An Enhanced Feature Selection Method to Predict the Severity in Brain Tumor

International Journal of Trend in Scientific Research and Development (IJTSRD)
Volume 3 Issue 5, August 2019 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470
@ IJTSRD | Unique Paper ID – IJTSRD26802 | Volume – 3 | Issue – 5 | July - August 2019 Page 1910
An Enhanced Feature Selection Method to
Predict the Severity in Brain Tumor
Parthiban J, Dr. B. Sathees Kumar
Research Scholar, Bishop Heber College, Trichy, Tamil Nadu, India
How to cite this paper: Parthiban J|Dr.B.
Sathees Kumar "An Enhanced Feature
Selection Method to Predict the Severity
in Brain Tumor"
Published in
International
Journal of Trend in
Scientific Research
and Development
(ijtsrd), ISSN: 2456-
6470, Volume-3 |
Issue-5, August 2019, pp.1910-1917,
https://doi.org/10.31142/ijtsrd26802
Copyright © 2019 by author(s) and
International Journal ofTrend inScientific
Research and Development Journal. This
is an Open Access article distributed
under the terms of
the Creative
Commons Attribution
License (CC BY 4.0)
(http://creativecommons.org/licenses/by
/4.0)
ABSTRACT
The level of severity of brain tumor is captured through MRI and then
assessed by the physician for their medical interpretation. The facts behind
the MRI images are then analyzed by the physician for further medicationand
follow-up activities. An MRI imagecomposedoflargevolumeoffeatures.Ithas
irrelevant, missing and information which is not certain. In medical data
analysis, an MRI image doesn’t express facts very clearly to the physician for
correct interpretation all the time. It also includes huge amount of redundant
information within it. A mathematical model known as rough-set theory has
been applied to resolve thisproblemby eliminating theredundancyinmedical
image data.
This paper uses a rough set method to find the severity level of the brain
tumor of the given MRI image. Rough set feature selection algorithms are
applied over the medical image data to select the prominent features. The
classification accuracy of the brain tumor can be improved to a better level by
using this rough set approach. The prominent features selected through this
approach deliver a set of decision rules for the classification task. A search
method based on the particle swarm optimizationis proposedinthis paperfor
minimizing the attribute set. This approach is compared with previously
existing rough set reduction algorithm for finding the accuracy. The reducts
originated from the proposed algorithm is more efficient and can generate
decision rules that will better classify the tumor types. Therule-basedmethod
provided by the rough-set method delivers classification accuracy in higher
level than other smart methods such as fuzzy ruleextraction,neuralnetworks,
decision trees and Fuzzy Networks like Fuzzy Min-Max Neural Networks.
KEYWORDS: Brain tumor; Malignancy Level; Rough sets; Particle swarm
optimization; prominent feature selection
1. INTRODUCTION
The level of severity in brain tumor decides the natureofthe
treatment to perform. For lower grade brain tumor, the
operation provides the higher survival rate. For other cases,
the risk in the surgery and the poor life quality are the
factors to consider before the operation. MRI technology
helps the physician in finding the real facts about the clinical
data before the surgery. The nature of brain tumor is severe,
but it happens infrequently. Enough experience is needed to
make correct judgments about the medical data. The neuro
radiologists have to make qualitative decisions out of their
rich experience. The relationship between the severity
degrees of the brain tumor can be captured through MRI
features and described clearly using rules.
C. Z. Ye et al. [9] focused on several constraints like
robustness, missing values, understandability and accuracy.
He framed a new rule extraction method based on Fuzzy
Min-Max neural networks (FRE-FMMNN) [10, 11]. The
proposed methodology uses a multi layer perceptron
network trained with a back propagation algorithm, and it
makes use of nearest neighborhood method. This method
(FRE-FMMNN) provides better predictions in finding the
malignancy degree than the other available intelligent
methods. C.Z. Ye concentrated only over the classification
task. Finding glioma MRI features and the degree of
malignancy is not very easy by having only two rules for
predictions.
Medical data composed ofirrelevant features,missingvalues
and uncertainties. These are all the factors that complicate
the process of finding the malignancydegreein medicaldata.
To find the real facts behind the medical data, it requires
how efficiently the analyst handles the incomplete and
inconsistent information, and with the various levels of
representation of data. The intelligent methods likedecision
tree, fuzzy theory and neural networks are based on such
strong assumptions (Sufficient number of experiments,
knowledge about dependencies, probability distributions).
Taking decisions fromincompleteknowledgeorinconsistent
information are the tedious task.
The Rough set theory [6] can effectively deals with
incompleteness and uncertainty in medical data analysis. It
is considered to have the good quality of knowledge to
efficiently differentiate the malignancy level inmedicaldata.
This algorithm removes the redundant informationfeatures
and selects a feature subset that has the same discernibility
features [6] as the original set. Identification of most
prominent subsets is the effective part that improves the
quality in decision making process. The rule induction
IJTSRD26802

