Received 3 March 2023, accepted 11 March 2023, date of publication 16 March 2023, date of current version 22 March 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3257988

Internet of Things and Deep Learning Enabled

Diabetic Retinopathy Diagnosis Using Retinal
Fundus Images
THANGAM PALANISWAMY , (Member, IEEE), AND MAHENDIRAN VELLINGIRI
Department of Electrical and Computer Engineering, Faculty of Engineering, King Abdulaziz University, Jeddah 21589, Saudi Arabia
Corresponding author: Thangam Palaniswamy ([email protected])
The authors extend their appreciation to ‘‘the Deputyship for Research & Innovation, Ministry of Education in Saudi Arabia for funding
this research work through the project number G: 693-144-1443’’ and King Abdulaziz University, Deanship of Scientific Research (DSR),
Jeddah, Saudi Arabia.

ABSTRACT Recently, the Internet of Things (IoT) and computer vision technologies find useful in different
applications, especially in healthcare. IoT driven healthcare solutions provide intelligent solutions for
enabling substantial reduction of expenses and improvisation of healthcare service quality. At the same
time, Diabetic Retinopathy (DR) can be described as permanent blindness and eyesight damage because
of the diabetic condition in humans. Accurate and early detection of DR could decrease the loss of damage.
Computer-Aided Diagnoses (CAD) model based on retinal fundus image is a powerful tool to help experts
diagnose DR. Some traditional Machine Learning (ML) based DR diagnoses model has currently existed
in this study. The recent developments of Deep Learning (DL) and its considerable achievement over
conventional ML algorithms for different applications make it easier to design effectual DR diagnosis model.
With this motivation, this paper presents a novel IoT and DL enabled diabetic retinopathy diagnosis model
(IoTDL-DRD) using retinal fundus images. The presented Internet of Things Deep Learning – Diabetic
Retinopathy Diagnosis (IoTDL-DRD) technique utilizes IoT devices for data collection purposes and then
transfers them to the cloud server to process them. Followed by, the retinal fundus images are preprocessed to
remove noise and improve contrast level. Next, mayfly optimization based region growing (MFORG) based
segmentation technique is utilized to detect lesion regions in the fundus image. Moreover, densely connected
network (DenseNet) based feature extractor and Long Short Term Memory (LSTM) based classifier is used
for effective DR diagnosis. Furthermore, the parameter optimization of the LSTM method can be carried out
by Honey Bee Optimization (HBO) algorithm. For evaluating the improved DR diagnostic outcomes of the
IoTDL-DRD technique, a comprehensive set of simulations were carried out. A wide ranging comparison
study reported the superior performance of the proposed method.

INDEX TERMS Computer aided diagnosis, deep learning, diabetic retinopathy, fundus images, honey bee
optimization.

I. INTRODUCTION umbrellas, Air Conditioners (AC), etc. [1]. IoT method will
IoT is defined as practice of molding and designing Internet- support cloud computing (CC) for enhancing the operation
connected Things by computer networks. The word‘Io’ indi- regarding higher resource application, memory, power along
cates instead of utilizing smaller effective devices such as with computational capability. Moreover, CC obtains advan-
smart phone, laptops, and tablets, it can be better to utilize a tages from IoT via rising its scope for managing real-time
number of minimal effectual gadgets like fridges, wrist bands, applications and providing many services in dynamic format
and distributed way [2]. The IoT-related CC is expanded for
The associate editor coordinating the review of this manuscript and advancing innovative services and applications in a health-
approving it for publication was Carmelo Militello . care platform. In the year 2017, nearly 451 million individuals

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

27590 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
T. Palaniswamy, M. Vellingiri: IoT and DL Enabled DR Diagnosis Using Retinal Fundus Images

