SPECIAL SECTION ON DEEP LEARNING

ALGORITHMS FOR INTERNET OF MEDICAL THINGS

Received March 6, 2020, accepted March 12, 2020, date of publication March 17, 2020, date of current version April 2, 2020.
Digital Object Identifier 10.1109/ACCESS.2020.2981337

Optimal Feature Selection-Based Medical Image

Classification Using Deep Learning Model in
Internet of Medical Things
R. JOSHUA SAMUEL RAJ1 , S. JEYA SHOBANA2 , IRINA VALERYEVNA PUSTOKHINA3 ,
DENIS ALEXANDROVICH PUSTOKHIN4 , DEEPAK GUPTA 5 , AND K. SHANKAR 6
1 Department of Information Science and Engineering, CMR Institute of Technology, Bengaluru 560037, India
2 Department of Computer Science, College of Science, Knowledge University, Erbil 446015, Kurdistan Region, Iraq
3 Department of Entrepreneurship and Logistics, Plekhanov Russian University of Economics, 117997 Moscow, Russia
4 Department of Logistics, State University of Management, 109542 Moscow, Russia
5 Department of Computer Science and Engineering, Maharaja Agrasen Institute of Technology, New Delhi 110086, India
6 Department of Computer Applications, Alagappa University, Karaikudi 630002, India

This work was supported in part by the RUSA-Phase 2.0 grant sanctioned vide Letter No. F. 24-51/2014-U, Policy (TNMulti-Gen),
Dept. of Edn. Govt. of India, Dt. 09.10.2018.

ABSTRACT Internet of Medical Things (IoMT) is the collection of medical devices and related applications
which link the healthcare IT systems through online computer networks. In the field of diagnosis, medical
image classification plays an important role in prediction and early diagnosis of critical diseases. Medical
images form an indispensable part of a patient’s health record which can be applied to control, handle
and treat the diseases. But, classification of images is a challenging task in computer-based diagnostics.
In this research article, we have introduced a improved classifier i.e., Optimal Deep Learning (DL) for
classification of lung cancer, brain image, and Alzheimer’s disease. The researchers proposed the Optimal
Feature Selection based Medical Image Classification using DL model by incorporating preprocessing,
feature selection and classification. The main goal of the paper is to derive an optimal feature selection model
for effective medical image classification. To enhance the performance of the DL classifier, Opposition-
based Crow Search (OCS) algorithm is proposed. The OCS algorithm picks the optimal features from
pre-processed images, here Multi-texture, grey level features were selected for the analysis. Finally, the opti-
mal features improved the classification result and increased the accuracy, specificity and sensitivity in the
diagnosis of medical images. The proposed results were implemented in MATLAB and compared with
existing feature selection models and other classification approaches. The proposed model achieved the
maximum performance in terms of accuracy, sensitivity and specificity being 95.22%, 86.45 % and 100%
for the applied set of images.

INDEX TERMS IoMT, classification, deep learning, medical image, features, Crow search algorithm,
optimization.

I. INTRODUCTION as Magnetic Resonance Imaging (MRI), Computer Tomogra-

In recent days, medical informatics has become a hot research phy (CT) scan as well as ultrasound B-scan. To achieve bet-
topic in which Information Technology meets the needs ter classification performance, both descriptiveness as well
of human healthcare requirements. Many research activities as discriminative power of the extracted features are crit-
have been conducted in the area of ‘investigation of medical ical when it comes to image classification problems. This
images’ with the intention of diagnosis as well as clinical machine learning section is necessary because of its wide
studies [1]. The images are retrieved from computer-aided range of applications in the field of data mining, forecasting
investigation analysis through various imaging systems such models, recovery of media information, etc.
The medical image databases are used for image classifica-
The associate editor coordinating the review of this manuscript and tion and for teaching purposes. It frequently contains images
approving it for publication was Wei Wei . with various qualities pictured under diverse conditions,

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
58006 VOLUME 8, 2020
R. J. S. Raj et al.: Optimal Feature Selection-Based Medical Image Classification Using DL Model in IoMT