International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
algorithm [4] followed by rough set methodology generates
decision rules, which potentially mine the profound medical
knowledge and provides a new medical insight from a new
dimension. The rules generated by these methods are very
useful for physicians to classify the malignancy level [15]
and provides the better understanding about theproblem in
their hand.
In this paper, we use rough sets to predict the malignancy
level of brain tumor. A rough set feature selection
methodology is used to select feature subsets that are
prominent (The feature set becomes a prominent one
because of eliminating the often repeating features and
describes the decisions as performed by the original feature
set to get improved level of accuracy). The selected
prominent features are those that influence the decision
concepts and it will be helpful in cause-effect analysis [16].
We introduce a new rough set attribute reduction method,
which works based on the PSO (search method). This
approach is compared with other rough set approaches.The
experimental results show thatthereducedsubsetgained by
this approach are more efficient and generates decision rule
with better classification performance. This rough set
approach can achieve higher classificationaccuracythanthe
other intelligent methods[3].
This paper is organized as follows. Section 2 describes the
concepts of rough sets. In Section 3 the proposed rough set
feature selection algorithm with Particle Swarm
Optimization is explained. In Section 4 the reducts using the
rough set rule induction algorithm and rule based
classification methods are explained. Section 5 describes
about the dataset. In Section 6 results are compared with
other methods for accuracy analysis. Conclusions are
described in Section 7.
2. Rough set methodology
Rough set theory [6] efficiently handles the uncertainty and
vagueness in data analysis part through the mathematical
approach. An object becomes indiscernible because of the
limited available information. A rough set theory is
characterized by two particle concepts known as lower and
upper level approximations generated using the objects
indiscernibilities. Here the important problem is the
reduction of attributes and the generation of decision rules.
Inconsistencies are not aggregated or corrected. Instead of
that the lower and upper approximations are computed and
rules are induced. The rules [4] are categorized into definite
and approximate rules depending on the lower and upper
approximations.
2.1 Basic theory of rough set concepts
An information system can be definedas I=(U,AU{d},where
U defined as universe with non-empty set of finite objects. A
is defined as non-empty set of finiteconditionattributes,and
d is a decision attribute. the function defined as fa:
UVa, where Va is the set of values of a. If p ⊆A, there is an
equivalence relation based on association
INDE(P) ={(x,y) U U| a P, fa (x) = fa (y)} (1)
The partition of universe U, produced by INDE(P) isdenoted
as U/P. If (x,y) IND(P), then x and y are indiscernible by
attributes from P. The equivalence classes of P-
indiscernibility relation are defined [x]p. Let X⊆U, the p-
lower approximation PX and X upper approximation of set
X can be defined as
P̠X = { x U|[x]p ⊆ X } (2)
PX = { x U|[x]p X ≠ } (3)
Let P, Q⊆A be equivalence relations over U, thenthepositive
and the negative regions can be defined as follows
POSp(Q) = U PX (4)
x U/Q
NEGp(Q) = U - U X (5)
x U/Q
BNDp(Q) = U PX -U X (6)
x U/Q x U/Q
The positive region of the partition U/Q with respect to P,
POSp(Q), is the set of all objects that can be classified into
blocks of the partition U/Q by means of P. Q depends on P in
a degree k(0 1) denoted P k(Q).
K= (7)
For k=1, Q depends on P, for 0<k<1 Q depends partiallyonP,
For k=0 P doesn’t dependent on Q. When P is a set of
condition attributes and Q is the decision, p (Q) denotes the
quality level of the classification task.
The goal of reducing the attribute[4,5] is to remove
redundant attributes, so that the reduced set provides the
same quality of classification as the original attribute set.
The set of all reducts is denoted as
Red={R⊆C|ᵧR(D)=ᵧC(D), } (8)
A dataset may contain many attribute reducts, The set of all
optimal reduct is as follows
Redmin = { R Red| Red, |R|≤ |R’|} (9)
2.2 Decision rules
The definition of the decision rule is as follows.
For an expression c:(a=v) where a and v is an
atomic formula for the decision rule which can be tested for
any x . A basic elementary condition c can be interpreted
as the following equationc:U .AconjunctionC
of q elementary conditions is defined as C=c1Ʌc2Ʌ…..Ʌcq.
The cover of a conjunction C, denoted as [C] or | C |A, is the
subset of examples that satisfy the conditions denoted by C,
[c]={x }, which is mentioned as supporting
descriptor. If k is the concept, the positive cover =
[C] denotes the set of possible examples covered by C.
A decision rule r for A is any expression in the form
where =c1 Ʌ c2Ʌ…..Ʌcq is a conjunction,
satisfying and v d, Where Vd is the set of values
of d.The attribute value pairs occurring in the left hand side
of the rule r is the condition part, Pred(r), and Succ(r) in the
right hand side is the decision part.