suffered from this disease [3]. A high blood sugar level method can measure the image features into different classes
severely weakens body organs and causes complications that provide the severity level of DR disease. The researchers
like retinopathy, coronary episode, cataracts, vision loss, in [12] devise a DL-enabled optimized feature selection tech-
dementia, and glaucoma. A rising population, irrespective nique for classifying the phase of DR severity from the fun-
of age, suffering from diabetic mellitus has issues in vision dus image. The candidate lesion regions are identified with
called DR [4]. Though DR is mostly symptomless in the the help of an Attention-based Fusion Network (AFU-Net).
initial levels, neural retinal damage and clinically unseen Afterward, texture and shape features were derived, and con-
micro vascular changes grow at the time of initial phases. sequently, optimum feature subset was chosen by making use
Therefore, regular eye screening is mandated for patients of the Improved Harris Hawk Optimizing technique. At last,
affected by diabetes, as prompt diagnosis of the condition a deep CNN categorizes the DR phases, and the model weight
becomes necessary [5]. Though earlier identification of DR can be updated through the same technique.
depends upon operational changes in Electro Retino-Graphy In [13], modelled an Autoregressive-Henry Gas Sailfish
(ERG), retinal blood vessel caliber, and retinal blood flow, in Optimization (Ar-HGSO)-related DL method for identifying
medical practices initial identification depends on fundus DR and severity level classifiers of DR and Macular Edema
analysis [6]. Fundus photography was a well-tolerated, (ME) related to color fundus images. The segmenting pro-
rapid, non-invasive, and broadly accessible that forms one cedure becomes highly essential for correct identification
method that is mostly utilized for assessing the extent of and classifying procedure that segregates the image into sev-
DR [7]. Using fundus images, ophthalmologists monitor eral subgroups. The DL technique can be used for effectual
retina lesions at higher resolution for diagnosing DR and detection of DR and severity classification of ME and DR.
predicting its severity. But manually DR identification from Further, the DL approach can be well-trained by the proposed
fundus images requires higher level of expertise and effort Ar-HGSO method to gain superior outcomes. Aljehane [14]
by a specialized ophthalmologist, particularly in remote areas presents an intelligent moth flame optimization with Incep-
or densely populated counties such as Africa and India [8], tion network-based DR detection and grading (IMFO-INDR)
wherein the number of people with DR and diabetes can approach. The aim of this IMFO-INDR algorithm was to
be projected to increase dramatically in the following years, identify the presence of lesions from the fundus image and
whereas the ophthalmologists are disproportionally low. This allot suitable class labels to it. Inception v4 method can be
has induced the research field for advancing computer-aided implemented as extracting feature and the hyper-parameters
diagnosis (CAD) system that minimizes the effort, cost, indulged in it were done optimal tuning by using MFO
and time required by a medical professional for identifying approach. Finally, SoftMax classifier can be employed for
DR [9]. New progressions in Artificial Intelligence (AI) and allocating class labels to the input fundus images.
the rise of computation resources and abilities have consti- Gunasekaran et al. [15] use RNN for the purpose of retriev-
tuted the opportunity for advancing deep learning applica- ing features from deep networks. So, utilizing computa-
tions for precise DR classification and detection [10]. tional techniques for identifying some ailments automatically
This paper presents a novel IoT and DL enabled diabetic becomes effective solution. The author tested and advanced
retinopathy diagnosis model (IoTDL-DRD) using retinal fun- numerous iterations of a DL structure for forecasting the
dus images. The presented IoTDL-DRD technique utilizes development of DR in diabetic persons who experienced tele-
IoT devices for data collection purposes and then transfers retinal DR assessment. A collection of 3-or 1-field color
them to the cloud server to process them. Followed by, fundus pictures is the input for both iterations. Using the
the retinal fundus images are preprocessed to remove noise presented DRNN method, advanced detection of the diabetic
and improve contrast level. Next, mayfly optimization based state has been carried out by HE identified in a blood vessel
region growing (MFORG) based segmentation technique is of eyes. Gupta et al. [16] project an Optimal Deep CNN
utilized to detect lesion regions in the fundus image. More- for Retinal Fundus Image Classification (ODCNN-RFIC)
over, densely connected network (DenseNet) based feature approach. The proposed method includes preprocessing in
extractor and long short term memory (LSTM) based classi- 2 stages they are Adaptive Median Filter (AMF) and Guided
fier is used for the effective DR diagnosis. Furthermore, the Filter (GF). The U-Net method was used for segmenting
parameter optimization of the LSTM model is carried out by images, permitting the detection of infected regions appro-
honey bee optimization (HBO) algorithm. For evaluating the priately. In [17], a new 2-phase Glaucoma Diagnosis Network
improved DR diagnostic outcomes of the IoTDL-DRD tech- (ODGNet) and Optic Disk localization were presented. In the
nique, a comprehensive set of simulations were carried out. initial stage, a visual saliency map merged with shallow
CNN can be employed for effective OD localizing from the
II. LITERATURE REVIEW fundus images. In the second stage, the TL-related pre-trained
Rajavel et al. [11] present a stochastic neighbor embedding methods were employed for glaucoma diagnosis.
(SNE) feature extracting technique with a view to removing
unnecessary noise and dimensional reduction from the fundus III. PROPOSED MODEL
images. Once the extraction of features is done, the pre- In this paper, a novel IoTDL-DRD method was projected
sented optimized deep belief network (O-DBN) classification for DR detection and classification on retinal fundus images.

VOLUME 11, 2023 27591

 T. Palaniswamy, M. Vellingiri: IoT and DL Enabled DR Diagnosis Using Retinal Fundus Images

The presented IoTDL-DRD method allows the IoT devices every grid point, the histogram of region H I1/2 nearby the
for data collection purposes and then transfers to the cloud image is defined. Then, the histogram is presented over the
server to process them. Then, the retinal fundus images are level of estimated clip limit is trimmed and the Cumulative
preprocessed for noise removal and contrast improvement. Distributive Function (CDF) is evaluated. The CDF is evalu-
Afterward, the MFORG technique was utilized for lesion seg- ated for pixel and the value based on 0 to 255. Following, for
mentation in the fundus images. For DR recognition and clas- all the pixels, the four nearby neighboring grid points should
sification, the HBO algorithm with LSTM model is exploited. be found. By employing the intensity rate, viz, the pixel index,
Figure 1 depicts the block diagram of IoTDL-DRD approach. and the mapping functions of the 4 grid points, CDF measure
is evaluated. However the procedure is ended, still it accom-
A. IMAGE PRE-PROCESSING plishes the concluding index of a window. Computer vision is
At the primary level, the IoT devices capture the retinal a subset of the Deep Learning model, enables the computer to
fundus images and forward them to the cloud server for possess an elevated level of understanding in the process of
further examination. Next, the median filtering (MF) tech- video and digital image processing. The proposed IoTDL –
nique is used to remove noise and CLAHE is applied to DRD method imparts the knowledge on the proposed subject
improve contrast level [18]. CLAHE algorithm is extremely matter to the computer vision, such that it can able to detect
effectual in biomedical image analyses and is chiefly utilized the diabetics retinopathy more effectively than the diagnosis
for improving the contrast level of an image. This method through the conventional methodologies.
needs a set of two input parameters such as dimension of clip
limit and sub window. In such a way, the presented algorithm B. IMAGE SEGMENTATION
recognizes the grid size of window, and also the grid value Once the images are preprocessed, they are passed into
originated from top left corner area of an image where the the MRORG technique to segment the fundus images.
computation starts from preliminary index of window. For Region growing (RG) is a pixel-based segmentation tech-
nique whereby the similarity constraint including intensity,
texture, and so on, are taken into account for grouping the
pixel into regions [19]. Initially, a collection of pixels is
integrated with the help of iteration methodology. Next, the
seed pixel is chosen alongside the regions and the group
is supported by integrating with neighboring pixel that is
corresponding and whereby it increases the region size.
The growth of region is ended whether the neighboring
pixel doesn’t satisfy the similarity constraint and another seed
pixel is carefully chosen. Repeat the process until all the
pixels in the image belongs to certain regions. In this work,
seed point and threshold selection take a decision regard-
ing the similarity condition because it plays a major part
in enhancing the segmentation accuracy. As mammogram
suffers from serious intensity variation, a constant threshold
selection doesn’t permit accurate segmentation. As a result,
this work emphasizes increasing the automatic DA method-
ology for generating an optimal seed point and threshold.
The step-wise procedure for region growing model is shown
below.
a. Input the abnormal image.
b. Now, ‘t’ characterizes the improved threshold generated
using MFO approach
c. Place t as seed point for region growing model.
d. Add four adjacent pixels.
e. Estimate the distance (d) between the neighboring pixels
and mean of region intensity.
f. Execute region growing if d ≤ t on four adjacent pixels
and add all if they are’t included before in the region and
save the coordinate of the new pixel.
g. Store the mean of novel region and return to step 2 along
with executing the region growing method until each
FIGURE 1. Block diagram of IoTDL-DRD approach. pixel is gathered.

