Health Care Management Science

https://doi.org/10.1007/s10729-019-09489-x

IoT with cloud based lung cancer diagnosis model using optimal
support vector machine
Dinesh Valluru 1 & I. Jasmine Selvakumari Jeya 2

Received: 8 April 2019 / Accepted: 25 June 2019

# Springer Science+Business Media, LLC, part of Springer Nature 2019

Abstract
In the last decade, exponential growth of Internet of Things (IoT) and cloud computing takes the healthcare services to the next
level. At the same time, lung cancer is identified as a dangerous disease which increases the global mortality rate annually.
Presently, support vector machine (SVM) is the effective image classification tool especially in medical imaging. Feature
selection and parameter optimization are the effective ways to improve the results of SVM and are conventionally resolved
individually. This paper presents an optimal SVM for lung image classification where the parameters of SVM are optimized and
feature selection takes place by modified grey wolf optimization algorithm combined with genetic algorithm (GWO-GA). The
experimentation part takes place on three dimensions: test for parameter optimization, feature selection, and optimal SVM. For
assessing the performance of the presented approach, a benchmark image database is employed which comprises of 50 low-
dosage and stored lung CT images. The presented method exhibits its superior results on all the applied test images under several
aspects. In addition, it achieves average classification accuracy of 93.54 which is significantly higher than the compared methods.

Keywords Classification . Feature selection . IoT . Lung cancer . Support vector machine

1 Introduction diagnosis is presented. On the other hand, ung cancer is an im-

portant reason behind the increased mortality rate [28]. It is re-
Recently, the growth of Internet of Things (IoT) based healthcare ported that 1.2 million people are diagnosed with lung cancer
services and sensing devices leads to significant research areas. annually. Researchers identified the survival rate is increased
Because of the expensive healthcare and existence of numerous when the cancer is identified at the earlier levels. But, the earlier
diseases, it is necessitate the conversion of hospital-centric sys- identification of the disease is a difficult process. Around one
tem to a person-centric environment immediately [25]. IoT and third of the people are correctly identified in the middle or ad-
cloud computing are greatly interlinked to one other and they can vanced stage of cancer. Computer-aided diagnosis models are
be used to monitor the patients in the remote areas. It is supported found to be effective to detect and diagnose the abnormalities
from virtual unlimited abilities and resources of the cloud to meet earlier and quickly. Presently, machine learning (ML) models
the technical limitations like memory, computational power and have been applied to detect the presence of lung cancer from
so on. The IoT centric-cloud model can be utilized to develop computer tomography (CT) images. The existing image classifi-
new applications and services in the healthcare section [27]. In cation approaches can be classified into supervised and unsuper-
this paper, an IoT and cloud based healthcare system for disease vised classification approaches.
An important supervised classification approach is support
vector machine (SVM). It is an effective tool for data classifi-
* Dinesh Valluru cation and regression issues on a small sample size with high
[email protected] dimensionality. It is commonly employed in diverse applica-
tions like network traffic detection, autonomous vehicles, and
I. Jasmine Selvakumari Jeya
enterprise credit risk assessment [1], and so on. In addition,
[email protected]
few enhanced SVM approaches are also presented such as
1
Anna University, Chennai -600025, India fuzzy decision tree (FDT)-SVM, manifold ranking (MR) -
2 SVM, spatial contextual (SC)-SVM, etc. The results of the
Department of Computer Science and Engineering, Hindusthan
College of Engineering and Engineering, Coimbatore 641032, Tamil SVM are highly based on the employed parameters and fea-
Nadu, India tures. Depending on the statistical theory, SVM is a ML
 Valluru D., Jeya I.J.S.