along with precision explanation. In medical diagnosis, it is module is preferred where most of the significant features are
important to recognize the most significant risk factor with selected using Opposition-based Crow Search optimization
disease identification. The fundamental feature of conven- algorithm i.e. OCS. The application of OCS algorithm in the
tional medical image classification algorithms are many, for selection of GLCM and GLRLM features show the novelty
example, color, texture and shape features. The subset of best of current work. Finally, the DL technique was utilized to
feature is given in a way for improving the classification categorize the given medical image as whether malignant or
performance or regression model, particularly when manag- benign. Besides, the consistent performance of the proposed
ing a high-dimensional feature space by optimization [2]. model was validated against brain, lung, and Alzheimer dis-
The optimized representation of the identified feature subset ease identification and classification processes.
represents new problems. Now, one of the newly proposed The paper is ordered as follows: Section 2 explains the
swarm intelligence-based algorithms is explained herewith. review of recent literature about medical image processing,
If the feature extraction algorithm defers a huge number of section 3 explains the current problems, section 4 provides
features, the manual selection of optimal feature becomes a brief explanation of methodology, section 5 discusses the
critical. Thus, the optimal feature selection is explained to implementation result of image classification models, and
computerize this process. Various researches have been car- finally, the conclusion section explains with possible future
ried out about lung cancer images classification. scope.
The computational complexity is comparatively high and
these techniques require clear information about the image II. LITERATURE REVIEW
structure. Majority of the classification processes find an In 2018, Singh and Singh et al. [10] created the engineered
efficient connection between the membership function [3] medical images with the help of newly proposed deep learn-
and the feature of the object. Moreover, the accurate iden- ing Generative Adversarial Networks (GANs). The classic
tification of this fundamental function is highly critical [4]. data expansion was performed by training the Convolutional
The proposed work i.e., Deep Neural Network (DNN) is Neural Network (CNN) whereas the synthetic data expan-
effectively applied to real world classification approaches, for sion and performance comparison were done. The results
example speech recognition, fault detection, medical diag- obtained were 78.6% sensitivity and 88.4% specificity for
nosis etc., [5]. It is a subfield of machine realization which the classification performance utilizing classic data expan-
utilizes a lot of calculations that endeavor to demonstrate sion. By adding the synthetic data expansion, the results got
the abnormal state in the present information by utilizing improved by 85.7% sensitivity and 92.4% specificity.
a [6] deep arrangement that has few layers with direct and A Computer Aided Diagnosis (CAD) was introduced in the
nonlinear change capacities [7]. Some of the advantages in research work by Frid-Adar et al. [11] for the categorization
deep learning model are reduction in the need for feature pro- of liver ultrasound image. The image features were separated
duction and it is deemed to be one of the most time-reducing using seven particular surface models to describe the surface
parts of machine learning approach. In case of huge data of Region of Interest (ROI). The Mutual Information (MI)
and images, it can easily recognize, understand the spoken feature selection strategy was utilized after it was chosen by
language, overcome issues and work more efficiently [8]. its exceptional segregation highlights. The proposed CAD
By using deep learning methods, it is possible to avoid framework was able to yield 95.55% precision and affectabil-
common problems which generally consumes too much of ity of 97.77% with 20 best features chosen by the MI includ-
time. Image classification is an important area and poses ing the determination method.
great challenge for an image to be classified with the medical Sharma et al. [12] conducted a research study in which the
expert knowledge. The main aim of this paper focuses only Firefly Algorithm (FA) was adjusted in selecting the com-
on the analysis of miming the medical image classification ponents of characterization as well as regression techniques
approach [9]. to support the fundamental forms with the help of data-
The authors have introduced an Optimal Feature Selection based learning models. To assess the effectiveness of the pro-
based Medical Image Classification using Deep Learning posed FA method, an aggregate of 29 characterizations and
model, by incorporating preprocessing, feature selection and 11 relapse benchmark informational indexes were utilized.
classification. The main goal of the article is to derive an The proposed FA adjustment offered a proficient strategy to
optimal feature selection model for effective medical image perceive ideal element subsets in classification as well as
classification. In order to achieve this goal, a set of Gray-level regression models for supporting the information-based basic
co-occurrence matrix (GLCM) and Gray-level Run Length leadership forms.
Matrix (GLRLM) were used. With the use of texture features The CT output of lung images was examined with the
like GLCM and GLRLM, the best subset of features are help of Optimal Deep Neural Network (ODNN) along with
extracted from pre-processed medical images. Traditionally, Linear Discriminate Analysis (LDA) by Zhang et al. [13].
the extracted features are provided to the classification pro- The profoundly highlighted sections were removed from CT
cess. But, in the proposed method, it is not directly given to lung pictures after which the dimensionality of the feature
the classification process since it consumes more computa- was reduced using LDR to categorize lung nodules as two
tional time to execute. Hence, the optimal feature selection types: malignant or benign. In order to recognize the lung