27592 VOLUME 11, 2023

T. Palaniswamy, M. Vellingiri: IoT and DL Enabled DR Diagnosis Using Retinal Fundus Images

MFO is a new heuristic approach to solving complicated the following: male and female population. Two off springs
nonlinear optimization problems that were developed [20]. (children) are generated in the following:
It was inclined by the reproduction process and behavior
of MF. In time t, the MF position in a 2D search space is Mchild1 = ⊖×Mmale + (1 − ⊖) × Mfemale (7)
initialized by a = (a1 , . . . ,ad )T and b = (b1 , . . . ,bd )T , Mchild2 = ⊖×Mfemale + (1 − ⊖) × Mmale (8)
correspondingly, and velocity v = (v1 , . . . ,vd )T is allo-
Male signifies the male and Mfemale female parents ⊖ denote
cated to every MF. The objective function (F) defines the
an arbitrary value within a provided range whereas the pri-
performance of MF. In every iteration, the model saves
mary velocity of the children is fixed as zero. Afterward
the global positions (gbest) and best personal (pbest). One
estimating the efficiency of the children, the mutation is
cycle (Cyc) is the period among the travel time for every
presented. The model is avoided getting a local minimum
visited node, sum of vacation charging, and two charging
by mutating the children. For inducing mutation in Eq. (8),
demands. The fitness function aim is to minimize system total
a uniform distribution arbitrary value is included for chosen
power consumption, the number of cycles, and MC overall
children.
distance traveled that maximizes mobile charging vacation
time. MChild ′ α = MChildα + σ Nα (0, 1) (9)
Cyc
F = ( 2⌊log Cyc⌋ − 1−8 )2 In Eq. (9), σ and Nα denoted the standard deviation and
10 10
uniform distributions, correspondingly. Also, the mutated
+(emin − 1−8 )2 +(ethresh − 1−8 )2 children are assessed in terms of efficiency. The mutated chil-
1 Dtotal dren are combined with the non-mutated children, afterward,
+( − 1−8 )2 +( − 1−8 )2
τvac 10 log10 Dtotal ⌋
2⌊ they are similarly distributed Parent and child population is
s.t Emin < emin < ethresh′ Emin = 0.05×Emax arranged by efficiency to choose the following generation of
emax < ethresh < Emax , Emax = 10.8Kj (1) MF for optimization.

Every MF location is attuned based on own and neighbor

proficiencies. The male MF velocity and position updating C. FEATURE EXTRACTION
are shown below: To derive optimal feature vectors, the DenseNet model is
exploited in this study. The simple DenseNet is created by
at+1
i = ati + Vit+1 (2) alternating transition and dense blocks followed by a Soft-
−βn2  
Max classifier and fully connected layer [21]. The dense
Vit+1 = g×Vijt + x1e p pbestij − atij
block is composed of two cascaded convolution layers com-
−β 2  
+ x2 e ng gbestj − atij (3) prised of batch normalization (BN), convolution, and leaky
rectified linear unit (LReLU) layers ‘‘BN-Conv-LReL’’).
Though it should remain in nuptial dance, the opti- Initially, the dense block makes use of 1 × 1 kernel size
mal male MF upgrades its velocity with the following for producing 4r output feature map whereas the next
expression generate r feature map with 3 × 3 kernel size, where
r is a pre-determined constant and set as 32. The dense
Vt+1
i = Vti +m × n (4) blocks concatenate r output feature map with the input map,
bt+1
i = b + Vit+1 (5) thus resulting in increased number of maps. The transi-
tion block comprises average pooling and convolution lay-
The female MF location and velocity are upgraded in the
ers with 2 × 2 pool sizes. Here, the output feature map
following:
count in the transition block is similar to input. In this
−βn2 (at −bti)

work,‘‘conv’’ implies a convolution uni;‘‘1 × 1 × 64 conv’’
t mf ij
if f (bi ) > f (ai ) (6)

