B. Ashok Int. Journal of Engineering Research and Applications www.ijera.

com
ISSN: 2248-9622, Vol. 6, Issue 1, (Part - 1) January 2016, pp.94-99

RESEARCH ARTICLE OPEN ACCESS

Comparison of Feature selection methods for diagnosis of cervical

cancer using SVM classifier
B. Ashok*, Dr. P. Aruna**
*(Department of Computer Science & engineering, Annamalai University, Annamalainagar, India)
** (Department of Computer Science & engineering, Annamalai University, Annamalainagar, India)

ABSTRACT
Even though a great attention has been given on the cervical cancer diagnosis, it is a tuff task to observe the
pap smear slide through microscope. Image Processing and Machine learning techniques helps the pathologist
to take proper decision. In this paper, we presented the diagnosis method using cervical cell image which is
obtained by Pap smear test. Image segmentation performed by multi-thresholding method and texture and shape
features are extracted related to cervical cancer. Feature selection is achieved using Mutual Information(MI),
Sequential Forward Search (SFS), Sequential Floating Forward Search (SFFS) and Random Subset Feature
Selection(RSFS) methods.
Keywords - Segmentation, Feature Extraction, Feature Selection, Multi-thresholding, Classification

I.INTRODUCTION time diagnosis of the disease could improve patients’

Cervical cancer is the fourth common cancer in treatment effect. One of the most important and
women society. In the low-income countries, the indispensable tasks in any pattern recognition system
acceptability of increased cervical cancer risk and is to overcome the curse of dimensionality problem,
cause of cancer death projected. Cervical screening which forms a motivation for using a suitable feature
programs has minimized the death rate in the selection method [4]. Random based feature
developed countries. Human Papilloma Virus (HPV) selection[5] using benchmark datasets also discussed
is found as the major cause for the cervical cancer. and a frame work of algorithmic randomness based
Although cervical cancer is one of the deadliest feature selection has been proposed to measure the
disease, it can be cured if detected in the earlier stage. feature importance score using a p-value that derives
Pap smear test is the non-invasive screening and from the combination of algorithmic randomness test
simple test to find out cervical cancer. Image and machine learning methods.
acquisition is obtained by applying the smear which Rough Set Theory (RST) and Particle Swarm
is taken by the pap smear test on the microscope Optimization (PSO) methods applied for medical
slide. From microscope attached digital camera diagnosis and achieved accuracy to certain extent [7].
capture the cell image. Image processing is the major Canonical Correlation Analysis (CCA) for uncertain
area that includes image enhancement, segmentation data streams and an uncertain canonical correlation
and feature extraction. Feature selection and analysis (UCCA) method discussed to reduce the
classification are the machine learning techniques. dimensionality of the features [8]. Extracting shape
In this paper we have applied image processing, data features, textural features and intensity features [9]
mining and machine learning techniques to diagnose was achieved by watershed transformation and k-
the cervical cancer. means spectral clustering supervised svm classifiers
applied for feature selection. Multi-Filtration Feature
II.RELATED WORK Selection (MFFS) [10], a approach which was
Feature selection is an important process. The applied to extract features, feature subset selection,
feature selection methods can be divided into filter, feature ranking and classification.
wrapper and embedded methods [1,2]. Filter methods A cosine similarity measure support vector
are computationally very fast, easy to interpret and machine CSMSVM [11] for feature selection and
are scalable for high- dimensional datasets like micro classification presented. Feature selection is
array data. However, most of them are univariate. performed by adding weights to features based on
Wrapper and embedded methods which always use margin verses cosine distance ratio. An unsupervised
machine learning algorithms to compute the feature learning (UFL) framework for basal cell
importance scores of the feature or feature subset, carcinoma image analysis[12] had been performed on
RF-ACC and RF-Gini based on Random Forest(RF) histopathology images.
[3] always achieve better performances than filter The K-SVM had been explained [13]. Missing of
methods on high-dimensional datasets. Premature important data and gathering unnecessary

www.ijera.com 94|P a g e
B. Ashok Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 6, Issue 1, (Part - 1) January 2016, pp.94-99