An object u U is matched by the decisionrule,
then we confirm that the rule classifies that the rule
classifies u to decision class v. Match(r) represents the
number of objects matched by the decision rule,
, which is equivalent to card (| |A). The support
factor of the rule, card( |A | |d=v|A) is the number of
objects supports the decision rule.
The accuracy and coverage of a decision rule is defined as
(D)= (10)
(D)= (11)
3. Feature selection using rough set [6] with particle
swarm optimization technique.
Feature selection through the rough set theory is valuable,
that can generate general decision rules [4,5] and provide
better classification quality for new samples. However, the
problem of finding a minimal reduct is coming undertheNP-
Hard Category [16]. Heuristic approximation methodology
has to be considered for solving the problem. Hu calculates
the significance of the attribute using heuristic ideas[14]
from discernibility matrices and proposes a heuristic
reduction algorithm [DISMAR]. A positive region based
method is taken as a significant factor by Hu. A conditional
information entropy reduction algorithm [CEAR] was
proposed by wang et.al.
We propose a new algorithm to find the minimal rough set
reducts using particle swarm optimization on brain tumor
data set. The proposed algorithm has been studied and
compared with the other rough set reduction algorithms for
analyzing accuracy. PSO algorithm performs better for
minimal rough set reduction in the experimental results.
Kennedy and Eberhart developed the PSO algorithm. It is an
evolutionary computation technique. The original idea was
to graphically produce the choreography of a collection of
birds. PSO was applied effectively to solve the optimization
problems. We apply PSO to find out minimal rough set
reducts [5].
3.1. PSO algorithm
PSO is initialized with a population of particles. Each and
every particle is taken as a point in an S-dimensional space.
The best previous position of any particle (Pbest – The
position giving the best fitness value) is Pi = (pi1,pi2…..pis).
The index of the global best particle is denotedas ‘gbest’.The
velocity for any particle i is Vi = (vi1,vi2…..vis). According to
the following formula the particles are manipulated in the
space S.
vid = wvid + c1 rand()(pid - xid) + c2 Rand()(pgd – xid) (12)
xid = xid + vid (13)
In the above equations is the inertia weight, suitable
selection of inertia weightprovides a betterbalance between
global and local exploration and it requires less number of
iterations on average to find the optimum value. Time
varying inertia will always provides betterperformance.The
acceleration constants are c1 and c2 in equation 12, which
represent the weighting of stochastic speeding terms that
pull each particle towards Pbest and gbest positions. Low
values allow particles to roam far from target regions. High
values result in abrupt movement toward target regions.
Rand() and rand() are two random functions in the
range[0,1] velocities of each particle on each dimension are
limited to a maximum velocity Vmax. If Vmax is smaller,
particles may not explore efficiently beyond locally good
regions. If Vmax too high paricles may fly past the good
solutions.
The first part of the eq-12 activates the“flyingparticles”with
able memory capacity and the ability to find a new search
space area. The second part is known as the “cognition
phase” which represents the private thinking of the particle.
“social phase” is the third part, which represents the
collaboration among the other particles.Equation-12isused
to update the velocity of the particle [16]. Then the particle
flies towards a new place according to the equation-13. The
fitness function is used to measure the performance of each
particle.
The implementation of PSO algorithm is as follows.
1. In S-dimensional problem space, a population of
particles with random positions and velocities are
initialized. Initialize Pi with a copy of Xi, and initializePg
with the index of the particle having the best fitness
function value among the population.
2. In S-dimensional problem space for each and every
particle present, evaluate the desired optimization
fitness function in d variables.
3. The particle’s fitness evaluation is compared with
particle’s pbest. If current value of the particle is better
than the pbest, then set the pbest value equal to the
current value, and the pbest location equal to the
current location in d dimensional space.
4. Comparision is done on fitness evaluation [22] with the
population’s overall previous bestvalue.Ifcurrentvalue
is better than the gbest, then reset gbest to the current
particle’s array index and value.
5. According to the formulas 12 and 13 changethevelocity
and position of the particle in the S-dimensional space.
6. Continue the looping until a criterion is met.
7. Usually a sufficiently good fitness value or a maximum
number of iteration for this process.
3.2. The encoding process
The position of the particle [3] is represented as a binary bit
of strings with length N. Where N representthetotalnumber
of attributes. Every bit is equivalent to anattribute,thevalue
‘1’ means the corresponding attribute is selected while ‘0’
means the attribute is not selected. Each position is an
attribute subset in the problem space.
3.3. Velocity representation
Particle’s velocity in the problem spaceis representedwitha
positive integer, which varies between 1 and Vmax. It means
that at any point in time how many of the particle’s [5] bit
should be changed to be the same as that of the global best
position, which means the velocity of the particle flying
toward the best possible position.
For example, pgbest=[1011101001] is the value and
Xi=[0100110101]. The difference between gbest and the
particle’s currentpositionis pgbest-Xi=[1-1110-11-100]. The
value ‘1’ means that, compared with the best position, this

