence, a more accurate and precise CAD system can
be developed by blending biomarkers and spleen and kidney information along with liver images.
Presently, to evaluate CAD system for diffuse liver disease, there is no publicly available
database. Even the researchers are using unevenly distributed database that results in biasness of
results in classification. A large publicly available database would attract even more research
groups into this field and help to develop a CAD system that combines multiple approaches and
outperform the available algorithms. Moreover, this database would further help catalyze research
by enabling reproducible research.
Combination of classifiers has shown promising results. To improve classification rate and
preciseness of CAD, hybrid classifier techniques could be used. Further research work on neural
networks, fuzzy logic, and evolutionary algorithm may prove beneficial, as they complement
each other.
Last but not the least, the focus of further research should be to decrease rate of false-negative
cases, so that probability of error in predicting disease case as normal get reduced. The critical
CAD testing by using robust test cases from large and diverse database could help in this cause.
Conclusion
Our study indicates that GLCM (correlation, contrast, energy, entropy) are commonly selected
features for calculating granularity and heterogeneity of diffuse liver. GLRLM has also shown
good results in measuring chaotic structure of liver tissues. Thus, we propose to study the behav-
ior of GLRLM and superior order GLCM, to determine specific features for each subtype. In
addition, we propose to determine multiresolution features by computing fractal dimensions at
different levels obtained from WPT. To improve training efficiency of training set in order to
achieve higher classification accuracy, we propose to divide training set into two groups. One
group would include values that purely separate diffuse liver classes, and the second group would
include overlapping values. As most support vectors are close to the border of diffuse liver
classes, it is highly likely that the support vectors are in overlapping values group. Thus, subse-
quent training set would consist of only the overlapping training set. In addition, as there are
some noises and outliers on the border of two classes of samples, the k-NN algorithm would be
used to remove noises and outliers. Finally, fuzzy SVM based on cluster hyperplane would be
used to train the final training set. This would improve training efficiency of the training set that
contains noises and outliers, thereby improving accuracy in classification.
Generally, ultrasound-based diagnosis of diffuse liver disease is subjective in nature, as to a
large extent, it depends on the skills and experience of the physicians. In this review, we estab-
lished the need for more reliable computer-based characterization and diagnosis systems by
reviewing ultrasound imaging findings of diffuse liver diseases and challenges to be solved.
Subsequently, we described techniques and algorithms used in feature extraction, feature selec-
tion, and classification stages of a CAD system. The different performance evaluation metrics
were also studied. Furthermore, the research work conducted in the direction of CAD systems for
diffuse liver diseases was summarized in tabular form for quick reference. This summary upholds
our view that ultrasound-based CAD systems for diffuse liver diseases can play a significant role
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 23
in reducing operator dependency in diagnosis. Also, there is a clear scope of further improvement
in CAD systems, especially speed, accuracy, and objectiveness of the diagnosis. We examine the
future trends. The paper will be useful for the researchers in ultrasound imaging of liver, image
processing, and computer vision.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or
publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
References
1. Ros PR, Mortele KJ. Diffuse liver disease. Clin Liver Dis. 2002;6(1):181-201.
2. World Health Organization. WHO welcomes global momentum on viral hepatitis. Available from
http://www.who.int/mediacentre/news/releases/2014/viral-hepatitis/en (accessed Jul 26, 2015).
3. World Health Organization. Hepatitis: “It’s closer than you think.” Available from http://www.who.
int/mediacentre/news/notes/2012/world_hepatitis_day_20120725/en/ (accessed Jul 26 2015).
4. Hart CL, Morrison DS, Batty GD, Mitchell RJ, Davey Smith G. Effect of body mass index and
alcohol consumption on liver disease: analysis of data from two prospective cohort studies. BMJ.
2010;340:1240.
5. Williams R. Global challenges in liver disease. Hepatology. 2006;44(3):521-6.
6. Merican I, Guan R, Amarapuka D, Alexander M, Chutaputti A, Chien R, et al. Chronic hepatitis B
virus infection in Asian countries. J Gastroenterol Hepatol. 2000;15(12):1356-61.
7. Gerstenmaier JF, Gibson RN. Ultrasound in chronic liver disease. Insights Imaging. 2014;5(4):441-
55.
