Automatic Recognition of the Hepatocellular Carcinoma from

Ultrasound Images using Complex Textural Microstructure

Co-Occurrence Matrices (CTMCM)

Delia Mitrea

, Sergiu Nedevschi

and Radu Badea

Computer Science Department, Technical University of Cluj-Napoca, Baritiu Str., No. 26-28, Cluj-Napoca, Romania

I. Hatieganu University of Medicine and Pharmacy, V. Babes Str., No. 8, Cluj-Napoca, Romania

Keywords: Complex Textural Microstructure Co-occurrence Matrices (CTMCM), Textural Model, Hepatocellular

Carcinoma, Ultrasound Images, Classification Performance.

Abstract: The hepatocellular carcinoma is one of the most frequent malignant liver tumours. The golden standard for

HCC detection is the needle biopsy, but this is a dangerous technique. We aim to perform the non-invasive

recognition of this tumour, using computerized methods within ultrasound images. For this purpose, we

defined the textural model of HCC, consisting of the relevant textural features that separate this tumour

from other visually similar tissues and of the specific values that correspond to these relevant features:

arithmetic mean, standard deviation, probability distribution. In this paper, we demonstrate the role that the

Complex Textural Microstructure Co-occurrence Matrices have in the improvement of the textural model of

HCC and in the increase of the recognition performance. During the experiments, we considered the

following classes: cirrhosis, HCC, cirrhotic parenchyma on which HCC evolved and hemangioma, a

frequent benign liver tumour. The resulted recognition accuracy for HCC was towards 90%.

1 INTRODUCTION

The hepatocellular carcinoma (HCC) is the most

frequent malignant liver tumour, occurring in 70%

of the liver cancer cases. It evolves from cirrhosis,

after a phase of liver parenchyma restructuring at the

end of which dysplastic nodules (future malignant

tumours) result (American Liver Foundation, 2015).

The most reliable method for HCC diagnosis is the

needle biopsy, but this technique is invasive,

dangerous, as it could lead to the spread of the

tumour inside the human body. We develop non-

invasive, computerized methods for the automatic

and computer assisted diagnosis of this tumour,

based on ultrasound images. Ultrasonography is a

reliable method for patient examination and

monitoring, being safe, non-invasive, inexpensive

and repeatable. Other examination techniques that

involve medical imaging, such as the computer

tomography, the magnetic resonance imaging, the

endoscopy, even the contrast enhanced

ultrasonography or elastography, are irradiating

and/or expensive. Concerning the aspect of the HCC

tumour in ultrasound images, in incipient phase, it

appears as a small lesion (2-3 cm), usually having a

hypoechogenic and homogeneous aspect. In more

advanced phases, mostly often, its aspect becomes

hyperechogenic and heterogeneous, due to the

interleaving of many tissue types (fibrosis, necrosis,

active growth tissue, fatty cells) and to the complex

vessel structure (Sherman, 2005). However, it is

difficult to distinguish this tumour from the

surrounding cirrhotic parenchyma, and also it

sometimes resembles the benign tumours. Texture is

an important property in this context, being able to

reveal subtle characteristics of the tissue, beyond the

perception of the human eye. In nowadays’ research,

there are many approaches involving the

combination between the texture analysis methods

and various classifiers, aiming to perform the

recognition of some severe pathologies, based on

ultrasound images (Sujana et al., 1996), (Yoshida et

al., 2003), (Madabhushi, 2005), (Chikui, 2005),

(Duda et al., 2013), (Masood, 2006). In our former

research, we defined the textural model of HCC,

consisting of the complete set of relevant textural

features that best characterize this tumour,

respectively of the specific values associated to the

178

Mitrea, D., Nedevschi, S. and Badea, R.

Automatic Recognition of the Hepatocellular Carcinoma from Ultrasound Images using Complex Textural Microstr ucture Co-Occurrence Matrices (CTMCM).

DOI: 10.5220/0006652101780189

In Proceedings of the 7th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2018), pages 178-189

ISBN: 978-989-758-276-9

relevant textural features: mean, standard deviation

and probability distribution. For texture analysis, we

previously employed classical methods, such as the

first order statistics of the grey levels (arithmetic

mean, maximum and minimum values), the

autocorrelation index, the grey level co-occurrence

matrix of order two and the associated Haralick

features, edge-based statistics, the frequency of the

textural microstructures derived by using the Laws’

convolution filters, the Hurst fractal index, the

Shannon entropy computed after applying the

Wavelet transform at two resolution levels (Meyer-

Base, 2009). We also developed more accurate

texture analysis techniques, such as the GLCM

matrix of superior order (Mitrea, D., et al., 2012),

(Mitrea, D., Mitrea, P. et al., 2012), the edge

orientation co-occurrence matrix of second and third

order (Mitrea, D., Mitrea, P. et al., 2012),

respectively the Textural Microstructure Co-

occurrence Matrices (TMCM) of second and third

order, all these techniques being experimented in the

context of abdominal tumour recognition based on

ultrasound images (Mitrea et al., 2014). In this work,

we highlighted the role that our newly defined

Complex Textural Microstructure Co-occurrence

Matrices (CTMCM) of order two and three, which

involved a set of Laws’ features, as well as gradient

features, had in the supervised recognition of the

HCC tumour.

(a.)

(b.)

(c.)

Figure 1: Representative US images for (a.) Cirrho.sis; (b.)

HCC; (c.) Hemangioma.

The Laws’ based CTMC was experimented

before in the context of the liver tumour recognition,

in the unsupervised detection of the HCC evolution

phases (Mitrea et al., 2016), and also for colo-rectal

tumour recognition, providing satisfying results

(Mitrea et al., 2015). In this paper, we compare the

role that the CTMCM based on Laws’ textural

microstructures, or gradient features, have in

increasing the supervised recognition accuracy of

the HCC tumour and liver cancer.

We assessed the improvement of the textural

model and the classification performance increase

by providing the new set of relevant textural features

at the inputs of some powerful classifiers and meta-

classifiers such as: Support Vector Machines

(SVM), Multilayer Perceptron (MLP), decision trees

(C4.5), AdaBoost combined with C4.5. We

compared the performance due to the newly defined

textural features with that provided by the formerly

existing textural features. The experimental dataset

consisted of B-mode ultrasound images representing

liver tissue affected by cirrhosis, hepatocellular

carcinoma, cirrhotic parenchyma surrounding the

HCC tumour, and also the hemangioma benign

tumour. Relevant images belonging to each of these

classes are illustrated in Figure 1. When taking into

account the newly defined textural features, the

obtained accuracy for HCC recognition was towards

90%.