bit(feature) should be selected but it is not because it
decreases the quality of classification. The value ‘-1’ means
that compared with the best position, this bit should not be
selected but it is selected. Redundant features will increase
the cardinality value of the subset. Both these cases willlead
to a lower fitness value. The number of ‘1’s is a and that of ‘-
1’ is b. The difference between a and b (a-b) is the distance
between the two taken particle’s position. The (a-b) value
may be positive or negative. Such a variety makes particles
possess ‘exploration ability’ in solution space.
In this example (a-b)=4-3=1, so Pg-Xi=1.
3.4. Strategies to update a particles position
After updating the velocity, a particle’s position will be
updated with the new velocity value. If the new velocity
value is V, the number of different bits between the gbest
[21] and the current particle is xg, there is two situations for
updating the position.
1. V . If a situation is like this, randomly change V bits
of the particle, which are different in form that of gbest.
The particle will move toward the global best while
keeping its ‘searching ability’.
2. V>xg. If a situation is like this, instead of changing all the
different bits to be same as that of gbest, we should
randomly change (V-xg) bits outside the different bits
between the particle and gbestt. After reaching the
global best position, it keeps on moving some distance
toward other possible directions, which gives it further
searching ability.
3.5. The velocity limit (maximum velocity, Vmax)
The particle’s velocity was initially limited to region [1,N]. It
was noticed that in some cases after several iterations, the
swarms [3] find a good solution (but not the real optimal
one), and in the following generations gbest remains
stationary. So, only sub optimal solution is obtained. This is
the condition in which the maximum velocity of the
particle[19,20] is too high and particle often ‘fly past’ the
optimal solution.
We set Vmax as(1/3)N and limit the range of the velocity in[1,
(1/3)N], which prevents this from being too large in value.
Limiting the maximum velocity, particles cannot fly too far
away [3,16] from the optimal solution. After finding a global
best position, other particles will adjust velocities and
positions and search around the best position. If V<1, then
V=1. If V>(1/3)N, V=(1/3)N. PSO [16] can find optimal
reducts [5] quickly under such limiting factors.
3.6. The fitness function
We apply the fitness function as given below:
fitness = (14)
Where decides the classification quality of the
condition attribute set R relative to decision D. |R| is the ‘1’
number of a position or the selected feature subset length.
|C| is the total number of features, are two
parameters that correspond to the importance of
classification quality and subset length, with and
=1- . In our experiment we set and . The
high value of assures that the best position [3] is at least a
real rough set reduct. The ultimate goal is to maximize the
fitness values.
3.7. Setting the parameters
In the given algorithm, the inertia weight decreases along
with the iterations according to the equation-15
W= Wmax - iter (15)
Where Wmax is the initial value of weighting coefficient, Wmin
the final value of the weighting coefficient, itermax the
maximum number of assignediterations [16]orgenerations,
and iter represents the current iteration or generation
number.
3.8. Measuring the time complexity
Let N be the number of features(representing conditional
attributes) and M the total number of objects. The time
complexity of POSAR is O(NM2), and that of the reduction
[17,18] based on conditional informationentropy(CEAR)[6]
is O(NM2)+O(N3), which is composed of computation of the
core and non-core attribute reduct. DISMAR has total time
complexity O((N+logM)M2).ForPSOREDUCT,thecomplexity
of the fitness function is O(NM2), the other impacts are
measured by the generation iterations. Time spent on
evaluating the particle’s position(that is fitness function).
4. Rough set rule induction algorithms
4.1. Algorithm for induction of minimum set of
decision rules.
LEMM algorithm was proposed to extract a minimum set of
decision rules. Let K be a non empty lower or upper
approximation of a concept, where c is a basic condition,and
C denotes the conjunction of such conditions being a
candidate for condition part of the decision rule. C[G]
represents the currently considered to be added to the
conjunction.
Procedure LEMM
(Input: a collection of objects K,
Output: Decision rules R derived out of the algorithm);
begin
G:=K;
R:=
while G≠ do
begin
C:=
C(G):={ c:[c] G≠ }
while(C= ) or (not [C] ⊆K) do
begin
select a pair c C(G) such that |[c] G| is minimum;
if match occur then select a join up c C(G) with the
biggest|[c]|;
if further ties occur then select the last join up from the list
C:= C {c};
G:= [c] G;
C(G):= {c:[c] G≠ }
C(G):= C(G) – C;
end {while}

