Breast Cancer Classiﬁcation by Artiﬁcial Immune Algorithm based

Validity Interval Cells Selection

Rima Daoudi and Khalifa Djemal

IBISC Laboratory, Evry Val dEssonne University, 91020 Evry, France

Keywords:

Breast Cancer, Classiﬁcation, Artiﬁcial Immune System, Validity Interval.

Abstract:

We present in this work an Artiﬁcial Immune System (AIS) algorithm for breast cancer classiﬁcation and diag-

nosis. The main contribution is to select memory cells according to their belonging to a validity interval based

on average similarity of training cells. The behaviour of these created memory cells preserves the diversity

of original cancer learning class. All these operations allow to generate a set of memory cells with a global

representativeness of the database which enables breast cancer classiﬁcation and recognition. Promising re-

sults have been obtained on both Wisconsin Diagnosis Breast Cancer Database (WDBC) and (DDSM) Digital

Database for Screening Mammography.

1 INTRODUCTION

Cancer is a disease in which cells become abnormal

and replicate forming more cells in an uncontrolled

way. Breast cancer begins in the tissues that makes

the breast, cancer cells may form a mass called a tu-

mor (Arora et al., 2008). Tumor can be benign or

malignant. The correct diagnosis of breast cancer has

become a major problem in the medical ﬁeld since

this famous type of cancer stills threatening the life

of most women. Indeed, approximately one in ten

women is affected by this disease in her lifetime. Al-

though some risk reduction can be achieved through

prevention, strategies in this direction can not allow

the elimination of the majority of breast cancers that

occur in low-income and middle-income countries.

Early detection remains the main way to ﬁght against

the disease, it improvesthe chances of survivalas well

as breast cancer outcome (Daoudi et al., 2015).

There is no doubt that the evaluation and decision-

making processes carried out by the experts are very

important factors. However, intelligent classiﬁcation

algorithms also help physicians, particularly by mini-

mizing errors of not experienced practitioners.

In this context, computerized diagnostic systems

are increasingly used to directly reach the ﬁnal di-

agnosis using artiﬁcial intelligence algorithms that

perform the role of classiﬁers such as Neural Net-

works (Hagan et al., 1996),(Marcano-Cede˜no et al.,

2011),(Neveset al., 2015), Genetic Algorithms (Yang

et al., 2013) or Support Vector Machines (Zemmal

et al., 2016),(Torrents-Barrena et al., 2015). In the

middle of the 90s, a new artiﬁcial intelligence ap-

proach inspired by immunology, has emerged, called

Artiﬁcial Immune Systems (AIS). Several concepts of

the immune response were extracted and applied as a

solution to real-world science and engineering prob-

lems (De Castro and Von Zuben, 2000).

Indeed, the Artiﬁcial Immune System (AIS) is

a distributed system that can perform classiﬁcation,

recognition and learning tasks using extraction, com-

munication and memorization processes. Efforts

made in this research focus have contributed to the

emergence of several algorithms that can be classi-

ﬁed into three large families as the natural mechanism

responsible for the implementation: negative selec-

tion (Dasgupta and Majumdar, 2002), clonal selec-

tion (De Castro and Von Zuben, 2000) and artiﬁcial

immune network (de Castro and Von Zuben, 2001)

algorithms.

The immune network theory was proposed by

Jerne in 1974 (Jerne, 1974). The hypothesis was that

the immune system maintains a network of intercon-

nected idiotypic antibodies to recognize an antigen.

Negative selection is a mechanism that protects the

body against the autoreactivelymphocytes. It uses the

ability of the immune system to detect unknown anti-

gens without affecting the self cells (Somayaji et al.,

1998). The clonal selection theory has been proposed

by Burnet in 1959 (Burnet et al., 1959). It explains

how an immune response is mounted as a model of

non-self antigen is recognized by the system. There

Daoudi, R. and Djemal, K.

Breast Cancer Classiﬁcation by Artiﬁcial Immune Algorithm based Validity Interval Cells Selection.