2 THE STATE OF THE ART

The texture-based methods were widely used during

nowadays’ research, in combination with classifiers,

with the purpose of the automatic recognition of the

tumours within medical images (Yoshida et al.,

2003), (Madabhushi, 2005), (Chikui, 2005), (Duda

et al., 2013), (Masood, 2006). Thus, methods like

the Wavelet and Gabor transforms were used in

combination with Artificial Neural Networks (ANN)

and Bayesian classifiers for the recognition of the

liver and prostate tumours from ultrasound (Sujana

et al., 1996), (Yoshida et al., 2003) and magnetic

resonance (MRI) images (Madabhushi, 2005), the

fractal-based methods were used for the recognition

of the salivary gland tumours in (Chikui, 2005),

while the run-length matrix parameters, in

combination with the Haralick features derived from

the GLCM matrix were used in conjunction with

ANN classifiers, Support Vector Machines and

Fisher Linear Discriminants, for the automatic

recognition of the liver lesions (Sujana et al., 1996).

In a more recent approach, the authors performed the

recognition of the liver tumours based on the

computation of the textural parameters from typical

and contrast enhanced computer tomographic (CT)

images. Feature selection was performed and then a

C4.5 classifier was applied, which provided a

recognition rate above 90% (Duda et al., 2013).

Automatic Recognition of the Hepatocellular Carcinoma from Ultrasound Images using Complex Textural Microstructure Co-Occurrence

Matrices (CTMCM)

179

The technique of Grey Level Co-occurrence

Matrix (GLCM) was frequently used with the

purpose of tumour characterization and recognition

within ultrasound images and proved its efficiency

in many situations. A representative approach is

described in (Khuzir et al., 2009), where the GLCM

method was implemented in conjunction with a

decision trees based classifier, in order to perform

the recognition of the breast lesions from

mammographic images. This matrix was computed

for the following directions: 0°, 45°, 90°, 135° and

the corresponding Haralick features were derived.

The accuracy, measured in terms of the area under

the ROC curve, was 84%. The generalized co-

occurrence matrices, defined by Davis (Davis,

1979), were recently used in the form of the Local

Binary Pattern (LBP) Co-occurrence Matrix (Sujatha

et al., 2012), respectively in the form of the texton

and texture orientation co-occurrence matrix

(Sujatha et al., 2013). In the context of the research

described in (Sujatha et al., 2012), the authors firstly

performed texton detection, by applying specific

convolution filters, then the Local Binary Pattern

(LBP) method was applied; the logically compact

LBP features were also determined, after applying

the OR operator among the neighbours of the current

pixel. At the end, the Logical Compact Local Binary

Pattern Co-occurrence Matrix (LCLBPM) was

determined and experimented on standard textures

from the VisTex database, providing a satisfying

classification accuracy, situated between 80%-98%,

better than that provided by the Gabor features

applied in the same context. In the work described in

(Sujatha et al., 2013), the authors applied

convolution filters for texton detection, then they

determined the texton co-occurrence matrix for the

following directions of the displacement vectors: 0º,

45º, 90º, 135º. The texture orientation co-occurrence

matrix was also determined, after computing the

orientation of the edges resulted after the application

of the Sobel and Canny specific methods. These

methods were experimented on the VisTex database

and provided recognition rates situated between 89%

- 98.8%. As it results from the above-described

approaches, there does not exist any significant

study involving a co-occurrence matrix based on

textural microstructures determined after applying

the Laws’ convolution filters, or multiple edge

detection filters. Also, there is no systematic study

of the relevant textural features that best characterize

HCC, in the context of the automatic recognition of

this tumour. In our previous research, we defined the

textural model of the malignant tumours (Mitrea et

al., 2012), useful in both automatic and computer

aided diagnosis. In this work, we analyse the

efficiency of the Complex Textural Microstructure

Co-occurrence Matrices (CTMCM) in the context of

HCC recognition from ultrasound images.

3 THE PROPOSED SOLUTION

3.1 The Imagistic Textural Model of

the Hepatocellular Carcinoma

The imagistic textural model of the hepatocellular

carcinoma (HCC), also defined in (Mitrea et al.,

2012), consists of:

• The complete set of the relevant textural

features able to differentiate this tumour from

visually similar classes

• The specific values associated to the relevant

textural features: arithmetic mean, standard

deviation, probability distribution.

In order to build the imagistic textural model, the

following phases are necessary: first, a preliminary

phase is performed, when the training set is built,

consisting of regions of interest selected inside the

tissue of interest, within the ultrasound images.

Then, an image analysis phase is due, when the

textural features are computed by applying specific,

texture analysis methods. The learning phase

consists of selecting the relevant textural features,

respectively of computing the specific values of the

relevant textural features (arithmetic mean, standard

deviation, probability distribution). At the end, the

validation phase is performed, when the values of

the relevant textural features are provided to

classifier inputs and supervised classification

methods are applied in order to assess the

classification performance due to the generated

imagistic textural model.

3.2 The Image Analysis Phase

During the image analysis phase, classical textural

features, as well as newly defined textural features,

derived from advanced texture analysis methods,

were computed. Thus, first order statistics of the

grey levels (the arithmetic mean, the standard

deviation and the probability distribution), edge-

based statistics (the edge frequency, edge contrast,

the average edge orientation), the density and

frequency of the textural microstructures obtained

after applying the Laws convolution filters, second

order statistics of the grey levels (the autocorrelation

index, the Gray Level Co-occurrence Matrix -

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

180

GLCM of order two), the Hurst fractal index, as

well as the Shannon entropy computed after

applying the Wavelet transform, recursively, twice,

were determined in the first phase. Then, the GLCM

of order 3, 5 and 7, as well as the Edge Orientation

Co-occurrence Matrix (EOCM) of order 2 and 3 and

the associated Haralick features were also computed.

In addition, a new type of superior order generalized

co-occurrence matrix, the Complex Textural

Microstructure Co-occurrence Matrix (CTMCM) of

order two and three was employed and analysed in

the current work, as described below.