8. Saleh HA, Abu-Rashed AH. Liver biopsy remains the gold standard for evaluation of chronic hepatitis
and fibrosis. J Gastrointest Liver Dis. 2007;16(4):425-6.
9. Tobkes AI, Nord J. Liver biopsy: review of methodology and complications. Dig Dis. 1995;13(5):267-
74.
10. Sahani DV, Kalva SP. Imaging the liver. Oncologist. 2004;9(4):385-97.
11. Tchelepi H, Ralls WP, Radin R, Grant E. Sonography of diffuse liver disease. J Ultrasound Med.
2002;21:1023-32.
12. Zwiebel WJ. Sonographic diagnosis of diffuse liver disease. Semin Ultrasound CT MR. 1995;16(1):8-
15.
13. Vicas C, Lupsor M, Badea R, Nedevschi S. Usefulness of textural analysis as a tool for noninvasive
liver fibrosis staging. J Med Ultrason. 2011;38(3):105-17.
14. Lo C, Shen YW, Huang CS, Chang RF. Computer-aided multiview tumor detection for automated
whole breast ultrasound. Ultrason Imaging. 2014;36(1):3-17.
15. Krishnan KR, Sudhakar R. Automatic classification of liver diseases from ultrasound images using
GLRLM texture features. Soft Computing Appl. 2013;195:611-24.
16. Miguel R, Silva DA, Ribeiro T. Computer-aided diagnosis in chronic liver disease based on ultra-
sound imaging [dissertation]. Instituto Superior Tecnico. 2013.
17. Badawi AM, Derbala AS, Youssef BM. Fuzzy logic algorithm for quantitative tissue characterization
of diffuse liver diseases from ultrasound images. Int J Med Inform. 1999;55(2):135-47.
18. Li G, Luo Y, Deng W, Xu X, Liu A, Song E. Computer aided diagnosis of fatty liver ultrasonic images
based on support vector machine. In Engineering in Medicine and Biology Society, 2008. EMBS
2008. 30th Annual International Conference of the IEEE 20 Aug 2008, pp. 4768-4771.
19. Stoitsis J, Valavanis I, Mougiakakou SG, Golemati S, Nikita A, Nikita KS. Computer aided diagnosis
based on medical image processing and artificial intelligence methods. Nucl Instruments Methods
Phys Res Sect A Accel Spectrometers Detect Assoc Equip. 2006;569:591-5.
by guest on April 20, 2016uix.sagepub.comDownloaded from
24 Ultrasonic Imaging
20. Al-Absi HRH, Abdullah A, Hassan MI. Soft computing in medical diagnostic applications: a short
review. In: National Postgraduate Conference, Kuala Lumpur, Malaysia, 19-20 September 2011,
pp. 1-5.
21. De Ridder D, De Ridder J, Reinders MJT. Pattern recognition in bioinformatics. Brief Bioinform.
2013;14(5):633-47.
22. Egmont M, Ridder D, Handels H. Image processing with neural networks—a review. Pattern Recognit.
2002;35(10):2279-301.
23. Ramesh AN, Kambhampati C, Monson JRT, Drew PJ. Artificial intelligence in medicine. Ann R Coll
Surg Engl. 2004;86(5):334-8.
24. Shi Z, He L. Application of neural networks in medical image processing. Symp Netw Secur. 2010;1:23-6.
25. Shiraishi J, Li Q, Appelbaum D, Doi K. Computer-aided diagnosis and artificial intelligence in clini-
cal imaging. Semin Nucl Med. 2011;41(6):449-62.
26. Doi K. Computer-aided diagnosis in medical imaging: historical review, current status and future
potential. Comput Med Imaging Graph. 2007;31(4-5):198-211.
27. Kadah YM, Farag A, Zurada JM, Badawi AM, Youssef ABM. Classification algorithms for quantita-
tive tissue characterization of diffuse liver disease from ultrasound images. IEEE Trans Med Imaging.
1996;15(4):466-78.
28. Mergo PJ, Ros PR, Buetow PC, Buck JL. Diffuse disease of the liver: radiologic-pathologic correla-
tion. Radiographics. 1994;14(6):1291-307.
29. Ishak K, Baptista A, Bianchi L, Callea F, De Groote J, Gudat F, et al. Histological grading and staging
of chronic hepatitis. J Hepatol. 1995;22(6):696-9.