Vit+1 = g × Vij + x2e denotes the convolution layer in the unit that makes use
g × V t + fl × n if f (bi ) ≤f (ai ) of the 1 × 1 kernel size for producing 64 output feature
ij
map;‘‘2 × 2 pool’’ represents an average pooling layer with
In Eq. (6), x1 and x2 indicates individual learning vari- 2 × 2 pool siz;θ represents the number of concluding out-
able. g denotes the inertia weight, β refers to distance sight put classe;‘‘[·] × ’’ indicates the structure‘‘[·]’’ is cascaded
coefficient, m denotes the nuptial dance, and fl character- repeatedly for more than one times. There exists one FC layer
izes random flight, correspondingly. The Cartesian distance and twenty three convolution layer that is much shallower
is signified as np and ng , and n shows an arbitrary value than that of the original DenseNet with 121 convolution
between −1 and 1. Male MF is thought to move at a lower layers.
velocity in the nuptial dance, while female MF moves at a
higher velocity in the arbitrary flight. Parent MF is selected
to mate according to the fitness value. Accordingly, the higher D. IMAGE CLASSIFICATION
the fitness values, the greater the probability of selection. For DR recognition and classification, the HBO algorithm
The crossover operator portrays the mating of two MFs in with LSTM model is exploited in this study. The most

VOLUME 11, 2023 27593

 T. Palaniswamy, M. Vellingiri: IoT and DL Enabled DR Diagnosis Using Retinal Fundus Images

important algorithm of DL methods is CNN. In the study, it = sigmoid(Wi [ht−1 , Xt ] + bi )

the hybrid amalgamation of LSTM together with CNN is ft = sigmoid Wf [ht−1 , Xt ] + bf


exploited [22]. Honey bees exhibit a variety of complex
ot = sigmoid(Wo [ht−1 , Xt ] + bo )
behaviors in nature, including mating, reproduction, and for-
aging. Several honey bee-based optimization algorithms have ct = ft · Ct−1 + it · zt
imitated these behaviors. Honey Bees Optimization is a well- ht = 0t · tanh(ct ) (10)
known algorithm that was inspired by the mating and breed-
At the end stage, the HBO algorithm is exploited for opti-
ing behavior of honeybees (HBO). The algorithm begins with
mal hyper parameter adjustment process. A honeybee (HB)
a single queen that has no family and progresses to the growth
colony could fly up to 11 km for exploiting food sources and
of a colony that has families that contain one or more queens.
might attempt to terminate larger areas with large sources of
A colony of honey bees can access numerous food sources
food [23]. The cycle of forager starts with the hunt by scout
in expansive fields and can fly up to 11 km to do so. About
bee for enticing flower patch. Once the scout bee discovers
one-fourth of the colony’s bees work as foragers. Scout bees
the flower patch, they should search in advance to determine
start the foraging process by looking for promising flower
the best patch. The scout bee is desperately searching to
patches. During the harvesting season, a portion of the scout
discover the optimal spots.
bees are kept by the colony. The scout bees will search farther
after discovering a flower patch in the hopes of discovering
one that is even better. Randomly, the scout bees look for
the best patches. An artificial neural network called Long
Short-Term Memory (LSTM) is employed in deep learning
and artificial intelligence. LSTM has feedback connections
as opposed to typical feedforward neural networks. Such
a recurrent neural network (RNN) can process entire data
sequences in addition to single data points (such as images).
In order to process the consecutive dataset, RNN is broadly
employed. The existing input and the previous output are
connected to each other through these RNN models. The FIGURE 2. Structure of LSTM.
tanh activation functionthereby controls it, such that the series
state has been taken into account. In time t, the derivative During the hive, the scout bee acts as the counterpart
RNN would communicate and spread to time t−1,t−2, . . . ,l, regarding the nature of food sources, on the basis of the
thus resulting in the presence of a multiplicative factor. Gradi- sugar content, amongst others. The scout bee injects its nectar
ent disappearance and explosion occur if there exists endless and later goes ahead of the hive for celebrating the ballet,
multiplication. In the forward procedure, the input of starting termed waggle dance, with others. The waggle dance is
series has negligible or small effects on the later happening modeled afterward the wagging run (the performer makes a
sequence, and consequently, it is regarded as a key challenge loud buzzing rhythm and moves the body simultaneously)
of loss distance dependency. By familiarizing numerous gat- that the scout be used to transmit the residual parts of the
ing models, LSTM could resolve the problem easily. For a community data on the food sources. The scout bee offers
sigmoid operation, the contemporary input in a function Xt this data via waggle dance; consistency and distance from
and the previous output ht−1 are provided as an input such the path and hive sources. The waggle direction has eight-
that values among zero and one are produced, thus defining shaped patterns. First, the scout bees make a noise using their
the existing novel data that could be effortlessly maintained. wing muscles which produces a loud buzz, and start to run
The comprehensive state Ct of the following moment is in a vertical direction in a straight line which indicates the
attained using input and forget gates, and it is employed path for sun azimuth at the square and food source location.
for the initiation of the hidden state ht of the subsequent Instead, scouts circle back and transform left and right.
layer, thus establishing the output of the existing units. The The HBO algorithm derives a Fitness Function (FF) for
description of the output can be performed through the output obtaining enhanced performance of the classification. It sets
gate regarding the data attained from the cell state. A sigmoid an optimistic numeral to characterize the good outcome of the
function similar to input gate that produces a value 0t among candidate solution. The minimization of the classifier error
zero and one, which illustrates the quantity of cell state data rate has been taken as the FF in this study, as demonstrated in
described for projecting them as output. Figure 2 illustrates Eq. (11). The worse solution accomplishes the highest error
the design of a standard LSTM layer using appropriate inputs rate and the optimum solution has the lowest error rate.
and outputs. For the LSTM, the respective alliances among
the different gates are arithmetically formulated in the subse- fitness (xi ) = Classifier Error Rate (xi )
quent expression: number of misclassified samples
= ∗ 100
Total number of samples
zt = tanh(Wz [ht−1 , Xt ] + bz ) (11)

27594 VOLUME 11, 2023

T. Palaniswamy, M. Vellingiri: IoT and DL Enabled DR Diagnosis Using Retinal Fundus Images

TABLE 1. Pseudocode of HBO algorithm. TABLE 2. Dataset details.