DOI: 10.5220/0006057202090216

In Proceedings of the 8th International Joint Conference on Computational Intelligence (IJCCI 2016) - Volume 1: ECTA, pages 209-216

ISBN: 978-989-758-201-1

209

are two processes: the recognition of the shape of the

antigen and selecting the antibody speciﬁc to it. The

idea is that only antibodies capable of recognizing the

antigen are activated for proliferation (cloning + mu-

tation).

Since 1959, there have been improvements of the

Burnet’s theory, especially on the way the antigens

are recognized. But the basic principles of clonal se-

lection and afﬁnity maturation by hypermutation are

sufﬁcient for the purposes of artiﬁcial clonal selection

algorithms.

In 2002, De Castro and Von Zuben (De Castro and

Von Zuben, 2000) proposed a clonal selection based

algorithm named CLONALG. The principle of this al-

gorithm is to build an initial memory cells population

and expose them to the antigens (training examples)

for a number generations to develop a population of

more speciﬁc memory cells to these antigens through

the cloning and mutation processes.

The main learning steps of CLONALG algorithm

are:

• Generation of initial memory cells by a random

selection of training examples.

• Memory cells evaluation and selection of the most

representative of the antigen.

• Cloning, mutation and re-selection of the best mu-

tated clone.

• Maintaining of diversity by the rejection of the

less good memory cells and their replacement by

randomly generated ones.

Compared to other clonal selection algorithms,

CLONALG has low complexity and requires fewer

parameters that can inﬂuence the classiﬁcation accu-

racy (Brownlee, 2005),(Zhang, 2011). It has been

successfully applied to solve various complex prob-

lems, and offers a promising precision in the ﬁeld of

pattern recognition.

Several studies have been published to improve

the potentially negative features in the learning of

CLONALG algorithm, including the exploration of

the information contained in the population of mu-

tated clones(White and Garrett, 2003), (Tasnim et al.,

2014), or the generalization of memory cells(Sharma

and Sharma, 2011). In (Daoudi et al., 2014), the work

is to improve two limitations observed on CLON-

ALG algorithm, the ﬁrst in the way it is initialized,

and the second in the selection of memory cells to be

cloned to avoid rejecting cells and their replacement

by randomly generated ones. The rejection step of the

less competent memory cells in CLONALG is used

to maintain diversity in the algorithm, but no check

is made that the added random cells are better than

those which are rejected. The improvements consist

Figure 1: Cells Clonal Selection Artiﬁcial Immune System

(CCS-AIS) Diagram (Daoudi et al., 2014).

in maintaining diversity by creating initial antibodies

(memory cells) and new memory cells from averaged

local subgroups. Diagram of CCS-AIS approach pro-

posed in (Daoudi et al., 2014) is given in ﬁgure 1. The

results signiﬁcantly improves the accuracy of CLON-

ALG. Nevertheless, we found that averaging cells op-

erations do not adequately represent the true diversity

of all examples of the class to learn. Indeed, even if

the initial memory cells (antibodies) are not randomly

selected, they do not often guarantee the overallrepre-

sentativeness of the class. In this work, we propose a

method to validate these cells by using cells selection

validity interval.

The rest of the paper is composed as follows: in

the next section we present the databases that we used

to evaluate our approach. Section 3 details the differ-

ent steps of the proposed Validity Interval Cells Se-

lection Artiﬁcial Immune System (VICS-AIS) algo-

rithm. The results of application and the work’s con-

clusion are given in sections 4 and 5 respectively.

2 USED DATABASES

To evaluate our proposed algorithm, we chose to ap-

ply it on two different databases the most used in the

ﬁeld of diagnosis of breast cancer: Wisconsin Diag-

nosis Breast Cancer Database (WDBC) and Digital

Database for Screening Mammography (DDSM).