30. Ribeiro R, Sanches J. Fatty liver characterization and classification by ultrasound. Pattern Recognition
and Image Analysis. 2009;5524:354-61.
31. Acharya UR, Faust O, Molinari F, Sree SV, Junnarkar SP, Sudarshan V. Ultrasound-based tissue char-
acterization and classification of fatty liver disease: a screening and diagnostic paradigm. Knowledge-
Based Syst. 2015;75:66-77.
32. Farrell GC, Larter CZ. Nonalcoholic fatty liver disease: from steatosis to cirrhosis. Hepatology.
2006;43:99-112.
33. Zhang L, Li Q, Duan Y, Yan G, Yang Y, Yang R. Artificial neural network aided non-invasive
grading evaluation of hepatic fibrosis by duplex ultrasonography. BMC Med Inform Decis Mak.
2012;12(1):55.
34. Lupsor M, Nedevschi S, Vica C. Estimating the fibrosis stage in the human liver tissue using image process-
ing methods on ultrasonographic images. Preliminary results. InProceedings of the Euro-Mediterranean
Medical Informatics and Telemedicine. 3rd International Conference 2007 May, pp. 311-8.
35. Mitrea ED, Eng P, Nedevschi S, Cenan EC. Exploring texture-based parameters for noninvasive detec-
tion of diffuse liver diseases and liver cancer from ultrasound images. In: 8th WSEAS International
Conference on Mathematical Methods and Computational Techniques in Electrical Engineering,
Stevens Point, WI, 16-17 October 2006, pp. 259-65.
36. Tchelepi H, Ralls PW, Radin R, Grant E. Sonography of diffuse liver disease. J Ultrasound Med.
2002;21(9):1023-32.
37. Badea R, Ioanitescu S. Ultrasound imaging of liver tumors-current clinical applications. INTECH
Open Access Publisher; 2012; 75-102.
38. Foster FS, Lockwood GR, Ryan LK, Harasiewicz KA, Berube L, Rauth AM. Principles and appli-
cations of ultrasound backscatter microscopy. IEEE Trans Ultrason Ferroelectr Freq Control.
1993;40(5):608-17.
39. Li G, Citrin D, Camphausen K, Mueller B, Burman C, Mychalczak B, et al. Advances in 4D medical
imaging and 4D radiation therapy. Technol Cancer Res Treat. 2008;7(1):67-81.
40. Colli A, Fraquelli M, Andreoletti M, Marino B, Zuccoli E, Conte D. Severe liver fibrosis or cirrhosis:
accuracy of US for detection—analysis of 300 cases. Radiology. 2003;227(1):89-94.
41. Shen L, Li JQ, Zeng MD, Lu LG, Fan ST, Bao H. Correlation between ultrasonographic and pathologic
diagnosis of liver fibrosis due to chronic virus hepatitis. World J Gastroenterol. 2006;12(8):1292-5.
42. Schneider AR, Teuber G, Kriener S, Caspary WF. Noninvasive assessment of liver steatosis, fibrosis
and inflammation in chronic hepatitis C virus infection. Liver Int. 2005;25(6):1150-5.
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 25
43. Allan R, Thoirs K, Phillips M. Accuracy of ultrasound to identify chronic liver disease. World J
Gastroenterol. 2010;16(28):3510-20.
44. D’Onofrio M, Martone E, Brunelli S, Faccioli N, Zamboni G, Zagni I, et al. Accuracy of ultrasound
in the detection of liver fibrosis in chronic viral hepatitis. Radiol Med. 2005;110(4):341-8.
45. Ferral H, Male R, Cardiel M, Munoz L, Quiroz y Ferrari F. Cirrhosis: diagnosis by liver surface analy-
sis with high-frequency ultrasound. Gastrointest Radiol. 1992;17(1):74-8.
46. Nishiura T, Watanabe H, Ito M, Matsuoka Y, Yano K, Daikoku M, et al. Ultrasound evaluation of the
fibrosis stage in chronic liver disease by the simultaneous use of low and high frequency probes. Br J
Radiol. 2005;78(927):189-97.
47. Aubé C, Oberti F, Korali N, Namour MA, Loisel D, Tanguy JY, et al. Ultrasonographic diagnosis of
hepatic fibrosis or cirrhosis. J Hepatol. 1999;30(3):472-8.