and 24 samples into PDR. In addition, with 20% of TS data,

IoTDL-DRD method has classified 181 samples into normal,
60 samples into mild, 77 samples into moderate, 17 samples
into severe, and 6 samples into PDR. Finally, with 70% of
TR data, the IoTDL-DRD approach has classified 699 sam-
ples into normal, 175 samples into mild, 224 samples into
moderate, 31 samples into severe, and 1 sample into PDR.
Finally, with 30% of TS data, the IoTDL-DRD method has
classified 296 samples into normal, 78 samples into mild,
93 samples into moderate, 13 samples into severe, and 3 sam-
ples into PDR.
The fundus images used in the proposed work possess
sharp margins with normal RGB pattern. The fundus images
have normal features including the optic disc, fovea and blood
vessels. The images have smaller central cup with a normal
background. Table 3 exhibits the DR classification outcomes
IV. RESULTS AND DISCUSSION of the IoTDL-DRD model on 80% of TR data and 20% of
The performance validation of the IoTDL-DRD method is TS data. These table values demonstrate the improved per-
validated utilizing MESSIDOR dataset [24]. It comprises formance of the IoTDL-DRD model. The dataset is fed with
1744 fundus images under five class labels as shown in the normal sample of 1017, 270 samples with mild diabetics,
Table 2. The presented IoTDL-DRD method can be simulated 347 samples with moderate illness, 75 samples with severe
utilizing Python tool. Some sample images are depicted in while 35 samples with PDR level. The Figure 4 depicts the
Figure 3. Figure 4 illustrates the confusion matrices derived confusion matrix of IoTDL – DRD approach with 80% of TR
by the IoTDL-DRD model on varying training (TR) and data, 20% of TS data, 70% of TR data and 30% of TS data
testing (TS) data. With 80% of TR data, the IoTDL-DRD respectively. From this confusion matrix, 58.28% were found
method has classified 813 samples into normal, 203 samples to be normal, 14.55% were reported with mild diabetics,
into mild, 267 samples into moderate, 52 samples into severe, 19.14% as a moderate level of diabetics and 3.73% with a
severe level of diabetics retinopathy having more chances of
loss of eyesight. The IoT captured images were stored in the
database and is employed for the analysis purposes.
Figure 5 offers a brief set of DR classification performance
of the IoTDL-DRD model on 80% of TR data. The method
has classified fundus image into normal class with accuy ,
sensy , specy , Fscore , AUCscore , and MCC of 98.14%, 97.72%,
98.76%, 98.43%, 98.24%, and 96.16% respectively. The high
level of performance in the aforementioned parameters is
possibly achieved by incorporating the Deep learning models
in the diagnosis of DR.
Similarly, the IoTDL-DRD technique has categorized fun-
dus image into mild class with accuy , sensy , specy , Fscore ,
AUCscore , and MCC of 98.78%, 97.13%, 99.07%, 95.98%,
98.10%, and 95.27% correspondingly. It has categorized
FIGURE 3. Sample images. fundus image into moderate class with accuy , sensy ,

VOLUME 11, 2023 27595

 T. Palaniswamy, M. Vellingiri: IoT and DL Enabled DR Diagnosis Using Retinal Fundus Images

TABLE 3. Result analysis of IoTDL-DRD approach with distinct class labels under 70:30 of TR/TS data.

TABLE 4. Result analysis of IoTDL-DRD approach with distinct class labels under 70:30 of TR/TS data.

specy , Fscore , AUCscore , and MCC of 98.92%, 98.89%, categorized fundus image into mild class with accuy , sensy ,
98.93%, 97.27%, 98.91%, and 96.62% correspondingly. specy , Fscore , AUCscore , and MCC of 99.14%, 98.36%,
Finally, it has categorized fundus image into severe classes 99.31%, 97.56%, 98.83%, and 97.04% correspondingly.
with accuy , sensy , specy , Fscore , AUCscore , and MCC of It has categorized fundus image into moderate class with
99.28%, 92.86%, 99.55%, 91.23%, 96.20%, and 90.87% accuy , sensy , specy , Fscore , AUCscore , and MCC of 99.14%,
correspondingly. 100%, 98.90%, 98.09%, 99.45%, and 97.56% correspond-
Figure 6 provides a detailed set of DR classification ingly. Finally, it has categorized fundus image into severe
performance of the IoTDL-DRD method on 20% of TS classes with accuy , sensy , specy , Fscore , AUCscore , and MCC
data. The IoTDL-DRD algorithm has categorized fun- of 99.43%, 89.47%, 100%, 94.44%, 94.74%, and 94.31%
dus image into normal class with accuy , sensy , specy , correspondingly.
Fscore , AUCscore , and MCC of 98.28%, 97.84%, 98.78%, Table 4 exhibits the DR classification outcomes of the
98.37%, 98.31%, and 96.56% correspondingly. Also, it has IoTDL-DRD model on 70% of TR data and 30% of TS data.

27596 VOLUME 11, 2023

T. Palaniswamy, M. Vellingiri: IoT and DL Enabled DR Diagnosis Using Retinal Fundus Images

FIGURE 4. Confusion matrices of IoTDL-DRD approach (a) 80% of TR data, (b) 20% of TS data, (c) 70% of
TR data, and (d) 30% of TS data.

FIGURE 5. Average analysis of IoTDL-DRD approach under 80% of

TR data. FIGURE 6. Average analysis of IoTDL-DRD approach under 20% of
TS data.

VOLUME 11, 2023 27597

 T. Palaniswamy, M. Vellingiri: IoT and DL Enabled DR Diagnosis Using Retinal Fundus Images

FIGURE 10. TRL and VLL analysis of IoTDL-DRD approach.

FIGURE 7. Average analysis of IoTDL-DRD approach under 70% of TR data.