information at the time of feature extraction should Image preprocessing includes noise removal and
be avoided. Multi resolution local binary pattern image enhancement. Image resize also performed.
(MRLBP) [14] variants based texture feature Cell image converted from RGB image to Gray scale
extraction techniques had been presented. Image image.
segmentation performed using discrete wavelet
transform (DWT) further Daubechies wavelet(db2) 3.1 Image segmentation
used as decomposition filter. Deep Sparse Support
Vector Machine (DSSVM) based feature selection Segmentation means split the input image into
method had been discussed [15] and further desired regions so as to get the features. There are lot
extraction of color and texture features was of methods to segment an image such as edge based
performed using super pixels. The importance of methods, region based, watershed method and
local texture(LTP) and local shape (MultiHPOG) thresholding methods. In this work, we used multi
features discussed [16] and the further improvement thresholding method to segment the input image.
in performance achieved by adding the global shape Nucleus and Cytoplasm of the cervical cell image are
feature (Gabor wavelets). Feature selection based on segmented.
information has the drawback such as the interaction RGB Gray scale Segmented Segmented
between the features and classifiers. This may lead to Image Image Nucleus Cytoplasm
the selection irrelevant features. To overcome this Image Image
problem, Mutual Information based techniques used
[17]. Even though new feature extraction, feature
selection and classification methods have been
introduced, there will be a requirement exists for
newer or better techniques. In our previous works
[18][19] various segmentation methods were
analyzed.

III.MATERIALS AND METHODS

The Block diagram of the proposed method

Input image
Preprocessing
Image Preprocessing

Segmentation using
multi-thresholding

Figure 2: Sample pap smear images & their

Feature Extraction Phase
corresponding segmented images
Textural Shape
3.2 Feature Extraction
Features Features 3.2.1 Texture features
Texture features are extracted using Gray Level
Co-occurrence Matrix (GLCM). The list of features
are as follows homogeneity, contrast, correlation,
Feature Selection Phase variance, inverse difference moment, sum average,
sum variance, sum entropy, entropy, difference
MI SFS SFFS RSFS
variance, difference entropy, information measure I,
information measure II and maximal correlation
coefficient. Totally 14 texture features are extracted.

Classification using SVM classifier 3.2.2 Shape features

The following list of features are the shape
Figure 1: Block Diagram of the proposed method for features which are extracted from nucleus and
diagnosis of cervical cancer cytoplasm of the cervical cell. Besides the shape
features, derived features are also included. Nucleus
related features are Area, Centroid, Eccentricity,

www.ijera.com 95|P a g e
B. Ashok Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 6, Issue 1, (Part - 1) January 2016, pp.94-99

Equiv Diameter, Major Axis Length, Minor Axis Random Subset Feature Selection methods are used
Length, Perimeter, roundness, position, brightness to select the features.
and nucleus cytoplasm ratio and cytoplasm related
features are Area, Centroid, Eccentricity, Equiv 3.3 Mutual Information (MI)
Diameter, Major Axis Length, Minor Axis Length, Basically Mutual information is the
Perimeter, brightness, roundness. Totally 30 shape measurement of similarity. The mutual
features are extracted. In this proposed work, by information of two random variables is a measure of
combining texture and shape features we extract the variable’s mutual dependence that is correlation
totally 44 features. between two variables. Increase in mutual
information which is often equivalent to minimizing
IV.Feature Selection conditional entropy. MI comes under filter method.
Feature selection is one of the optimization Generally filter methods provide ranking to the
process to select a minimum number of effective features. Choosing the selection point based on
features. Feature selection is required to reduce curse ranking is performed through cross-validation. In
of dimensionality due to irrelevant or overfitting, cost feature selection, MI is used to characterize the
of feature measurement and computational burden. In relevance and redundancy between features.
order to select a feature subset, a search method and The mutual information between two discreet
an objective function is required. Small sample size random variables X, Y jointly distributed according