3.2.1 The Complex Textural Microstructure

Co-Occurrence Matrix (CTMCM) of

Second and Superior Order

The Complex Textural Microstructure Co-

occurrence Matrix (CTMCM) was determined

through the methodology described below,

consisting of the following steps:

(1) First, we associated feature vectors to the pixels

in the region of interest, consisting of:

• the results obtained after applying the 2D

Laws’ convolution filters for detecting levels,

edges, spots, waves, ripples and also combined

microstructures (L5L5, E5E5, S5S5, W5W5,

R5R5, S5R5, R5S5), in the case of the Laws’

based CTMCM (Laws, 1980).

• the results obtained after applying edge detection

filters: the Sobel filters for identifying horizontal

and vertical edges (Meyer-Base, 2009), the

Kirsch Compass filters for finding edges with

different orientations (Kirsch, 1971), as well as

the Laplacian convolution filter, in the case of

the edge-based CTMCM (Meyer-Base, 2009).

(2) Then, we applied an improved k-means

clustering method, in the following manner: we

started from a minimum number of centres (k=50);

this number was increased by splitting the

corresponding centres; a centre was split in two

other centres, if the standard deviation of the items

within the corresponding class (cluster) overpassed

the threshold equal with ¾ of the average standard

deviation of all the existing classes. The newly

resulted centres were computed as being 1/2 of the

old centre, respectively 3/2 of the old centre.

(3) All the labels of the pixels from the region of

interest (ROI) were re-assigned after splitting the old

centres, then the step (2) was performed again. The

condition for the algorithm to finish was the

convergence, the maximum number of centres being

also established to 200. The solution of the

algorithm (the optimal solution) corresponded to the

minimum value of WCSS (Within Cluster Sum of

Squared Errors) (Duda, 2003). Thus, the definition

of the Complex Textural Microstructure Co-

occurrence Matrix (CTMCM) was provided in (1):

)}sgn()))(sgn((,

),..sgn()))(sgn((

|,||||,..,||||,|||

,),(,..,),(,),(

:)),(),..,,(),,(),,{((#),..,,(

1111

111212

11213112

222111

33221121













nnnn

nnn

nnnD

ydxdyyxx

ydyyydyyydyy

xdxxxdxxxdxx

tyxAtyxAtyxA

yxyxyxyxtttC





(1)

In (1), the # symbol represents the number of

elements of the set specified between the braces and

(2)

are the displacement vectors. „A” stands for the

attribute associated to each pixel, while t

, t

,..., t

are the values of the textons (cluster labels) obtained

after the application of the improved k-means

clustering algorithm. Thus, each element of the

CTMCM matrix represents the number of the n-

tuples of pixels, with the coordinates (x

, y

), (x

),…, (x

, y

) having the cluster labels t

, t

,..., t

and

being in a spatial relationship defined by the

displacement vectors. We computed the CTMCM

matrix of order two and three and we determined the

corresponding Haralick parameters, in a similar way

as described in (Mitrea et al., 2015). For the

CTMCM of order two, the following directions were

considered: 0°, 90°, 180°, and 270°. For the

CTMCM of order three, the current pixel was

considered in the central position and together with

the other two pixels, they were either collinear, or

formed a right angle triangle (the current pixel being

in the position of the right angle). We considered the

following orientations for the two displacement

vectors: (0°, 180°), (90°, 270°), (45°, 225°), (135°,

315°) for the case of collinear pixels; (0°, 90°), (90°,

180°), (180°, 270°), (0°, 270°), (45°, 135°), (135°,

225°), (225°, 315°), (45°, 315°), for the right angle

triangle case. The displacement vectors had the

absolute value 2, in both cases. We determined the

CTMCM matrices for all the considered direction

combinations of the displacement vectors, the final

features resulting as an average between the

Haralick features of the individual matrices, just as

in the case of the superior order GLCM, respectively

EOCM. We also extended the cluster shade and

cluster prominence features at order n, and we

computed the maximum area of the intersection of

the corresponding histogram with a horizontal plane,

)),(),..,,(),,((

112211 



ydxdydxdydxdd



Automatic Recognition of the Hepatocellular Carcinoma from Ultrasound Images using Complex Textural Microstructure Co-Occurrence

Matrices (CTMCM)

181

as described in (Mitrea et al., 2016); (Nanni et al.,

2013).

3.3 The Learning Phase

3.3.1 Feature Selection Methods

In order to derive the set of the relevant textural

features, the following methods were considered,

which provided the best results in our former

experiments: Correlation based Feature Selection

(CFS) combined with genetic search, respectively

Gain Ratio Attribute Evaluation combined with the

Ranker method. The method of Correlation-based

Feature Selection (CFS) assigned a merit to each

group of features with respect to the class parameter

(Hall, 2003), as described by the formula (3). In (3)

represents the average correlation of the features

to the class,

is the average correlation between

the features, while k is the number of the elements in

the subset.

(3)

The second technique assessed the individual

features by assigning them a gain ratio with respect

to the class (Hall, 2003), as provided in (4). In (4),

H(C) is the entropy of the class parameter, H(C|A

)

is the entropy of the class after observing the

attribute A

, while H(A

) is the entropy of the

attribute A

GainR(A

) = (H(C) - H(C|A

)) / H(A

)

(4)

The final relevance score for each feature was

obtained by computing the arithmetic mean of the

individual relevance values provided by each

method. Only those features that had a significant

value for the final relevance score (above a

threshold) were taken into account.

3.3.2 The Specific Values for the Relevant

Textural Features

The specific values of the relevant textural features

(the arithmetic mean, the standard deviation,

respectively the probability distribution) were

computed using the Weka library (Weka, 2017). In

order to determine the probability distribution for

each feature, with respect to the class, the method of

Bayesian Belief Networks (Duda, 2003) was

adopted. The technique of Bayesian Belief Networks

detects influences between the features, by

generating a dependency network, which is

represented as a directed, acyclic graph (DAG). In

this graph, the nodes represent the features, while

the edges stand for the causal influences between

these features, having associated the values of the

corresponding conditional probabilities. Every node

X in this graph has a set of parents, P and a set of

children, C. Within a Bayesian Belief Network, each

node has associated a probability distribution table,

indicating the specific intervals of values for that

node, given the values of its parents.