VOLUME 8, 2020 58007

 R. J. S. Raj et al.: Optimal Feature Selection-Based Medical Image Classification Using DL Model in IoMT

malignant growth arrangement, ODNN was connected to CT • Existing feature determination models display chal-
scans and was upgraded using Modified Gravitational Search lenges i.e., optimization [13], [14] and dimension
Algorithm (MGSA). reduction joining features from various information
To enhance the analysis of Parkinson’s ailment, the models [17]. Further, the potential challenges of per-
advanced variant of Crow Search Algorithm was intro- forming feature selection in little sample estimate situa-
duced (OCSA) by Lakshmanaprabu et al. [14]. The outcomes tions need to be exhibited.
were against the primary Chaotic Crow Search Algo- • By analyzing the existing papers, it was inferred that sev-
rithm (CCSA) whereas the execution of OCSA was estimated eral researchers took efforts to develop methodologies
for minimum standard datasets. The proposed nature- so as to overcome these challenges. Yet, some specific
motivated calculation found an ideal subset of features from problems related to optimization algorithm still exist
the test result and expanding the exactness and diminishing such as high computational time, misdetection because
various features were chosen for further study. of small size, etc., The motivation to overcome these
In the research conducted by Gupta et al. [15], in the issues paved the way to propose a new method.
instance of medical data possessing bigger hierarchal connec-
tions in the information, the upside of machine learning can
be found through algorithms without relentless hand-making
features. Lately, the astounding accomplishment of machine
learning calculations in image recognition intersects with a
period of significantly expanded usage of electronic medical
records and image diagnosis.
A deep learning technique was introduced by Ker et al. [16]
in 2018 to classify the lung images. Here, the dataset utilized
was VGG-16 CNN learning design with convolutional chan-
nel of (3×3) that was actualized on mammograms’ ROIs from
IRMA. For the main completely-associated layer, the deep
element grid was processed. The outcome shows that 10 folds
on SVM was utilized with two-fold trees, straightforward
coordination and KNN (with k = 1, 3, 5) classifiers. The
final result was achieved with 100% classification accuracy
with AUC 1.0.
To overcome these problems, novel meta-heuristic opti-
FIGURE 1. The overall process of medical image classification.
mizer, namely CCSA was proposed by Gardezi et al. [17].
The CSCA performance was compared with other known
techniques like metaheuristic algorithms. In order to discover IV. METHODOLOGY
an optimal feature subset, to increase the performance of clas- In the domain of medical image classification, diagnosing
sification and to minimize the number of selected features, the individual’s disease from medical dataset is a promis-
the study was conducted and the experimental result was ing computation. The current research work considered four
analyzed with the ability of CSCA. datasets for image classification analysis and the main aim
is to achieve maximum accuracy in disease prediction. The
III. PROBLEM IDENTIFICATION three datasets considered were brain, lung, breast cancer and
After analyzing the existing literature related to the field of Alzheimer’s disease. From these datasets, the stages of each
image classification and feature selection, the authors found patient’s image were classified as Benign and Malignant at
several issues in the existing classifiers which are detailed as an early stage by the proposed classification (DL) technique.
below: This medical image classification consists of three stages
• The explanatory power of ANN is not satisfactory
such as pre-processing, feature selection, and classification.
whereas optimization of the network can be challenging, Figure 1 depicts the diagrammatic representation of medical
especially in avoiding the overtraining of data image classification.
• In case of an SVM classifier, it leads to omissions and • Preprocessing: The purpose of image preprocessing is
misclassification when identifying small and irregular- to get a better quality of images. Here, the histogram
sized cancers equalization is used to enhance the input image quality.
• Tumor diagnosis is a basic criteria that needs to be met. • Optimal Feature Selection (OCS): One of the significant
In imaging test, finding the right tumor part is a mind- image processing steps is feature selection; this makes
blowing test. In addition, the image mining systems are the image classification, a simple process.
present only to find the exact tumor cells by following • Classification: According to the selected best features,
numerous techniques. it classifies medical images into benign and malignant.

58008 VOLUME 8, 2020

R. J. S. Raj et al.: Optimal Feature Selection-Based Medical Image Classification Using DL Model in IoMT

TABLE 1. Features of GLCM and GLRLM.

A. PRE-PROCESSING Wimage and Himage indicate the image width and height value
This section consist of Histogram Equalization (HE) which respectively. The image gray level is denoted as G. Table 1
is used to enhance the image quality. This is performed by shows the pre-processing results of the medical image.
leveling the image pixel gray-levels in order to reorganize
them constantly in the spatial space. The general execution B. FEATURE SELECTION MODULE
steps of HE [18] is depicted in the figure 1. The histogram of Feature selection is a dimensionality reduction method used
the input image is evaluated, normalized for sum evaluation in datasets; it eliminates irrelevant or redundant features
and at last, the input image is transformed to an output from the pre-processed image. The feature selection strat-
image while the sample HE image is shown in the figure 2 egy implies less data transmission. Usually, several feature-
below. extraction models are applied in data mining process. In the
proposed study, two techniques were used to extract the
desired features i.e., GLCM and GLRLM.
• Gray-level Co-occurrence Matrix (GLCM): It is a
measurable technique of examining the texture that
considered spatial correlation of pixels in GLCM or else
it is also termed as gray-level spatial dependence matrix.
The GLCM [19] portrays the ascertaining of the pixel
FIGURE 2. HE for medical image.
with explicit qualities in a predefined spatial relation-
ship that take place in an image, making it a GLCM.
The HE formula, for transforming an input image into an After this process, the factual measures are separated
output image, is depicted in the following equation (1). from the matrix. The matrix component m(a, b|d1 , d2 )
signifies the proportionate separated by a pixel distance
 
cdf (image)−cdfmin
HE(image) = round ×(G−1) (d1 and d2).
(Wimage ×Himage ) − cdfmin
• Gray-level Run Length Matrix (GLRLM): The GLRLM
(1)
technique is a method for removing higher order sta-
The expansion of equation (1) is: Cumulative Distribution tistical texture data. In maximum gray dimensions G,
Function of the medical image is abbreviated as cdf , the min- the image is frequently reduced by prior re-quantizing in
imum non-zero value of cdf is represented as cdfmin while order to aggregate the network. The features which are

VOLUME 8, 2020 58009

 R. J. S. Raj et al.: Optimal Feature Selection-Based Medical Image Classification Using DL Model in IoMT

TABLE 2. Performance measures.

isolated with the help of GLRLM and GLCM are listed

in the table 1. GLRLM is created as follows:

K (θ ) = (r (u, v)/θ ), 0 ≤ u ≤ Nr , 0 ≤ v ≤ K max (2)

where Nr implies maximum gray level values, K max denotes

more length as well as u, v indicate the sizes of the matrix
values.

C. OPTIMAL FEATURE SELECTION USING

OCS ALGORITHM
By using texture features like GLCM and GLRLM, the best
subset of features (i.e. depicted in the tables 2 and 3) were
extracted from the pre-processed medical image. The
extracted features were not directly provided to the classifi-
cation process since it consumes more computational time to
execute. So, the optimal feature selection module was chosen
where most of the significant features are selected using the
optimization algorithm i.e. OCS.