method that can search an optimal solution for classification tuning of SVM parameters should be considered concurrently.
outcomes with constrained data about a smaller sample In this paper, we introduce an optimal SVM for lung image
dataset. SVM employs the kernel function concepts to change classification. Since the parameter selection and feature selec-
the problem of non-linearity into problem of linearity and tion plays a vital role in the results of SVM, an integration of
decreases the mapping complexity [2, 26]. The optimal prob- GWO with GA called (GWO-GA) takes place. A set of ex-
lem can be transformed to quadratic convex function problem periments has been performed for investigating the results
by employing the kernel function optimization and transfor- interms of feature selection results and classifier performance.
mation. SVM classifier clearly reaches the global optimal so- The simulation outcome clearly defines the enhanced perfor-
lution in theory and is shown in Fig. 1. mance of the optimal SVM among the compared methods.
Presently, swarm intelligence algorithms are being applied The succeeding sections of the paper are arranged as fol-
to enhance the results of the SVM. At the same time, optimal lows. An overview of related works is given in Section 2. The
SVM model can be attained when the parameters and features presented GWO-GA is given in Section 3. The application of
employed are assumed and optimized simultaneously. An im- presented method for medical image classification is ex-
proved grey wolf optimization (IGWO) method has been pre- plained in Section 4. The results are validated in 5 and con-
sented to filter the significant features and then SVM is ap- cluded in section 6.
plied to classify data using the features chosen by IGWO
algorithm. In this study, grey wolf optimization (GWO) algo-
rithm is integrated with GA and is employed to classify med-
2 Related works
ical images.
The contribution of the paper is summarized here.
In the SVM classification performance, the kernel function pa-
Nowadays, e-healthcare services become more common with
rameter δ and penalty factor C has much influence over it. At the
the use of IoT devices and cloud technologies. Several studies
same time, while SVM is used for classification, feature selection
were presented to develop an automatic disease classification
is important for high-dimensional features dataset. Generally, in
model to identify the presence of lung cancer. Presently, sup-
SVM, feature selection and parameter optimization plays a major
port vector machine (SVM) is the effective image classifica-
role and has influence on one another. And, it is difficult to gain
tion tool especially in medical imaging. But, the results of
optimal SVM classifier if the SVM parameters and feature subset
SVM are solely based on its parameters. In case of medical
are optimized distinctly. Hence, feature selection, parameter op-
images with high-dimensional features, feature redundancy
timization and optimal classifier of SVM shall be concurrently
has a great impact on classification results. Feature selection
optimized; this is known as optimal SVM for convenience. In
and parameter optimization are the effective ways to improve
common, both feature selection and parameter optimization can
the results of SVM. The choice of optimum feature subset and
be known as an issue of combinatorial optimization and might be
managed through swarm intelligence algorithms and evolution-
ary algorithms like differential evolution (DE) algorithm, genetic
algorithm (GA), GWO, particle swarm optimization (PSO) algo-
rithm, etc.,

2.1 Parameter optimization

For bankruptcy prediction, [3] employed GA to improve the

SVM parameters and simulation outcomes showed that the
presented techniques find highly useful for hybrid SVM mod-
el by means of generalization ability and predictive accuracy.
Through the PSO algorithm, PSO-SVM model has been built
[4] to find the SVM parameters for the purpose of target rec-
ognition. To find out the optimal parameters with SVM com-
bination, [5] used DE algorithm and simulation outcomes
shows that prediction accuracy of the projected method is
improved over others methods. A hybrid GA-PSO algorithm
[6] attains SVM global optimal parameters and image classi-
fication in wetland remote sensing and the outcomes exhibited
that the hybrid method can give enhanced accuracy when
Fig 1 SVM classifier compared to PSO and GA methods.
IoT with cloud based lung cancer diagnosis model using optimal support vector machine

