https://doi.org/10.31449/inf.v46i2.3906 Informatica 46 (2022) 205–221 205
Exploring the Parametric Impact on a Deep Learning Model and Proposal of a
2-Branch CNN for Diabetic Retinopathy Classification with Case Study in
IoT-Blockchain based Smart Healthcare System
Manaswini Jena
1
, Smita Prava Mishra
1
, Debahuti Mishra
1
, Pradeep Kumar Mallick
2
and Sachin Kumar
3
*
E-mail: manaswini.jena88@gmail.com, smitamishra@soa.ac.in, debahutimishra@soa.ac.in,
pradeepmallick84@gmail.com, sachinagnihotri16@gmail.com
1
Department of Computer Science and Engineering, Siksha ‘O’ Anusandhana Deemed to be University, Bhuvaneswar
Odisha, India
2
School of Computer Engineering, Kalinga Institute of Industrial Technology (KIIT) Deemed to be University, Bhu-
vaneswar, Odisha, India
3
Department of Computer Science, South Ural State University, Chelyabinsk, Russia
Keywords: healthcare system, 2-branch CNN, diabetic retinopathy, fundus images, medical diagnosis, internet of things
Received: January 9, 2022
Smart healthcare has changed the way how the patient interacts with the specialists for treatment. How-
ever, security and support for various diseases are still the concern for such smart automated systems. One
of the critical diseases namely Diabetic Retinopathy (DR), is a major concern for the person with pro-
longed diabetes and may lead to complete blindness irrespective of age groups. Moreover, in recent years
blockchain has gained popularity in providing secure communication between sender and receiver. Hence,
this work focus on designing a blockchain-based smart healthcare system for the early detection of diabetic
retinopathy. However, early detection of DR impose complexities and requires expert diagnosis, which is
not available everywhere. Hence, the proposed smart healthcare model contains a Computer-Aided Diag-
nosis (CAD) assistance for early detection of symptoms of the disease. The CAD model may assist the
ophthalmologists in the early detection of DR, which requires intensive research in developing an efficient
and accurate model that can operate without human interaction. This study provides an empirical analysis
of these factors to design the best model for early detection of DR. The best model can be used to develop
IoT based smart devices to detect DR in diabetic patients. The study also explains the importance of IoT
and blockchain-based technology for the development of smart healthcare systems. The values of the pa-
rameters and type of hyperparameters choosen from the study is used in a proposed 2-branch CNN model,
and the model is validated using the Kaggle fundus image set. Analysis of various parameters and using
their best values gives an outstanding performance in the proposed 2-branch CNN model.
Povzetek: Predstavljena je bloˇ cna metoda za varno komunikacijo med zdravnikom in pacienti, pri tem je
uporabljena metoda globokih nevronskih mrež za problem diabetiˇ cne slepote.
1 Introduction
Healthcare systems are an essential part of modern society
to save people’s lives and provide services remotely. How-
ever, the concern of the healthcare system is the need for
the protection of information of the patient and correctly
recognizing the symptoms for early detection. Diabetic
Retinopathy(DR) is one of the fatal diseases where early di-
agnosis is very crucial. Computer-Aided Diagnosis(CAD)
assists experts in the early diagnosis of the disease. In re-
cent years, researchers have developed many CAD models
to detect DR effectively. However, a complete healthcare
model is not yet developed. This research attempts to pro-
pose a CAD model with security to protect patient data and
a deep learning model to detect DR accurately.
*
Corresponding author
Blockchain, IoT, and AI have shown their potential in
almost every domain of our life. Smart healthcare is one
of the major sectors that has been influenced by IoT in-
frastructure and solutions [1-2]. IoT-based healthcare sys-
tems have immensely added value to our lifestyle and
health monitoring with the use of portable wearable de-
vices. However, there are few complex healthcare areas
where Blockchain and IoT have to do some evolutionary
steps to protect the healthcare data to ensure patient confi-
dentiality.
Similarly, Convolutional Neural Network (CNN) is
widely used for accurate CAD in the healthcare domain.
These automatic diagnostic systems can improve the qual-
ity and productivity of healthcare services. Moreover, CAD
can help in the early detection of diabetic retinopathy, one
of the primary causes of blindness in the world [3]. The
limited resources of healthcare facilities lead to difficult ac-

206 Informatica 46 (2022) 205–221 M. Jena et al.
cess to expert doctors for diabetic retinopathy analysis [4].
Currently, CAD includes convolutional networks to iden-
tify the early signs of diabetic retinopathy disease. The
CNN is a type of deep feed-forward artificial neural net-
work. This architecture is mainly designed and adopted to
classify images using a multilayer architecture [5-7]. Var-
ious types of CNNs are proposed and their efficiencies are
proved in the complicated image classification techniques
[8-11]. Prolonged diabetics result in diabetic retinopathy
leading to the significant cause of vision loss worldwide
[10-11]. Several technical implementations have done us-
ing CNN for diabetic retinopathy identification [12-17].
The performance of the CNN model depends on the num-
ber of convolution layers, the number of nodes in convolu-
tion layers with several parameters, and hyper-parameters
such as learning rate, activation functions, and pooling [18-
19]. To our best knowledge, no study has been carried out
that can justify the choices made in designing a CNN for
diabetic retinopathy classification.
