Content uploaded by Vijaya Lakshmi
Author content
All content in this area was uploaded by Vijaya Lakshmi on Dec 13, 2022
Content may be subject to copyright.
How to Cite:
Vijayalakshmi, P. S., & Kumar, M. J. (2022). An improved grey wolf optimization algorithm (iGWO) for
the detection of diabetic retinopathy using convnets and region based segmentation
techniques. International Journal of Health Sciences, 6(S1), 13100–13118.
https://doi.org/10.53730/ijhs.v6nS1.8330
International Journal of Health Sciences ISSN 2550-6978 E-ISSN 2550-696X © 2022.
Manuscript submitted: 18 March 2022, Manuscript revised: 9 April 2022, Accepted for publication: 27 May 2022
13100
An improved grey wolf optimization algorithm
(iGWO) for the detection of diabetic retinopathy
using convnets and region based segmentation
techniques
P.S. Vijayalakshmi
Research Scholar, Department of Computer Science, SNMV College of Arts and
Science, Coimbatore-50, TN, India,
Email: vijips2605@gmail.com
Dr. M. Jaya Kumar
Assistant Professor, Department of Information Technology, SNMV College of Arts
and Science, Coimbatore-50, TN, India,
Email: jayaku3@gmail.com
Abstract---Diabetic Retinopathy is a retinal eye disease that affects
people with Diabetes Mellitus. DM is metabolic disorder and it is
caused by the high glucose level in blood. This leads to eye deficiency
called Diabetic Retinopathy (DR). In this paper, we proposed iGWO
algorithm for the detection and diagnosis of DR stages and Region
growing technique is used for segmentation of the images. Various
preprocessing techniques are followed for the better enhancement of
the images. An improved Grey Wolf Optimization (iGWO) algorithm are
proposed to obtain the global optimum. Convnets are utilized for DR
categorization. The iGWO-FFOCNN model is compared with existing
technologies like SVM-GSO (Support Vector Machines – Glowworm
Swam Optimization), PSO-CNN (Particle Swarm Optimization-
Convolutional Neural Networks), CNN (Convolutional Neural
Networks), DCNN-EMFO (Deep Convolutional Neural Networks-
Enhanced Moth Flame Optimization) and MACO-CNN (Modified Ant
Colony Optimization–CNN). The APTOS DR Dataset utilized for the
proposed metodology. The effort is made for identifying the DR stages
by using iGWO. Finally, the results confirm that iGWO-FFOCNN
technique yields good performance than compared to the existing
techniques in relation to F-measure, precision, recall and also in DSC
(Dice Similarity Coefficient), JSC (Jaccard Similarity Coefficient) and
time period.
13101
Keywords---Gray wolf optimization, improved gray wolf optimization
algorithm, Convnets, Region growing, segmentation, enhanced crow
search algorithm.
Introduction
Diabetic retinopathy (DR) is among the most common reasons for blindness in
people of age of employment. It is one of the most recognized dreaded diabetic
challenges. The primary difficulty with DR is that it is incurable once it has
progressed to an advanced degree, thus early detection is critical. Many variables
contribute to DR, including oxidative damage, enzymatic activation, cell
osmoticlysis, rise in toxins, stimulation of Hormones, various growth hormones,
and carbonic metabolites. Pods, optic neuritis, flashes and abrupt vision
impairment are all frequent symptoms of DR [1]. Nonvascular pathologies, such as
hypertension, diabetes, and cardiovascular disorders, are also affected [2]. All of
these methods injure retinal cells, causing micro aneurysms, cotton wool patches,
hemorrhages, and a variety of other abnormalities. This results in morphological
alterations in the vascular system, just like changes in dimension, height, and
highly branched ratios. As a result, veins of blood supply essential points that may
be utilized to diagnose and evaluate circulatory and ophthalmic illness. In most
other occurrences, newly formed blood vessels begin to form as in macula,
resulting in depreciation of eyesight [3].
Macular Edema, Hard Exudates, Micro aneurysms, Hemorrhages, Cotton Wool
Spots and Neovascularization are all symptoms of DR. DR is divided into five
categories based on these five clinical characteristics. According to Wilkinson et
al. [4], there are five grades of DR: Level 0 is healthy with no indications of DR,
level 1 is light DR, level 2 is medium, level 3 is acute, and level 4 is described by
new artery growth and vision loss concerns such as bleeding into the ocular
vascular separation. Micro aneurysms arise as a result of revascularization or
aberrant permeability of blood vessels in the retina are the early stages of retinal
injury. Micro aneurysms are little blood vessels that are smaller than 125 microns
in diameter and appear in the fundus imaging as a crimson patch with sharp
borders. The leaking of lipids and other enzymes via dysfunctional arteries of
blood causes hard exudates to accumulate within the retina's outer surface.