2.2 Feature selection 2.3 The optimal SVM

At present times, numerous unsupervised methods had been A GA method had been projected [13] for optimal SVM and
projected for feature selection. Few unsupervised feature simulation outcomes demonstrate that the projected GA-based
selection methods have been projected [7, 8] depending consistent parameter optimization and feature selection tech-
on structural learning through combining feature learning nique had enhanced accuracy with some features. A PSO-
and image understanding into a framework of joint learn- based method in [14] showed that the simulation outcomes
ing. For image understanding tasks, over revealing latent is good at removing insignificant or unnecessary features,
subspace, the simulation outcomes denote that the projected and it determined the parameter rates efficiently by means of
method was efficient and the classification accuracy was enhancing the entire classification outcomes. GA for the opti-
enhanced over other methods. For embedded feature selec- mal parameters detection of SVM and the discriminative fea-
tion, unsupervised learning methods had been projected [9] ture subset is presented [15] for remote sensing image dataset.
using kernel K-means, that had been employed to reduce For a given classification problem, the simulation outcomes
violation in contrast to initial cluster structure. For future denote that the projected technique can find better feature
analysis, simulation outcomes showed that the projected subset and gain enhanced classification accuracy. For remote
technique had superior phenomena understanding. In image sensing images, [16] examined a PSO algorithm for determin-
retrieval scope, unsupervised methods are used commonly ing the parameter, feature selection in a SVM-based classifi-
for every pixel classification in an image and choose the cation. It has reached low sensitivity and enhanced classifica-
optimal feature subset simultaneously. This method is tion accuracy due to the dimensionality curse. Some other
employed for every pixel classification and to choose the related works are also developed and found in the literature
optimal feature subset. For an image, pixel is the minimal [17–21].
unit and size of the data is comparatively large; it may take
a huge time to estimate the data correlation of various 2.4 Characteristics of the proposed method
features that will tend to poor efficacy.
The major feature selection aim is to gain proper classifica- Though the PSO and GA algorithm be implied for optimal
tion accuracy with some possible features. As the SVM fea- SVM, GA and PSO might get trapped simply in local opti-
tures are chosen through swarm intelligence algorithms or evo- mum and it is hard to search global optimal solution. By the
lutionary algorithms, SVM is employed as supervised classifier use of presented GWO-GA algorithm, the projected method
which is a prominent one among classifiers. A model of new attempts to find a best solution and it can attain classification
walker assisted gait has been projected through Martins in this accuracy using the feature subset and optimal parameters.
view that merges SVM and GA-based feature selection method
to differentiate among non-assisted and assisted gait with fore-
arm supports that had been successful at feature count restric- 3 GWO-GA method
tion. A diagnosis model had been build [10] depending on
SVM and PSO method to choose the optimal feature subset. GWO algorithm is based on the hunting procedure of leader-
It founds optimal while diagnosing erythemato-squamous dis- ship hierarchy. The grey wolves are considered as the top-
eases and actual feature set. The simulation results show that level predators and it usually lives in the numbers of 5–12
classification precision of the projected method was superior wolves. Based on the hunting behavior of wolves, it is cate-
when compared to given outcomes. And, PSO method can gorized into 4 kinds such as alpha (α), beta (β), delta (δ), and
efficiently decide the SVM parameter rates that appropriately omega (ω). The most dominating one called α wolves which
enhanced the entire classification accuracy. A hybrid feature take care of the place to hunt, sleep and so on. These decisions
selection method depending on SVM (GA-SVM) and GA will be followed the group and the members of the group
was projected [11] for remote sensing image classification acknowledges the α words by keeping their tails in a down-
and the simulation outcomes shows that the projected method ward direction. The subsequent level of grey wolves is β
usually gives enhanced classification accuracy with some which are subordinate wolves that help α to make decisions.
employed features when comparing with methods of non- The β wolves are assumed as α wolves in situations when the
feature selection. A new classification and feature selection α wolves dies. The minimum ranking grey wolves are ω
method has been projected [12] for remote sensing images by wolves that play the role of scapegoat. Approaching,
merging PSO algorithm global optimization capability. The encircling and attacking the prey are the different functions
simulation method demonstrates that the feature selection per- involved in GWO algorithm. It has exploration as well as
formance can efficiently decrease the dimensionality of data exploitation phase. The first phase looks for the optimum so-
and attained enhanced accuracy when comparing with all the lutions in the local search space. The swarm of grey wolf
method. encircles and attacks the prey while exploring the optimal
 Valluru D., Jeya I.J.S.

solutions in a local search space. The prey will be searched in GA [22] is an important evolutionary algorithm that pro-
the exploration phase where the prey will be searched in the vides the algorithms, which follows the process of natural
entire search area. During the process of encircling prey, the selection and evolution. It is commonly employed for gener-
wolf recognizes the prey location and encircles them. The ating useful solutions to optimize and search problems partic-
position vectors of the prey are found and the search agents ularly when there is no heuristic information of the problem to
alter the position based on the attained best solution. The be managed. It is assumed that the initial heuristic information
encircling prey can be defined as follows for GWO algorithm is arbitrarily created and it does not use
the data based on the managed problem.
! !
! ! In fortunate, GA does not require any heuristic data in the
D ¼ j C : Xp ðtÞ− X ðtÞj ð1Þ
task of evolution and it can attain acceptable solution in most
!
! !! cases. On the other hand it is helpless to premature conver-
X ðt þ 1Þ ¼ X p ðkÞ− A : D ð2Þ
gence and has very low convergence speed later in the evolu-
! ! tion process. Keeping these features, GWO-GA algorithm is
Where t denotes the present round, A and C are coefficient