The fundamental units for this convolution network are
shared weight, bias, local perceptive, and pooling. The ac-
tivation function and pooling are critical for model perfor-
mance. The activation function is an elementary part of ev-
ery neural network architecture. Moreover, it is responsible
for transforming the summed weighted input into its output
[20-21]. The activation function and threshold are the ba-
sic terminology for every neural network for execution and
flow of network as the activation of neurons depend upon it
[22-23]. Non-linear transformations and back-propagation
are made using activation functions. Several different ac-
tivation functions are utilized for NN [24-27]. Pooling is
applied as a downsampling approach to reduce the input
signal resolution without affecting the representable, essen-
tial, and large feature elements [24-25]. In addition to ac-
tivation and pooling, the learning rate and optimizer used
for back-propagation in the network has also a great im-
pact on the performance of the neural network. So, all four
features are going to be analyzed for the classification of
diabetic retinopathy using the CNN model.
The learning rate is a hyper-parameter used to control the
learning speed of the neural network model. The key role
of the learning rate is to scale the magnitude of weight to re-
duce the network’s loss function. The training or learning
is generally performed through updating weights to map
the input with the output. These updated weights are cal-
culated by analytical methods via empirical optimization
techniques called stochastic gradient descent (SGD). The
stepping amount needed to change the weights during the
training process is called the learning rate. This is the value
used for adjusting the weights according to loss gradient
through back-propagation [26].
The value of the learning rate is a small positive value in
the range of 0.0 to 1.0. It may vary differently for differ-
ent models and we can find a good learning rate value via
the trial and error method. Though the optimal learning rate
calculation is analytically not possible yet performance im-
provement can be done by achieving a good learning rate.
During training, the learning rate can be adjusted to im-
prove the performance and this is called adaptive learning
[27-28]. Between the epochs or iterations, the value can be
increased or decreased, instead of having a fixed value. The
change in value is made is called a learning rate schedule.
Here, an attempt is made to visualize the change in accu-
racy and performance of the model according to the change
in learning rate values.
Another feature considered here is the types of opti-
mizers used for back-propagation in the neural network.
Back-propagation is a weight updater in neural networks
used to update the weights by optimizing the values to get
the target result. For this optimization, SGD is generally
[29-30] used. SGD is an iterative optimization method
that uses shuffled samples for the evaluation of the gra-
dient values. Sometimes the random probabilistic nature
of SGD for choosing samples results in a noisy path to
reach the destination from typical gradient descent algo-
rithms as it may take a higher number of iterations due to
randomness. But, the computationally inexpensive nature
of this algorithm makes it suitable for usage. Another type
of optimizer available to be used instead of this SGD is
Adam. The memory requirement is very less for this opti-
mizer as well as it can be used for large problems in terms
of data/parameters and appropriate for non-stationary ob-
jectives along with a problem with noisy gradients [31].
AdaDelta is another optimizer based on a moving window
of gradient updates. It does not accumulate the past gra-
dients rather continues learning from updates [32]. Here
these three optimizers are chosen for the experimentation.
The objective of this study is to analyze the impacts of
these four features i.e. activation function, pooling, learn-
ing rate, and optimizer parameters on a CNN model for
diabetic retinopathy classification. The study is conducted
in phases to analyze the impact of each parameter. Finally,
the best specifications are used in a proposed CNN model
for diabetic retinopathy classification.
The rest of the article is organized as follows: Section
II represents the state-of-the-art literature review. Section
III and IV provides the methodolgy with model description
and an overview of the dataset respectively. Section V pro-
vides the experiment details. The results and discussion are
in Section VI . Finally, Section VII concludes the study.
2 Literature review
IoT, Blockchain and Cloud technologies are integrated in
the medical environment for offering healthcare and tele-
medical laboratory services along with AI. For the classifi-
cation of diabetic retinopathy through fundus image CNN
is widely used and proved as a great classifiers, but some-
times the data taken directly from the patient may not give
the expected accurate report. More features and parameters
need to be tuned very minutely and this is the main objec-
tive here. This section represents some studies where CNN
is used for the classification of DR. In [33] transfer learn-

Exploring the Parametric Impact on a Deep. . . Informatica 46 (2022) 205–221 207
ing is used on pre-trained CNN models and an accuracy of
74.5% is achieved for binary classification model. Tertiary
and quaternary classification models are also tested for the
classification the study. A set of pre-processing technique
and data augmentation is followed by transfer learning for
the experiment. In [34] the CNN used with augmentation
for the classification of diabetic retinopathy gives better
result compared to the CNN used without augmentation
i.e. 94.5% and 91.5% respectively. Here augmentation im-
proves the performance. DIARETDB1 data is used in [35]
for classification of DR by N. Gharaibeh et al. Here, in the
first phase feature extraction is done followed by classifi-
cation of features acquiring 98.4% accuracy using CNN. In
[39] retinal images are transformed into entropy images to
improve the classification. The complexity of original fun-
dus images are represented by quantifying the information
of the image. Then classification is done using deep learn-
ing which improves the accuracy to 86.0% from 81.8% that
was achieved using normal fundus images. Image enhance-
ment techniques like histogram equalization is used in [57]
along with deep leaning to improve the classification of DR
and an accuracy of 97% is achieved. In [37] the segmented
image classification using CNN for DR achieves an accu-
racy of 99.18% accuracy results. Another work in [36] uses
CNN along with Support vector machine (SVM) as classi-
fier and results with 86.17% accuracy. Here max-pooling is
replaced by fractional max-pooling in the CNN for extrac-
tion of features before using it in SVM. Following Table 1
highlights the findings of the above discussion:
However, a study [56] states that although high accu-
racy is achieved for detection of DR using various models
of machine learning, but the requirements of real time de-
ployment and clinical validation is not yet fulfilled com-
pletely. It also indicates to solve the problem arising due
to the diversity of datasets in terms of ethnicities and cam-
era used for capturing the images in real practical clinical
field. So, fine-tuning of every parameters and features of
the network is required to get a perfect model. However,
this study of various features and parameters and their ef-
fects will help in development of prediction models. Fur-
thermore, IoT/Blockchain based devices provides portable
and secure technology for healthcare monitoring, however
it is a challenging task to achieve such facility with high se-
curity. In [41], authors proposed an approach to deal with
these challenges for Internet of Medical Things [IoMT].