48. Simonovský V. The diagnosis of cirrhosis by high resolution ultrasound of the liver surface. Br J
Radiol. 1999;72(853):29-34.
49. Colli A, Colucci A, Paggi S, Fraquelli M, Massironi S, Andreoletti M, et al. Accuracy of a pre-
dictive model for severe hepatic fibrosis or cirrhosis in chronic hepatitis C. World J Gastroenterol.
2005;11(46):7318-22.
50. Yang PM, Chen CM, Lu TW, Yen CP. Computer-aided diagnosis of sonographic liver cirrhosis: a
spleen-reference approach. Med Phys. 2008;35(4):1180-90.
51. Hultcrantz R, Gabrielsson N. Patients with persistent elevation of aminotransferases: investigation
with ultrasonography, radionuclide imaging and liver biopsy. J Intern Med. 1993;233(1):7-12.
52. Mathiesen UL, Franzen LE, Åselius H, Resjö M, Jacobsson L, Foberg U, et al. Increased liver echo-
genicity at ultrasound examination reflects degree of steatosis but not of fibrosis in asymptomatic
patients with mild/moderate abnormalities of liver transaminases. Dig Liver Dis. 2002;34(7):516-22.
53. Choong CC, Venkatesh S, Siew EY. Accuracy of routine clinical ultrasound for staging of liver fibro-
sis. J Clin Imaging Sci. 2012;2(1):58.
54. Viganò M, Visentin S, Aghemo A, Rumi MG, Ronchi G. US features of liver surface nodularity as a
predictor of severe fibrosis in chronic hepatitis C. Radiology. 2005;234(2):641.
55. Paggi S, Colli A, Fraquelli M, Viganò M, Del Poggio P, Facciotto C, et al. A non-invasive algorithm
accurately predicts advanced fibrosis in hepatitis C: a comparison using histology with internal–exter-
nal validation. J Hepatol. 2008;49(4):564-71.
56. Hamer OW, Aguirre DA, Casola G, Lavine JE, Woenckhaus M, Sirlin CB. Fatty liver: imaging pat-
terns and pitfalls. Radiographics. 2006;26(6):1637-53.
57. Berzigotti A, Castera L. Update on ultrasound imaging of liver fibrosis. J Hepatol. 2013;59(1):180-2.
58. Das C, Baruah M, Singh D. Imaging of non alcoholic fatty liver disease: a road less travelled. Indian
J Endocrinol Metab. 2013;17(6):990.
59. Gonçalves VM, Delamaro ME, Nunes FD. A systematic review on the evaluation and characteristics
of computer-aided diagnosis systems. Rev Bras Eng Biomédica. 2014;30(4):355-83.
60. Van GB, Schaefer-Prokop CM, Prokop M. Computer-aided diagnosis: how to move from the labora-
tory to the clinic. Radiology. 2011;261(3):719-32.
61. Mittal D, Kumar V, Saxena SC, Khandelwal N, Kalra N. Enhancement of the ultrasound images by
modified anisotropic diffusion method. Med Biol Eng Comput. 2010;48(12):1281-91.
62. Contreras SH, Chiu T, Fox MD. Ultrasound image enhancement: a review. Biomed Signal Process
Control. 2012;7(5):419-28.
63. Latifoğlu F. A novel approach to speckle noise filtering based on Artificial Bee Colony algorithm: an
ultrasound image application. Comput Methods Programs Biomed. 2013;111(3):561-9.
64. Achim A, Bezerianos A, Tsakalides P. Novel Bayesian multiscale method for speckle removal in
medical ultrasound images. IEEE Trans Med Imaging. 2001;20(8):772-83.
65. Bhutada GG, Anand RS, Saxena SC. Edge preserved image enhancement using adaptive fusion of
images denoised by wavelet and curvelet transform. Digit Signal Process A Rev J. 2011;21(1):118-30.
66. Shrimali V, Anand R, Kumar V. Comparing the performance of ultrasonic liver image enhancement
techniques: a preference study. IETE J Res. 2010;56(1):4-10.