FIGURE 11. Precision-recall analysis of IoTDL-DRD approach.

FIGURE 8. Average analysis of IoTDL-DRD approach under 30% of TS data.

FIGURE 12. ROC analysis of IoTDL-DRD approach.

Figure 7 presents a complete set of DR classification per-

formances of the IoTDL-DRD approach on 70% of TR data.
The IoTDL-DRD algorithm has classified fundus image into
FIGURE 9. TRA and VLA analysis of IoTDL-DRD approach. normal class with accuy , sensy , specy , Fscore , AUCscore , and
MCC of 96.23%, 97.90%, 93.87%, 96.81%, 95.89%, and
92.23% correspondingly.
These table values demonstrate the improved performance of Similarly, it has categorized fundus image into mild class
the IoTDL-DRD model. with accuy , sensy , specy , Fscore , AUCscore , and MCC of

27598 VOLUME 11, 2023

T. Palaniswamy, M. Vellingiri: IoT and DL Enabled DR Diagnosis Using Retinal Fundus Images

FIGURE 13. Comparatives analysis of IoTDL-DRD approach (a) Accuy , (b) Sensy , (c) Specy , (d) Fscore ,
and (e) AU Cscore .

97.46%, 93.58%, 98.16%, 91.86%, 95.87%, and 90.38% image into moderate class with accuy , sensy , specy , Fscore ,
correspondingly. Simultaneously, it has categorized fun- AUCscore , and MCC of 96.18%, 92.08%, 97.16%, 90.29%,
dus image into moderate class with accuy , sensy , specy , 94.62%, and 87.94% individually. Eventually, it has catego-
Fscore , AUCscore , and MCC of 95.41%, 91.06%, 96.51%, rized fundus image into severe class with accuy , sensy , specy ,
88.89%, 93.78%, and 86.04% correspondingly. Meanwhile, Fscore , AUCscore , and MCC of 96.76%, 59.09%, 98.41%,
the IoTDL-DRD approach has categorized fundus image into 60.47%, 78.75%, and 58.79% correspondingly.
severe classes with accuy , sensy , specy , Fscore , AUCscore , and The training accuracy (TRA) and validation accuracy
MCC of 97.79%, 58.49%, 99.57%, 69.66%, 79.03%, and (VLA) gained by IOTDL-DRD methodology on test dataset
69.94% correspondingly. is shown in Figure 9. The experimental outcome denoted
Figure 8 demonstrates a detailed set of DR classification the IOTDL-DRD approach has exhibited maximal values of
performance of the IoTDL-DRD approach on 30% of TS TRA and VLA. Seemingly the VLA is greater than TRA. The
data. The method has categorized fundus image into normal training loss (TRL) and validation loss (VLL) attained by the
class with accuy , sensy , specy , Fscore , AUCscore , and MCC of IOTDL-DRD methodology on test dataset are exhibited in
95.42%, 97.69%, 92.31%, 96.10%, 95%, and 90.62% corre- Figure 10. The experimental result represented the IOTDL-
spondingly. Also, it has categorized fundus image into mild DRD approach has shown least values of TRL and VLL.
class with accuy , sensy , specy , Fscore , AUCscore , and MCC Particularly, the VLL is lesser than TRL. An obvious
of 98.28%, 93.98%, 99.09%, 94.55%, 96.53%, and 93.53% precision-recall analysis of the IOTDL-DRD system on test
correspondingly. In the meantime, it has categorized fundus dataset is shown in Figure 11. The figure stated that the

VOLUME 11, 2023 27599

 T. Palaniswamy, M. Vellingiri: IoT and DL Enabled DR Diagnosis Using Retinal Fundus Images

TABLE 5. Result analysis of IoTDL-DRD approach with distinct class labels under 70:30 of TR/TS data.