Several research studies focused on the effectiveness of IoT
and bloackchain for the healthcare ecosystem [42-43]. In
addition, cloud-server can be utilized to provide a global
secure network to monitor a large number of petients in
several regions with cloud based IoT network in health care
domain [44-47].
It is observed that, image features are also a very impor-
tant factor [48], to obtain a high classification performance.
Various techniques have been used for accurate classifica-
tion of the image by extracting the discriminative features.
The amalgamation of both global and local features can be
fed to a classifier for classification of the image. For this a
2 branch CNN can be designed having the ability to extract
features in an optimized way for the classification in a uni-
fied optimized framework [49]. A multi-domain CNN per-
forms better than a single frequency domain CNN and spa-
tial domain CNN for detection of image compression [50].
Hyperspectral image classifications show the dominance of
a multi-branch CNN in many studies [51-53]. These kinds
of architecture extracts feature from each source of the im-
age at their native pixel of resolution and it is done by ex-
ploiting both spatial and spectral information from the im-
age. Keeping this in mind, the workflow is designed to
create a better model for diabetic retinopathy detection as
compared to the existing models using the analysis.
3 Methodology
At first a two class classification model using several CNN
configurations were developed to understand the impor-
tance of several parameters for the classification of DR in
diabetic patients. These configurations are mentioned in
Table 2 as follows:
The result of the empirical analysis is used in a 5 class
classification model. A two-branch CNN is proposed for
the classification of DR using fundus image. The proposed
flow diagram of a two-branch CNN model is given in Fig-
ure 1. Two separate feature sets collected from two sepa-
rate branches of CNN are assembled and used in the fully
connected neural network for classification of DR.
Figure 1: A two branch CNN model.
Some basic pre-processing steps are applied here for the
classification as follows:
(i) Cropping: Extracting the part of the re-
gion of interest (ROI) from the image is
done through cropping. Most of the fundus
images have a black border which needs
to be reduced as possible to concentrate on
the region of interest.
(ii) Resizing of image: For any machine
learning technique or deep learning tech-
nique, resizing of the image is required.
All the images of data need to be standard-
ized. Here, all the images are converted
into one size i.e. 256×256×3.

208 Informatica 46 (2022) 205–221 M. Jena et al.
Table 1: Comparison of existing studies for DR classification.
Model used Dataset used Accuracy
CNN with pre-processing, feature extraction and augmentation (used
for binary classification) [33]
Messidor
74.50%
MildDR
CNN with augmentation [34] Kaggle 94.50%
CNN without augmentation [34] Kaggle 91.50%
Feature extraction and classification using CNN [35] DIARETDB1 98.40%
Transformed entropy images and classification using CNN [36] Kaggle 86.00%
Histogram Equalization and classification using CNN [57] Messidor 97%
Segmentation and pre- processing with CNN [37] DiaretDB0, 99.17%
DiaretDB1 98.53%
DrimDB 99.18
CNN with SVM as classifier [38] Kaggle 86.17%
Figure 2: First row represents diseased fundus images and
second row represents normal fundus images.
(iii) Contrast Enhancement: It is a process
of making the image features more promi-
nent by changing the range of pixel val-
ues in the image. Contrast Limited Adap-
tive Histogram Equalization (CLAHE) is
the method which improves contrast in dig-
ital images. The results of medical image
enhancement using this technique are bet-
ter than other histogram equalization tech-
niques [54]. Here, this CLAHE is used for
the contrast enhancement to make the hid-
den features visible.
(iv) Normalization: Max-min normalization
is applied here for normalizing the pixel
values.
4 Dataset description
The dataset under study have been obtained with the sup-
port of the administration of Ruby Eye hospital in Orisha,
India. The fundus images have been captured with a high
resoulion cameras. The database shows a male and female
ratio of 50:26 with diabetic retinopathy patients’ in 30-65
age group. For diabetic retinopathy, the separated fundus
images of diseased and non-diseased patients are collected.
The samples of the infected and non-infected fundus im-
ages are shown in Fig 2.
The dataset consists of the fundus images of 102 dia-
betic patients with the presence of both healthy and dis-
eased images with equal distribution. The two classes of
these fundus images are either diabetic retinopathy (DR)
i.e. the diseased or normal i.e. non-diseased. Normalized
of the images are performed with max-min normalization
in the range [0-1] in order to achieve zero mean and unit
variance. This dataset is available online publically for re-
search purpose [55].
Kaggle retina fundus image dataset is also used here for
experiment. It contains the images based on five sever-
ity levels such as 0=No DR, 1=Mild NPDR, 2=Moderate
NPDR, 3=Severe NPDR, 4=PDR. For both the data, the
training and testing ratio is taken as 80:20 here.