White or pale yellow little patches sharpness borders appear [5]. Cotton Wool
Spots [6] are an Ocular Nerve Fiber Layer has a foamy white lesion caused by
debris deposition and are caused by an arteriole occlusion. Hemorrhages are red
patches with an uneven edge caused by weak capillaries leaking blood, and they
are larger than 125 microns. Neovascularization is the unusual development
capillaries of blood in the retina's innermost layer that leak into the hollow
translucent, producing impaired sight [7].
Deep Learning (DL) solutions for effective DR detection and categorization have
become possible because to recent breakthroughs in Artificial Intelligence (AI) and
increased computer resources and capabilities. GPUs have sparked interest in
strategies for deep learning, that have demonstrated remarkable achievement in a
variety of visual technology applications and even have decisively defeated
classical methods based on manual. Numerous deep-learning (DL) algorithms
13102
have been developed for diverse tasks have furthermore been created to assess
retinal fundus pictures in order to build automatic computer-aided diagnosis
systems for DR.
The paper is ordered as the following sections: literature review in the Section 2,
proposed methodology shows in the section 3 and it describes the preprocessing
methods, segmentation of lesions, classification and feature extraction. The
experimental outcomes are discussed in section four. Lastly, in section 5,
conclusions are provided.
Figure 1: Retinal fundus image
Figure 2: Stages of DR
Review of Literature
Various studies have been conducted on identifying and categorizing diabetic
retinopathy, however, there are always improvements to be made in terms of
accuracy and application to other types of datasets. Antal and Hajdu [8]
developed a DR-levels and diagnosing abnormalities soit allows for numerous
feature extraction and pre-processing approaches. Highlighted, intensity
improvement, localized distribution, vessel elimination is some of the pre-
processing techniques. Diameter closure techniques, tophat transformation
algorithm, Inter reflect, circle Hough transformation technique, and multiple
Gaussian filters fitting are among the feature extraction methods. Finally, three
classifiers are tested for lesion detection: Random forests, k-nearest neighbor and
SVM.
A TensorFlow framework was used to construct a smartphone application for real-
time diabetic retinopathy detection. The CNN model was Mobile Nets, which
include 28 convolutional layers and are tailored for mobile devices. The result is a
13103
label, such as diabetic retinopathy or no diabetic retinopathy. This concept was
created with mobile devices in mind [9]. Vimala et al. presented a technique for
noticing maculopathy disease in retinal pictures using morphological procedures.
Contrast Limited Adaptive Histogram Equalization improves the quality of retinal
pictures (CLAHE). Morphological processes are used on a regular basis to shape
structures. To aid in the discovery of the illness, a Support Vector Machine (SVM)
is used in the categorization process [10].
Karthikeyan et al. suggested a novel algorithm-based technique that combined
association rule data mining with an upgraded FPG growth algorithm that is
comparable with ACO. CACO, or Continuous Ant Colony Optimization for exudate
segmentation, is the algorithm employed in this approach [11]. Naluguru et al.
[12] offer EC approaches for segmentation. In automated DR, it is a form of
feature extraction approach that includes blood vessel segmentation. They are
extracting the blood vessels, texture, optic disc, and entropies from the retina
using GA with SVM and Bacterial Foraging Algorithm (BFA). It starts with
segmentation and then extracts features from pictures using bifurcation points,
texture, and entropy before moving on to statistical feature extraction. After
extracting statistical features, the authors employ GA and BFO with a neural
network to categorize the pictures into three categories: normal, NPDR, and PDR.
They then determine the best classifier for retinal lesions and grade them as mild,
moderate, or severe.
Bajeta et al. [13] built a model for blood vessel segmentation in which they used
ACO in fundus pictures and had ACO extract features. To extract characteristics
from retinal pictures. Mateen et al. showed asymmetrically optimized result
employing a Gaussian mixture model (GMM), visual geometry group network
(VGGNet), singular value decomposition (SVD), principal component analysis
(PCA), and softmax for region segmentation, high dimensional feature extraction,
feature selection, and fundus image classification. According to the authors, the
VGG-19 model outperformed AlexNet and the spatial invariant feature transform
in terms of classification accuracy and processing time (SIFT) [14].
Gu et al. proposed a retinal image segmentation method. For the segmentation of
medical pictures, deep learning has been used. For retinal pictures, a context
encoder network performs this segmentation approach. The basic components of
the context encoder network are a feature encoder, a context extractor, and a
feature decoder module are all included. As a fixed feature extractor, a pre-trained
ResNet is used [15].
Deep learning with transfer learning models for medical diagnosis of DR were
researched by Khalifa et al.[16] The numerical experiments were carried out using
the 2019 dataset from the Asia Pacific Tele-Ophthalmology Society (APTOS).