! ! introduced which make use of binary encoding, which makes
vectors and position vector of the prey is indicated as Xp ðkÞ; X cooperation easier. The solution of GA is applied to the initial
is the position vector, || is the absolute value and . is element population of the GWO algorithm. When the fitness value of
! ! present individual is effective when compared to GA, the in-
multiplication. The vectors A and C are indicated by:
dividual will be employed for replacing the respective GA and
!
A ¼ 2!a :!
r −!
a ð3Þ GA restarts to execute, hence, GA and GWO gets executed till
! the stopping criteria is satisfied.
C ¼ 2:!r ð4Þ

where components of ~a are linearly reduced from 2 to 0 for

several rounds and r1, r2 are random vectors in [0, 1]. The
grey wolf has its own nature of identifying the positions of the
4 Proposed optimal SVM for lung cancer
prey and encircles it. The hunting process is led by α wolves
prediction
whereas the β and δ may contribute in some situations. In
The overview of the presented method is shown in Fig. 2.
addition, for mathematical simulation of the hunting nature
Initially, the features will be extracted by the optimal SVM and
of grey wolves, it is considered that α, β and δ holds more
then classification of images takes place using SVM classifier.
facts about the significant locations of the prey. So, the three of
With a pair of n instance-labels, assume a problem of classifica-
the first solutions attained will be saved and make the other
tion, instance vector is S = (xi, yi), (i = 1, 2, …., n) xi ∈ R and class
search agents to update the locations based on the location of
label is referred by yi ∈ {−1 + 1}. To search a hyperplane, classi-
the best search agents, which can defined as given in Eq. (5).
fier training is employed which distinguishes the negative (−1)
! ! ! ! samples and positive (+1) samples. Optimization is included in
D α ¼ j C a * X α − X j*
! ! ! ! ð5Þ
training procedure as given in Eq. (7).
D β ¼ j C b* X β− X j
! ! ! !
D δ ¼ j C c* X δ− X j 1
∅ðw; εÞ ¼ kwk2 þ C∑ni¼1 εi ð7Þ
! ! ! 2
Where D α , D β and D δ are the altered distance vector be-
! !
tween the α, β and γ positions to the other wolves and C a , C b ,
! With respect to the constraints:
and C c are three coefficient vector used in distance vector.
i
! ! ! ! yi ½w:xi Þ þ b ≥ 1−εi ; i ¼ 1; 2; …; n ð8Þ
X a ¼ X α− A a* D α
! ! ! !
X b ¼ X β− A b* D β ð6Þ
! ! ! ! Let hyper-plane normal vector is ω, slack variable is εi ≥ 0
X c ¼ X δ− A c* D δ for measuring classification errors, for the error term ∑ni¼1 εi , C
is a penalty factor or positive constant, and ∅ is a function that
!
where X a is the newly obtained position vectors of α position maps the input space to a high-dimensional feature space. The
! ! ! entire procedure is a simple linearly separable problem in high
X α and distance vector D α , X b indicates the new position
! dimensions, and the transformation is based on kernel func-
vector attained by the use of β position X β and distance vector
! ! tion of SVM which is described as:
D β , X v indicates the new position vector determined by the use    
! ! K xi ; x j ¼ ∅ðxi ÞT :∅ x j ð9Þ
of δ position X δ and distance vector D δ.
IoT with cloud based lung cancer diagnosis model using optimal support vector machine

Fig. 2 Block diagram of

proposed method

4.1 Parameter optimization and kernel function The functional form is highly simple for linear kernel
of SVM function. To deal with the problem in linear separable
ones, it is highly useful. The functional form is highly
For SVM, the kernel function is highly essential that creates difficult for polynomial kernel function that might cost
the actual data as linear separable, and dimensionality curse a huge sum of computing time [23]. It can recognize
has been avoided efficiently. In SVM, there exist numerous non-linear mapping for RBF kernel function and the
kernel functions types such as radial basis function (RBF) needed count of RBF kernel function parameters is
kernel function, linear kernel function and polynomial kernel small comparatively and it comprise low number of
functions and are represented as complexities. When comparing with other generally
  employed kernel functions [24], numerous simulation
K xi ; x j ¼ xTi :x j ð10Þ outcomes exhibit that RBF kernel function has enhanced
   δ performance. Therefore, for classification, RBF kernel
K xi ; x j ¼ xTi :x j þ 1 ð11Þ
function is employed for every dataset to find optimal
 2 !
  −xi −x j  penalty factor C combination and kernel function pa-
K xi ; x j ¼ exp ð12Þ
δ2 rameter δ.