5 Experimental analysis
The detailed evaluation of hyper-parameters on a large
model is a computationally intensive task. Hence, the study
uses simple model to evaluate various hyper-parameter set-
tings like pooling, activation, learning rate, and optimiza-
tion. Finally, the proposed 2-branch CNN model uses the
best combination of hyper-parameters learnt from the sim-
ple model.
To analyse the significance of pooling, activation, learn-
ing rate, and optimization speciation of the CNN model,
multiple experiments have been conducted, and these are
divided into four phases.
In the first phase (Phase 1), the performance of the pa-
rameters or factors are analyzed to be selected for further
testing. Following this in the second phase (Phase 2), a
selective combination of factors are evaluated to establish
the relationship of these to the model performance. In next
two phases (Phase 3 and Phase 4), more factors are evalu-
ated by incremental tests to obtain maximum performance
for fundus image classification using this dataset.
Here, data augmentation is used for virtually increasing
the number of samples in the dataset during training. It
is a technique used to manufacture or generate own data

Exploring the Parametric Impact on a Deep. . . Informatica 46 (2022) 205–221 209
Table 2: Used CNN model and its layer description.
Type Layer Size Learnables Total Learnables
Image_Input 256× 256× 3 - 0
Convolution 242× 242× 32 Weights 15× 15× 3× 32 21632
Bias 1× 1× 32
Batch_Normalization 242× 242× 32 Offset 1× 1× 32 64
Scale 1× 1× 32
ReLU 242× 242× 32 - 0
Max_Pooling 121× 121× 32 - 0
Convolution 115× 115× 64 Weights 7× 7× 32× 64 100416
Bias 1× 1× 64
Batch_Normalization 115× 115× 64 Offset 1× 1× 64 128
Scale 1× 1× 64
ReLU 115× 115× 64 - 0
Max_Pooling 57× 57× 64 - 0
Convolution 53× 53× 128 Weights 5× 5× 64× 128 204928
Bias 1× 1× 128
Batch_Normalization 53× 53× 128 Offset 1× 1× 128 256
Scale 1× 1× 128
ReLU 53× 53× 128 - 0
Max_Pooling 26× 26× 128 - 0
Convolution 24× 24× 128 Weights 3× 3× 128× 128 147584
Bias 1× 1× 128
Batch_Normalization 24× 24× 128 Offset 1× 1× 128 256
Scale 1× 1× 128
ReLU 24× 24× 128 - 0
Max_Pooling 12× 12× 128 - 0
Keras Flatten 1× 1× 18432 - 0
Fully_Connected 1× 1× 100 Weights 100× 18432 1843300
Bias 100× 1
ReLU 1× 1× 100 - 0
Fully_Connected 1× 1× 2 Weights 2× 100 Bias 2× 1 202
ReLU 1× 1× 2 - 0
Classification_Output - - 0

210 Informatica 46 (2022) 205–221 M. Jena et al.
with the existing data [56]. Data augmentation have been
used to deform the labelled data without semantic infor-
mation loss[57]. Here five online data augmentations are
used such as horizontal and vertical shifts, horizontal and
vertical flips and rotation. The horizontal shift and vertical
shifts are of 2% and rotation of 90% is applied. The modi-
fied input images are introduced during the training process
automatically. For experiment, the dataset is divided in the
ratio of 80:20 for the training and testing data.
5.1 Evaluation of the parameters
For evaluation various matrices like accuracy, sensitivity,
specificity, precision, recall and F
1
-Score [58-62] are taken
into consideration.The analysis of results is done step by
step through different phases as described below:
Phase 1: As described before in the experimental setup,
different pooling layers, activation functions, optimization
parameters and learning rates are evaluated to understand
the behaviour of the model. The obtained results for vari-
ation in parameter combination are presented pictorially in
the following graphs. The 3-dimensional bar graph repre-
sentation is used to visualize the behaviour of the model
with respect to testing accuracies achieved for different
combinational approaches made for the CNN through the
process of classification of diabetic retinopathy data.
Four pooling types such as max pooling, min pooling,
average pooling and maxmin pooling is considered here.
For each pooling, the variation in activation function, learn-
ing rate are presented in x-axis and y-axis respectively. Fur-
thermore, four widely used activation functions such as,
LReLu, ReLU, sigmoid and tanh are chosen to be studied
with combination of learning rate of 0.1, 0.01, 0.001 and
0.0001. Finally, the observed testing accuracy is given in
the z-axis. The learning ability of the model with respect to
various optimizers such as SGD, ADAM, Adadelta is also
observed by the colour indicated in graph respectively. Fig-
ure 3 shows the changes in accuracy value for max pooling
with different activation functions, learning rate and differ-
ent optimizer for backpropagation in the CNN model.
The Figure 3 (a), (b), (c), and (d) are the 3-dimensional
graphical representation of accuracy values measured using
max pooling at epoch 25, 50, 75, and 100 respectively.
From the Figure 3(a) it is observed that, at epoch 25,
a maximum accuracy result of 0.686 is obtained by us-
ing LReLu and ReLu activations at 0.01 learning rate with
Adadelta optimizer. The same accuracy is observed at
0.0001 learning rate with ADAM optimizer at 100 epochs.
The results show that the increase of epoch does not ensure
better accuracy here. Moreover, the accuracy can also de-
crease. This occurs due to overtraining which affects the
results or due to improper tuning of the learning rate.