1) CROW SEARCH (CS) ALGORITHM

The researchers [20] developed CS algorithm on the assump-
tions of nature of crow with respect to hiding and grabbing
the food. By seeking the crow’s nature, the features of CSA is
illustrated as follows (a) It is in the form of a flock, (b) it saves
food hiding places, (c) for the purpose of stealing, they follow
each other along with (d) by probability, and they guard their
caches against being pilfered. The original and new positions
of the two crows are described in figure 3. FIGURE 3. Inspiration of CS algorithm.

2) OPPOSITION BASED CS (OCS) ALGORITHM

To develop the efficiency of traditional CS technique, the con- crows (features) are arbitrarily located in a search space.
trast operation was added. For every initiated solution,
the adjacent opposite operation also starts working. If the
solutions are compared, better solution can be selected to Fi = F1 , F2 , . . . . . . .Fn , where i = 1, 2, 3 . . . n. (3)
obtain the optimal solution.
The implementation procedure of OCS algorithm is 4) OPPOSITION PROCESS
described in the steps below: Generally, the metaheuristic optimization models start with
few initial solutions and try to improve it by simultane-
3) INITIALIZATION OF CROW’S POPULATION ously monitoring the contrast solution. While comparing the
Assign the population of crows (features extracted from the solutions, the optimal one could be selected as an initial solu-
medical image) with respect to Fi whereas the initialized tion. For example, let f ∈ (g, h) is a real number. By applying

58010 VOLUME 8, 2020

R. J. S. Raj et al.: Optimal Feature Selection-Based Medical Image Classification Using DL Model in IoMT

TABLE 3. Preprocessed results.

the opposite point definition, it can be written as the given dataset images.
OFi = MAX (Accuracy) (5)
f˜j = gj + hj − fj (4)
6) GENERATE NEW POSITION
5) FITNESS EVALUATION Assume a crow to form a novel position in a random man-
The fitness function of OCS algorithm is calculated on the ner by choosing one of the flock crows in such a way that
basis of objective function of the research work. Here, the the crow ‘j’ has its own position as well as storage space.
contribution of optimization is to attain optimal features from The innovative position of the crow Pi, iter is attained by the

VOLUME 8, 2020 58011

 R. J. S. Raj et al.: Optimal Feature Selection-Based Medical Image Classification Using DL Model in IoMT

following equation (6).


i,iter + r × fl i,iter × (memj,iter − Pi,iter )
P
 i
i,iter+1
P = if rj ≥ APj,iter (6)

rand P otherwise


The expansion of equation (6) is explained here: ri and rj indi-

cate random number of crows i and j respectively, between
[0 and 1], fl i, iter represents the flight length of crow i, P sym-
bolizes the position of crow, memj, iter denotes the memory
location of jth crow and APj,iter indicates the attentiveness
probability of crow j at iteration.

7) MEMORY RENEWAL
The currently upgraded crow’s place and memory value are
updated with the application of equation (7).
(
i,iter+1 Pi,iter f (Pj,iter+1 ) > f (memj,iter )
mem = (7)
memi,iter otherwise

8) TERMINATION CONDITION OF OCS

It is monitored that the fitness value for the new location of
a crow is superlative in sustained position. The crow often
updates the storage space by novel position. If more than
one iteration is attained, the optimal location of the memory,
which corresponds to objective, is addressed as the best solu-
tion of filtered set of features. The well-defined patterns of
OCS technique are denoted in the figure 4.
FIGURE 4. Flowchart of OCS algorithm.
D. IMAGE CLASSIFICATION
Followed by optimal feature selection, DL technique was
In addition, Restricted Boltzmann Machine (RBM) is estab-
employed to divide the image into two levels:
lished to resolve the problems of viable activation function
• Benign
production.
• Malignant
The alleviated feature set was induced for classification 2) RESTRICTED BOLTZMANN MACHINE
method. Hence the accomplished image, including the best RBM is a sort of Markov arbitrary domain which consists of
features, designs the precision of classification operation single layer of (commonly Bernoulli) stochastic hidden units
while comparing the classification performance of actual including a single layer of (ordinarily Bernoulli or Gaussian)
images [24]–[26]. stochastic visible units.
Step 1: Initiate the clear units ν to train the RBM vector
1) DEEP LEARNING (DL)
K X
L K L
Artificial Neural Network (ANN) strategy is generally mod- X X X
F (v, h) = − Wkl vk hl − αk vk − βl hl (8)
eled using various layers of hidden units as well as outputs
k=1 l=1 k=1 l=1
which are named as DL. It consists of two levels, namely
• Pre-training
The expansion of equation (8) is as follows; Wkl denotes
• Fine-tuning stages
the symmetric communication among the visible unit vk
and the hidden unit hl , α, β are the bias terms, K , L are the
a: PRE-TRAINING STAGE number of visible and hidden units. The subordinate of the
log likelihood of a preparation vector concerning a weight is
Being a deep structure as well as crucial feed forward net-
inconsistently simple. From the hidden units of RBM, there
work, the Deep Belief Network (DBN) is employed where
is no sudden response that tends to obtain simple impartial
the instance gets originated from input layer to output layer
sample from (Vk , hl )data
through maximum number of hidden layers along with addi-
tional layers. Based on the application of DBN [13] technique K
!
X
and in accordance to the hidden units which differentiate ρ (hl = 1 |v ) = δ Wkl vk + αl (9)
the network itself, the network produces activation functions. k=1