Fig. 3 Sample CT lung images

 Valluru D., Jeya I.J.S.

4.2 Process of feature selection based on SVM 5 Performance validation

In CT lung images, high-dimensional features are re- The presented method is simulated using MATLAB
quired for attaining high classification accuracy. But, 2014b. The parameter settings used for experimentation
few correlated and repetitive features are also involved are given as follows: Kernel type: Polynomial, Kernel
in these features that will decrease classification accura- degree:2, Cost: 0.1 and Gamma:0.01.The presented
cy and computational efficiency. A procedure of choos- method concentrates on the optimization capability of
ing relevant features subset from entire features for con- the GWO-GA and the classification accuracy of the op-
struction of model and removes duplicate features is timal SVM. The experimentation part takes place on
known as feature selection. Selection of feature is three dimensions: test for parameter optimization, fea-
known as a combinatorial optimization problem and is ture selection, and optimal SVM. For assessing the per-
solved by the GWO-GA algorithm as discussed in the formance of the presented approach, a benchmark image
previous section. For classification, SVM is used and database is employed which comprises of 50 low-dosage
the objective function depends on the classification ac- and stored lung CT images (http://www.via.cornell.edu/
curacy gained through SVM number of features were lungdb.html). Each image is 1.25 mm slice thickness
chosen by the presented method. and is generated in a single breath. The position of

Table 1 Results of parameter optimization for different algorithms

Image Optimal
CT Images Type Measures GA BPSO BDE
name SVM
I1 Fiv(%) 78.1 80.3 80.8 82.9

Normal
Std(%) 4.22 3.83 3.03 2.75

Time 0.75 1.09 0.99 0.65

I2 Fiv(%) 84.3 84.8 85.8 86.50

Benign Std(%) 2.31 2.18 1.80 1.40

Time 2.88 4.49 3.08 2.62

I3 Fiv(%) 80.0 80.1 80.2 80.9

Std(%) 2.06 2.06 1.69 1.77
Malignant
Time 2.72 3.64 3.16 2.57

I4 Fiv(%) 94.4 95.0 95.3 96.1

Std(%) 1.47 1.33 1.21 0.97
Normal
Time 2.54 3.87 3.04 2.36

I5 Fiv(%) 81.8 82.2 82.7 84.0

Std(%) 2.10 1.88 1.57 1.44

Malignant

Time 1.61 2.46 2.01 1.47

IoT with cloud based lung cancer diagnosis model using optimal support vector machine

the nodules is recognized by the expert are also whereas the GA, BPSO and BDE achieved lower fitness
provided in the dataset. Some of the sample test values of 84.3, 84.8 and 85.8 respectively.
images are shown in Fig. 3. Likewise, on the test image I3, the higher fitness value of
Tables 1, 2 and 3 provides the fitness values of pa- 80.9 is attained by the presented method whereas the lower
rameter optimization, feature selection, and optimal fitness values of 80, 80.2 and 80.2 is achieved by the GA,
SVM that undergone optimization using diverse algo- BPSO and BDE methods. In the same way, for all the applied
rithms. It is shown in Table 1 that the fitness value of test images, maximum fitness value is obtained by the optimal
optimized by different algorithms, respectively. It is SVM compared to other methods. From the Table 1, it is also
displayed in Table 1 that the fitness value of GA and observed that the presented optimal SVM takes minimum
BPSO algorithm is noticeably poor than the compared computation time than the other methods.
BDE and the presented method. For instance, for the Table 2 provides the optimal optimization capability of the
image I1, a maximum fitness value of 82.9 is obtained optimal SVM, its average fitness value, standard deviation, num-
by the optimal SVM whereas the GA, BPSO and BDE ber of chosen features (Fn) and time. From this table, it is evident
achieved fitness values of 78.1, 80.3 and 80.8 respec- that the presented method chooses less number of features and
tively. Similarly, on the applied image I2, the optimal higher fitness value for all the applied images. For instance, for
SVM achieved a maximum fitness value of 86.50 the applied image I1, the presented method has chosen a