IOTDL-DRD technique has resulted in higher values of comparison study reported the superior performance of the
precision-recall values in all classes. A brief ROC study of presented model compared to recent state of art approaches.
the IOTDL-DRD method on test dataset was depicted in In upcoming years, deep instance segmentation models were
Figure 12. The outcomes represented by the IOTDL-DRD derived to boost the classification results.
method has revealed its capability in categorizing distinct
classes on test dataset. ACKNOWLEDGMENT
For assuring the enhanced DR classification performance The authors extend their appreciation to ‘‘the Deputy-
of the IoTDL-DRD model, a wide ranging comparative anal- ship for Research & Innovation, Ministry of Education in
ysis is made in Table 5 and Figure 13 [25]. The fallouts Saudi Arabia for funding this research work through the
highlighted the IoTDL-DRD model has shown enhanced out- project number G: 693-144-1443’’ and King Abdulaziz Uni-
comes over other models [26]. Based on accuy , the IoTDL- versity, Deanship of Scientific Research (DSR), Jeddah,
DRD model has offered higher accuy of 99.08% whereas Saudi Arabia.
the Inception-v3, GoogLeNet, AlexNet, and ResNet mod-
els have attained lower accuy of 94.59%, 94.34%, 94.02%, REFERENCES
and 94% respectively. Similarly, according to sensy , the
[1] D. S. W. Ting, H. Lin, P. Ruamviboonsuk, T. Y. Wong, and D. A. Sim,
IoTDL-DRD approach has obtainable superior sensy of ‘‘Artificial intelligence, the Internet of Things, and virtual clinics: Ophthal-
94.28% whereas the Inception-v3, GoogLeNet, AlexNet, and mology at the digital translation forefront,’’ Lancet Digit. Health, vol. 2,
ResNet techniques have attained lower sensy of 94.12%, no. 1, pp. e8–e9, Jan. 2020.
[2] K. Parthiban and K. Venkatachalapathy, ‘‘Internet of Things and cloud
93.16%, 92.56%, and 94.22% correspondingly. enabled hybrid feature extraction with adaptive neuro fuzzy inference
Likewise, based on Fscore , the IoTDL-DRD technique has system for diabetic retinopathy diagnosis,’’ J. Comput. Theor. Nanosci.,
existing superior Fscore of 94.84% whereas the Inception-v3, vol. 17, no. 12, pp. 5261–5269, Dec. 2020.
GoogLeNet, AlexNet, and ResNet systems have reached [3] J. T. Kelly, K. L. Campbell, E. Gong, and P. Scuffham, ‘‘The Internet of
Things: Impact and implications for health care delivery,’’ J. Med. Internet
lesser Fscore of 92.35%, 90.83%, 91.76%, and 93.28% corre- Res., vol. 22, no. 11, Nov. 2020, Art. no. e20135.
spondingly. Therefore, the experimental results [27] verified [4] R. J. Kavitha, T. Avudaiyappan, T. Jayasankar, and J. Selvi, ‘‘Industrial
that the IoTDL-DRD model has accomplished enhanced DR Internet of Things (IIoT) with cloud teleophthalmology-based age-related
macular degeneration (AMD) disease prediction model,’’ in Smart Sensors
classification performance over other DL models [28]. The for Industrial Internet of Things. Cham, Switzerland: Springer, 2021,
computational time for the proposed process is computed and pp. 161–172.
is observed to possess a reduced duration of 11.21 msec. This [5] A. B. Tufail, I. Ullah, W. U. Khan, M. Asif, I. Ahmad, Y.-K. Ma, R. Khan,
and M. S. Ali, ‘‘Diagnosis of diabetic retinopathy through retinal fun-
observed value is proved to be lesser than the other computer dus images and 3D convolutional neural networks with limited number
vision processes [29]. of samples,’’ Wireless Commun. Mobile Comput., vol. 2021, pp. 1–15,
Nov. 2021.
[6] T. J. Jebaseeli, J. David, and V. Jegathesan, ‘‘Machine learning and Internet
V. CONCLUSION
of Things techniques to assist the type I diabetic patients to predict the
In this paper, a novel IoTDL-DRD model was developed for regular optimal insulin dosage,’’ in Internet of Medical Things. Cham,
DR detection and classification on retinal fundus images. Switzerland: Springer, 2021, pp. 159–174.
The presented IoTDL-DRD method allows the IoT devices [7] V. Ramalingam, D. B. Mariappan, R. Gopal, and K. M. Baalamurugan,
‘‘An effective social Internet of Things (SIoT) model for malicious node
for data collection purposes and then transfers to the cloud detection in wireless sensor networks,’’ in Artificial Intelligence Tech-
server to process them. Then, the retinal fundus images are niques in IoT Sensor Networks. Boca Raton, FL, USA: CRC Press, 2020,
preprocessed for noise removal and contrast improvement. pp. 181–195.
[8] E. A. Refaee and S. Shamsudheen, ‘‘A computing system that integrates
Afterward, the MFORG technique can be employed for deep learning and the Internet of Things for effective disease diagnosis in
lesion segmentation in the fundus images. For DR recognition smart health care systems,’’ J. Supercomput., vol. 78, no. 7, pp. 9285–9306,
and classification, the HBO algorithm with LSTM model is May 2022.
exploited in this study. For evaluating the improved DR diag- [9] S. M. Zobaed, M. Hassan, M. U. Islam, and M. E. Haque, ‘‘Deep learn-
ing in IoT-based healthcare applications,’’ in Deep Learning for Inter-
nostic outcomes of the IoTDL-DRD technique, a compre- net of Things Infrastructure. Boca Raton, FL, USA: CRC Press, 2021,
hensive set of simulations were carried out. A wide ranging pp. 183–200.