for each c C do
if [ C-c ]⊆ K then C:= C – {c};
Create rule r based on C and add it to rule set R;
G:= K –
end{ while };
for each r R do
if = K then R:= R-r;
end {procedure};
4.2 Classification based on Decision rules
The LEMM algorithm is mainly used for classification
purpose. The induced set of rules is employed to [12]
successfully classify a new object. If new object matches
more than one rule, it needs to resolve the conflict between
sets of rules classifying tested objects to different decision
[16] classes.
In Ref [16], additional coefficients characterizing rules are
taken in to account: the strength of matched or partially
matched rules( the total number of cases correctly classified
by the rule during training phase), the number of non-
matched conditions, the rule specificity(i.e. length of the
condition parts). Allthesecoefficients arecombinedtogether
and the strongest decision wins. If no rule is matched, the
partially matched rules are taken and the most probable
decision is chosen from that.
The global strength defined in Ref. [16] for rule negotiation
is a number in the range [0,1] representing the importance
of sets of decision rules relative to the considered tested
object. Let us assume that T= (U,AU{d}) is a given decision
table, Consider ut is a test object, Rul(Xj) is the set of all
calculated decision rules[18] for the equation T, classifying
objects to the decision class
Xj( = d), MRul(Xj,ut)⊆Rul(Xj) matching tested object ut.
The global strength of decision rule set MRul(Xj,ut)isdefined
as follows
Glstrength (Xj,ut)= (16)
To classify a new case, rules are first selected matching the
new case. The strength of the selected rule sets is calculated
for any decision class, and then the decision class with
maximal strength is selected, with the new case being
classified to this current class. The quality of the complete
set of rules on a dataset with size n is evaluated by the
classification accuracy: nc/n, where nc is the number of
examples that have been correctly classified.
5. Brain tumor data set
The brain tumor data set [15] contains 14 condition
attributes and one decision attribute, as shown in Table 1.
The decision attribute ‘clinical grade’, is the actual grade of
glioma obtained from the surgery. Except the attributes like
‘gender’, ‘age’ and ‘clinical grade’, other attributes are
derived from the MRI of the patient and described with
uncertainty to various extent levels. The numericalattribute
‘age’ is discretized into three degrees, 1-30,31-60,61-90,
represented by 1,2,3 respectively.
Table1: Attribute description of brain tumor dataset
No Label Attributes Description
1 a1 Sex 0:female; 1:male
2 a2 Age 1:[<=30]; 2:[31-60];3:[>60]
3 a3 Type of the Shape 1:round;2:ellipse;3:irregular
4 a4 Nature of Contour 1:clear;2:partially clear;3:blur
5 a5 Tumor capsule 1:uninjured;2:partially injured;3:absent
6 a6 Edema 0: absent; 1: light; 2: middle; 3: heavy
7 a7 Presence of Mass 0: absent; 1: light; 2: middle; 3: heavy
8 a8 Post-contrast enhancement -1: unknown; 0: absent; 1:homogenous; 2: heterogeneous
9 a9 Blood supply level 1: normal; 2: middle; 3:affluent
10 a10 Premature death of cells 0: absent; 1: present;
11 a11 Calcification of the tumor 0: absent; 1: present;
12 a12 Nature of Hemorrhage 0: absent; 1: acute; 2: Chronic
13 a13 T1- weighted image’s signal intensity
1: hyper intense only; 2: isointense or accompanied by hyper
intense; 3: hyperintense or accompanied by isointense
14 a14 T2- weighted image’s signal intensity
1: hyper intense only; 2: isointense or accompanied by
hyperintense; 3: hyperintense or accompanied by isointense
15 a15 Lesions -1:unknown 1:present 2:absent
16 a16 Clinical grade 1:low grade 2:high grade
In total, 290 cases of brain glioma [15] are collected and seperatedintotwoclasses: lowergradeandhighergrade,in which174
are of low-grade glioma and 116 are of high-grade. There are 126 cases containing missing values on “post-contrast
enhancement”. By removing the incomplete 126 cases, the remaining subset of 164 complete cases contains 90 low-grade
glioma and 74 high-grade. Investigations are conducted on both the 290 cases and the 164completecases withoutthemissing
values. The quality of classification for both the 290 and 164 cases data are equal to 1. i.e the positive regions contains all the
cases.