2.1 Wisconsin Diagnosis Breast Cancer

Database (WDBC)

To evaluate our work, we chose to apply it on the Wis-

consin Diagnosis Breast Cancer Database (WDBC). It

was supported by Dr. William H Wolberg et al. (Wol-

berg and Mangasarian, 1990). WDBC consists of data

from 569 breast ﬁne needle aspirate (FNA) cases con-

ECTA 2016 - 8th International Conference on Evolutionary Computation Theory and Applications

210

taining 32 descriptive features, where the two ﬁrst

features correspond to a unique identiﬁcation number

and the diagnosis status (benign or malignant). The

rest 30 features are computed from a digitized image

of a FNA of a breast mass by obtaining a small drop of

ﬂuid from a breast tumor using a ﬁne needle. With an

interactiveinterface, active contour models are initial-

ized near the boundariesof a set of different cells. The

customized snakes are deformed to the exact shape of

cells, ten features are computed for each cell and the

mean value, largest value and standard error of each

feature are computed for each image. The case dis-

tribution includes 357 cases of benign breast changes

and 212 cases of malignant breast cancer. The de-

scriptive features are recorded with four signiﬁcant

digits including:

1. Radius; 2. Texture; 3. Perimeter; 4. Area;

5. Smoothness; 6. Compactness; 7. Concavity; 8.

Concave points; 9. Symmetry; 10. FracDim = fractal

dimension The features are recorded with four signif-

icant digits, and since they are measured in different

scales, the error function will be dominated by large-

scale variables. Thus, to eliminate the effect of differ-

ent levels, standardization is needed before learning.

In our work the WDBC database is normalized in

the range [0, 1] according to the following equation:

− x

min

max

− x

min

(1)

Where x

min

is the minimum of the data X for all

i, x

max

is the maximum of the data X for all i, x

the original i

data of data X, and x

the normalized

feature value.

2.2 Digital Database for Screening

Mammography (DDSM)

The digital database for screening mammography

(DDSM) was assembled by a group of researchers

from the University of South Florida and was com-

pleted in 1991(Heath and Bowyer, 2000). It com-

prises 2620 cases collected from the hospital, ”Mas-

sachusetts General Hospital” (MGH), the Univer-

sity ”Wake Forest University” (WFU) and the hospi-

tal ”Washington University of St. Louis School of

Medicine” (WUSTL). DDSM has been widely used

by the scientiﬁc community in the ﬁeld of breast

cancer diagnosis; it has the advantage of using the

same standardized lexicon by the American Col-

lege of Radiology (ACR) in the BI-RADS (Breast

Imaging-Reporting And Data System). Different pa-

tient records were made as part of breast cancer

screening and were classiﬁed into three cases: normal

cases (no lesions) benign, and malignant cases. Each

ﬁle is composed of four views that contain the ex-

ternal oblique (MLO) and craniocaudal (CC) of each

breast. These ﬁles are also provided with data anno-

tations by radiologists. These annotations are used

to describe the various lesions present in the images

such as the number and type of anomalies (microcal-

ciﬁcations / masses), the biopsy result (Benin / ma-

lignant), the location of lesions, etc. In this work, we

only deal with the case of masses (not microcalciﬁca-

tions). Sub-database of DDSM was created consisting

of 242 masses: 128 benign and 114 malignant. These

examples will be partitioned (in the same way that the

WDBC database) into training and test examples.

The description of breast masses is a very im-

portant step, three new descriptors: the Skeleton

End Point (SEP), Protuberance selection (PS) and the

Spiculated Mass Descriptor (SMD) were proposed

in ((Sellami-Masmoudi et al., 2009), (Cheikhrouhou

et al., 2011) and (Kachouri et al., 2012)) respectively,

which were compared to 19 other features proposed

in the literature ((Daoudi et al., 2014)). In this work,

all of the 22 features are used to evaluate the perfor-

mance of the proposed VICS-AIS classiﬁer.