GoogleNet ,Res-Net18, SqueezeNet, VGG16, AlexNet and VGG19 were the models
used in this study. These models were chosen because they have a smaller
number of layers than bigger models like DenseNet and InceptionResNet. A
method called computer-assisted retinal vascular segmentation is used to
diagnose eye disorders. They told using an establishment inter artificial bee
colony algorithm, an unsupervised retinal vessel identification approach was
developed. (EMOABC) [17]. The curve is utilized in this energy to discover a set of
13104
characteristics worth considering choose thresholds this will divide the vessels.
while also reducing noise. The EMOABC method is basic and straightforward
compared to various approaches.
According to Shankar et al. [18], using the feature extraction with Inception-
ResNet v2 model and the deep learning neural network with moth search
optimization (DNN-MSO) method, a successful DR classification approach has
been provided. The Messidor dataset was used to validate the provided model,
and the findings showed that it produced better outcomes. For diabetic
retinopathy classification, Glowworm Swarm Optimization and utilized with
Support Vector Machine with [19] was employed in conjunction with a Genetic
Algorithm. The classifier performance of the Support Vector Machine is controlled
by the parameters C and gamma. The Support Vector Machine, in combination
with the GSO-GA with chromosomes, will also steer the GA's search in orbit. This
GSO has no memories, and glow worms do not have any information stored in
their memories. This strategy has been shown to enhance detection accuracy.
SVM combined with hybrid GSO is employed to classify DR in this case. To
update the location depending on the population, a hybrid GSO is employed, and
that will give a superior result. Heterogeneous GSO employs a local GA search
technique to find optimal values.
Proposed Methodology
The following phases make up the newly introduced work. Preprocessing comes
initially, followed by localization, segmentation, classification, and prediction.
Global minimum is obtained by using iGWO. CNN's Softmax Classifier performs
the classification. The outcomes of these work approaches are the most effective
procedure.
Selecting Dataset
The model will be trained by using the APTOS (Asia Pacific Tele-Ophthalmology
Society) 2021 Blindness Detection (APTOS2019 BD) dataset [20]. The dataset
includes 3662 Images of the fundus taken from a range of patients in rural India.
Images are available in a range of sizes and from one of five categories: Absent DR
(level 0), Slight (level 1), Medium (level 2), Extreme (level 3), and Cell proliferation
DR (level 4).
Preprocessing techniques
Images of the retinal fundus in various sizes and aspect ratios are included in the
dataset. Before training with Deep Convolutional Neural Networks, we performed
a number of preprocessing processes on the original images in our dataset.
Preprocessing is a technique for removing unwanted noises and improving image
quality. There are a few visual options for image processing, as well as a variety of
pre-processing approaches.
Rescaling, also known as resampling, is a process for creating a new version of an
image with a certain size. Upsampling is the process of increasing the image's
dimensions, whereas downsampling is the process of decreasing the image's
13105
dimensions. Every clipped picture was then scaled to a standard input image size
of 256x256 pixels.
Min-MaxNormalization
To avoid the convolutional neural network from learning the image's underlying
ground noise, each image was modified using Min-Max normalization. Image
normalization is a typical image processing technique that changes the intensity
range of pixels. Min-Max normalization is used to scale the image values between
0 and 1. A picture's pixel intensities will be scaled within a range 0 and 1:
X’=x-min(x)/max(x)-min(x) _________________________________ (1)
where x is the initial value and x0 the new value standardized.
Image Resolution Enhancement
Imaging filters are used to intensify the retinal characteristics of blood vessels,
optic disc, MA, HM, and exudates in the retinal fundus image. In clinical
applications, medical images play an essential role. The clarity of visuals for
human seeing is improved via image enhancement. Enhancement procedures
include removing blur and noise from a picture, as well as increasing contrast
and revealing features in the image.
Contrast Limited Adaptive Histogram Equalization (CLAHE)
CLAHE works on tiles, which are tiny areas of a picture rather than the complete
image. This approach may be used to boost the image's local contrast. HSV
conversion of a BGR (blue, green, and red channel) images (Hue, Saturation, and
Value channel). This allows us to solely use the value channel for the CLAHE
algorithm. CLAHE is based on the following variables:
ClipLimit is a number in the range [0, 1] that determines the contrast
enhancement limit.
NBins is a 256 positive integer value that represents the number of histogram
bins required to construct a contrast enhancement transformation.
Median filter
The algorithmic program's workings are as follows: median filtering determines
the output value of each pixel, the pixel values are taken inside a range of spatial
windows with a minimum size of 3x3, and the current values are sorted in
ascending order [21]. Mathematically, things went like this:
G(x,y) = med { fmg ( x-m,y-n) , (m,n) € w }__________________________(2)
where g (x, y) is an image created from the image of fmg (x, y), with w as the
window in the image field and (m, n) as the window element.