The graphical Figure 3(b) shows increased accuracy val-
ues as compared to Figure 3(a). The two activation func-
tions ReLu and LReLu at 0.1 learning rate with Adadelta
optimizer present accuracies of 0.725 and0.705, respec-
tively.
Figure 3(c) shows that the Adadelta optimizer presents
an accuracy of 0.725 with tanh activation at 0.01 learning
rate for epoch 75. Moreover, the ADAM optimizer also
presents reliable results with an accuracy value 0.745 with
LReLu at 0.0001 learning rate. Figure 3(d) shows the accu-
racy of the model by using max pooling layer at epoch 100.
The best accuracy of 0.686 is achieved with ReLu activa-
tion at 0.01 learning rate and SGD optimization and also
by the tanh activation function at learning rate of 0.1 and
ADAM optimization. The LReLu activation achieves an
accuracy of 0.647 with SGD and ADAM optimizers at 0.1
and 0.0001, respectively. The sigmoid function presents
the lowest accuracy results.
Figure 4 represents the 3-dimensional graph for the sec-
ond type of pooling taken min pooling at different epochs
and features. The Figure 4(a) is the plot for accuracy val-
ues at epoch 25 using min pooling in the model. Here, the
ADAM optimizer at 0.0001 learning rate with ReLu acti-
vation function achieves an accuracy of 0.764 and the SGD
optimizer at 0.01 with LReLu activation gets a 0.66 accu-
racy. In Figure 4(b) for epoch 50, the ADAM optimizer
at 0.0001 learning rate with ReLu activation gets an accu-
racy of 0.686. In this same condition LReLu activation in-
creases the accuracy to 0.725. In this specific situation, the
LReLu performs better than the ReLu activation. Again,
the Adadelta optimizer at 0.0001 learning rate with LReLu
gives the same accuracy as 0.725. The Figure 4(c) is the
accuracy plot taking epoch value as 75, where only one
combination gives a marked accuracy result of 0.686 by
Adadelta optimizer at 0.0001 learning rate and tanh activa-
tion function. The Figure 4(d) gives the bar graph represen-
tation for epoch 100. In this point, the LReLu activation at
0.01 learning rate and ADAM optimizer in the model gets
an accuracy of 0.705 and tanh function at 0.1 learning rate
and Adadelta optimizer gets an accuracy of 0.686. The ob-
served results shows that, the min pooling and max pool-
ing in the model combined with ADAM or Adadelta opti-
mizer performs better than other optimizer combinations.
The following Figure 5 represents the plotting for another
type of pooling called average pooling taken in the model.
The Figure 5(a) is for epoch 25 which yields two impor-
tant accuracy observation as 0.686 and 0.705 using ADAM
and Adadelta optimizers. In the Figure 5(b), the ADAM
optimizer achieves an accuracy of 0.666 with ReLu activa-
tion at 0.0001 learning and 0.745 with LReLu at 0.01 learn-
ing rate. From the graph it also can be said that the average
pooling method in the network model along with SGD op-
timizer and learning rate of 0.01works well with tanh func-
tion resulting an accuracy value of 0.705 compared to ReLu
function that yields 0.666 accuracy. The Figure 5(c) is plot-
ted for 75 epochs and yields accuracies of 0.666 and 0.686
by using SGD optimizer at 0.1 and 0.01 learning rate with
ReLu and LReLu activations. The ADAM optimizer also
reaches an accuracy of 0.764 with tanh activation function
at 0.0001 learning rate. But with increased epoch of 100 at
Figure 5(d) the accuracy values does not increase further,
rather staying at value of 0.666 by SGD and ADAM opti-

Exploring the Parametric Impact on a Deep. . . Informatica 46 (2022) 205–221 211
(a)
(b)
(c)
(d)
Figure 3: Plotting of accuracy values by taking max pool-
ing in the model. (a) is for epoch =25, (b) is for epoch =50
(c) is for epoch =75 (d) is for epoch =100.
(a)
(b)
(c)
(d)
Figure 4: Plotting of accuracy values by taking min pooling
in the model. (a) is for epoch =25, (b) is for epoch =50, (c)
is for epoch =75 and (d) is for epoch =100.

212 Informatica 46 (2022) 205–221 M. Jena et al.
(a)
(b)
(c)
(d)
Figure 5: Plotting of accuracy values by taking average
pooling in the model. (a) is for epoch =25, (b) is for epoch
=50, (c) is for epoch =75 and (d) is for epoch =100.
mizers. Till now with these three pooling layers ReLu and
LReLu activation functions performs well on most of the
cases, while tanh performs in few and the sigmoid function
seems to be a poor performer at this point.
In this similar procedure the accuracy values achieved
by using max-min pooling in the model along with differ-
ent parameters are also analyzed. By taking 25 epochs in
the model, it reaches to an accuracy value of 0.803 using
ADAM optimizer at 0.0001 learning rate with tanh activa-
tion function. This accuracy does not increase in further
increase of epochs to 50, 75 or 100, rather falls to 0.666 in
some combinational approaches. The experimental results
clearly conveys that every condition, every combination of
parameter is effecting the model’s behaviour. Each combi-
national approach gives dissimilar results to analyze. From
these results, selective combinations, which gives the high-
est result among all, are chosen to test with more increased
epoch and taken to phase two.