 p( x, y) log
and what objective function to use are some of the to p(x, y) is given by
potential difficulties in feature selection process.
I(X; Y) = (1)
Number of features is directly proportional to the
x, y
feature subset that is as in the case of an increase in
features there may be an increase in feature subset = H(X) − H(X|Y) (2)
also.
There is a large number of search methods, which = H(Y) − H(Y|X) (3)
can be grouped in three categories. Exponential
algorithms, Sequential algorithms and Randomized = H(X) + H(Y) − H(X, Y) (4)
algorithms. Exponential algorithms evaluate a
number of subsets that grows exponentially with the Where H(X) and H(Y) are the individual entropy and
dimensionality of the search space. Exhaustive H(X|Y) and H(Y|X) are conditional entropy. H(X,Y)
Search, Branch and Bound, Beam Search comes Joint entropy. p(x,y) joint distribution and p(x)
under exponential algorithms. Sequential algorithms probability distribution
add or remove features sequentially, but have a In this work, all 44 features are ranked by the MI
tendency to become trapped in local minima. algorithm. The first top ranked 10 features are
Representative examples of sequential search include selected from the ranking list.
Sequential Forward Selection, Sequential Backward
Selection, Sequential Floating Selection and 3.4 Sequential Forward Selection (SFS)
Bidirectional Search. Randomized algorithms Sequential forward selection is one of the
incorporating randomness into their search procedure deterministic single-solution method. Sequential
to escape local minima. Examples are Random forward selection is the simplest and fastest greedy
Generation plus Sequential Selection, Simulated search algorithm start with a single feature and
Annealing and Genetic Algorithms. iteratively add features until the termination criterion
Objective function is divided into three types. One is is met. The following algorithm explain the concept
filter method, second one is wrapper method and of SFS
another is embedded method. Filter methods are
defined as independent evaluation of the
classification algorithm and the objective function
evaluates feature subsets by their information
content, typically interclass distance, mutual
information or information-theoretic measures. In
wrapper method, the objective function is a pattern
classifier, which evaluates feature subsets by their
predictive accuracy (recognition rate on test data) by
statistical re-sampling or cross validation. Embedded
methods are specific methods. In this paper filter
based Mutual Information (MI) and wrapper based
methods such as Sequential Forward Selection,
Sequential Floating Forward selection(SFFS) and

www.ijera.com 96|P a g e
B. Ashok Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 6, Issue 1, (Part - 1) January 2016, pp.94-99

subset is now better than it was before step 1.

Algorithm Continue to step 3.
Step 3: Continuation of conditional exclusion. Again
Input: the set of all features, F = find the least significant feature in feature subset. If
f1,f2,...,fn its exclusion will leave feature subset with at least 2
features, then remove it and repeat step 3. Whenever
Initialization: X0 =”null set”, k = 0 this condition finish, return to step 1.only 6 features
are selected by the sequential floating forward
Addition: add significant feature selection method and it is given in table 3.
Xj = argmax [J(Xk + Xj)], Xj ∈ F - Xk
where Xj is the new feature to be added Table 3 Selected Feature Numbers by SFFS
Xk+1 = Xk + X+k = k + 1 Serial number 1 2 3 4 5 6
Go to Addition Selected feature 5 4 6 29 3 9
Termination: stop when ‘k’ equals the number numbers
of desired features
Output : a subset of features, Xk 3.6 Random subset feature selection (RSFS)
Features are selected by classifying the data with
a classifier while applying randomly chosen subsets.
According to the classification performance the
In this work, 8 features are selected by the SFS relevance of each feature is computed. Based on the
method out of 44 features. The selected feature average usefulness each feature is evaluated in
subset is given in table 2. random subset feature selection. 12 features are
selected by the random subset feature selection
Table 2: Selected Feature Numbers by SFS method and it is given in table 4.
Serial
1 2 3 4 5 6 7 8
number Table 4 Selected Feature Numbers by RSFS
Selected 1 2 3 4 5 6 7 8 9 10 11 12
feature 3 5 4 6 29 18 44 20
3 5 28 20 9 44 12 33 14 24 34 18
numbers