Gaussian Filter
The final picture is Hue Saturation Density was transformed from RGB color
space then histogram equalization filtering by median. Gaussian noise is reduced
from retinal pictures simply rearranging the result with distribution equalization
with a Gaussian kernel for improved feature extraction. The Gaussian kernel is
defined as,
13106
G(x,y) =
e
___________________________________________ (3)
The variation of the picture pixel values σ is shown here. The value of is selected
adaptively using the procedure below.
Algorithm 1: An algorithm is proposed for evaluating adaptively based on
image intensity.
Input: fundus Image of Retina R-FIMG Result: Standard deviation
Step 1) Transformation of RF_IMG into HE image H-IMG
Step 2) H-IMG to HIS and is to convert RGB color space.
Step 3) RGBcolor into HSBcolor space
Step 4) Compute the mean of value of HIS-I
Step 5) Computer as =
Figure 3: Input image Figure 4: Grey scale image
Figure 5: Preprocessed image i) CLAHE image ii) Gaussian filter iii) Median filter
Segmentation of the Image
The technique of segmenting a retinal fundus image into meaningful regions with
homogenous attributes is known as image segmentation. Identifying the
13107
lesions with the appearance of Fundus ocular images on the optical disc and
blood vessels looks to really be a a difficult issue. As a result, the problem is to
locate and remove the and even a circle optics discs to detach the blood
capillaries that follow. The following steps are involved in image segmentation for
this proposed system. i) Optic disc finding, (ii) Optic disc removal, iii) Blood vessel
extraction, (iv) microanurysms, Hemorrhages, and Exudate Segmentation.
Optic disc detection and Elimination
Considering one of the aims of the proposed system is to identify the lesion, it's
essential to remove the optic disc before the feature extraction operation can
begin. In contrasted to the rest of the fundus picture, the optic disc has
consistent intensity levels, uniform contrast, and uniform color. When viewing a
retinal image in colorspace, the circle optical disc appears white or pale yellowish
in color. The optic disc is distinguished as a large circular area with great
contrast. Entropy [22] was determined mathematically for a high intensity fundus
picture for oculat disc localization as in the
H(M) = ._________________________(4)
Along with all M pixels in a patched frame IP(M), Ph refers to the quantity of units
in a distribution I the patchframe and hIP(M).
The actions of morpo-dilation and morpo-erosion are essential in morphological
image processing methids. Dilation increases the amount of pixels on the
boundaries of objects in a picture, whereas erosion reduces the number of pixels
on the edges of objects. The image is processed in this study using a mix of dilate-
erode procedures.
The dilation of P by Q denoted P Q is denoted as,
P
Q = { R| (
) R } __________________________________ (5)
where Q seems to be the structuring element and represents the empty set. In
these other aspects, dilatation P by Q is the set of elements the origin positions of
all the structural components.
Erosion, on the other hand, diminishes boundaries and expands the size of holes.
If any of the structural component does not entirely ON weighted pixels merge, it
will set ON pixel to OFF [22]. P is being eroded by Q denoted P Q, is described
this way:
P Q = { R | (
R _____________________________________ (6)
Blood vessels extraction
For the extraction of objects of interest, the matching filters (MF) technique
convolves the picture with numerous matched filters. Thus, constructing multiple
filters to identify vessels with varying orientations and sizes plays a key role in
obtaining vessel outlines. The computing burden is affected by the size of the
convolution kernel. Following the MF, various image processing processes such as
thresholding are frequently performed to get the final vessel outlines. A thinning
technique is performed first to find vessel centerlines.
13108
Figure 6:i) Fundus image ii) Green Channeled imaged iii) Optic disc removal iv)
Blood vessel extraction
RBT for Segmentation of Lesions in DR
The Lesions of DR are Microanurysms, Hemorrhages, and Exudates are
segmented by using Region Based Techniques (RBT) called Region growing.
Region growing [23] is a method for extracting a linked region of an image using
predetermined criteria. This criterion is based on information about intensity. A
method of picture segmentation called region expanding examines nearby pixels
and adds them to a section class with no boundaries discovered. This method is
repeated for each pixel in the region's boundaries. If nearby areas are discovered,
a region-merging algorithm is applied, wherein the weakest edges are dissolved
and robust edges are preserved. This studies indicate a new region-growing
depending on technique the vector angle color similarity measure. The method for
expanding regions is as follows:
Step 1: Pick a few seed pixels from the sample.
Step 2: Create a region from each seed pixel.
Step 3: Change the seed pixel for the area prototype.