Phase 2: In the second phase, the accuracy and other
evaluation parameters are analyzed for the chosen combi-
nations. The following Table 3 shows the selected accu-
racy values for different pooling with combination of fea-
tures and parameters chosen along with its sensitivity (Sn),
specificity (Sp), precision (Pr), recall (Re) and F
1
-score (F)
values. Above table contains evaluation parameters for the
combinations through which the maximum accuracy val-
ues are achieved in phase 1. It shows that, the max pool-
ing combined with Adadelta optimizer at 0.1 learning rate
goes well for achieving a good accuracy values. Finally,
the ADAM optimizer with LReLu and 0.0001 learning rate
attains the maximum accuracy among this for max pooling.
At this accuracy value, the sensitivity is 0.76 and the speci-
ficity the precision are 0.73 with recall and F
1
-score to be
0.76 and 0.745 respectively. From the combinational ap-
proaches for min pooling the ADAM optimizer gives bet-
ter results as compared to other optimizers and the learning
rate 0.0001 can be selected for proper tuning here. The
LReLu function also proved to be more performable than
ReLu and other activation functions in this situation. The
actual positive cases that are correctly identified is mea-
sured by sensitivity and the maximum of it is 0.79 for
the accuracy 0.76 and specificity of 0.74 while the preci-
sion, recall and F
1
-score are 0.73,0.8 and 0.76 respectively.
These are the best values obtained in this spot.
While taking average pooling in this model at various
epochs the tanh activation seems to perform better at this
state. At learning rate 0.01 accuracies are increased but
the maximum accuracy yield here is at 0.0001 learning rate
which is 0.764. The sensitivity and specificity are 0.791
and 0.740 while, the precision, recall and F
1
-score are 0.73,
0.8 and 0.76 respectively. Here the ADAM optimizer gives
the maximum result compared to all other optimizers.
In this case of maxmin pooling, the ADAM optimizer
performs very well compared to other optimizers again. It
gains an accuracy of 0.803 and a sensitivity of 0.9, which
is the maximum among all the previous results. Till now
this is the best combinational approach for the model us-

Exploring the Parametric Impact on a Deep. . . Informatica 46 (2022) 205–221 213
ing this dataset. The specificity, precision, recall and F
1
-
score are 0.741, 0.692, 0.9 and 0.782 respectively. Here it
is visualized how the change in pooling layer along with
its parameters and features effects the behaviour of model.
The LReLu activation works well for min pooling and tanh
function for average pooling, but in other pooling they are
not the same. The ADAM optimizer gives better result in
min pooling and maxmin pooling, however AdaDelta opti-
mizes well in max pooling for this model with this data.
The graphs and tables demonstrates the clear conceptu-
alization of the independent behaviour of parameters and
conditions in a particular scenario. Now, all of the above
combinations of all four pooling layers are taken to phase
3 for further experiment.
Phase 3: The above combinations are the repeated with
increased iteration up to 1500 and the acquired outputs are
analyzed. The following tables compares the accuracy cal-
culated at 100 and less than 100 iterations with accuracy
values at 1500 iterations. In this the actual set of com-
binations are selected i.e. the set of states which gives
higher results are finally selected at top best combinations
for this particular model. The accuracy at 1500 iterations
are shown in Table 4 for max pooling. From the above
table, it is marked that the accuracy value which is actu-
ally the maximum among the four values is not increased
more with the increase in epoch. Another new combina-
tion of features has acquired an increased highest accuracy
of 0.84 with the increased epoch of 1500. The sensitivity
is 0.87 while the specificity is 0.81. The precision, recall
and F
1
-score are 0.8, 0.88 and 0.84 respectively. So for
max pooling this can the best combination of parameters
and features to get a good result using this model. It can be
seen in table 4 that in min pooling the expected accuracy
value is not achieved, rather the accuracies decreased with
the increase in number of iterations. It may happen due to
overtraining which affects the result in opposite way.
Table 4 also illustrates the accuracy values and other
evaluation parameter values for the average pooling in
the model. Here also the increased accuracy value is
not the further increased value from the previous phase.
The Adadelta optimizer performs well here compared to
ADAM as before for this average pooling at learning rate
0.01 rather than 0.0001. It achieves an accuracy of 0.862
with a sensitivity of 0.85 which is smaller than the speci-
ficity 0.875. Here some more tuning may increase the ac-
curacy even more by increasing the sensitivity. The preci-
sion, recall and F
1
-score are 0.888, 0.84 and 0.86 respec-
tively (Table 4). In this table of accuracy representation
for max-min pooling layer, the highest accuracy value from
the previous phase increased to 0.86 from 0.80. Though the
sensitivity was high before, still it maintains a value of 0.88
and specificity of 0.84.
The precision, recall and F
1
-score are 0.85, 0.88 and 0.86
respectively. We can see that some accuracy values are in-
creased due to increase in iteration and some of them give
a decreased value as they might got stuck in local minima
or may be due to overtraining. Only a few among these
yields refinement of accuracies after increase in iteration
and is marked with bold font. Further, those highlighted
accuracy’s combinational condition can be taken for exper-
imentation with a more increased value of iteration.
Phase 4: Here the three pooling techniques i.e. max
pooling, average pooling and max-min pooling and their
combinational features are finally selected in Table 5
which are working well for this model and dataset. Fur-
ther this combinations can be used either with optimization
technique or with some hybridization to achieve more suit-
able model having higher accuracy values.