3.3.3 Textural Model Assessment through

Supervised Classification

The following supervised classifiers and classifier

combinations, well known for their performance,

which provided the best results in the context of our

experiments, were taken into account: the Multilayer

Perceptron (MLP), the Support Vector Machines

(SVM), the Random Forest (RF) classifier, as well

as the AdaBoost meta-classifier combined with the

C4.5 method for decision trees. For the Multilayer

Perceptron, multiple architectures were

experimented and the best one was adopted in each

case.For classification performance assessment, we

considered the following parameters: the

classification accuracy (recognition rate), the TP rate

(sensitivity), the TN rate (specificity), as well as the

area under the ROC curve (AUC) (Duda, 2003).

4 EXPERIMENTS AND

DISCUSSIONS

For the experiments, we used 100 cases of cirrhosis,

300 cases of HCC, together with the cirrhotic

parenchyma on which HCC had evolved,

respectively 100 cases of hemangioma. All these

patients underwent biopsy, for diagnostic

confirmation. For each patient, 3 images were

considered, for various orientations of the

transducer. The images were acquired by our

research collaborators, the medical specialists from

the 3

Medical Clinic of Cluj-Napoca, using a Logiq

7 ultrasound machine, at the same settings:

frequency of 5.5 MHz, the gain of 78 and the depth

of 16 cm. On each image, a region of interest,

having 50x50 pixels in size, was selected manually

inside the corresponding class of tissue, this process

being supervised by the medical specialists in

rkkk

Merit

)1( 



ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

182

radiology. A number of 90 textural features were

computed on each ROI independently on orientation,

illumination and region of interest size, using our

Visual C++ software modules. The values of the

textural features followed by the class specification

represented an instance of the training set. In the

context of binary classification, the classes were

combined in equal proportion inside the training set

(the two considered classes always had the same

number of instances). The feature selection methods

and the classification techniques were implemented

on the dataset using the Weka library (Weka3,

2017). For feature selection, we used Correlation

based Feature Selection (CFS) in combination with

genetic search, respectively Gain Ratio Attribute

Evaluation in combination with the ranker method,

as all these techniques provided the best results in

our experiments. For classification, we used the

following algorithms of the Weka library: John

Platt’s Sequential Minimal Optimization (SMO) for

implementing SVM; in this case, we used one of the

second or third degree polynomial kernels, the input

data being normalized. The J48 method, standing for

the C4.5 algorithm, was adopted as well; also, the

AdaBoost meta-classifier with 100 iterations, in

conjunction with the J48 method was employed. For

the RF method, we used the Weka’s algorithm with

the same name, with 100 trees. The

MultilayerPerceptron (MLP) classifier was also

adopted. Multiple architectures of this classifier

were experimented, in the following manner: the

total number of nodes was always equal with the

arithmetic mean between the number of attributes

and the number of classes

(no_of_attributes+no_of_classes)/2, while the

number of layers was one, two or three, the nodes

being equally distributed among the layers. For the

MLP classifier, the learning rate was 0.2 and the

value of α parameter was fixed to 0.8, in order to

achieve both high speed and high accuracy for the

learning process (Weka, 2017). After all the

parameters of these classifiers were assigned, the

strategy of cross-validation with 5 folds was

implemented on the dataset, for classification

performance assessment, using the Weka library.

4.1 The Role of the CTMCM Matrices

in the Differentiation between

Cirrhosis and Cirrhotic

Parenchyma around HCC

• The relevant textural features derived from

the Laws' based CTMCM matrix

The set of the textural features which were relevant

concerning the differentiation between cirrhosis and

the cirrhotic parenchyma on which HCC had

evolved are provided in (5):

{Laws_CTMCM3_Energy, Laws_CTMCM3

_Entropy, Laws_CTMCM_Entropy, Laws_

CTMCM_Homogeneity, Laws_CTMCM3

_MaxArea, Wavelet_Entropy1, Wavelet_

Entropy4, Laws_CTMCM_Correlation,

Laws_CTMCM3_ClusterPromminence,

Laws_CTMCM3_Homogeneity, Wavelet

_Entropy5_hl, Wavelet_Entropy5_ll,

Wavelet_Entropy5_lh, Wavelet_Entropy5 _hh,

Wavelet_Entropy6_ll, Wavelet_ Entropy6_hl,

Wavelet_Entropy6_hh, Wavelet_Entropy6_lh,

Wavelet_Entropy2, Wavelet_Entropy3,

Min_Grey_Level, GLCM_Contrast, Edge_

Orientation_Variability, Directional_Grad

_Variabiliy, Laws_CTMCM_ClShade,

Laws_Spot_Frequency, Laws_CTMCM

_ClPromminence, Laws_Spot_Mean,

Laws_CTMCM3_ClusterShade}

(5)

We can notice the importance of the textural

features derived from the second and third order

Laws' based CTMCM matrix. Most of these features

were situated on the first seven positions within the

feature ranking. From this set, the second and third

order CTMCM Energy, the second order CTMCM

Entropy, as well as the second order CTMCM

Homogeneity denoted differences in homogeneity

between liver cirrhosis and the cirrhotic parenchyma

which surrounds HCC and also the more accentuated

chaotic character of the cirrhotic parenchyma

structure, due to a more evolved restructuring

process. The CTMCM maximum area parameter

emphasizes the more increased structural complexity

of the cirrhotic parenchyma on which HCC had

evolved, in comparison with cirrhosis. We can also

notice the relevance of the second and third order

CTMCM Cluster Promminence parameter and of the

third order CTMCM Cluster Shade feature. Another

important relevant feature was the CTMCM

Correlation, denoting differences in granularity

between the two considered classes, due to the

evolving restructuring process. While the features

derived from the CTMCM matrices together with

the Laws Spot Mean and with the Laws Spot

Frequency express differences in homogeneity,

structural complexity and granularity in terms of

textural microstructures, the other relevant textural

features, derived from the GLCM matrix or referring

to edge-based statistics, emphasize these differences

in terms of grey levels or edges. The entropy

Automatic Recognition of the Hepatocellular Carcinoma from Ultrasound Images using Complex Textural Microstructure Co-Occurrence

Matrices (CTMCM)

183

computed at multiple levels, after applying the

Wavelet transform twice, is also present at a great

extent among the relevant features, expressing the

fact that the increasing chaotic character of the tissue

structure, which is due to the restructuring process,

can be visualized at multiple resolutions.