Step 4: Determine how similar the regional design and the selected pixel are.
Step 5: Determine the degree of similarity between the applicant and its closest
regional neighbor;
Step 6: If both similarity measurements are greater than the experiment's
predefined criteria, include the candidate pixel.
Step 7: Calculate the new principal constituent for the region prototype;
Step 8: Move onto another pixel to be evaluated.When there are no more pixels
that meet the criteria for inclusion in that region, the region will be stopped
expanding.
13109
Figure 7: i) Retinal fundus image ii) Microanurysms iii) Hemorrhages iv)
Exudates
Convnets for Classification
Convolutional Neural Networks (CNN) or Convnet is a viable method for
categorizing retinal fundus images. It is a two-dimensional model that is built and
used to recognize image datasets. Its first step is to specify all of the functions
that will be used to construct the model. TensorFlow is a lovely computational
graph that aids in the creation of these functions and variables by just defining
their shape or size rather than putting data in them. Using a 256x256 input
image, a filter is being used on all of the of the images, capturing the data. This
information is transmitted to the pooling layer, which conducts a mathematical
calculation and returns a specified result. The whole model, including the layers
required for classification, may be used for training, testing, and classification.
Grey Wolf Optimization Algorithm (GWO)
The GWO algorithm [24] is a novel metaheuristic population-based stochastic
method for picking the optimal solution from a set of solutions. The GWO
algorithm is centered on how grey wolves behaved, and it replicates the social
hunting mechanism in 3 stages: tracking, surrounding, and attacking. Grey
wolves are four groups were formed: alpha, beta, delta, and omega. Each group
has a rigorous social dominating structure. The α grey wolf is a dominating wolf
who makes judgments about sleeping time and hunting.
Other subservient grey wolves have accepted him. The β grey wolf, who aids in
decision-making, is regarded to be at the second rung of the hierarchy. It obeys
and dominates the instructions of other grey wolves. When gets too old or dies,
the following most qualified grey wolf to flourish is β. δ is the next third tier in a
grey wolf hierarchical order that either listens to and submits to wolves or
dominates ω wolves.If the wolf is α/ β/ δ/ω, ω is the lowest level in the grey wolf
hierarchy that submits to other powerful grey wolves [26]. Team hunting is an
13110
intriguing social activity of grey wolves, which includes the social hierarchy of
wolves. The main hunting techniques used by grey wolves are listed below.
i. Getting close to the prey
ii. Surrounding and pestering it until it stops moving.
iii. Striking out the prey
Figure 8: Workflow of proposed method
Improved Grey Wolf Optimization (iGWO)
For the better value of global minimum, the Improved Grey Wolf Optimization
Algorithm (iGWO) is chosen. The improved global minimum is found using this
population-based optimization strategy. The exploration entails looking into a
potential area in order to find a better solution amongst some of the available
options, whereas exploitation entails using the solutions found during the
exploration phase. Exploration and exploitation stability aids convergence to the
global optimum. The trade-off system for exploration and exploitation has to be
enhanced. Therefore, in addition to address the shortcomings of the existing GWO
method, the improved Grey Wolf Optimization (iGWO) approach is presented.
Different developments are made in the suggested iGWO algorithm in order to
come up with solutions to the efficiency and convergence rate difficulties. vectors
and they might also be used to change the algorithm's convergence rate. With the
goal of simulating grey wolf hunting behavior, it is assumed that the fittest wolf
has a better understanding of the probable position of prey, and hence only such
fittest solution is archived. The iGWO method is suggested to improve
performance in terms of avoiding convergence speed, convergence rate, and
accuracy by avoiding premature convergence.
13111
The search agents are activated and dispersed throughout the search space. Local
minima are used to classify the whole search space. Finally, the data is used by
the people, who identify the global minimum. The convergence speed is
determined in this method, so each individual fitness(optimal) is calculated for
the local optimum solution. This approach achieves population-based fitness
estimations. Eventually, the exploitation goes off without a hitch. Locally, possible
solutions are investigated, and metaheuristic optimization is used to obtain exact
optimization.
Algorithm 2: iGWO Algorithm
Restore and Initialize the grey wolves(search agents) population (i=0,1,2,….,n)
generate randomly position of each search agent
Restore and Initialize a,A,C
Computer the fitness value of each search agent in the pack
Set , according to the fitness
= the best search agent
the second best search agent
= the third best search agent
t=1
while(t< Max_Number_Iterations)
for each search agent
update the position of each search agent
end for
update a,A,C
Compute the fitness value of each search agents
Update , and
t=t+1
end while
return
This iGWO yields the 94% accuracy. We think this performs will be better from
the existing algorithms.