6. Experiment results
In our experiments, we firstly use the rough set feature selection algorithm to select prominent feature subsets fromthebrain
glioma data [15]. Then, the selected feature subset are applied to generate decision rules to helptheneuro radiologists predict
the degree of malignancy in brain glioma.
6.1 Feature selection and rule set based classification
Matlab 2015a environment is used for implementing the PSOREDUCT [4] algorithm and other types of rough set algorithms.
The computer is intel (R) core (TM), 2.20 GHz CPU, 2 GB RAM and the system is windows 7 professional.
We applied the rough set feature selection algorithm to select the prominent features [4]subsetfromthebrainglioma data set.
Then the subset of selected features is applied to generate different decisionrules toassisttheneuroradiologists to predictthe
malignancy degree [8] in brain tumor.
We use 10-fold cross validation to evaluate the accuracy of classification of the rule set induced from the data set. All casesare
re-ordered randomly [16] and then the set of all cases is divided into 10 disjoint subsets approximately in equal size. For each
subset, all remaining cases are used for training, i.e., for rule induction, while the other remaining subset is used for testing
purpose. Different reordering result in different error rates. So, for each test we perform 10 times 10-fold cross validationand
the results are averaged.
The experimental results are listed in Table 2 and Table 3. Theparametersettingsandtheirvalues forPSOREDUCTareinTable
4. We performed experimentation on 290 brain glioma dataset [15] and the 176 case data. For both datasets, decision rules
generated from reducts produce a higher classification accuracy than those with the full 15 condition attributes. So, this
method of feature selection can improve the accuracy effectively.
The proposed rough set feature selection algorithm is compared against other rough set reduction algorithms. The reducts
[4,5] found by our algorithm are more efficient and can generate decision rules with better classification performance.
Compared with the other methods provided in the tables, the roughsetrule-basedclassificationmethodachievehigherlevelin
classification accuracy.
Table2: Results of 290 glioma cases
Algo. Name
Rough set Reduction algorithms FRE
FMMNNAll DISMAR POSR CEAR PSOREDUCT
Reduct 1-15 2,3,6-9,13,14,15 1-4,6 9,11,13,14,15 2,4,6,7,9,10,12,13,14,15 2,3,5,6,8,9,13,14,15 2,6-9,12,15
Rules 51 54 52 50 51 2
Avg.
accuracy (%)
82.80 84.70 83.58 85.49 87.67 83.21
High (%) 96.45 94.48 96.48 100 100 89.39
Low (%) 67.85 60.71 67.85 67.85 64.38 75
STD 7.23 7.11 5.87 6.09 6.63 5.35
Table3: Experimental results of 164 complete glioma
Algo. Name
Rough set Reduction algorithms
FRE FMMNN
All DISMR POSR CEAR PSOREDUCT
Reduct 1-15 2,3,6-9,13,14,15 2,3,5-9,11-14,15 1-3,6-9,13,14,15 2,3,5,6,8,9,13,14,15 2,6,8,9,11,13,15
Rules 32 32 26 34 30 2
Avg.accuracy
(%)
78.26 82.77 86.77 86.10 87.67 86.37
High (%) 93.63 100 100 100 100 100.00
Low (%) 53.63 53.63 70 80 73.36 73.36
STD 8.78 10.24 8.48 5.48 9.39 8.59
Table4: PSOREDUCT parameter values
Parameters PSO
Population Size 25
Max generation 100
Intial value of c1 2.0
Initial value of c2 2.0
Weight 1.4 - 0.4
Maximal velocity 1-(1/3)N
Features like age, edema, post contrast enhancement, blood supply and signal intensity [15] of the T1-weighted image arethe
most important factors for predicting the malignancy degree. These results are in accord with the experiences of experts and
other researcher’s contributions, and are very useful to neuro radiologists.