27600 VOLUME 11, 2023

T. Palaniswamy, M. Vellingiri: IoT and DL Enabled DR Diagnosis Using Retinal Fundus Images

[10] X. Ma, T. Yao, M. Hu, Y. Dong, W. Liu, F. Wang, and J. Liu, ‘‘A survey [26] O. M. Al-Hazaimeh, A. Abu-Ein, N. Tahat, M. Al-Smadi, and
on deep learning empowered IoT applications,’’ IEEE Access, vol. 7, M. Al-Nawashi, ‘‘Combining artificial intelligence and image processing
pp. 181721–181732, 2019. for diagnosing diabetic retinopathy in retinal fundus images,’’ Int. J. Online
[11] R. Rajavel, B. Sundaramoorthy, K. Gr, S. K. Ravichandran, and Biomed. Eng., vol. 18, no. 13, pp. 131–151, Oct. 2022.
K. Leelasankar, ‘‘Cloud-enabled diabetic retinopathy prediction system [27] N. Gharaibeh, O. M. Al-Hazaimeh, B. Al-Naami, and K. M. Nahar,
using optimized deep belief network classifier,’’ J. Ambient Intell. Human- ‘‘An effective image processing method for detection of diabetic retinopa-
ized Comput., vol. 2022, pp. 1–9, Jul. 2022. thy diseases from retinal fundus images,’’ Int. J. Signal Imag. Syst. Eng.,
[12] A. M. Dayana and W. R. Emmanuel, ‘‘Deep learning enabled opti- vol. 11, pp. 206–216, Jan. 2018.
mized feature selection and classification for grading diabetic retinopathy [28] N. Gharaibeh, O. M. Al-hazaimeh, A. Abu-Ein, and K. M. Nahar,
severity in the fundus image,’’ Neural Comput. Appl., vol. 34, pp. 1–21, ‘‘A hybrid SVM Naïve-Bayes classifier for bright lesions recognition in eye
Jun. 2022. fundus images,’’ Int. J. Electr. Eng. Informat., vol. 13, no. 3, pp. 530–545,
[13] J. G. R. Elwin, J. Mandala, B. Maram, and R. R. Kumar, ‘‘Ar-HGSO: Sep. 2021.
Autoregressive-Henry gas sailfish optimization enabled deep learning [29] V. Conti, C. Militello, L. Rundo, and S. Vitabile, ‘‘A novel bio-inspired
model for diabetic retinopathy detection and severity level classification,’’ approach for high-performance management in service-oriented net-
Biomed. Signal Process. Control, vol. 77, Aug. 2022, Art. no. 103712. works,’’ IEEE Trans. Emerg. Topics Comput., vol. 9, no. 4, pp. 1709–1722,
[14] N. O. Aljehane, ‘‘An intelligent moth flame optimization with inception Oct. 2021.
network for diabetic retinopathy detection and grading,’’ in Proc. 2nd Int.
Conf. Comput. Inf. Technol. (ICCIT), Jan. 2022, pp. 370–373.
[15] K. Gunasekaran, R. Pitchai, G. K. Chaitanya, D. Selvaraj, S. A. Sheryl, THANGAM PALANISWAMY (Member, IEEE)
H. S. Almoallim, S. A. Alharbi, S. S. Raghavan, and B. G. Tesemma, received the B.E. degree in computer hardware and
‘‘A deep learning framework for earlier prediction of diabetic retinopa- software engineering from Avinashilingam Uni-
thy from fundus photographs,’’ BioMed Res. Int., vol. 2022, pp. 1–15, versity, India, in 2001, and the M.E. degree in com-
Jun. 2022. puter science and engineering and the Ph.D. degree
[16] I. K. Gupta, A. Choubey, and S. Choubey, ‘‘Mayfly optimization with in information and communication engineering
deep learning enabled retinal fundus image classification model,’’ Comput. from Anna University, India, in 2007 and 2013,
Electr. Eng., vol. 102, Sep. 2022, Art. no. 108176. respectively. She has a total teaching experience
[17] J. Latif, S. Tu, C. Xiao, S. U. Rehman, A. Imran, and Y. Latif, ‘‘ODGNet: of 15 years in various reputed engineering colleges
A deep learning model for automated optic disc localization and glaucoma
in Tamil Nadu. She is currently an Associate Pro-
classification using fundus images,’’ Social Netw. Appl. Sci., vol. 4, no. 4,
fessor with the Department of Electrical and Computer Engineering, King
pp. 1–11, Apr. 2022.
Abdulaziz University, Saudi Arabia. Her research interests include databases,
[18] N. Al-Saiyd and S. Talafha, ‘‘Mammographic image enhancement
tachniques—A survey,’’ Int. J. Comput. Sci. Inf. Technol. Secur., vol. 5, data processing and mining, medical image analysis, image processing, cryp-
pp. 200–206, Jan. 2015. tography, embedded systems, and the Internet of Things. Her contributions in
[19] M. A. Mohammed, M. K. Abd Ghani, R. I. Hamed, D. A. Ibrahim, and professional societies include IEEE, International Association of Engineers,
M. K. Abdullah, ‘‘Artificial neural networks for automatic segmentation Indian Society for Technical Education, and International Association of
and identification of nasopharyngeal carcinoma,’’ J. Comput. Sci., vol. 21, Computer Science and Information Technology.
pp. 263–274, Jul. 2017.
[20] M. A. M. Shaheen, H. M. Hasanien, M. S. E. Moursi, and
A. A. El-Fergany, ‘‘Precise modeling of PEM fuel cell using improved
chaotic MayFly optimization algorithm,’’ Int. J. Energy Res., vol. 45,
no. 13, pp. 18754–18769, Oct. 2021. MAHENDIRAN VELLINGIRI received the B.E.
[21] Z. Huang, X. Zhu, M. Ding, and X. Zhang, ‘‘Medical image classifica- degree in electrical and electronics engineer-
tion using a light-weighted hybrid neural network based on PCANet and ing from Maharaja Engineering College, Avinasi,
DenseNet,’’ IEEE Access, vol. 8, pp. 24697–24712, 2020. affiliated to Bharathiyar University, Coimbatore,
[22] J. Lei, C. Liu, and D. Jiang, ‘‘Fault diagnosis of wind turbine based on Tamil Nadu, in 2000, the M.E. degree in power
long short-term memory networks,’’ Renew. Energy, vol. 133, pp. 422–432,
electronics and drives from K. S. Rangasamy Col-
Apr. 2019.
lege of Technology, Tiruchengode, affiliated to
[23] S. Nayak, C. Kumar, S. Tripathi, N. Mohanty, and V. Baral, ‘‘Regression
Anna University, Chennai, Tamil Nadu, in 2006,
test optimization and prioritization using honey bee optimization algorithm
with fuzzy rule base,’’ Soft Comput., vol. 25, no. 15, pp. 9925–9942, and the Ph.D. degree in electrical engineering from
Aug. 2021. Anna University, India, in 2014. He is currently
[24] X. Zhang, G. Cazuguel, B. Lay, B. Cochener, C. Trone, and P. Gain, ‘‘Feed- an Assistant Professor with the Department of Electrical and Computer
back on a publicly distributed image database: The Messidor database,’’ Engineering, King Abdulaziz University, Jeddah, Saudi Arabia. His research
Image Anal. Stereol., vol. 33, pp. 231–234, Aug. 2014. interests include soft computing applications in the control of power elec-
[25] A. Bilal, L. Zhu, A. Deng, H. Lu, and N. Wu, ‘‘AI-based automatic tronics drives, control systems, electrical machines, solar energy, and power
detection and classification of diabetic retinopathy using U-Net and deep systems.
learning,’’ Symmetry, vol. 14, no. 7, p. 1427, Jul. 2022.

VOLUME 11, 2023 27601