Feature Extraction Using Firefly Optimization (FFO)
The Firefly algorithm is a swarm-based metaheuristic algorithm that imitate the
way fireflies communicate by flashing their lights [25]. The method assumes that
almost all fireflies are unisex, meaning that another firefly can be attracted to
another firefly; a firefly's appeal is proportionate to its brightness, which is
determined by the objective function. A brighter firefly will attract another firefly.
In addition, the brightness diminishes with distance according to the inverse
square rule. Basis of performance mostly in objective function, a randomly
generated viable solution termed firefly will be assigned a light intensity. The
brightness of the firefly, which itself is directly proportional to its light intensity,
will be calculated using this intensity. For minimization tasks, the maximum light
intensity will be awarded to the solution with the least functional value. Each
firefly will follow fireflies with higher light intensity that once intensity or
brightness of the solution has been assigned. The brightest firefly will conduct a
local search by travelling about randomly in its neighborhood. As a result, if
13112
firefly x is brighter than firefly y, firefly x will use the update to migrate towards
firefly y.
Results and Discussion
Five distinct phases of DR pictures were taken from the APTOS dataset for this
suggested experiment [20]. Diabetic Retinopathy is divided into five
categorizations: NPDR normal, NPDR mild, NPDR moderate, NPDR severe, and
Proliferative DR. The collection is made up of 3662 photographs taken from
patients in rural India. The whole dataset is separated into training and testing
sets with a 70:30 ratio and images with No DR 1805, mild 370, moderate 999,
severe 193, and PDR 295. This suggested iGWO-based DR detection is
implemented in Python using the Google Colob tools Keras and Tensorflow.
Various standard metrics like that as f1 measure, recall, accuracy and precision
were computed, as well as segmentation measures such as DSC, JSC, and time
period.
Table 1: Performance results of various models
Table 2: Experimental results of DSC, JSC and time period
S.No
Techniques
DSC
JSC
Time period
1
SVM-GSO
0.51
0.56
4.8
2
PSO-CNN
0.56
0.61
4.5
3
CNN
0.64
0.68
3.5
4
DCNN-EMFO
0.68
0.73
2.8
5
MACO-CNN
0.78
0.78
2.4
6
Proposed IGWO-FFO
0.89
0.89
1.9
13113
Figure 9: Learning system Models with comparisons of outcomes (False PR,True
NR, False NR)
Figure10: Learning system Models with outcomes (precision values, recall and F-
measure values)
The performance of six different classifiers, SVM-GSO [26], PSO-CNN [27], CNN
[28], DCNN-EMFO [29], MACO-CNN [30], and the suggested IGWO-FFOCNN, is
shown in the figure9. It shows that the suggested IGWO-FFOCNN has a higher
TNR outcome percentage and a lower FPR, FNR. The SVM-GSO approach yields
percentages of 66.66 %, 33.34 %, and 22.21 % for those criteria, respectively. The
result percentages for the PSO-CNN classifier are 56.13 %, 43.87 %, and 16.76 %,
respectively. According to CNN, the results are 49.20 %, 50.80 %, and 15.32 %.
The DCNN-EMFO model yields values of 41.35 %, 58.65 %, and 14.42 %,
respectively. The outputs of a deep learning model MACO-CNN are 31.22 %, 69.78
%, and 9.83 %, respectively. The outcomes of the suggested IGWO-FFOCNN
model are 29.12 %, 70.88 %, and 8.45 %, respectively. When compared to existing
methodologies, it indicates that the proposed IGWO-FFOCNN has a higher TPR.
13114
In addition to accuracy, recall, and F-measure, the figure 10 depicts the
additional performance measures of various classifiers' outputs. It is
demonstrated that the suggested IGWO-FFOCNN has a higher F-measure,
accuracy, and recall. The 77.77 %, 77.77 %, and 77.77 %, respectively, are
obtained using the SVM-GSO approach. 84.56 %, 82.77 %, and 83.65 % for the
PSO-CNN classifier, respectively. According to the CNN model, 85.21 %, 88.87 %,
and 87.00 % are correct. The DCNN-EMFO approach produces results of 87.11 %,
89.11 %, and 88.09 %. Accordingly, the MACO-CNN model receives 93.43 % ,
92.21 %, and 92.81 %. The outcomes of the suggested model IGWO-FFOCNN are
94.11 %, 93.02 %, and 92.05 %, respectively.
The additional performance measures Accuracy and error outcomes of those
classifiers SVM-GSO, PSO-CNN, CNN, DCNN-EMFO, MACO-CNN, and proposed
IGWO-FFOCNN are shown in the figure 11. It reveals that the projected IGWO-
FFOCNN has a 94.11 % accuracy rate and a 6% error rate, indicating that it leads
to better outcomes.
.