This table contains the three maximum accuracy values
of 0.843, 0.862 and 0.865 achieved in this experimental
approach with other evaluation parameters. These can be
taken as the best three combination of activation function,
learning rate, optimizer and pooling layer to achieve a good
accuracy for this model. In this paper only two class clas-
sification of diabetic retinopathy data is considered which
is either DR or normal. DR is taken as a five class problem
i.e. a normal or non DR class and four stages such as, mild
DR, moderate DR, severe DR, and proliferative DR classes
[39-40]. This empirical analysis will help us in the further
deep classification of DR. Some observations are made in
this experiment are stated below:
– Among various types of the pooling used here, max-
min pooling gives comparatively better result. The fu-
sion of max and min values of the pixels during the
selection through pooling generates a good selection
of pixel values for the model’s training and testing for
classification.
– A combined pooling helps in improving the perfor-
mance value of model. This fused pooling technique
also works well in [62].
– Unlike the ReLU function, the LReLu does not trans-
form all the negative values to zero. Here, the LReLu
performs finer than other activation functions.
– The combinations which give rise to good accuracy
results, amid of them, 50% of are having LReLu as
their activation function.
– Particularly, for this experiment if we grade the activa-
tion functions, LReLu takes the first position resulting
as the best activation function for image classification
using CNN.
– The tanh function comes in the second place and fi-
nally the ReLU comes in the third place by providing
good result in few combinations. But with increase in
iteration, the tanh function wins over the LReLu by
obtaining more accurate results.
– Going through the accuracy results, it is noticed that
more than 50% good accuracy values are seen in the
cases using Adam as an optimizer here. The Adadelta
optimizer also give results in less than 40% combina-
tional cases.

214 Informatica 46 (2022) 205–221 M. Jena et al.
Table 3: Selective combinations for various pooling.
Pooling Activation Optimizer LR Iter. Acc. Sn Sp Pr Re F
Max ReLU Adadelta 0.1 50 0.72 0.71 0.73 0.76 0.68 0.74
LReLu Adadelta 0.1 50 0.7 0.67 0.75 0.8 0.6 0.73
LReLu ADAM 0.0001 75 0.74 0.76 0.73 0.73 0.76 0.74
Tanh Adadelta 0.1 75 0.72 0.71 0.74 0.769 0.68 0.74
Min ReLU Adadelta 0.1 50 0.72 0.71 0.73 0.76 0.68 0.74
LReLu Adadelta 0.1 50 0.7 0.67 0.75 0.8 0.6 0.73
LReLu ADAM 0.0001 75 0.74 0.76 0.73 0.73 0.76 0.74
Tanh Adadelta 0.1 75 0.72 0.71 0.74 0.769 0.68 0.74
Avg ReLU Adadelta 0.01 50 0.745 0.74 0.75 0.769 0.72 0.754
LReLu Adadelta 0.01 25 0.7 0.789 0.656 0.57 0.89 0.666
LReLu ADAM 0.01 50 0.7 0.761 0.666 0.615 0.8 0.68
Tanh Adadelta 0.0001 75 0.764 0.791 0.74 0.73 0.8 0.76
Max-Min LReLu ADAM 0.001 50 0.72 0.8 0.677 0.615 0.84 0.695
Tanh ADAM 0.0001 25 0.803 0.9 0.741 0.692 0.92 0.782
Table 4: Combinations for several pooling selected from phase 1.
Pooling Activation Optimizer LR Acc.
(<=100)
Acc.
(at 1500)
Sn Sp Pr Re F
Max ReLU Adadelta 0.1 0.72 0.666 0.645 0.7 0.769 0.56 0.701
LReLu Adadelta 0.1 0.705 0.843 0.875 0.814 0.807 0.88 0.84
LReLu ADAM 0.0001 0.745 0.607 0.625 0.592 0.576 0.64 0.6
Tanh Adadelta 0.1 0.72 0.588 0.608 0.571 0.538 0.64 0.571
Min ReLU Adadelta 0.0001 0.764 0.549 0.565 0.535 0.5 0.6 0.53
LReLu Adadelta 0.1 0.725 0.549 0.56 0.538 0.538 0.56 0.549
LReLu ADAM 0.0001 0.725 0.627 0.684 0.593 0.5 0.76 0.577
Tanh Adadelta 0.0001 0.705 0.588 0.619 0.566 0.5 0.68 0.553
Avg ReLU Adadelta 0.01 0.745 0.647 0.63 0.66 0.73 0.56 0.678
LReLu Adadelta 0.01 0.7 0.862 0.85 0.875 0.884 0.84 0.867
LReLu ADAM 0.01 0.7 0.647 0.66 0.629 0.615 0.68 0.64
Tanh Adadelta 0.0001 0.764 0.607 0.615 0.6 0.615 0.6 0.615
Max-Min LReLu ADAM 0.001 0.72 0.686 0.66 0.714 0.769 0.6 0.714
Tanh ADAM 0.0001 0.803 0.865 0.88 0.846 0.85 0.88 0.867
Table 5: Combinations selected in phase 3 and their accuracy.
Pooling Activation Optimizer LR Acc.
(<=100)
Acc.
(at 1500)
Sn Sp Pr Re F
Max LReLu Adadelta 0.1 0.705 0.843 0.875 0.814 0.807 0.88 0.84
Avg Tanh Adadelta 0.01 0.7 0.862 0.85 0.875 0.884 0.84 0.867
Max-Min Tanh ADAM 0.0001 0.803 0.865 0.88 0.846 0.85 0.88 0.867

Exploring the Parametric Impact on a Deep. . . Informatica 46 (2022) 205–221 215
Figure 6: Proposed customized 2-branch CNN model for
DR classification.