67. Rathore S, Iftikhar MA, Hussain M, Jalil A. Texture analysis for liver segmentation and classification:
a survey. InFrontiers of Information Technology (FIT), 19 Dec 2011, pp. 121-6.
by guest on April 20, 2016uix.sagepub.comDownloaded from
26 Ultrasonic Imaging
68. Mala K, Sadasivam V. Automatic segmentation and classification of diffused liver diseases using
wavelet based texture analysis and neural network. In: Annual IEEE INDICON, Chennai, 11-13
December 2005, pp. 216-9.
69. Lee WL, Chen YC, Chen YC, Hsieh KS. Unsupervised segmentation of ultrasonic liver images by
multiresolution fractal feature vector. Inf Sci. 2005;175(3):177-99.
70. Doerfler N, Seidl B, Shevchenko N, Stenzel R, Lueth TC. Blood vessel detection in navigated ultra-
sound: an assistance system for liver resections. In: IEEE/ICME International Conference on Complex
Medical Engineering, Harbin, China, 22-25 May 2011, pp. 445-50.
71. Noble JA, Boukerroui D. Ultrasound image segmentation: a survey. IEEE Trans Med Imaging.
2006;25(8):987-1010.
72. Wang D, Fang Y, Hu B, Cao H. B-scan image feature extraction of fatty liver. In: Sixth International
Conference on Internet Computing for Science and Engineering, Henan, China, 21-23 April 2012,
pp. 188-92.
73. Mitrea D, Nedevschi S, Badea R. The role of the superior order GLCM and of the generalized cooccur-
rence matrices in the characterization and automatic diagnosis of the hepatocellular carcinoma, based
on ultrasound images. In: IEEE International Conference on Intelligent Computer Communication
and Processing, Cluj-Napoca, 25-27 August 2011, pp. 197-204.
74. Chen CC, Daponte JS, Fox MD. Fractal feature analysis and classification in medical imaging. IEEE
Trans Med Imaging. 1989;8(2):133-42.
75. Singh M, Singh S, Gupta S. A new measure of echogenicity of ultrasound images for liver classifi-
cation. In: 24th Canadian Conference on Electrical and Computer Engineering, Niagara Falls, ON,
Canada, 8-11 May 2011, pp. 317-20.
76. Smitha P. A review of medical image classification techniques. Int Conf on VLSI, Commun Instru.
2011;11:34-8.
77. Lee WL, Hsieh KS, Chen YC. A study of ultrasonic liver images classification with artificial neural
networks based on fractal geometry and multiresolution analysis. Biomed Eng Appl Basis Commun.
2004;16(02):59-67.
78. Wang S, Summers RM. Machine learning and radiology. Med Image Anal. 2012;16(5):933-51.
79. Kyriacou E, Pavlopoulosl S, Konnis G, Koutsourisl D, Zounipoulis P, Theotokai I. Computer assisted
characterization of diffused liver disease using image texture analysis techniques on B-scan images.
In: IEEE Nuclear Science Symposium, Albuquerque, NM, 9-15 November 1997;2:1479-83.
80. Stoean C, Stoean R, Lupsor M, Stefanescu H, Badea R. Feature selection for a cooperative coevolu-
tionary classifier in liver fibrosis diagnosis. Comput Biol Med. 2011;41(4):238-46.
81. Wu CM, Chen YC, Hsieh KS. Texture features for classification of ultrasonic liver images. IEEE
Trans Med Imaging. 1992;11(2):141-52.
82. Sun YN, Horng MH, Lin XZ, Wang JY. Ultrasonic image analysis for liver diagnosis. IEEE Eng Med
Biol. 1996;15(6):93-101.
83. Wu YH, Huang JY, Cheng SC, Yang CK, Lin CL. Evolutionary feature construction for ultrasound
image processing and its application to automatic liver disease diagnosis. In: International Conference
on Complex, Intelligent and Software Intensive Systems, Seoul, South Korea, 30 June-2 July 2011,
pp. 565-70.
84. Singh M. Investigations on ROI selection for liver classification. In: IEEE 27th Canadian Conference
on Electrical and Computer Engineering, Toronto, ON, Canada, 4-7 May 2014, pp. 1-6.
85. Haralick R, Shanmugan K, Dinstein I. Textural features for image classification. IEEE Transactions
on Systems, Man and Cybernetics. 1973;SMC-3(6):610-21.