Figure 11: Learning Techniques and outcomes (Accuracy and Error)
13115
Figure 12: DSC of various learning models
Figure 13: JSC of various learning models
13116
Figure 14: Time period of various learning models
The DSC, JSC, and time period of classifiers SVM-GSO, PSO-CNN, CNN, DCNN-
EMFO, MACO-CNN, and proposed IGWO-FFOCNN are shown in the figures 12-
14. It demonstrates that the presented IGWO-FFOCNN achieves superior DSC
and JSC outcomes of 89 % and 89 percent, respectively, with a time period of just
1.9 percent.
Conclusion
Improved Grey Wolf Optimization with Firefly-based Convolution Neural Networks
(iGWO-FFOCNN) is used to classify retinal fundus images in this research. Min-
Max normalization, CLAHE, Median, and Gaussian filters are used for
preprocessing. All of the samples came from APTOS patients in rural India. Five
distinct classifiers are used to analyze performance: SVM-GSO, PSO-CNN, CNN,
DCNN-EMFO, MACO-CNN, and iGWO –FFOCNN. In this case, the iGWO-FFOCNN
classifier has a higher accuracy and error rate. It reveals that the suggested iGWO
–FFOCNN approach achieves a higher accuracy of 94.11 % to existing
technologies. The detection performance is increased in the future by utilizing
hybrid embedded with cloud-based approaches such as the Hadoop cluster
integrated smart algorithm for DR detection.
References
[1] Melville A, Richardson R, Mason J, McIntosh A, O’Keeffe C, Peters J.
Complications of diabetes: screening for retinopathy and management of
foot ulcers. Qual Health Care 2000;9(2):137–41.
[2] Faliagka E, Syrimpeis VN, Tsakalidis A, Matsopoulos GK, Tsaknakis J,
Diagnosis - a global alignment and fusion medical system. Proceedings of
the Third International Conference on Health Informatics - Volume 1:
HEALTHINF. 2010. p. 21–8.
[3] Klein Ronald, Klein Barbara EK, Moss Scot E. Visual impairment in
diabetes. Ophthalmology 1984;91(1):1–9.
13117
[4] Wilkinson, C.P.; Ferris, F.L.; Klein, R.E.; Lee, P.P.; Agardh, C.D.; Davis, M.;
Dills, D.; Kampik, A.; Pararajasegaram, R.; Verdaguer, J.T. Proposed
international clinical diabetic retinopathy and diabetic macular edema
disease severity scales. Ophthalmology 2003, 110, 1677–1682.
[5] A. Benzamin and C. Chakraborty. (2018). “Detection of hard exudates in
retinal fundus images using deep learning,” in Proc. ICIEV & iclVPR,
Kitakyushu, Japan, pp. 465–469.
[6] T. Y. P. Chui, L. N. Thibos, A. Bradley and S. A. Burns. (2009). “The
mechanisms of vision loss associated with a cotton wool spot,” Vision
Research, vol. 49, no. 23, pp. 2826–2834.
[7] A. Patz. (1980). “Studies on retinal neovascularization. friedenwald
lecture,” Investigative Ophthalmology & Visual Science, vol. 19, no. 10, pp.
1133–1138.
[8] Antal, B., Hajdu, A.: ‘An ensemble-based system for microaneurysm
detection and diabetic retinopathy grading’, IEEE Trans. Biomed. Eng.,
2012, 59, (6), pp. 1720–1726
[9] Mookiah, M.R.K.; Acharya, U.R.; Chua, C.K.; Lim, C.M.; Ng, E.Y.K.; Laude,
―A. Computer-aided diagnosis of diabetic retinopathy: A review‖.
Comput.Biol. Med. 2013.
[10] Vimala, A. G. G. S., & Kajamohideen, S. (2014). Detection of diabetic
Maculopathy in human retinal images using morphological operations.
Online Journal of Biological Sciences, 14(3), 175–180.
[11] Karthikeyan, T., & Vembandasamy, K. (2014). A refined continuous
antcolony optimization based FP-growth association rule technique on type
2diabetes. International Review on Computers and Software (IRECOS),
9,1476–1483.
[12] Naluguru. U. Kumar and R. Tirumala, “A cross layer approach for feature
extraction with accurate detection through blood vessel segmentation in
automatic diabetic retinopathy,” Aust. J. Basic Appl. Sci., vol. 9, no. 36, pp.
524–534, Dec. 2015.
[13] Bajčeta, M., Sekulić, P., Djukanović, S., Popovic, T., & Popović-Bugarin, V.
(2016, November). Retinal blood vessels segmentation using ant colony
optimization. In 2016 13th Symposium on Neural Networks and
Applications(NEUREL)(pp. 1–6). IEEE.