3.5 Sequential Floating Forward Search (SFFS) V.Classification

The sequential floating forward search method is Support Vector Machine (SVM) is supervised
one of the most effective feature selection technique. learning method and one of the kernel- based
This algorithm starts with a null feature set. Then the techniques in machine learning algorithms. The
best feature in the feature set that satisfies some SVM is developed mainly by Vapnik at Bell
criterion function is added with the current feature laboratories. Basically SVM constructs a high-
set. This is one process of the sequential forward dimensional space and then separates according to
selection. Further any improvement of the criterion is the training data. In figure 3, the SVM classification
searched if some feature is excluded. By this way the process is illustrated. The rounded objects are support
worst feature according to the criterion is eliminated vectors.
from the feature set. This is one process of sequential
backward selection. In this method the algorithm
continues by increasing or decreasing the number of
features until the desired feature subset is obtained.
The word floating denotes the increase or decrease in
dimensionality of features in the selected feature
subset.

Algorithm
Step 1: Inclusion. Use the basic sequential feature Figure 3, SVM classifier
selection method to select the most significant feature
with respect to feature set and include it in the
VI.Results and Discussion
feature subset. Stop if desired features have been
In this work, 150 images of pap smear test are
selected, otherwise go to step 2.
collected from Rajah Muthiah Medical College,
Step 2: Conditional exclusion. Find the least
Annamalainagar, out of 150 images, 100 images are
significant feature k in feature subset. If it is the
used for training the SVM classifier and 50 images
feature just added, then keep it and return to step 1.
are used for testing.
Otherwise, exclude the feature k. Note that feature

www.ijera.com 97|P a g e
B. Ashok Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 6, Issue 1, (Part - 1) January 2016, pp.94-99

The combination of texture and shape features VII.CONCLUSION

are extracted from each images. It has 44 features. From the table 1 to table 4, it is found that
The optimal features are selected using feature Sequential floating forward selection method
selection methods like Mutual Information, outperforms the other method. That is the feature
Sequential forward selection, Sequential floating selected from Sequential floating forward selection
forward selection and Random subset feature method gives the most optimal features to diagnosis
selection. The cervical cancer affected images are the cervical cancer. Because accuracy 98.5%,
found using SVM classifier. Different feature sensitivity 98% and specificity 97.5% obtained from
selection methods are compared to find the best sequential floating forward selection method which is
feature selection method suited for the diagnosis of higher than other methods.
cervical cancer.
REFERENCES
1)Accuracy: Accuracy is obtained by correctly [1] Isabelle Guyon, AndreElisseeff, An
classified images divided by the total classified introduction to variable and feature
images. selection, J.Mach. Learn. Res. 3 (2003)
Accuracy = (6) 1157–1182.
[2] YvanSaeys, InakiInza, PedroLarranaga, A
Where TN is true negative,TP is true positive, FP is review of feature selection techniques in
false positive and FN is false negative. bioinformatics, Bioinformatics 23(19)
(2007) 2507–2517.
Table 5 Comparison of Accuracy of feature [3] L. Breiman, Random forests, Mach. Learn.
selection methods 45 (1 ) (2001) 5–32.
Slno Method Accuracy % [4] N. Rami, N. Khushaba, A. Al-Ani, A. Al-
1 MI 92 % Jumaily, Feature subset selection using
2 SFS 93 % differential evolution and a statistical repair
3 SFFS 98.5% mechanism, in: Expert Systems with
4 RSFS 94% Applications, Elsevier, 2011, pp. 11515–
11526.
2) Sensitivity: Sensitivity is obtained as correctly [5] Huazhen Wang, BingLv, FanYang ,
classified true positive rate divided by true positive KaiZheng, XuanLi, XueqinHu, Algorithmic
and false negative samples. Inconclusive results that randomness based feature selection for
are true positives are treated as errors for calculation. traditional Chinese chronic gastritis
Sensitivity = (7) diagnosis, Neuro computing, Elsevier 140
(2014) pp. 252–264.
[6] Maria Martinsa, Lino Costab, Anselmo
Table 6 Comparison of Sensitivity of feature Frizerac,Ramón Ceresd, Cristina Santos,
selection methods Hybridization between multi-objective
Slno Method Sensitivity % genetic algorithm and support vector
1 MI 89 % machine for feature selection in walker-
2 SFS 94.5 % assisted gait , computer methods and
3 SFFS 98% programs in biomedicine, Elsevier, 113
4 RSFS 91% (2014) pp.736–748
[7] H. Hannah Inbarania, Ahmad Taher Azarb,
3) Specificity: Specificity is calculated as correctly G. Jothi, Supervised hybrid feature selection
classified true negative rate divided by the true based on PSO and rough sets for medical
negative and false positive samples. Results that are diagnosis, computer methods and programs
true negatives are treated as errors. in biomedicine, Elsevier, 113 (2014)
pp.175–185
Specificity = (8)
[8] Wen-PingLi , JingYang , Jian-PeiZhang ,
Uncertain canonical correlation analysis for
Table 7 Comparison of Specificity of feature multi-view feature extraction from uncertain
selection methods data streams, Neuro computing, Elsevier,
Sl no Method Specificity % 149 (2015) pp.1337–1347
1 MI 91 % [9] Marina E. Plissiti, Christophoros Nikou,
2 SFS 92 % Antonia Charchanti, Combining shape,
3 SFFS 97.5% texture and intensity features for cell nuclei
4 RSFS 93 % extraction in Pap smear images, Pattern