• The relevant textural features derived from

the edge based CTMCM matrix

In order to assess the role of the edge based

CTMCM matrix in the differentiation between

cirrhosis and the cirrhotic parenchyma on which

HCC had evolved, we considered the set formed by

the previously existing textural features, respectively

edge based CTMCM features. The relevant textural

features resulted after applying the feature selection

methods described in 3.3.1 are depicted in (6).

{Edge_CTMCM_Max_Area, Wavelet_

Entropy4, Edge_CTMCM3_Max_Area,

Min_Grey_Level,

Edge_CTMCM_Entropy,

Wavelet_Entropy1,Wavelet_entropy5_lh,

Wavelet_Entropy5_hl, Wavelet_Entropy5_

hh, Wavelet_entropy5_ll,

Wavelet_Entropy6 _hh,

Wavelet_Entropy6_ll,

Wavelet_Entropy6_hl,

Wavelet_Entropy2, Edge_CTMCM3

_Entropy, Edge_CTMCM_Homogeneity,

GLCM_Contrast, Edge_Contrast, GLCM_

Variance, Wavelet_Entropy_3, Edge_

CTMCM_Energy, Directional_gradient_

variability, Laws_Spot_Frequency, Laws

_Spot_Mean, Laws_Ripple_Mean}

(6)

From (6), it results the importance of the newly

defined features Maximum Area derived from the

second and third order edge based CTMCM matrix,

denoting differences in structural complexity

between the two considered classes (cirrhosis and

cirrhotic parenchyma on which HCC had evolved).

The entropy derived from the second and third

order CTMCM matrix, as well as the homogeneity

computed from the CTMCM matrix of order two

are also met among the relevant textural features,

emphasizing, in terms of edges, differences in

homogeneity between the liver affected by cirrhosis

and the cirrhotic parenchyma on which HCC had

evolved. The other part of the set containing the

relevant textural features, illustrated in (6), has an

extended intersection with the relevant textural

feature set described in (5), the entropy computed

on various components resulted after applying the

Wavelet transform at multiple resolutions being

met in this case as well.

• Class differentiation accuracy, due to the

CTMCM matrices

Figure 2 illustrates the comparison among the

classification accuracies resulted when considering

various CTMCM features in combination with the

already existing (old) textural features. The

following feature sets were taken into account: the

old textural features, including classical textural

features, superior order GLCM features and EOCM

features; the old textural features combined with the

edge-based CTMCM features; the old textural

features combined with the Laws' based CTMCM

features. Thus, we notice that the feature sets

including the newly defined textural features,

derived from the Laws' based CTMCM matrix and

from the edge-based CTMCM matrix led to an

increase in accuracy, for each considered classifier,

in comparison with the set containing only the old

textural features. The highest recognition rate, of

86.71%, resulted in the case of the AdaBoost meta-

classifier combined with the J48 method

(AdaBoost+J48), when considering the set

containing the old textural features combined with

the edge based CTMCM features. In this case, the

best structure of the MLP classifier, that provided

the highest values for the performance parameters,

consisted of three layers.

Figure 2: The increase in accuracy due to the newly

defined textural features when differentiating between

cirrhosis and cirrhotic parenchyma on which HCC had

evolved.

Concerning the arithmetic mean of the recognition

rates, for all the classifiers, in the case of each

textural feature set, the highest mean value, of

81.83%, resulted for the set containing the old

textural features combined with the edge based

CTMCM features, followed by that resulted when

combining the Laws' based CTMCM features with

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

184

the old textural features (80.075%), and then by the

mean accuracy value which corresponded to the old

textural feature set (68.89%). Concerning the other

classification performance parameters, the highest

sensitivity (TP Rate) of 87.8% and the highest AUC

of 88.2%, resulted in the case of the AdaBoost meta-

classifier combined with the J48 technique, while

the highest specificity (TP Rate) of 85.2% resulted

in the case of the SMO classifier.

4.2 The Role of the CTMCM Matrices

in the Differentiation between HCC

and the Cirrhotic Parenchyma

around HCC

• The relevant textural features derived from

the Laws' based CTMCM matrix

The relevant textural features obtained in this

case, when considering the set formed by the

previously existing textural features and the Laws’

based CTMCM features, are illustrated in (7).

{EOCM3_Energy, EOCM3_ Entropy,

GLCM7_entropy, Laws_CTMCM3

_Contrast, Laws_CTMCM_Contrast,

Laws_CTMCM_Correlation, GLCM5

_Entropy, EOCM3_Correlation,

GLCM3_Energy, GLCM7_Energy,

Wavelet_Entropy4, EOCM3_Contrast,

EOCM3_Variance, Laws_CTMCM

_Variance, Laws_CTMCM3 _Variance,

Wavelet_Entropy6_lh, Wavelet_

Entropy6_hl, Wavelet_Entropy6_ll,

Wavelet_entropy6_hh, Wavelet

_entropy5_ll, Wavelet_entropy5_lh,

Wavelet_entropy5_hl, Wavelet

_entropy5_hh, Wavelet_Entropy1,

Wavelet_Entropy2, Wavelet_Entropy3,

EOCM_Variance, GLCM7_Correlation,

GLCM7_Variance, Min_grey_level,

Laws_CTMCM_Homogeneity,

GLCM_Variance}

(7)

We remark that the contrast derived from the

second and third order CTMCM matrices, as well as

the correlation derived from the second order

CTMCM matrix are situated on the top of the

relevant feature ranking in this case. The variance

derived from the second and third order CTMCM

and the homogeneity computed from the second

order CTMCM are relevant as well. These features

emphasize the complex structure of the tumour

tissue (the contrast and the variance computed from

the second and third order CTMCM), the differences

in granularity which exist between HCC and the

cirrhotic parenchyma on which HCC had evolved

(the CTMCM Correlation) and also the

heterogeneous structure of the HCC malignant

tumour (the CTMCM Homogeneity). Concerning

the classical textural features, we notice the

importance of the textural features derived from the

second and third order EOCM (energy, entropy,

correlation, contrast and variance), from the GLCM

of order two, three, five and seven and also of the

Shannon entropy computed after applying the

Wavelet transform, at multiple resolutions. All these

features confirm the echogenicity increase in the

cases of HCC, as well as the complex and chaotic

structure of this malignant tumour.

• The relevant textural features derived from

the edge based CTMCM matrix

The set of the relevant textural features obtained in

this situation is depicted in (8).