3 PROPOSED AIS ALGORITHM

BASED VALIDITY INTERVAL

CELLS SELECTION (VICS-AIS)

We propose in this work a method for validating the

memory cells by using a validity interval for improv-

ing breast cancer recognition. This validity interval

is based on the standard deviation of average simi-

larities of all training cells, more particularly of the

relevant class. As we mentioned in section 1, the cre-

ated memory cells in (Daoudi et al., 2014) which are

used in classiﬁcation do not often ensure a good rep-

resentation of all original cells of the different training

classes (benign and malignant). To improve the rep-

resentativeness of theses training cells, we propose to

determine a validity interval allowing the selection of

more efﬁcient memory cells. By using a validity inter-

val we garantee a better deversity for the set of mem-

ory cells, and we have less identical cells that may

slow down the training process. This proposed solu-

tion is composed of three main steps. The ﬁrst one

consist in determining the validity interval of selec-

tion, the second step concerns the selection of initial

memory cells using validity interval presented in the

ﬁrst step. The last step is the global AIS training sys-

tem.

Breast Cancer Classiﬁcation by Artiﬁcial Immune Algorithm based Validity Interval Cells Selection

211

0 100 200 300

0.1

0.2

0.3

0.4

0.5

(a)

Similarities

OAC(Benign)

0 50 100 150 200

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

(b)

Similarities

OAC (Malignant)

Figure 2: Histogram of similarities between the overall av-

erage cell (OAC) and training examples of Benign class (a)

and Malignant class (b).

3.1 Validity Interval of Selection

From training cells, we calculate an overall average

cell (OAC), then we determine its similarity (Sim)

with all the examples of the training class. We calcu-

late the statistical characteristics: the mean, variance

and standard deviation.

Similarity is the measure of match between the

antigen (training example) and a memory cell. In our

work, and as the attribute values of the used databases

are real, it is determined using the Euclidean distance

(ED) calculated by:

ED(OAC, Training

Example) =

∑

i=1

− y

)

(2)

with OAC = x

, .., x

and Training

Example =

.., y

, and n is the database dimension.

Sim(OAC, Training

Example) = 1 − ED (3)

The average similarity of the class (Sim

moy) is

the average of similarities of OAC with all instances

of the class. Figure 2 shows an example of the sim-

ilarities between the overall average cell (OAC) and

training examples of each class (left: Benign,right:

Malignant) of WDBC database. The ﬁgure illustrates

the diversity of training examples of each class, which

we must select the right memory cells representative

of the same diversity.

Subsequently, we calculate the standard deviation

σ of the similarities of each class by the equations:

σ =

∑

i=1

− ¯x)

(4)

With ¯x = The average similarity of the class

(Sim

moy), and M: the total number of training data.

The Validity Interval (VI) of each training class is

determined by:

VI = Sim

moy± σ = [Sim moy − σ, Sim moy+ σ]

(5)

This interval will be used to validate the selected

clones to reach all ﬁnal memory cells pool.

3.2 Initial Memory Cells Selection using

Validity Interval (VI)

In this step, the creation of the initial memory cells for

each training class is performed in the same manner

as in (Daoudi et al., 2014), i.e. from averaged local

subgroups of training examples. The aim is to create

initial cells representing all the data to learn, instead

of randomly selecting cells not necessarily represen-

tative of the class.

After creating the initial memory cells

(MC

, .., MC

), we compute the average simi-

larity of each one with all the Training examples

(TE

, .., TE

with N: the number of initial memory cells

(NB

Sim

moy(MC

) =

∑

j=1

Sim(MC

, TE

) (6)

• If Sim

moy(MC

) ∈ VI ; MC

is kept in the initial

memory cells set.

• Otherwise, it will not be considered.

3.3 AIS Training System

In this step, the training of the artiﬁcial immune sys-

tem is made. Average cells are created and added to

the set of the ﬁnal memory cell as has been described

in (Daoudi et al., 2014), but with an additional condi-

tion.

Indeed, the memory cells (medium or mutated

clones) will be added to the memory cells only if their

average similarities are in the validation interval (VI)

of the relevant class.

• If Sim moy(Cell) ∈ VI ; add Cell to ﬁnal memory

cells. ( Cell= Average cell or mutated clone).

• Otherwise, reject Cell.