58012 VOLUME 8, 2020

R. J. S. Raj et al.: Optimal Feature Selection-Based Medical Image Classification Using DL Model in IoMT

where δ (x) is the logistic sigmoid function. (1+exp(x)) , vk , hl

1
The parameter setting of the proposed model is given here;
is the unbiased sample. Number of input nodes: 100, hidden layer 1 node: 50, aver-
age activation: 0.5, weight decay is 0.0001, M: 30, upper
3) UPDATING PROCESS bound: 1, lower bound: 0 and maximum iteration count:
The hidden unit is updated whereas the visible units are 60. Here multi-medical image database was considered for
assumed to be concurrent from the provided visible as well as analysis such as brain, lung, and Alzheimer disease as shown
hidden units. It leads to route a complex procedure as follows. in the figure 6 [21]–[23]. The proposed optimal features,
with DL model, were compared with other optimizations and
1Wkl θ (vk hl )data − (vk hl )reconstruction (10)
classifiers such as CS [14], FA [12], Support Vector Machine
At this point, RBM undergoes training. A divergent RBM (SVM), CNN [10] and NN with different performance mea-
could be stacked across a frame using multilayer technique. sures as listed in the able 2.
Every time, there is a distinct RBM available which is Table 3 explains the pre-processed result using HE. It was
stacked. While the input visible layer is arranged as a vector, performed with three types of medical image databases such
the qualities for units, which are efficiently deployed with as brain, lung and Alzheimer disease. In all the three types of
RBM layers, have been shared with the application of shared the image database, some of the sample images were taken
models in current weights as well as biases. Hence, the finish- for pre-processing. First, the original image was taken in
ing layer, which is trained officially, is locked to be in novel which the HE technique was applied as a result of which the
RBM. Therefore, the accomplished DNN weights are filled pre-processed image is attained. HE is a technique used to
with fine-tuning phase. The architecture of DL along with enhance the contrast of the image. At the end of this step,
n-hidden units are shown in the figure 5. a clear medical image was obtained for further processing.
Figure 7 demonstrates the accuracy analysis of various
medical images, analyzed with different features. The con-
sidered features were F1, F2, F3, F4, F5, F6 etc., whereas
the considered algorithms were CS, FA and OCA. For each
algorithm, different optimal feature values were obtained.
In feature F1, the CS algorithm attained 72%, 65% for FA
algorithm and the OCA algorithm attained 79.9% accuracy.
In feature F2, 79% was attained for CS algorithm, 75.5%
for FA algorithm and the OCA algorithm achieved the maxi-
mum accuracy at 90%. Similarly, the entire feature provided
various optimal values and the OCS algorithm yielded better
accuracy compared to existing algorithms.
FIGURE 5. Structure of DL. Figure 8 explains the feature selection process with various
optimization algorithms. In CS algorithm, the number of fea-
tures was 15 in which the selected features obtained were 10.
a: FINE TUNING PHASE The number of features for FA was 15 and the selected feature
The fine tuning level is performed according to back was 13. Here, the number of features was the same for all
propagation technique. the algorithms and the selected features were different. In the
In order to segregate the medical images into two stages, proposed OCS algorithm, the number of features was 15 and
output layer is organized from top of DNN. Also, N number the selected feature was 6.
of input neurons as well as three hidden layers are applied Table 4 explains the image classification outcome for
in DL method. The enhanced weight should be obtained the proposed DL algorithm. For all pre-processed images,
during training phase, along with the help of training data set, the accuracy (%), sensitivity (%) and specificity (%) were
in such a way that the back propagation is initialized with the calculated. In this study, nine sample images were consid-
weights that are gained from pre-training level. Finally, using ered. Using different medical images, the disease has been
equation (8), the minimum error value was calculated; also predicted as benign or malignant. Here, some of the pre-
the maximum accuracy of DL classification was achieved by processed images are shown and the class of such images was
the optimized weight. At last, based on this optimal weight predicted to be benign or malignant for all the pre-processed
(optimal feature sets), the medical images such as brain, images. The analysis predicted the class as malignant for
lung, breast cancer and Alzheimer were classified as two brain image with 92.23% accuracy, 86.45% sensitivity with
categories: Benign and Malignant. 100% specificity for the first image. The first lung image
proved to be benign and the predicted class was benign
V. RESULT ANALYSIS with 91.22% accuracy, 83.1sensitivity and 94.5% specificity.
The proposed medical image classification and the imple- The second and third images were classified to be malignant
mentation results are discussed in this section. It was sim- and the predicted class was benign. In this case, different
ulated in MATLAB 2017a with i7 processor and 4GB RAM. accuracy values were obtained. In 1stAlzheimer disease,

VOLUME 8, 2020 58013

 R. J. S. Raj et al.: Optimal Feature Selection-Based Medical Image Classification Using DL Model in IoMT

FIGURE 6. Sample database medical images.

FIGURE 7. Accuracy for feature selection.