www.ijera.com 98|P a g e
B. Ashok Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 6, Issue 1, (Part - 1) January 2016, pp.94-99

Recognition Letters, Elsevier 32 (2011) pp.

838–853
[10] S. Sasikala, S. Appavu alias Balamurugan,
S. Geetha, Multi Filtration Feature Selection
(MFFS) to improve discriminatory ability in
clinical data set, Applied Computing and
Informatics (2014)
[11] Gang Chen, JinChen, A novel wrapper
method for feature selection and its
applications, Neuro computing, Elsevier
159(2015) pp. 219–226
[12] John Arevalo, Angel Cruz-Roa, Viviana
Arias, Eduardo Romero, Fabio A. González,
An unsupervised feature learning framework
for basal cell carcinoma image analysis,
Artificial Intelligence in Medicine, Elsevier
64 (2015) pp. 131–145
[13] Bichen Zheng, Sang Won Yoon , Sarah S.
Lam, Breast cancer diagnosis based on
feature extraction using a hybrid of K-means
and support vector machine algorithms,
Expert Systems with Applications 41 (2014)
1476–1482
[14] Arvind R. Yadav, R.S. Anand, M.L. Dewal,
Sangeeta Gupta, Multi resolution local
binary pattern variants based texture feature
extraction techniques for efficient
classification of microscopic images of
hardwood species, Applied Soft Computing,
Elsevier 32 (2015) pp. 101–112
[15] YangCong, ShuaiWang , JiLiu , JunCao ,
YunshengYang , JieboLuo Deep sparse
feature selection for computer aided
endoscopy diagnosis, Pattern Recognition,
Elsevier, 48 (2015) pp. 907–917
[16] Fengyi Song, XiaoyangTan, XueLiu,
Songcan Chen Eyes closeness detection
from still images with multi-scale
histograms of principal oriented gradients,
Pattern Recognition 47 (2014) 2825–2838
[17] Mohamed Bennasar, YuliaHicks, Rossitza
Setchi, Feature selection using Joint Mutual
Information Maximisation, Expert Systems
With Applications, Elsevier 42 (2015)
8520–8532
[18] B.Ashok , P.Aruna, Prognosis of Cervical
Cancer using k-Nearest Neighbor
Classification, International Journal of
Applied Engineering Research ISSN 0973-
4562 Volume 10, Number 9 (2015) 6676-
6680
[19] B.Ashok, Dr.P.Aruna, Analysis of Image
Segmentation Methods for Detection of
Cervical Cancer, International Journal of
Applied Engineering Research, ISSN 0973-
4562 Vol. 10 No.75 (2015), 81-85

www.ijera.com 99|P a g e