Table 2 Results of feature selection for different algorithms

Image Optimal
CT Images Type Measures GA BPSO BDE
name SVM
I1 Fiv(%) 0.78 0.78 0.79 0.82
Std(%) 0.02 0.02 0.02 0.015
Normal
Fn 6 6 5 4

Time 0.26 0.44 0.36 0.23

I2 Fiv(%) 0.26 0.44 0.36 0.45

Std(%) 0.81 0.82 0.82 1.40
Benign
Fn 5 4 3 3
Time 1.69 2.61 1.88 1.41
I3 Fiv(%) 0.61 0.61 0.64 0.66
Std(%) 0.03 0.02 0.02 0.01
Malignant
Fn 8 6 6 5

Time 1.50 2.43 1.87 1.37

I4 Fiv(%) 0.74 0.75 0.77 0.84

Std(%) 0.03 0.02 0.02 0.01
Normal
Fn 12 11 10 9

Time 1.27 1.64 1.39 1.24

I5 Fiv(%) 0.68 0.68 0.71 0.84

Std(%) 0.02 0.02 0.02 0.01
Malignant
Fn 15 14 12 11
Time 0.66 0.85 0.70 0.54
 Valluru D., Jeya I.J.S.

Table 3 Results of the optimal SVM for different algorithms

Optimal
CT Images Type Meas. GA BPSO BDE
SVM

Fiv(%) 0.77 0.78 0.83

0.78

Std(%) 0.02 0.03 0.014

0.02
Normal
Fn 8 7 6
7

Time 0.47 0.80 0.41

0.70

Fiv(%) 0.82 0.82 0.86

0.83

Std(%) 0.01 0.01 0.01

0.01
Benign
Fn 7 6 5
6

Time 2.27 3.67 1.82

2.67

Fiv(%) 0.76 0.76 0.81

0.77

Std(%) 0.02 0.01 0.01

0.01
Malignant
Fn 10 10 8
9

Time 1.79 2.72 1.48

2.01

Fiv(%) 0.90 0.90 0.94

0.91

Std(%) 0.01 0.01 0.01

0.01
Normal
Fn 19 19 16
18

Time 1.64 2.56 1.45

1.96

Fiv(%) 0.78 0.79 0.86

0.80

Std(%) 0.01 0.01 0.01

0.01
Malignant
Fn 21 20 18
19

Time 0.95 1.71 0.86

1.33

minimum of 4 features whereas the other methods have chosen a Likewise, on the test image I3, the higher fitness value of
number of 5 and 6 features. Interms of fitness value, for the image 0.66 is attained by the presented method whereas the lower
I1, a maximum fitness value of 0.82 is obtained by the optimal fitness values of 0.61, 0.61 and 0.64 is achieved by the GA,
SVM whereas the GA, BPSO and BDE achieved fitness values BPSO and BDE methods. In line with, for the applied image
of 0.78, 0.78 and 0.79 respectively. Similarly, on the applied I4, the presented optimal SVM showed superior results with
image I2, the optimal SVM achieved a maximum fitness value the fitness value of 0.84. However, the compared GA, BPSO
of 0.45 whereas the GA, BPSO and BDE achieved lower fitness and BDE methods depicted comparative lower results of 0.74,
values of 0.26, 0.44 and 0.36 respectively. 0.75 and 0.77 respectively. The values depicted that the
IoT with cloud based lung cancer diagnosis model using optimal support vector machine