FIGURE 8. Feature selection based Comparison.

the image was benign and the predicted class was benign being 96.12% for 90% training and 10% testing. In case
with 92.22 % accuracy, 86.45% sensitivity and 83.12 speci- of 80% training and 20% testing, the accuracy obtained was
ficity. The 2nd image was malignant and the predicted class 83.22%, specificity was 85.22% with 51.2% sensitivity. The
was malignant with 93.11% accuracy, 79.45% sensitivity and training image was 70% and the testing image was 30% with
89.11% specificity. Similarly, the third image was benign and 90.1% accuracy, 72.2% specificity and 53.22% sensitivity.
the predicted class was also benign with different accuracy The training image was 60% and the testing image was
values obtained from the analyses. 40%. The accuracy was 69.22% with specificity being 79.2%
Table 5 explains the training and testing result for the and 51.2% sensitivity. For 40% training and 60% testing,
proposed (OCS-DL) algorithm. Here, the sample lung images the accuracy obtained was 56.11%, specificity was 43.4 %
were considered for analysis. The image was trained and and sensitivity was 46.2.2%.
the accuracy, specificity, and sensitivity were calculated. The Figure 9 (a) represents the performance analysis for image
accuracy was 86.2% with 91.22% specificity and sensitivity classifier. The performance was compared with various

58014 VOLUME 8, 2020

R. J. S. Raj et al.: Optimal Feature Selection-Based Medical Image Classification Using DL Model in IoMT

TABLE 4. Proposed (DL) medical image classification results.

FIGURE 9. Feature selection based comparison.

performance measures such as FPR, FNR and FDR. For DL

algorithm, the FPR and FDR attained 0, FNR attained 0.14.
The FPR value for CNN was 0.23, FNR was 0.23, and FDR
was 0.2. For SVM classification, FNR attained the value 0,
FPR attained 0.34 and FDR achieved 0.19. Similarly, for NN
classification, various performance values were obtained.

algorithms. The compared algorithms were DL, CNN, SVM

and NN. The performance is represented in (%). The accu- FIGURE 10. Classification rate analysis.
racies obtained for DL, CNN, SVM and NN were 92, 8,
85 and 79. The sensitivity values for DL, CNN, SVM and Figure 10 demonstrates the classification rate analysis for
NN were 83.2, 79, 100 and 60 respectively. For specificity various medical images. This section discusses whether the
analysis, the respective values for DL, CNN, SSVM and NN image classification was correct or wrong using optimization
were 100, 82, 75 and 90 respectively. Figure 9(b) explains the algorithms and classification techniques. The selected images

VOLUME 8, 2020 58015

 R. J. S. Raj et al.: Optimal Feature Selection-Based Medical Image Classification Using DL Model in IoMT

TABLE 5. Sample training and testing result for the proposed model.