In this way we enable the generation of memory

cells with an overall representativeness of each class

to learn. General diagram of VICS-AIS proposed al-

gorithm is given in Figure 3

At the end of a deﬁned number of iterations (ﬁxed

by the user), we dispose of a set of global memory

cells for each training class. These cells will be used

in the test phase to determine the class of each sam-

ple to be classiﬁed. The application results of our ap-

proach for breast cancer diagnosis are presented in the

following section.

ECTA 2016 - 8th International Conference on Evolutionary Computation Theory and Applications

212

Figure 3: Diagram of VICS-AIS composed of Step I: Validity interval Creation for selection (VI), Step II: initial memory

cells selection using (VI) and Step III: AIS training system.

4 EXPERIMENTAL RESULTS

The overall results of this research are given in this

section. First we present the parameters used in eval-

uation, then all achieved tests are given and discussed

in subsequent sections.

4.1 Algorithm Parameters

We present in this section all of the parameters used

in our evaluations of the VICS-AIS approach on both

WDBC and DDSM databases.

4.1.1 Number of Initial Memory Cells

The initial memory cells population must be con-

sistent enough to properly represent the learning

database. At the same time, the number of these cells

should not be very great because it affects the speed of

the algorithm. The number of initial memory cells of

each class (NB

) which is also the number of local

sub-groups is randomly determined by the equation:

= round(rand(

)) (7)

With X the total number of training examples of a

given class.

4.1.2 Number of Clones

The number of clones of each memory cell is calcu-

lated proportionally to its similarity value by the fol-

lowing equation:

Clones

= round(β ∗ Sim) (8)

Where β is a cloning factor ﬁxed by user. In our

work, β was set to 5 experimentally.

4.1.3 Range of Mutation

The mutation interval of each clone is inversely pro-

portional to its similarity value. I.e, the larger similar-

ity value is, the less the interval of change is wide.

Thus, the mutation value is randomly selected be-

tween [Sim-1,1-Sim]. Table 1 summarizes all the pa-

rameters used in the evaluation of the proposed VICS-

AIS approach:

Table 1: Used parameters in evaluation.

Parameter Value

Similarity 1- Euclidean distance

Clones

β∗ Sim

β (Cloning factor) 5

Mutation Rand(sim− 1,1− sim)

rand(

)

k (Local subgroup size)

Breast Cancer Classiﬁcation by Artiﬁcial Immune Algorithm based Validity Interval Cells Selection

213

After presenting the parameters used for the eval-

uation of the approach that we proposed to improve

the clonal selection, we present in the next section

the application results of VICS-AIS (Validity Inter-

val Clonal Selection Artiﬁcial Immune System). The

algorithm VICS-AIS aims to enhance the diversity of

the CCS-AIS algorithm (Daoudi et al., 2014). Indeed,

to improve the overall representativeness of the train-

ing data and preserve good diversity in the algorithm,

we proposed to use a validity interval for each class

to learn based on the statistical characteristics of the

latter.

4.2 Classiﬁcation Results

Based on the parameters set in the previous section,

this work uses four-fold cross validation to evaluate

the performance of the proposed approach. We shared

our databases into four equal parts, and used three

parts, for training and one for testing at each evalu-

ation. After 1,2 5 and 10 generations (iterations) of

the algorithm, the memory cells generated at the end

of training are used in evaluation. The average of 10

successive runs is taken as the end result of an eval-

uation. Tables 2 and 3 present the classiﬁcation re-

sults obtained by the VICS-AIS method on WDBC

and DDSM databases are respectively, for the differ-

ent number of generations.

Table 2: Classiﬁcation results on WDBC database.

Generations (%) Train ±σ (%) Test ±σ

1 96.96 ±0.85 94.18 ±1.32

2 97.80 ±0.30 94.75 ±1.05

5 98.55 ±0.32 95.48 ±0.90

10 98.95 ±0.16 97.58 ±0.22

Table 3: Classiﬁcation results on DDSM database.