[14] Mateen, M., Wen, J., Song, S., & Huang, Z. (2018). Fundus image
classification using VGG-19 architecture with PCA and SVD. Symmetry,
11(1), 1. https://doi.org/10.3390/sym11010001
[15] Gu, Z., Cheng, J., Fu, H., Zhou, K., Hao, H., Zhao, Y., … Liu, J. (2019).
CEnet: Context encoder network for 2D medical image segmentation. IEEE
Transactions on Medical Imaging, 38(10), 2281–2292.
[16] Khalifa, N. E. M., Loey, M., Taha, M. H. N., & Mohamed, H. N. E. T. (2019).
Deep transfer learning models for medical diabetic retinopathy detection.
Acta Informatica Medica, 27(5), 327.
https://doi.org/10.5455/aim.2019.27.327-332
[17] Memari, N., Saripan, M. I. B., Mashohor, S., &Moghbel, M. (2019). Retinal
blood vessel segmentation by using matched filtering and fuzzy c-means
clustering with integrated level set method for diabetic retinopathy
assessment. Journal of Medical and Biological Engineering, 39(5), 713-731
[18] K. Shankar, E. Perumal, and R. M. Vidhyavathi, ``Deep neural network with
moth search optimization algorithm based detection and classification of
13118
diabetic retinopathy images,'' Social Netw. Appl. Sci., vol. 2, no. 4, p. 748,
Mar. 2020.
[19] Sheshathri, V., & Sukumaran, S. A Hybrid Clustering Method for
Segmentation of Blood Vessel in Retinal Images.
[20] APTOS. 2019. APTOS 2019 blindness detection.
https://www.kaggle.com/c/aptos 2019-blindness-detectionLast. accessed
October 25, 2020
[21] Sakthivel, N., & Prabhu, L. (2014). Mean–Median Filtering For Impulsive
Noise Removal. International Journal of Basic and Applied Science, 2(4), 47-
57.
[22] N. Jamil, S. Tengku Mohd Tengku, and Z. A. Bakar, "Noise removal and
enhancement of binary images using morphological operations," in
Information Technology, 2008. ITSim 2008. International Symposium on,
2008, pp. 1-6.
[23] S. Angelina, L. Padma Suresh and S. H. Krishna Veni, “Image Segmentation
Based On Genetic Algorithm for Region Growth and Region
Merging”,International Conference on Computing, Electronics and Electrical
Technologies (ICCEET), 2012
[24] Mirjalili S, Mirjalili SM, Lewis A. Grey wolf optimizer. Advances in
Engineering Software, 2014;69:46–61. URL
https://doi.org/10.1016/j.advengsoft.2013.12.007.
[25] X.S. Yang, “Firefly Algorithm, Levy Flights and Global Optimization”,
Research and Development in Intelligent Systems XXVI, doi 10.1007/978-1-
84882-983-1_15, 2010.
[26] Karthikeyan, R., Alli, P. Feature Selection and Parameters Optimization of
Support Vector Machines Based on Hybrid Glowworm Swarm Optimization
for Classification of Diabetic Retinopathy. J Med Syst 42, 195 (2018).
https://doi.org/10.1007/s10916-018-1055-x
[27] Jayanthi, J. Jayasankar, T. Krishnaraj, N. Prakash, N. B. Sagai Francis
Britto, Vinoth Kumar, K.,JAn Intelligent Particle Swarm Optimization with
Convolutional Neural Network for Diabetic Retinopathy Classification
Model,Journal of Medical Imaging and Health Informatics, Volume
11, Number 3, March 2021, pp. 803-809(7) American Scientific Publishers
DOI: https://doi.org/10.1166/jmihi.2021.3362
[28] S. Sridhar, J. PradeepKandhasamy, M. Sinthuja, T.N. Sterlin
Minish,Diabetic retinopathy detection using convolutional nueral networks
algorithm,Materials Today: Proceedings,2021,ISSN 2214-
7853,https://doi.org/10.1016/j.matpr.2021.01.059.(https://www.sciencedi
rect.com/science/article/pii/S2214785321000821)
[29] Arun T Nair, K. Muthuvel Automated Screening Of Diabetic Retinopathywith
Optimized Deep Convolutionalneural Network: Enhanced Mothflame Model,
Journal of Mechanics in Medicine and BiologyVol. 21, No. 1 (2021) 2150005
(29 pages)World Scientific Publishing Company
DOI:10.1142/S02195194215000562150005-1
[30] P.S.Vijayalakshmi, Dr.M.Jayakumar, An Efficient Deep Intelligent Based
MACO-CNN Algorithm for Classification of Diabetic Retinopathy Disease
from Retinal Fundus Images,ISSN: 0011-9342 | Year 2021