were classified correctly using the DL algorithm with [5] C. Affonso, A. L. D. Rossi, F. H. A. Vieira, and
classification rate being 89.9% whereas the classification A. C. P. de Leon Ferreira de Carvalho, ‘‘Deep learning for biological
image classification,’’ Expert Syst. Appl., vol. 85, pp. 114–122, Nov. 2017.
rate for being wrongly classified was 0.3%. In CNN, [6] A. Harouni, A. Karargyris, M. Negahdar, D. Beymer, and
the image was correctly classified at 68.9% and wrongly clas- T. Syeda-Mahmood, ‘‘Universal multi-modal deep network for
sified at 20.12%. In SVM, the classification rate was 60.56% classification and segmentation of medical images,’’ in Proc. IEEE
15th Int. Symp. Biomed. Imag. (ISBI), Apr. 2018, pp. 872–876.
for correctly classified and 20.35% for wrongly classified. [7] A. Kharrat, K. Gasmi, M. B. Messaoud, N. Benamrane, and M. Abid,
The classification rate was 59.9% for correctly classified and ‘‘Medical image classification using an optimal feature extraction algo-
40% as wrongly-classified in case of NN. rithm and a supervised classifier technique,’’ Int. J. Softw. Sci. Comput.
Intell., vol. 3, no. 2, pp. 19–33, 2011.
[8] P. Gnanasekar, A. Nagappan, S. Sharavanan, O. Saravanan,
VI. CONCLUSION D. Vinodkumar, T. Elayabharathi, and G. Karthik, ‘‘Investigation
A novel classification method was proposed to classify the on feature extraction and classification of medical images,’’ Ph.D.
dissertation, Dept. Comput. Sci. Eng., Vinayaka Missions Univ., Salem,
medical images by choosing the best features from the India, 2016.
images. Thus, a novel system of clinical image classifica- [9] S. Singh and N. Singh, ‘‘Object classification to analyze medical imaging
tion was presented on the basis of soft set to accomplish data using deep learning,’’ in Proc. Int. Conf. Innov. Inf., Embedded Com-
mun. Syst. (ICIIECS), Mar. 2017, pp. 1–4.
better execution regarding accuracy, precision and compu- [10] M. Frid-Adar, I. Diamant, E. Klang, M. Amitai, J. Goldberger,
tational speed. The performance measures of the proposed and H. Greenspan, ‘‘GAN-based synthetic medical image augmenta-
DL achieved 95.22% accuracy, 86.45% sensitivity and 100% tion for increased CNN performance in liver lesion classification,’’
2018, arXiv:1803.01229. [Online]. Available: http://arxiv.org/abs/1803.
specificity in DL classifier with optimal features (OCSA) 01229
model. The present strategies can be modified by including [11] V. Sharma and K. C. Juglan, ‘‘Automated classification of fatty and normal
other measurable features so as to get expanded unwavering liver ultrasound images based on mutual information feature selection,’’
IRBM, vol. 39, no. 5, pp. 313–323, Nov. 2018.
quality in the classification of difficult medical images. The [12] L. Zhang, K. Mistry, C. P. Lim, and S. C. Neoh, ‘‘Feature selection
technique presented here incurred maximum cost. Though it using firefly optimization for classification and regression models,’’ Decis.
yielded the optimal result, the precision must be improved Support Syst., vol. 106, pp. 64–85, Feb. 2018.
[13] S. K. Lakshmanaprabu, S. N. Mohanty, K. Shankar, N. Arunkumar, and
under the application of eliminating massive features and G. Ramirez, ‘‘Optimal deep learning model for classification of lung
prolonging the training dataset. In future investigations, the cancer on CT images,’’ Future Gener. Comput. Syst., vol. 92, pp. 374–382,
segmentation systems and some automatic classification Mar. 2019.
techniques should be used to identify the tumor part in med- [14] D. Gupta, S. Sundaram, A. Khanna, A. E. Hassanien, and
V. H. C. de Albuquerque, ‘‘Improved diagnosis of Parkinson’s disease
ical images and furthermore feature reduction strategy is using optimized crow search algorithm,’’ Comput. Electr. Eng., vol. 68,
likewise to be considered. pp. 412–424, May 2018.
[15] J. Ker, L. Wang, J. Rao, and T. Lim, ‘‘Deep learning applications
in medical image analysis,’’ IEEE Access, vol. 6, pp. 9375–9389,
CONFLICT OF INTEREST 2018.
No author expresses any conflict of interest about the publi- [16] S. J. S. Gardezi, M. Awais, I. Faye, and F. Meriaudeau, ‘‘Mammogram
classification using deep learning features,’’ in Proc. IEEE Int. Conf. Signal
cation of this paper. Image Process. Appl. (ICSIPA), Sep. 2017, pp. 485–488.
[17] G. I. Sayed, A. E. Hassanien, and A. T. Azar, ‘‘Feature selection via a
REFERENCES novel chaotic crow search algorithm,’’ Neural Comput. Appl., vol. 31, no. 1,
pp. 171–188, Jan. 2019.
[1] R. Bharath, P. K. Mishra, and P. Rajalakshmi, ‘‘Automated quantification [18] P. Jagatheeswari, K. S. Suresh, and M. Rajaram, ‘‘Contrast enhancement
of ultrasonic fatty liver texture based on curvelet transform and SVD,’’ for medical images based on histogram equalization followed by median
Biocybernetics Biomed. Eng., vol. 38, no. 1, pp. 145–157, 2018. filter,’’ in Proc. Int. Conf. Man-Mach. Syst. (ICoMMS), Batu Ferringhi,
[2] S. B. Park, J. W. Lee, and S. K. Kim, ‘‘Content-based image classifi- Malaysia, Oct. 2009.
cation using a neural network,’’ Pattern Recognit. Lett., vol. 25, no. 3, [19] T. Vo-Van, H. Che-Ngoc, and T. Nguyen-Trang, ‘‘Textural features selec-
pp. 287–300, Feb. 2004. tion for image classification by Bayesian method,’’ in Proc. 13th Int. Conf.
[3] P. Smitha, L. Shaji, and M. G. Mini, ‘‘A review of medical image classi- Natural Comput., Fuzzy Syst. Knowl. Discovery (ICNC-FSKD), Jul. 2017,
fication techniques,’’ in Proc. Int. Conf. VLSI, Commun. Intrumrnataiom, pp. 139–733.
2011, pp. 34–38. [20] S. Sannigrahi and P. Acharjee, ‘‘Implementation of crow search algorithm
[4] A. Qayyum, S. M. Anwar, M. Awais, and M. Majid, ‘‘Medical image for optimal allocation of DG and DSTATCOM in practical distribution
retrieval using deep convolutional neural network,’’ Neurocomputing, system,’’ in Proc. Int. Conf. Power, Instrum., Control Comput. (PICC),
vol. 266, pp. 8–20, Nov. 2017. Jan. 2018, pp. 1–6.