Table 4 Comparative of various methods interms of accuracy

Image FCM-SVM Parameter Feature selection Optimal SVM

optimization

I1 90.96 94.98 90.67 96.54

I2 89.85 93.95 89.87 95.58
I3 87.17 92.45 88.11 92.98
I4 59.91 89.58 82.94 92.15
I5 68.75 88.74 83.95 90.45

presented SVM is superior to other methods under several 0.77 is achieved by the GA, BPSO and BDE methods. In line
aspects. From the Table 2, it is observed that the presented with, for the applied image I4, the presented optimal SVM
optimal SVM takes lower computation time than the other showed superior results with the fitness value of 0.94.
methods. At the end, based on the data present in the However, the compared GA, BPSO and BDE methods
Table 3, the values depicted the superior performance of the depicted comparative lower results of 0.90, 0.90 and 0.91
optimal SVM compared to the existing methods in several respectively. These values clarified that the presented SVM
ways. The presented algorithm achieved an average fitness showed enhanced results over the compared methods in a
value of 0.828 on all the applied test images which signifies significant way.
the optimization capability of presented algorithm than the Finally, an analysis of the compared methods and proposed
other existing algorithms. Particularly, interms of chosen fea- method is made interms of classification accuracy and the values
tures, the presented method has chosen less number of features are provided in Table 4. From the table and Fig. 4, it is evident
with a minimum of 6 features for the image I1 and a maximum that the maximum accuracy is attained by the presented method
of 18 features for the image I5. In addition, the computation with the classification accuracy of 96.54 on the image I1 whereas
time required by the presented method is significantly lower the FCM-SVM attained a lower classification accuracy of 90.96.
than the time needed by the compared methods. For instance, Similarly, on all the applied test images, the presented method
interms of fitness value, for the image I1, a maximum fitness exhibited better performance with the maximum classification
value of 0.83 is obtained by the optimal SVM whereas the results. On average, FCM-SVM, parameter optimization, feature
GA, BPSO and BDE achieved fitness values of 0.77, 0.78 and selection and optimal SVM attains the classification accuracy of
0.78 respectively. Similarly, on the applied image I2, the op- 79.328, 91.94, 87.108 and 93.54 respectively. From these values,
timal SVM achieved a maximum fitness value of 0.86 where- it is evident that the optimal SVM exhibits superiority with an
as the GA, BPSO and BDE achieved lower fitness values of average classification accuracy of 93.54 over the compared
0.82, 0.82 and 0.83 respectively. Likewise, on the test image methods. In short, the presented method easily converges to the
I3, the higher fitness value of 0.81 is attained by the presented optimum solution for parameter optimization, feature selection,
method whereas the lower fitness values of 0.76, 0.76 and and the optimal SVM on all the applied test images.

Fig. 4 Comparative of various

methods interms of accuracy FCM-SVM Parameter optimization Feature selection Optimal SVM

100
90
80
70
60
Accuracy

50
40
30
20
10
0
I1 I2 I3 I4 I5
Images
 Valluru D., Jeya I.J.S.