Generations (%) Train ±σ (%) Test ±σ

1 94.55 ±1.40 94.64 ±0.87

2 94.70 ±1.60 94.91 ±1.12

5 95.54 ±0.56 95.18 ±0.80

10 96.66 ±0.20 95.76 ±0.39

On WDBC database, proposed VICS-AIS algo-

rithm obtained 97.58% classiﬁcation accuracy after

10 generations, compared to CCS-AIS which pro-

vided 96.80%. On DDSM database, the ﬁnal classiﬁ-

cation result after 10 generations of VICS-AIS algo-

rithm was 95.76%, while CCS-AIS algorithm reached

94.98%. We also noticed a rapid learning of the

VICS-AIS algorithm with respect to CCS-AIS, and

this is thanks to the good selection of appropriate

memory cells, and elimination of redundant or iden-

tical cells. So we can say that the proposed approach

has contributed to the improvement of CCS-AIS al-

gorithm.

Since the VICS-AIS algorithm is an improvement

CCS-AIS algorithm, we present in Figure 4 a compar-

ison between the two approaches in terms of average

similarities of both Benign and Malignant classes of

WDBC database (because it is more consistent than

the base DDSM). The average similarities of ﬁnal

memory cells obtained by the two approaches are cal-

culated and compared with the average similarity of

the original cells of the database. we also present a

comparison between these values and between the va-

lidity intervals of each approach in table 4.

The validity intervals of ﬁnal memory cells ob-

tained by CCS-AIS algorithm on both benign and ma-

lignant classes are narrower than those of the mem-

ory cells obtained by our approach (table4). From

ﬁgure 4, we can observe that average similarities of

ﬁnal memory cells of VICS-AIS are nearest to the

average similarities of training WDBC database for

benign and malignant classes, unlike average simi-

larities of CCS-AIS memory cells which are smaller.

From these obtained results, we can say that the pro-

posed VICS-AIS method is effective. Indeed, the

introduction of validity interval in the algorithm al-

lowed a global representation of the diversity of train-

ing data, which is important to avoid local minima and

ensures a more accurate classiﬁcation.

4.3 Comparative Study

In this section, we provide a comparative study be-

tween the approach we proposed, and some clonal se-

lection algorithms of the literature (including CCS-

AIS) that we have implemented. We applied each al-

gorithm on the two databases (WDBC and DDSM)

using the same parameters listed in table 1. The re-

sults of each application for 10 generations are listed

in table 5.

By comparing the results, it is easily noticeable

that the VICS-AIS algorithm has the best classiﬁca-

tion rate on both WDBC and DDSM databases. An-

other important point which illustrates the effective-

ness of improvements in CCS-AIS algorithm is the

standard deviation (σ). Indeed, the VICS-AIS algo-

rithm has smaller values of σ, which means that it

is the most accurate approach. It was reduced by

0.3% on WDBC database and 0.39 % on the DDSM

database. Learning the 10 generations of VICS-AIS

is also the fastest among the approaches listed in Ta-

ble 5, thanks to the good selection of memory cells,

and elimination of repetitive cells due to the cloning

and the mutation operators.

ECTA 2016 - 8th International Conference on Evolutionary Computation Theory and Applications

214

Figure 4: Comparison between average similarities values of WDBC database and ﬁnal memory cells of CCS-AIS algorithm

and our approach.

Table 4: Comparison of Validation Intervals and average similarities values of WDBC database with ﬁnal memory cells of

CCs-AIS algorithm and our approach.

(%) VI (B) VI (M) Sim moy (B) Sim moy (M)

WDBC [0.08,0.19] [0.08,0.20] 0.13 0.14

CCS-AIS(Daoudi et al., 2014) [0.06,0.12] [0.06,0.13] 0.09 0.09

VICS-AIS [0.08,0.12] [0.08,0.14] 0.10 0.11

Table 5: Comparison of classiﬁcation results.