58016 VOLUME 8, 2020

R. J. S. Raj et al.: Optimal Feature Selection-Based Medical Image Classification Using DL Model in IoMT

[21] Accessed: Aug. 18, 2018. [Online]. Available: https://www.oasis- DENIS ALEXANDROVICH PUSTOKHIN recei-
brains.org/ ved the Ph.D. degree in logistics and supply
[22] Accessed: Aug. 18, 2018. [Online]. Available: http://www.via.cornell. chain management from the State University of
edu/lungdb.html Management, Moscow, Russia. He is currently
[23] Accessed: Aug. 18, 2018. [Online]. Available: http://adni.loni.usc.edu/ an Associate Professor with the State University
[24] K. Shankar, A. R. W. Sait, D. Gupta, S. K. Lakshmanaprabu, of Management. He has published over 30 con-
A. Khanna, and H. M. Pandey, ‘‘Automated detection and classifica- ferences and journal articles. His research inter-
tion of fundus diabetic retinopathy images using synergic deep learning
ests include enterprise logistics planning, artificial
model,’’ Pattern Recognit. Lett., vol. 133, pp. 210–216, May 2020, doi:
intelligence, big data, the Internet of Things, and
10.1016/j.patrec.2020.02.026.
[25] S. K, L. S. K., A. Khanna, S. Tanwar, J. J. P. C. Rodrigues, and N. R. Roy, reverse logistics network design.
‘‘Alzheimer detection using group grey wolf optimization based features
with convolutional classifier,’’ Comput. Electr. Eng., vol. 77, pp. 230–243,
Jul. 2019. DEEPAK GUPTA received the Ph.D. degree from
[26] M. Elhoseny, G.-B. Bian, S. K. Lakshmanaprabu, K. Shankar, A. K. Singh, Dr. A. P. J. Abdul Kalam Technical University.
and W. Wu, ‘‘Effective features to classify ovarian cancer data in He has completed his first Postdoctoral Research
Internet of medical things,’’ Comput. Netw., vol. 159, pp. 147–156, with Inatel, Brazil. He is an eminent academician,
Aug. 2019. plays versatile roles and responsibilities juggling
between lectures, research, publications, consul-
tancy, community service, and Ph.D. degree and
R. JOSHUA SAMUEL RAJ received the B.E. post-doctorate supervisions. With 12 years of rich
degree in computer science and engineering and expertise in teaching and two years in indus-
the M.E. degree in computer science and engi- try, he focuses on rational and practical learning.
neering from Anna University, Chennai, India, He has contributed massive literature in the fields of human–computer
in 2005 and 2007, respectively, the Ph.D. degree interaction, intelligent data analysis, nature-inspired computing, machine
from the Faculty of Information Technology, learning, and soft computing. He is working as an Assistant Professor
Kalasalingam University, India, in 2015, and the with the Maharaja Agrasen Institute of Technology (GGSIPU), New Delhi,
M.B.A. and the M.E. (literature) through the Dis- India. He has served as the Editor-in-Chief, the Guest Editor, an Associate
tance Education Program from Manonmaniam Editor of SCI and various other reputed journals (Elsevier, Springer, Wiley
Sundaranar University, Tirunelveli, India. He is a and MDPI). He has actively been a part of various reputed international
Life Member of the Indian Society for Technical Education and The Indian conferences. He is not only backed with a strong profile, but his innovative
Science Congress. He is currently serving as a Professor and the Research ideas, research’s end-results, and the notion of implementation of technology
Head of the Department of Information Science and Engineering, CMR in the medical field is by and large contributing to the society significantly.
Institute of Technology, Bengaluru. He has held key positions such as the He is currently a Postdoctoral Researcher with University of Valladolid,
vice principal, the principal, and the director of engineering colleges. He has Spain. He has authored/edited 33 books with national/international level pub-
authored over 50 articles in refereed international and domestic journals lisher (Elsevier, Springer, Wiley, and Katson). He has published 101 scien-
and conferences, and holds five patents. His areas of interests include grid tific research publications in reputed international journals and conferences
computing, cloud computing, data mining, deep learning, scheduling, image including 49 SCI-indexed journals of the IEEE, Elsevier, Springer, Wiley,
processing, and artificial intelligence. and many more. He is also associated with various professional bodies like
ISTE, IAENG, IACSIT, SCIEI, ICSES, UACEE, Internet Society, SMEI,
IAOP, and IAOIP. He was Invited as a Faculty Resource Person/Session
S. JEYA SHOBANA received the B.E. degree in Chair/Reviewer/TPC Member of different FDP, conferences, and journals.
computer science and engineering from Manon- He is the convener of ICICC conference series. He is the Editor-in-Chief
maniam Sundaranar University, Tirunelveli, India, of OA Journal-Computers, an Associate Editor of Expert Systems (Wiley),
in 2003, the M.E. degree in computer science and Intelligent Decision Technologies (IOS Press), the Journal of Computational
engineering from Anna University, Chennai, India, and Theoretical Nenoscience, an Honorary Editor of ICSES Transactions
in 2006, and the Ph.D. degree in computer science on Image Processing and Pattern Recognition. He is also a Series Editor
and engineering from Manonmaniam Sundaranar of Smart Sensors Technology for BioMedical Engineering (Elsevier), Intelli-
University, in 2016. She has 12 years of teaching gent Biomedical Data Analysis (De Gruyter, Germany), and Computational
experience in various institutions. She is currently Intelligence for Data Analysis (Bentham Science).
working as an Assistant Professor with the Depart-
ment of Computer Science, College of Science, Knowledge University, Erbil,
Kurdistan Region, Iraq. She has published more than seven articles in various K. SHANKAR is currently a Postdoctoral Fellow
international journals. Her fields of interests are wireless sensor networks, of Alagappa University, Karaikudi, India. Collec-
deep learning methods, artificial neural networks, information security, and tively, he has authored/coauthored over 41 ISI
data mining. journal articles (with total impact factor 107.749)
and over 150scopus-indexed articles. He has guest-
edited several special issues at many journals pub-
IRINA VALERYEVNA PUSTOKHINA received lished by Inderscience and MDPI. He has served
the M.B.A. degree in logistics and supply chain as a Guest Editor, an Associate Editor of SCI,
management and the Ph.D. degree in logistics and of Scopus indexed journals like Elsevier, Springer,
supply chain management from the State Univer- Wiley, and MDPI. He authored/edited conference
sity of Management, Moscow, Russia. She is cur- proceedings, book chapters, and two books published by Springer. He has
rently an Associate Professor with the Plekhanov been a part of various seminars, paper presentations, and research paper
Russian University of Economics, Moscow. She reviews, and a convener and a session chair of the several conferences.
has published over 40 conferences and journal arti- He displayed vast success in continuously acquiring new knowledge and
cles. Her research interests include supply chain applying innovative pedagogies, and he has always aimed to be an effective
management, regional logistics development, sus- educator and has a global outlook. His current research interests include
tainable urban development, city logistics, intelligent logistics systems, big healthcare applications, secret image sharing scheme, digital image security,
data technology and applications, information management, and the Internet cryptography, the Internet of Things, and optimization algorithms.
of Things.

VOLUME 8, 2020 58017