6.2 Decision rules generated using brain glioma data
The results based on the full 290 cases are more useful to neuro radiologists. In Table 6 we present part of the rules extracted
from the 290-case brain glioma data. The rules are generated by the rough set rule induction algorithm and include certain
rules and possible rules. Rules 1-3 are possible rules and others are certain rules.
Table 5: Selected feature subset
Features selected for our approach
By experiment(experts) 5-10, 12-15
Ye 2,6-9,11-14
PSOREDUCT 2,3,5,6,8,9,13,14,15
Intersection 2,6,8,9,13,15
The three possible rules have the higher accuracy and coveragelevel.Rule1,if(post-contrast enhancementis absent)thenthat
denotes the (low-grade brain glioma), covers 55 of 169 low-grade cases and has anaccuracyof98.2%.Rule2,If(affluentblood
supply) Then (high-grade brain glioma), covers 80 of 111 high grade cases and has an accuracy of 81.6%. Rule 3 shows that
hypointense only of signal intensity of the T1- and T2-weighted image always leads tolow-gradebrainglioma.This rulecovers
114 of 169 low grade cases and has an accuracy of 72.61%.
Table6: Decision rules with measures
No Rules Measures(supp,acc,cov)
1 (a8=0) ≥ (d=1) (56,98.2%, 32.5%)
2 (a9 = 3) ≥ (d=2) (81, 81.6%,72.0%)
3 (a13 = 1)&(a14 = 1) ≥ (d=1) (115, 72.61%, 67.46%)
4 (a8 = 0)&(a9 = 1) ≥ (d=1) (55, 100%, 30.18%)
5 (a6 = 1)&(a9 = )&(a13 = 1) ≥ (d = 1) (49,100%, 27,81%)
6 (a2 = 1,2)&(a6 = 0)&(a9 = 1) = > (d =1) (48, 100%, 24.85%)
7 (a3 = 2)&(a6 = 0)&(a13 = 1) ≥ (d = 1) (23, 100%, 12.43%)
8 (a3 = 2)&(a6 = 1)&(a9 = 1) ≥ (d = 1) (35, 100%, 20.71%)
9 (a2 = 1)&(a5 = 1) ≥ (d = 1) (19, 100%, 10.65%)
10 (a2 = 2)&(a3 = 1)&(a9 = 1)&(a13 = 1) = > (d = 1) (22, 100%, 11.24%)
11 (a2 = 2)&(a6 = 2)&(a8 = 2)&(a9 = 3) ≥(d=2) (19, 100%, 10.65%)
12 (a9 = 3)&(a14 = 3) ≥(d=2) (19, 100%, 10.65%)
13 (a3 = 3)&(a9 = 3)&(a13 = 2)&(a14 = 1) ≥(d = 2) (9,100%,4.73%)
Rules starting from 4-13 are considered as certain rules,
where rules from 4-10 are for low grade glioma and rules
from 11-13 are for high-grade brain glioma. The following
conclusions can be drawn from these set of rules.
1. If age is young AND the shape is regular AND edema is
absent or light ANDpost contrastenhancement is absent
AND blood supply is normal AND the signal intensity is
hypointense for T1 and T2 weighted images then the
result is possibly brain glioma with lower grade.
2. If age is old AND the shape is irregular AND edema is
heavy AND post contrast enhancement is homogeneous
or heterogeneous AND the blood supply is affluentThen
the result is most possibly brain glioma with higher
grade.
Light edema or the absence of edema often represent the
lower grade brain glioma [15], if the edema is heavy in
nature, it represent the higher grade brain glioma. If the
shape is round or ellipse (Regular in shape) that represent
the lower grade case. The normal blood supply and the
absence of post contrast enhancement always indicates the
lower grade. The affluent supply of blood always indicates a
high-grade brain glioma case.
Experimental results are also matches with the medical
expert’s experiences and other researcher’s contributions.
These conditional also provides meaningful medical
explanations for the physician.
7. Conclusions
The rough set theory is applied in this paper to predict the
malignancy degree of brain glioma and satisfactory results
were obtained. In this paper, attributereductionusingrough
set concept is applied with PSO to select the more efficient
feature subset. Decisionrules werederivedfromtheselected
prominent subsets for predicting the degree level of brain
glioma. The proposed approach performs well, while
comparing with the other rough set methodologies. The
reducts found by the proposedapproachweremoreefficient
and decision rules were generated based on these reducts
provides the better classification accuracy. The degree
prediction is based on the features like shape, age, edema,
blood supply, post-contrast enhancement and the signal
intensity of the T1 and T2 weighted image. Theclassification
accuracy can be improvised by the appropriate feature
selection. The rough set based method can achieve good
classification accuracy compared with other intelligent
techniques.
The decision rule generated by the rough set rule induction
approach is useful for both classification and medical
knowledge discovery process. Theserules effectively expose
the interpretable patterns of relations between glioma MRI
features and the degree of malignancy. The outcomes of
rules are helpful for the medical experts.