Algorithm (%) WDBC (%) DDSM

CLONALG (De Castro and Von Zuben, 2002) 89.86± 4.45 85.15± 6.46

AIRS (Watkins et al., 2004) 90.30± 1.36 89.15± 1.88

CLONAX (Sharma and Sharma, 2011) 93.40± 2.23 92.25± 1.47

MF-AIS (Daoudi et al., 2013) 95.03± 0.50 94.91± 0.61

CCS-AIS (Daoudi et al., 2014) 96.80± 0.52 94.98± 0.78

VICS-AIS (This Study) 97.58± 0.22 95.76± 0.39

5 CONCLUSION

The objective of this work is to provide help to the

experts for a second opinion for breast cancer diagno-

sis. In this aim, we presented in this paper an artiﬁcial

immune algorithm based validity interval for memory

cells selection (VICS-AIS). The proposed approach

selects the memory cells according to their belonging

to a validity interval based on average similarity to en-

sure global representation of the diversity of the data

to learn. The main contributions that we brought and

that justify the improved results are :

• Improvement of initialization, by creating speciﬁc

initial memory cells for each learning class, and

selection of the most representative among these

cells.

• Creation of potential memory cell for cloning and

mutation.

• Use of a validity interval of selection to improve

the overall representativeness and preserve good

diversity of learning data.

The obtained results on WDBC and DDSM

databases show a great performance of the classiﬁer.

Based on this work, we can say that the introduction

of validity intervals in the training of AIS algorithms

is effective to properly represent the true diversity of

the training data, and guarantee good quality memory

cells.

REFERENCES

Arora, N., Martins, D., Ruggerio, D., Tousimis, E., Swistel,

A. J., Osborne, M. P., and Simmons, R. M. (2008).

Breast Cancer Classiﬁcation by Artiﬁcial Immune Algorithm based Validity Interval Cells Selection

215

Effectiveness of a noninvasive digital infrared thermal

imaging system in the detection of breast cancer. The

American Journal of Surgery, 196(4):523–526.

Brownlee, J. (2005). Clonal selection theory & clonalg-the

clonal selection classiﬁcation algorithm (csca). Swin-

burne University of Technology.

Burnet, S. F. M. et al. (1959). The clonal selection theory of

acquired immunity. University Press Cambridge.

Cheikhrouhou, I., Djemal, K., and Maaref, H. (2011). Pro-

tuberance selection descriptor for breast cancer diag-

nosis. In Visual Information Processing (EUVIP),

2011 3rd European Workshop on, pages 280–285.

IEEE.

Daoudi, R., Djemal, K., and Benyettou, A. (2013). An

immune-inspired approach for breast cancer classi-

ﬁcation. In International Conference on Engineer-

ing Applications of Neural Networks, pages 273–281.

Springer Berlin Heidelberg.

Daoudi, R., Djemal, K., and Benyettou, A. (2014). Dig-

ital database for screening mammography classiﬁ-

cation using improved artiﬁcial immune system ap-

proaches. In 6th International Conference on Evolu-

tionary Computation Theory and Applications (ECTA

2014) Part of the 6th International Joint Conference

on Computational Intelligence (IJCCI 2014), pages

244–250.

Daoudi, R., Djemal, K., and Benyettou, A. (2015). Improv-

ing cells recognition by local database categorization

in artiﬁcial immune system algorithm. application to

breast cancer diagnosis. In Evolving and Adaptive

Intelligent Systems (EAIS), 2015 IEEE International

Conference on, pages 1–6. IEEE.

Dasgupta, D. and Majumdar, N. S. (2002). Anomaly detec-

tion in multidimensional data using negative selection

algorithm. In wcci, pages 1039–1044. IEEE.

De Castro, L. N. and Von Zuben, F. J. (2000). The clonal

selection algorithm with engineering applications. In

Proceedings of GECCO, volume 2000, pages 36–39.

de Castro, L. N. and Von Zuben, F. J. (2001). ainet: an arti-

ﬁcial immune network for data analysis. Data mining:

a heuristic approach, 1:231–259.

De Castro, L. N. and Von Zuben, F. J. (2002). Learn-

ing and optimization using the clonal selection prin-

ciple. Evolutionary Computation, IEEE Transactions

on, 6(3):239–251.

Hagan, M. T., Demuth, H. B., Beale, M. H., and De Jes´us,

O. (1996). Neural network design, volume 20. PWS

publishing company Boston.