86. Mojsilović A, Popović M, Marković S, Krstić M. Characterization of visually similar diffuse dis-
eases from B-scan liver images using nonseparable wavelet transform. IEEE Trans Med Imaging.
1998;17(4):541-9.
87. Virmani J, Kumar V, Kalra N, Khandelwal N. SVM-based characterization of liver ultrasound images
using wavelet packet texture descriptors. J Digit Imaging. 2013;26(3):530-43.
88. Ahmadian A, Mostafa A, Abolhassani M, Salimpour Y. A texture classification method for diffused
liver diseases using Gabor wavelets. Conf Proc IEEE Eng Med Biol Soc. 2005;2:1567-70.
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 27
89. Yamaguchi T, Hachiya H. Proposal of a parametric imaging method for quantitative diagnosis of liver
fibrosis. J Med Ultrason. 2010;37(4):155-66.
90. Wilson LS, Robinson DE, Griffiths KA, Manoharan A, Doust BD. Evaluation of ultrasonic attenua-
tion in diffuse diseases of spleen and liver. Ultrason Imaging. 1987;9(4):236-47.
91. Thijssen JM. Ultrasonic speckle formation, analysis and processing applied to tissue characterization.
Pattern Recognit Lett. 2003;24:659-75.
92. Gaitini D, Baruch Y, Ghersin E, Veitsman E, Kerner H, Shalem B, et al. Feasibility study of ultrasonic
fatty liver biopsy: texture vs. attenuation and backscatter. Ultrasound Med Biol. 2004;30(10):1321-7.
93. Wan YL, Tai DI, Ma HY, Chiang BH, Chen CK, Tsui PH. Effects of fatty infiltration in human livers
on the backscattered statistics of ultrasound imaging. Proc Inst Mech Eng H. 2015;229(6):419-28.
94. Ribeiro RT, Marinho RT, Sanches JM. Classification and staging of chronic liver disease from multi-
modal data. IEEE Trans Biomed Eng. 2013;60(5):1336-44.
95. Toyoda H, Kumada T, Kamiyama N, Shiraki K, Takase K, Yamaguchi T, et al. B-mode ultrasound
with algorithm based on statistical analysis of signals: evaluation of liver fibrosis in patients with
chronic hepatitis C. AJR Am J Roentgenol. 2009;193(4):1037-43.
96. Hashem AM, Rasmy MEM, Wahba KM, Shaker OG. Prediction of the degree of liver fibrosis
using different pattern recognition techniques. In: 5th Cairo International Biomedical Engineering
Conference, Cairo, Egypt, 16-18 December 2010, pp. 210-4.
97. Krawczyk B, Niak M, Orczyk T, Porwik P, Musialik J, Ska-Fajfrowska B. Classification techniques
for non-invasive recognition of liver fibrosis stage. J of Med Inf Technol. 2012;20, pp.121-6.
98. Jiang Z, Yamauchi K, Yoshioka K, Aoki K, Kuroyanagi S, Iwata A, et al. Support vector machine-
based feature selection for classification of liver fibrosis grade in chronic hepatitis C. J Med Syst.
2006;30(5):389-94.
99. Belciug S, Lupsor M, Badea R. Features selection approach for non-invasive evaluation of liver fibro-
sis. Ann Uni Craiova-Maths Com Sci Ser. 2008;35:15-20.
100. Carey E, Carey WD. Noninvasive tests for liver disease, fibrosis, and cirrhosis: is liver biopsy obso-
lete? Cleve Clin J Med. 2010;77(8):519-27.
101. Fallatah HI. Noninvasive biomarkers of liver fibrosis: an overview. Adv Hepatol. 2014;2014(8):1-15.
102. Raoufy MR, Vahdani P, Alavian SM, Fekri S, Eftekhari P, Gharibzadeh S. A novel method for diag-
nosing cirrhosis in patients with chronic hepatitis B: artificial neural network approach. J Med Syst.
2011;35(1):121-6.
103. Stoean R, Stoean C, Lupsor M, Stefanescu H, Badea R. Evolutionary-driven support vector machines
for determining the degree of liver fibrosis in chronic hepatitis C. Artif Intell Med. 2011;51(1):53-65.
104. Lin RH, Chuang CL. A hybrid diagnosis model for determining the types of the liver disease. Comput
Biol Med. 2010;40(7):665-70.