References
[1] R. B. Bhatt and M. Gopal, “FRID: Fuzzy-RoughInteractive
Dichotomizers,” IEEEInternational ConferenceonFuzzy
Systems (FUZZ-IEEE’04), pp. 1337–1342, 2004.
[2] Xiangyang wang, Jie yang, Richard jenson, Xiaojun liu,
Rough set feature selection and rule induction for
prediction of malignancy degree in brain glioma,
Computer methods and programs in biomedicine 83
(2006) 147-156
[3] E. Bonabeau, M. Dorigo, and G. Theraulez, Swarm
Intelligence: From Natural to Artificial Systems, Oxford
University Press Inc., 1999.
[4] I. Cloete and J. Van Zyl, “Fuzzy rule induction in a set
covering framework,” IEEE Transactions on Fuzzy
Systems, vol. 14, no. 1, pp. 93–110, 2006.
[5] W. W. Cohen, “Fast effective rule induction,” In Machine
Learning: Proceedings of the 12th International
Conference, pp. 115–123, 1995.
[6] C. Cornelis, M. De Cock and A. Radzikowska, “Vaguely
Quantified Rough Sets,”, Proc. 11th Int. Conf. on Rough
Sets, Fuzzy Sets, Data Mining and Granular Computing
(RSFDGrC2007), pp. 87–94, 2007.
[7] C. Cornelis, G. Hurtado Mart´ın, R. Jensen, and D. Sle¸zak,
“Feature ´ Selection with Fuzzy Decision Reducts”, Proc.
3rd Int. Conf. on Roughset.
[8] M. Bredel, L.F. Pollack, The p21-Ras signal transduction
pathway and growth regulation in human high-grade
gliomas, Brain Res. Rev. 29 (1999) 232–249.
[9] C. Z. Ye, J. Yang, D.Y. Geng, Y. Zhou, N. Y. Chen, Fuzzy
rules to predict degree of malignancy in brain glioma,
Med. Biol. Comput. Eng. 40 (2) (2002) 145–152.
[10] P. K. Simpson, Fuzzy Min-Max Neural Networks. Part
2.Clustering, IEEE Trans. Fuzzy Syst. 1 (1993) 32–45.
[11] J. R. Quinlan, Induction of decision trees, Mach. Learn.
1(1986) 81–106.
[12] J. M. Zurada, Introduction to Artificial Neural Systems,
West Publishing Co., New York, 1992.
[13] W. Andrew, Statistical Pattern Recognition, Oxford
University Press Inc., Oxford, 1999.
[14] M. A. Lopez-Gonzalez, J. Sotelo, Brain tumors in Mexico:
characteristics and prognosis of glioblastoma, Surg.
Neurol. 53 (2000) 157–162.
[15] Z. Pawlak, Rough Sets: Theoretical Aspects ofReasoning
About Data, Kluwer Academic Publishers, Dordrecht,
1991.
[16] A. E. Hassanien, Rough set approach for attribute
reduction and rule generation: a case of patients with
suspected breast cancer, J. Am. Soc. Inform. Sci. Technol.
55 (11). (2004) 954–962.
[17] Gunther, G., and D. Ivo (2000). Statistical Techniquesfor
Rough Set Data Analysis in Rough Sets: New
Developments. Physica–Verlag.
[18] Kryszkiewicz, M., and H. Rybinski (1996b). Reducing
information systems with uncertain real value
attributes. In Sixth International Conferences,
Informatio Processing and Management of Uncertainty
in Knowledge Based Systems, Proceedings (IPMU’96),
Vol. II. July 1–5, Grenada. pp. 1165.
[19] Millonas, M. M. (1994). Swarms, phase transitions, and
collective intelligence. In C. G. Langton, Ed.,ArtijicialLife
III. Addison Wesley, Reading, MA.
[20] Reeves, W. T. (1983). Particle systems - a technique for
modeling a class of fuzzy objects. ACM Transactions on
Graphics, 2(2):91-108.
[21] Reynolds, C. W. (1987). Flocks, herds and schools: a
distributed behavioral model. Computer Graphics, 2 1
(4):25-34.
[22] Wilson, E.O. (1975). Sociobiology: The new synthesis.
Belknap Press, Cambridge, MA.

An Enhanced Feature Selection Method to Predict the Severity in Brain Tumor

Recommended

More Related Content

What's hot (16)

Similar to An Enhanced Feature Selection Method to Predict the Severity in Brain Tumor (20)

More from ijtsrd (20)

Recently uploaded (20)

An Enhanced Feature Selection Method to Predict the Severity in Brain Tumor