Heath, M. D. and Bowyer, K. W. (2000). Mass detection

by relative image intensity. In Proceedings of the

5th International Workshop on Digital Mammography

(IWDM-2000), pages 219–225.

Jerne, N. K. (1974). Towards a network theory of the im-

mune system. In Annales d’immunologie, volume

125, pages 373–389.

Kachouri, I. C., Djemal, K., and Maaref, H. (2012). Charac-

terisation of mammographic masses using a new spic-

ulated mass descriptor in computer aided diagnosis

systems. International Journal of Signal and Imag-

ing Systems Engineering, 5(2):132–142.

Marcano-Cede˜no, A., Quintanilla-Dom´ınguez, J., and An-

dina, D. (2011). Wbcd breast cancer database clas-

siﬁcation applying artiﬁcial metaplasticity neural net-

work. Expert Systems with Applications, 38(8):9573–

9579.

Neves, J., Guimar˜aes, T., Gomes, S., Vicente, H., Santos,

M., Neves, J., Machado, J., and Novais, P. (2015).

Logic programming and artiﬁcial neural networks in

breast cancer detection. In Advances in Computa-

tional Intelligence, pages 211–224. Springer.

Sellami-Masmoudi, D., Maaref, H., Cheikhrouhou, I., Dje-

mal, K., and Derbel, N. (2009). Empirical descriptors

evaluation for mass malignity recognition. In First

International Workshop on Medical Image Analysis

and Description for Diagnosis Systems (MIAD 2009),

pages 91–100.

Sharma, A. and Sharma, D. (2011). Clonal selection algo-

rithm for classiﬁcation. In Artiﬁcial Immune Systems:

10th International Conference, ICARIS 2011, Cam-

bridge, UK, July 18-21, 2011. Proceedings, volume

6825, page 361. Springer.

Somayaji, A., Hofmeyr, S., and Forrest, S. (1998). Princi-

ples of a computer immune system. In Proceedings of

the 1997 workshop on New security paradigms, pages

75–82. ACM.

Tasnim, M., Rouf, S., and Rahman, M. S. (2014). A

clonalg-based approach for the set covering problem.

Journal of Computers, 9(8):1787–1795.

Torrents-Barrena, J., Puig, D., Melendez, J., and Valls, A.

(2015). Computer-aided diagnosis of breast cancer via

gabor wavelet bank and binary-class svm in mammo-

graphic images. Journal of Experimental & Theoreti-

cal Artiﬁcial Intelligence, pages 1–17.

Watkins, A., Timmis, J., and Boggess, L. (2004). Artiﬁ-

cial immune recognition system (airs): An immune-

inspired supervised learning algorithm. Genetic Pro-

gramming and Evolvable Machines, 5(3):291–317.

White, J. and Garrett, S. (2003). Improved pattern recogni-

tion with artiﬁcial clonal selection? Artiﬁcial Immune

Systems, pages 181–193.

Wolberg, W. H. and Mangasarian, O. L. (1990). Multisur-

face method of pattern separation for medical diag-

nosis applied to breast cytology. Proceedings of the

national academy of sciences, 87(23):9193–9196.

Yang, C.-H., Lin, Y.-D., Chaung, L.-Y., and Chang, H.-

W. (2013). Evaluation of breast cancer suscepti-

bility using improved genetic algorithms to generate

genotype snp barcodes. IEEE/ACM Transactions on

Computational Biology and Bioinformatics (TCBB),

10(2):361–371.

Zemmal, N., Azizi, N., Dey, N., and Sellami, M. (2016).

Adaptive semi supervised support vector machine

semi supervised learning with features cooperation for

breast cancer classiﬁcation. Journal of Medical Imag-

ing and Health Informatics, 6(1):53–62.

Zhang, Y. (2011). Distributed Intrusion Detection System in

A Multi-Layer Network Architecture of Smart Grids.

PhD thesis, The University of Toledo.

ECTA 2016 - 8th International Conference on Evolutionary Computation Theory and Applications

216