105. Tan S, Feng F, Wu Y, Wu Y. Study of aided diagnosis of hepatic carcinoma based on artificial neural
network combined with tumor marker group. Phys Procedia. 2012;33:172-8.
106. Addison JFD, Wermter S, Arevian GZ. A comparison of feature extraction and selection techniques.
Int Conf Artif Neural Networks. 2003;2714:212-5.
107. Gorunescu F, Belciug S, Gorunescu M, Badea R. Intelligent decision-making for liver fibrosis stadial-
ization based on tandem feature selection and evolutionary-driven neural network. Expert Syst Appl.
2012;39(17):12824-32.
108. Stoean R, Stoean C, Lupsor M, Stefanescu H. Evolutionary conditional rules versus support vector
machines weighted formulas for liver fibrosis degree prediction. Ann Uni Craiova-Maths Com Sci
Ser. 2010;37(1):43-54.
109. Khashei M, Zeinal Hamadani A, Bijari M. A novel hybrid classification model of artificial neural
networks and multiple linear regression models. Expert Syst Appl. 2012;39(3):2606-20.
110. Pavlopoulos S, Kyriacou E, Koutsouris D, Blekas K, Stafylopatis A, Zoumpoulis P. Fuzzy neural
network-based texture analysis of ultrasonic images. IEEE Eng Med Biol Mag. 2000;19(1):39-47.
111. Horng MH, Sun YN, Lin XZ. Texture feature coding method for classification of liver sonography.
Comput Med Imaging Graph. 2002;26(1):33-42.
112. Lee WL, Chen YC, Hsieh KS. Ultrasonic liver tissues classification by fractal feature vector based on
M-band wavelet transform. IEEE Trans Med Imaging. 2003;22(3):382-92.
by guest on April 20, 2016uix.sagepub.comDownloaded from
28 Ultrasonic Imaging
113. Cao G, Shi P, Hu B. Liver fibrosis identification based on ultrasound images captured under varied
imaging protocols. J Zhejiang Univ Sci B. 2005;6(11):1107-14.
114. Cao G, Shi P, Hu B. Ultrasonic liver discrimination using 2-D phase congruency. IEEE Trans Biomed
Eng. 2006;53(10):2116-9.
115. Horng MH. An ultrasonic image evaluation system for assessing the severity of chronic liver disease.
Comput Med Imaging Graph. 2007;31(7):485-91.
116. Lu HH, Chen C, Huang Y. Computer-aided diagnosis of liver cirrhosis by simultaneous comparisons
of the ultrasound images of liver and spleen. J Data Sci. 2008;6:429-48.
117. Sriraam N, Roopa J, Saranya M, Dhanalakshmi M. Performance evaluation of computer aided diag-
nostic tool (CAD) for detection of ultrasonic based liver disease. J Med Syst. 2009;33(4):267-74.
118. Lee WL, Hsieh KS. A robust algorithm for the fractal dimension of images and its applications to the
classification of natural images and ultrasonic liver images. Signal Processing. 2010;90(6):1894-904.
119. Mitrea D, Mitrea P, Nedevschi S, Badea R, Lupsor M, Socaciu M, et al. Abdominal tumor charac-
terization and recognition using superior-order cooccurrence matrices, based on ultrasound images.
Comput Math Methods Med. 2012. doi: 10.1155/2012/348135.
120. Vicas C, Lupsor M, Socaciu M, Nedevschi S, Badea R. Influence of expert-dependent variability over
the performance of noninvasive fibrosis assessment in patients with chronic hepatitis C by means of
texture analysis. Comput Math Methods Med. 2012. doi: 10.1155/2012/346713.
121. Wu CC, Lee WL, Chen YC, Lai CH, Hsieh KS. Ultrasonic liver tissue characterization by feature
fusion. Expert Syst Appl. 2012;39(10):9389-97.
122. Moldovanu S, Moraru L, Bibicu D. Computerized decision support in liver steatosis investigation. Int
J Biol Biomed Eng. 2012;6(1):69-76.
123. Singh M, Singh S, Gupta S. A new quantitative metric for liver classification from ultrasound images.
Int J Comput Electr Eng. 2012;4(4):605-7.