{EOCM3_Energy, EOCM3_ Variance,

Wavelet_Entropy1, GLCM5_Entropy,

GLCM7_Energy, Edge_CTMCM_MaxArea,

Edge_ CTMCM_Entropy, GLCM3_Energy,

Wavelet_Entropy4, EOCM3_Contrast,

GLCM7_Entropy, Wavelet_Entropy6_ll,

Wavelet_Entropy6_hh, Wavelet_ Entropy6_hl,

Wavelet_Entropy6_lh,

Wavelet_Entropy5_lh,Wavelet_Entropy5_ll,

Wavelet_Entropy5_hh, Wavelet_Entropy1,

EOCM3_Correlation, EOCM_Entropy,

EOCM3_Entropy, Wavelet_Entropy2,

GLCM7_Correlation, Min_Grey_Level,

Wavelet_Entropy4, GLCM_Variance,

Wavelet_Entropy3, EOCM3_Homogeneity,

GLCM_Contrast, Edge_CTMCM _Correlation,

Edge_CTMCM_Contrast}

(8)

We notice that the Maximum Area and the

Entropy derived from the second order CTMCM

matrix are situated on the top of the corresponding

ranking, while the CTMCM Correlation and

Contrast are also present in the relevant feature set.

Thus, the chaotic structure (CTMCM Entropy),

respectively the structural complexity (CTMCM

Maximum Area and Contrast) of HCC, as well as

the differences in granularity between HCC and the

cirrhotic parenchyma on which HCC had evolved,

are emphasized in this case, as well. Concerning the

old textural features, we remark the minimum grey

level, expressing the hyperechogenic nature of the

HCC tissue in most of the cases; the Energy

Automatic Recognition of the Hepatocellular Carcinoma from Ultrasound Images using Complex Textural Microstructure Co-Occurrence

Matrices (CTMCM)

185

computed based on the third and seventh order

GLCM, respectively the Entropy derived from the

fifth order GLCM, expressing the differences in

echogenicity between HCC and the surrounding

cirrhotic parenchyma, respectively the chaotic

structure of the malignant tumour; the variance

derived from the second and seventh order GLCM

characterizing the complexity of the HCC tissue, the

correlation derived from the seventh order GLCM

standing for the differences in granularity between

HCC and the neighbouring cirrhotic parenchyma.

The features derived from the third order EOCM

matrix and the Shannon entropy computed after

applying the Wavelet transform at multiple

resolutions are also met at a great extent within the

relevant feature set, denoting the complex, chaotic

structure of the malignant tumour, in comparison

with the cirrhotic parenchyma on which this tumour

had evolved.

• Class differentiation accuracy due to the

CTMCM matrices

Figure 3 depicts the comparison of the classification

accuracies obtained for each considered classifier for

the same types of feature sets as in the case of the

cirrhosis and cirrhotic parenchyma comparison. We

can notice that the classification accuracy is always

better for the feature sets containing the newly

defined textural features than in the case of the old

textural feature set. The best classification accuracy,

of 84.09% was obtained in the case of the AdaBoost

metaclassifier combined with the J48 method for the

set of the old textural features combined with the

Laws’ based CTMCM features. Concerning the

arithmetic mean value for all the considered

classifiers, the highest value, 79.68%, resulted in the

case of the old textural features combined with the

edge based CTMCM features, followed by the value

of 79.25% obtained in the case of the combination

between the old textural features and the Laws’

based CTMCM features, respectively by the value of

77.72% resulted in the case when only the old

textural features were considered. The MLP

classifier provided the best result when the

corresponding architecture had one single hidden

layer, in all the cases. Regarding the other

classification performance parameters, the highest

specificity (TN Rate) of 82.7% and the highest AUC

of 87.3%, resulted in the case of the adaboost meta-

classifier combined with J48, while the highest

sensitivity (TP Rate) of 82.7% resulted in the case of

the RF classifier.

Figure 3: The increase in accuracy due to the newly

defined textural features when differentiating between

HCC and the cirrhotic parenchyma on which HCC had

evolved.

4.3 The Role of the CTMCM Matrices

Concerning the Differentiation

between HCC and Hemangioma

• The relevant textural features derived from

the Laws' based CTMCM matrix

The set of the relevant textural features resulted in

this case is illustrated in (9):

{Laws_CTMCM_Homogeneity,

Laws_CTMCM_Correlation,

GLCM7_Entropy, GLCM5_Contrast,

Laws_CTMCM3_Energy,

Laws_CTMCM3_Contrast,

Directional_Gradient_Variability,

GLCM5_Entropy, GLCM3_Energy,

GLCM3_Homogeneity, Wavelet_

Entropy5_hl, Wavelet_Entropy6_hh,

Wavelet_entropy6_hl, GLCM_ Homogeneity,

Wavelet_Entropy7_hl,

GLCM5_Homogeneity}

(9)

The CTMCM Homogeneity and the CTMCM

Correlation are in the top of the relevant feature

ranking. They emphasize differences in

homogeneity (CTMCM Homogeneity) and in

granularity (CTMCM Correlation) between the

malignant and the benign tumour. The third order

CTMCM Energy and third order CTMCM Contrast

are also important, standing for the differences in

homogeneity, echogenicity (third order CTMCM

Energy) and tissue structure complexity (third order

CTMCM Contrast) between HCC and hemangioma.

Among the relevant textural features obtained in this

case we can also notice the features computed based

on the second and superior order GLCM, the

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

186

directional gradient variability and the entropy

computed after applying the Wavelet transform at

the second level, on the third (high-low) and fourth

(high-high) components, containing the horizontal,

respectively the diagonal edges. All these features

denote the heterogeneous, chaotic, complex structure

of HCC, compared with the more homogeneous

structure of the benign tumour.

• The relevant textural features derived from

the edge based CTMCM matrix

The set of the relevant textural features resulted in

this situation is provided in (10):

{EOCM3_Energy, GLCM7_Entropy,

EOCM3_Entropy, GLCM7_Energy,

GLCM3_Entropy, Edge_orientation

_Variability, Directional_Gradient

_Variability, GLCM5_Energy, GLCM3

_Homogeneity, Edge_CTMCM _MaxArea,

Edge_CTMCM3_Contrast,

GLCM_Homogeneity, Wavelet_Entropy7

_hl, GLCM5_Homogeneity,

GLCM7_Variance, Wavelet_Entropy3,

GLCM5_Contrast, Edge_CTMCM3

_Correlation, Directional_Gradient

_Variance, GLCM_Entropy, Laws

_Ripple_Mean, Edge_CTMCM_Contrast,

Autocorrelation_Index, Mean_Gray_Level}

(10)

We can notice, from (10), that the CTMCM

features are present within the set of the relevant

textural features in this case (the CTMCM

Maximum Area, the third order CTMCM Contrast,

respectively the second order CTMCM Contrast).