6 Conclusion 12. Yu N, Yang H (2007) An improved multi-class support vector ma-

chine for hyperspectral remote sensing imagery classification.
Remote Sensing Information 5:7–11
In this paper, we have devised an optimal SVM for lung image 13. Huang CL, Wang CJ (2006) A GA-based feature selection and
classification. Due to the difficulty in the process of parameter parameters optimization for support vector machines. Expert Syst
selection and feature selection in SVM, we have presented the Appl 31:231–240
14. Lin SW, Ying KC, Chen SC, Lee ZJ (2008) Particle swarm optimi-
GWO-GA. The experimentation part takes place on three di-
zation for parameter determination and feature selection of support
mensions: test for parameter optimization, feature selection, vector machines. Expert Syst Appl 35:1817–1824
and optimal SVM. The presented method exhibits its superior 15. Bazi Y, Melgani F (2006) Toward an optimal SVM classification
results on all the applied test images under several aspects. In 900 system for hyperspectral remote sensing images. IEEE Trans
Geosci Remote Sens 44:3374–3385
addition, it achieves average classification accuracy of 93.54
16. S. Ding, L. Chen, Classification of hyperspectral remote sensing
which is significantly higher than the compared methods. The 925 images with support vector machines and particle swarm opti-
detailed experimental analysis on the test images ensured that mization, in: 2009 International Conference on Information
the presented method can be applicable in real time data anal- Engineering and Computer Science, pp. 1–5
17. Wang Y, Liu H, Zheng W, Xia Y, Li Y, Chen P, Guo K, Xie H
ysis in hospitals, telemedicine and so on. In future, the pro-
(2019) Multi-objective workflow scheduling with deep-Q-
posed method can be improved by the inclusion of deep lean- network-based multi-agent reinforcement learning. IEEE Access
ing models. 7:39974–39982
18. Wu W, Pirbhulal S, Zhang H, Mukhopadhyay SC (2019)
Quantitative assessment for self-tracking of acute stress based on
Compliance with ethical standards triangulation principle in a wearable sensor system. IEEE journal of
biomedical and health informatics 23(2):703–713
Conflict of interest The authors declare no conflict of interest. 19. Han J, Ji X, Hu X, Zhu D, Li K, Jiang X, Cui G, Guo L, Liu T
(2013) Representing and retrieving video shots in human-centric
brain imaging space. IEEE Trans Image Process 22(7):2723–2736
20. Zhang D, Meng D, Han J (2016) Co-saliency detection via a self-
References paced multiple-instance learning framework. IEEE Trans Pattern
Anal Mach Intell 39(5):865–878
1. Wang G, Ma J (2012) A hybrid ensemble approach for enterprise 21. Han J, Zhang D, Cheng G, Guo L, Ren J (2014) Object detection in
credit risk assessment based on support vector machine. Expert Syst optical remote sensing images based on weakly supervised learning
Appl 39:5325–5331 and high-level feature learning. IEEE Trans Geosci Remote Sens
2. B.E. Boser, I.M. Guyon, V.N. Vapnik, A training algorithm for 53(6):3325–3337
optimal margin classifiers, in: Proceedings of the fifth annual work- 22. Srinivas M, Patnaik LM (1994) Genetic algorithms: a survey.
shop on Computational learning theory, pp. 144–152 Computer 27:17–26
3. Wu CH, Tzeng GH, Goo YJ, Fang WC (2007) A real-valued 1040 23. Scholkopf B, Smola AJ (2002) Learning with kernels: support vec-
genetic algorithm to optimize the parameters of support vector ma- tor ¨ machines, regularization, optimization, and beyond. MIT press
chine for predicting bankruptcy. Expert Syst Appl 32:397–408 24. Yin S, Yin J (2016) Tuning kernel parameters for SVM based on
4. Zhai S, Jiang T (2014) A novel particle swarm optimization trained expected square distance ratio. Inf Sci 370-371:92–102
support vector machine for automatic sense-through-foliage target 25. Lakshmanaprabu SK, Shankar K, Ilayaraja M, Nasir AW,
recognition system. Knowl-Based Syst 65:50–59 Vijayakumar V, Chilamkurti N (2019) Random forest for big data
5. Cheng MY, Hoang ND, Wu YW (2013) Hybrid intelligence approach classification in the internet of things using optimal features.
based on LS-SVM and differential evolution for construction cost index International journal of machine learning and cybernetics, 1-10.
estimation: a Taiwan case study. Autom Constr 35:306–313 https://doi.org/10.1007/s13042-018-00916-z
6. Cheng L, Bao W (2014) Remote sensing image classification based 26. Shankar, K., Lakshmanaprabu, S. K., Gupta, D., Maseleno, A., &
on optimized support vector machine. Indonesian Journal of de Albuquerque, V. H. C. (2018). Optimal feature-based multi-ker-
Electrical Engineering and Computer Science 12:1037–1045 nel SVM approach for thyroid disease classification. The journal of
7. Z. Li, J. Liu, J. Tang, H. Lu, Robust structured subspace learning for supercomputing, 1-16
data representation, IEEE transactions on pattern analysis 965 and 27. Lakshmanaprabu SK, Shankar K, Khanna A, Gupta D, Rodrigues
machine intelligence 37 (2015) 2085–2098 JJ, Pinheiro PR, De Albuquerque VHC (2018) Effective features to
8. Tang J, Li Z, Wang M, Zhao R (2015) Neighborhood discriminant classify big data using social internet of things. IEEE Access 6:
1030 hashing for large-scale image retrieval. IEEE Trans Image 24196–24204
Process 24:2827–2840 28. Lakshmanaprabu SK, Mohanty SN, Shankar K, Arunkumar N,
9. Maldonado S, Carrizosa E, Weber R (2015) Kernel penalized Ramirez G (2019) Optimal deep learning model for classification
Kmeans: a feature selection method based on kernel K-means. Inf of lung cancer on CT images. Futur Gener Comput Syst 92:374–
Sci 322:150–160 382
1 0 . A b d i M J , G i v e ki D ( 2 0 1 3) Au t o m a t i c d e t e c t i o n o f
erythematosquamous diseases using PSO-SVM based on associa-
tion rules. Eng Appl Artif Intell 26:603–608
11. Pedergnana M, Marpu PR, Dalla Mura M, Benediktsson JA,
Bruzzone L (2013) A novel technique for optimal feature selection
in attribute profiles based on genetic algorithms. IEEE Trans Geosci Publisher’s note Springer Nature remains neutral with regard to
Remote Sens 51:3514–3528 jurisdictional claims in published maps and institutional affiliations.