124. Andrade A, Silva JS, Santos J, Belo-Soares P. Classifier approaches for liver steatosis using ultra-
sound images. Procedia Technol. 2012;5:763-70.
125. Minhas FUAA, Sabih D, Hussain M. Automated classification of liver disorders using ultrasound
images. J Med Syst. 2012;36(5):3163-72.
126. Içer S, Coşkun A, Ikizceli T. Quantitative grading using grey relational analysis on ultrasonographic
images of a fatty liver. J Med Syst. 2012;36(4):2521-8.
127. Virmani J, Kumar V, Kalra N, Khandelwal N. Prediction of liver cirrhosis based on multiresolution
texture descriptors from B-mode ultrasound. Computing C. 2013;1(1):19-37.
128. Wu C, Lee W, Chen Y, Hsieh K. Evolution-based hierarchical feature fusion for ultrasonic liver tissue
characterization. IEEE J Biomed Health Inf. 2013;17(5):967-76.
129. Singh M, Singh S, Gupta S. An information fusion based method for liver classification using texture
analysis of ultrasound images. Inf Fusion. 2013;19:91-6.
130. Lee WL. An ensemble-based data fusion approach for characterizing ultrasonic liver tissue. Appl Soft
Comput. 2013;13(8):3683-92.
131. Kalyan K, Jakhia B, Lele RD, Joshi M, Chowdhary A. Artificial neural network application
in the diagnosis of disease conditions with liver ultrasound images. Advs Bioinfor. 2014. doi:
10.1155/2014/708279.
132. Santos J, Silva JS, Santos AA, Belo-Soares P. Detection of pathologic liver using ultrasound images.
Biomed Signal Process Control. 2014;14:248-55.
133. Owjimer M, HDanyali H, Helfroush MS. An improved method for liver diseases detection by ultra-
sound image analysis. J Med Signals Sens. 2015;5(1):21-9.
134. Kim KB, Kim CW. Quantification of hepatorenal index for computer-aided fatty liver classification
with self-organizing map and fuzzy stretching from ultrasonography. BioMed Res Int. 2015. doi:
10.1155/2015/535894.
135. Jiang J, Trundle P, Ren J. Medical image analysis with artificial neural networks. Comput Med
Imaging Graph. 2010;34(8):617-31.
136. Ceylan R, Ceylan M, Özbay Y, Kara S. Fuzzy clustering complex-valued neural network to diagnose
cirrhosis disease. Expert Syst Appl. 2011;38(8):9744-51.
by guest on April 20, 2016uix.sagepub.comDownloaded from
Bharti et al. 29
137. Huang YHY, Wang LWL, Li CLC. Texture analysis of ultrasonic liver image based on wavelet trans-
form and probabilistic neural network. Int Conf Biomed Eng Informatics. 2008;2:1172-6.
138. Ding S, Han Y, Yu J, Gu Y. A fast fuzzy support vector machine based on information granulation.
Neural Comput Appl. 2013;23:139-44.
139. Mitrea D, Nedevschi S, Lupsor M, Socaciu M, Badea R. Experimenting various classification tech-
niques for improving the automatic diagnosis of the malignant liver tumors, based on ultrasound
images. 3rd Int Congr Image Signal Process. 2010;4:1853-8.
140. SiRui Z, Jiuqing W. A survey of algorithms for the analysis of diffused liver disease from B-mode
ultrasound images. In: 9th International Conference on Electronic Measurement & Instruments,
Beijing, China, 16-19 August 2009, pp. 2576-82.
141. Zhou G, Wang Y, Wang W. Diagnosis of hepatic fibrosis by ultrasonic image analysis. Int Conf
Biomed Eng Biotechnol. 2012;775:775-6.
142. Salca VE, Gordan M, Vlaicu A. A multi-level voting scheme of kernel fisher discriminant classifiers
for liver fibrosis identification in ultrasound images. Proc Applied Maths Mech. 2011;11(1):861-2.
143. Castellino RA. Computer aided detection (CAD): an overview. Cancer Imaging. 2005;5:17-9.
144. Mitrea D, Nedevschi S, Socaciu M, Badea R. The role of the superior order GLCM in the characteriza-
tion and recognition of the liver tumors from ultrasound images. Radioengineering. 2012;21(1):79-85.
by guest on April 20, 2016uix.sagepub.comDownloaded from