These features denote differences in structural

complexity between HCC and the hemangioma

benign tumor. The other features which are among

the most relevant textural parameters in this case are

those derived from the GLCM matrix of second and

superior order, which occupy a large portion of the

diagram, some features computed from the EOCM

matrix of order three (the Energy and Entropy), the

edge orientation variability, the autocorrellation

index, the mean grey level, the arithmetic mean of

the pixels values after applying the Laws'

convolution filter for ripple detection, respectively

the entropy computed on both first and second levels

after applying the Wavelet transform twice. All

these features confirm the chaotic, inhomogeneous,

complex character of the malignant tumor tissue.

The specific values of the edge based CTMCM

maximum area parameter, for each class (HCC and

hemangioma), obtained after applying the technique

of Bayesian Belief Networks in Weka 3.6, are

depicted within Table 1. It results that the

distribution of the values of the edge based textural

microstructures is more balanced in the case of

hemangioma, while in the case of HCC, these values

are probably grouped towards higher values, as it is

assumed that the edges are more emphasized in this

case, due to the more complex structure of the

malignant tumor.

Table 1: The probability distribution table for the

CTMCM Maximum Area.

(-∞, 64658.5]

(64658.5, ∞)

Hemangioma

0.38

0.62

HCC

0.77

0.22

• Class differentiation accuracy due to the

CTMCM matrices

Figure 4 illustrates the comparison of the

classification accuracies obtained when considering

the feature sets mentioned before. We can notice a

classification accuracy increase in all cases, due to

the newly defined textural features. The best

classification accuracy, of 88.41%, resulted in the

case of the AdaBoost meta-classifier combined with

the J48 classifier, corresponding to the feature set

containing the old textural features and the Laws’

based CTMCM features. In the case of the

arithmetic mean (average) of the recognition rates,

the most increased value (84.32%) resulted for the

feature set that contained the Laws’ based CTMCM

features, followed by the value of 83.76% obtained

in the case when taking into account the edge based

CTMCM features, then by the value of 82.005%

resulted when taking into account only the old

textural feature set. Concerning the MLP classifier,

it provided the best results when a one layer

architecture was adopted, for the feature set

containing only the old features, as well as for the

feature set containing the old textural features and

the edge based CTMCM features, respectively,

when a two layers architecture was adopted, for the

feature set containing the old textural features and

the Laws’ based CTMCM features. The other

classification performance parameters resulted as

follows: the highest sensitivity (TP Rate), of 84.9%

was obtained in the case of the AdaBoost meta-

classifier combined with the J48 method; the highest

sensitivity (TN Rate) of 88.6% was obtained in the

case of SMO; the highest AUC of 89.9%, resulted in

the case of MLP.

Automatic Recognition of the Hepatocellular Carcinoma from Ultrasound Images using Complex Textural Microstructure Co-Occurrence

Matrices (CTMCM)

187

Figure 4: The increase in accuracy due to the newly

defined textural features when differentiating between

HCC and Hemangioma.

4.4 Discussions

The newly defined CTMCM textural features always

resulted among the most relevant textural features in

our experiments, being part of the imagistic textural

model in the considered cases. These relevant

textural features highlighted the increase of the

echogenicity, heteorogeneity, chaotic structure and

structural complexity of the tissue during the

restructuring process, from cirrhosis to HCC and

also the differences between the malignant and the

benign tumoral tissue (HCC versus hemangioma).

Another aspect to be noticed is that the features

referring to the skewness and kurtosis derived from

the superior order histograms (the CTMCM Cluster

Shade and the CTMCM Cluster Promminece),

respectively the maximum intersection area with a

horizontal plane, usually occupied a place in the top

of the relevant feature ranking, highlighting, once

again, the complex structure of the malignant

tumour. We can notice that the edge based CTMCM

textural features led to a better classification

performance than the Laws’ based CTMCM textural

features, when differentiating between cirrhosis and

cirrhotic parenchyma on which HCC had evolved,

respectively between cirrhotic parenchyma and

HCC. It results the importance of the edges

concerning the refined differentiation between the

different phases of liver parenchyma restructuring.

In the case of differentiation between HCC and

hemangioma, the Laws’ based CTMCM textural

features overpassed the edge based CTMCM

textural feature from classification performance

point of view. Thus, in this case, the Laws’ based

textural microstructures better emphasized the

differences between the malignant and benign liver

tumours. The best obtained classification

performance was above 88% in the case of the

differentiation between the malignant and the benign

liver tumours, respectively above 84% when

differentiating between HCC and the cirrhotic

parenchyma. The classifier that provided the best

classification accuracy was AdaBoost combined

with the J48 technique, this combination scheme

being well known for its performance.

• Comparison with the state of the art results

The classification accuracy due to the newly

defined textural features, in combination with the

formerly existing textural features always

overpassed the classification accuracy which was

due only to the formerly existing textural features, as

shown in the experiments above. Also, this accuracy

is comparable with that of the state of the art

algorithms, in the case of the differentiation between

malignant and benign liver tumors: 88.41% in the

case when the CTMCM matrices were employed;

above 80% when using classical textural features

and classifiers (Sujana et al., 1996); about 90%

when using hierarchical wavelet-based features

(Yoshida et al., 2003). In the case when

distinguishing between the malignant liver tumours

and the cirrhotic parenchyma, a recognition rate of

84.09% was obtained when employing the CTMCM

matrices within B-mode ultrasound images (in our

case), while in the case of employing textural

features derived from CEUS images, in combination

with classifiers, an accuracy of 90% resulted (Duda

et al., 2013). As it can be noticed, the accuracy is

lower in the case of differentiation between HCC

and the cirrhotic parenchyma, due to the fact that,

especially during the intermediate evolution phases,

HCC resembles sometimes the surrounding cirrhotic

parenchyma. Thus, our solution can be further

improved by using more advanced multi-resolution

techniques, as well as CEUS images, and also by

explicitly taking into account different HCC

evolution phases.

5 CONCLUSIONS

The CTMCM matrices demonstrated an obvious

contribution concerning the increase of the

classification performance and diagnosis accuracy in

the case of the HCC malignant tumour and of the

pathological phases that precede this form of liver

cancer (cirrhosis). The corresponding features were

always among the most relevant textural features,

considerably improving the imagistic textural model

of the considered pathologies. At the end, the

resulted classification accuracy was above 84%, in

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

188

all cases. In our future research, we aim to further

increase the HCC automatic diagnosis accuracy by

employing the multi-resolution versions of the

CTMCM matrices and of the corresponding

Haralick features. We will also consider larger

datasets in order to improve the validation procedure

and deep learning methods in order to increase the

classification performance. We take into account the

possibility of using other types of ultrasound images,

as well, such as contrast enhanced ultrasound images

(CEUS), respectively elastographic images.

REFERENCES

American liver foundation, 2015. Online: http://www.

liverfoundation.org/abouttheliver/info/

Sherman, M. 2005. Approaches to the Diagnosis of

Hepatocellular Carcinoma. In Current

Gastroenterology Reports, vol. 7, no. 1, pp.11-18.

Sujana, H., Swarnamani, S., 1996. Application of

Artificial Neural Networks for the classification of

liver lesions by texture parameters. In Ultrasound in

Medicine & Biology, vol. 22, no. 9, pp.1177 –1181.

Yoshida, H., Casalino, D., 2003. Wavelet packet based

texture analysis for diferentiation between benign and

malignant liver tumors in ultrasound images. In

Physics in Medicine & Biology, vol. 48, pp. 3735-

3753.

Madabhushi, A., 2005. Automated Detection of Prostatic

Adenocarcinoma from High-Resolution Ex Vivo MRI.

In IEEE Transactions on Medical Imaging, pp.1611-

1626.

Chikui, T., 2005. Sonographic texture characterization of

salivary gland tumors by fractal analysis, In

Ultrasound in Medicine & Biology, vol. 31, no. 10,

pp.1297- 1304.

Duda, D., Kretowski, M., 2013. Computer-Aided

Diagnosis of Liver Tumors Based on Multi-Image

Texture Analysis of Contrast-Enhanced CT. Selection

of the Most Appropriate Texture Features. In Studies

in Logic, Grammar and Rhetoric, vol. 35, no. 1, pp.

49–70.

Khuzir, A., Besar, R., 2009. Identification of masses in

digital mammogram using gray level co-occurrence

matrices. In Biomedical Imaging Intervention Journal,

vol. 5, no. 3, pp. e17-e30.

Masood, K., 2006. Co-occurrence and morphological

analysis for colon tissue biopsy classification. In

Proceedings of the 4th International Workshop on

Information Technology, 2006, Pakistan, pp. 211-216.

Meyer-Base, A., 2009. Pattern recognition for medical

imaging, Elsevier, San Diego, California.

Mitrea, D., Mitrea, P., Nedevcshi, S., Lupsor, M.,

Socaciu, M, Golea, A., Hagiu, C., Ciobanu, L., 2012.

Abdominal tumor characterization and recognition

using superior order co-occurrence matrices, based on

ultrasound images. In Computational and

Mathematical Methods in Medicine, 17 pages.

Mitrea, D., Nedevcshi, S., Socaciu, M., Badea, R., 2012.

The Role of the Superior Order GLCM in the

Characterization and Recognition of the Liver Tumors

from Ultrasound Images. In Radioengineering, vol.

21, no. 1, pp. 79-85.

Mitrea, D., Nedevschi, S., Badea, R., 2014. The Role of

the Textural Microstructure Co-occurrence Matrices

in the Classification of the Abdominal Tumors. The

10th IEEE ICCP 2014 Conference, Technical

University of Cluj-Napoca , pp. 187-190.

Mitrea, D., Nedevschi, S., Abrudean, M., Badea, R., 2015.

Colorectal cancer recognition from ultrasound images,

using complex textural microstructure co-occurrence

matrices, based on Laws’ features, In proccedings of

the 38

IEEE International Conference on

Telecommunications and Signal Processing, 2015

(TSP 2015), Praga, pp. 458-462.

Mitrea, D., Nedevschi, S, Badea, R., 2016. The Role of the

Complex Extended Textural Microstructure Co-

occurrence Matrix in the Unsupervised Detection of

the HCC Evolution Phases, based on Ultrasound

Images. In Proceedings of International Conference on

Pattern Recognition Applications and Methods

(ICPRAM 2016), Roma, pp. 698-705.

Davis, L.S., Clearman, M., Aggarwal, J., 1979. A

Comparative Texture Classification Study Based on

Generalized Cooccurence Matrix. Proceedings of the

18-th IEEE Conference on Decision and Control, pp.

71-78.

Sujatha, B., Kumar, V., Harini, P., 2012. A new logical

compact LBP Co-Occurrence Matrix for Texture

Analysis. In International Journal of Scientific &

Engineering Research, vol. 3, no. 2, pp. 1-5, 2012.

Sujatha, B., Sekhar Reddi, C., 2013. Texture classification

using texton co-occurrence matrix derived from

texture orientation. In International J. Soft Computing

and Eng, vol. 2, no. 6, pp. 18-23.

Laws, K.I., 1980. Rapid texture identification. In SPIE

Vol. 238 Image Processing for Missile Guidance,

pp.76-380.

Nanni, L., Brahnam, S., Ghidoni, S., Menegatti, E.,

Barrier, T., 2013. “Different Approaches for

Extracting Information from the Co-Occurrence

Matrix”. In PloS One Journal, Vol.8, No. 12, 2013, pp.

1-9.

Kirsch, R., 1971. Computer determination of the

constituent structure of biological images. In

Computers and Biomedical Research, Vol.4, No.3, pp.

315–328.

Hall, M, 2003. Benchmarking attribute selection

techniques for discrete class data mining. In IEEE

Trans On Knowledge and Data Eng., vol. 15, no. 3,

pp. 1-16.

Weka3, Data Mining Software in Java, 2017. Online:

http://www.cs.waikato.ac.nz/ml/weka.

Duda, R., 2003 Pattern Classification, Wiley Interscience,

Hoboken, New Jersey.

Automatic Recognition of the Hepatocellular Carcinoma from Ultrasound Images using Complex Textural Microstructure Co-Occurrence

Matrices (CTMCM)

189