– It can be concluded that, the Adam and Adadelta
works well for image in CNN compared to the SGD
optimizer. Many of the cases here shows SGD to be
very less powerful optimizer than others.
– The lower learning rate like 0.0001 and 0.001 obtains
superior results in half of the cases in this experimen-
tal analysis for CNN.
5.2 Evaluation of the model
From the above study of parameters, max-min pooling,
LReLu activation function, Adam optimizer and learning
rate as 0.0001 are chosen to be applied in the proposed 2-
branch CNN model for the five class classification of DR
using Kaggle dataset. Each CNN branch of the 2 branch
CNN model is inspired by the VGG16 network model [63].
Detailed structure of the proposed 2-branch CNN model is
shown in following Figure 6.
The proposed model is having two branches to which
inputs are supplied. These two branches are trained indi-
vidually with the Kaggle dataset for thorough learning of
features. After that, the output features of each branch are
fused and handed over to a fully connected neural network
for classification.
After pre-processing, the image of size 256×256×3 is
ready to be used in the network for training. Patch size
of the image given is 128×128.
For the 1
st
branch, for extraction of more possible fea-
tures locally, the image is divided into various patches. For
this patch division, functional layer Lambda is used. It is
included in the TensorFlow library. These received patches
are in one dimension format. So, to rearrange the extracted
patches into normal input shape, the Reshape layer is used
before going to VGG network through a convolution layer.
Here, the convolution layer is used to make the image size
suitable to be the input for VGG16 network. The pre-
trained VGG16 model from the library is used without its
fully connected layers. The pre-trained weights of the net-
work help the network to achieve better accuracy. The out-
put features are stored for further use.
In the 2
nd
branch, the pre-processed image is directly
given as input without patch division. The VGG16 network
is used without its fully connected layers and the output
features of the branch are stored. Both the output features
from the two branches are combined and fed to the fully
connected layer. The combined features are of size 1024
(i.e. 512+512 from each branch), from which the five class
predictions are performed.
Following is the steps used in the proposed 2-branch
model for the experiment -
Model flow Steps:
Step 1: Pre-processing of the Fundus Images
– Cropping
– Resizing of image
– Contrast Enhancement
– Normalization
Step 2: Loading the pre-processed data
Step 3: Processing for the Branch 1 CNN
– Patch division and resizing of the input im-
ages
– Input to the VGG16 network
– The output of the network is stored as A
Step 4: Processing for the Branch 2 CNN
– Input to the VGG16 network
– The output of the network is stored as B
Step 5: The output features of branch 1 and branch 2
are concatenated
Step 6: The total concatenated output is given to a
fully connected Neural Network for classification
6 Results and discussion
The results of the proposed model are discussed and com-
pared with state-of-the-art models. Along with various as-
sessment parameters such as accuracy, sensitivity, speci-
ficity, precision, recall, F1-score, AUC [64] and ROC [65]
are used here.
6.1 Analysis of results
Following Table 6 shows the classification result of the five
classes of diabetic retinopathy classification in the form of
confusion matrix. The above discussed assessment param-
eters are calculated from the confusion matrix and shown
in Table 7. To analyze the result in a proper way and to
access its performance, a result comparison table has been
shown in Table 8. The model proposed by S. Qummar et
al. [66] is also considering the same dataset and performing
the five-class classification. So, the comparison between
the models can be done.
The accuracy of the model is calculated as 0.966. The
highest sensitivity, specificity, precision, recall and F1-
score calculated are 0.992, 0.994, 0.978, 0.998 and 0.985.
The higher values of sensitivity and specificity show the

216 Informatica 46 (2022) 205–221 M. Jena et al.
Table 6: Confusion matrix from the proposed model.
Predicted class levels
No DR Mild Moderate Severe PDR
NPDR NPDR NPDR
Actual class
Label
No DR 137 1 3 0 1
Mild NPDR 5 135 1 0 1
Moderate NPDR 3 0 138 0 1
Severe NPDR 0 0 0 139 3
PDR 1 2 0 1 138
Table 7: Assessment parameter values from the confusion matrix.
Sensitivity Specificity Precision Recall F1-Score
No DR (0) 0.964789 0.984155 0.938356 0.991134 0.951389
Mild NPDR (1) 0.950704 0.994718 0.978261 0.987762 0.964286
Moderate NPDR (2) 0.971831 0.992958 0.971831 0.992957 0.971831
Severe NPDR (3) 0.992857 0.994737 0.978873 0.998239 0.985816
PDR (4) 0.958333 0.992933 0.971831 0.989436 0.965035
proportion of actual positives and effected subjects are clas-
sified accurately. The high precision rates of 0.978, 0.971,
0.978 and 0.971 in class 1, class 2, class 3 and class 4
shows a low false-positive rate in the classification model.
The higher recall values of 0.991, 0.992 and 0.998 for class
0, class 2 and class 3 shows very low false-negative rate.
The other two classes also have a false-negative rate in the
classification with a recall value of 0.987 and 0.989. The
weighted average of precision and recall is the F1-score
which is remarkable here. All the five values of F1-score
for five classes are more than 0.95 and the maximum is
0.98, which is an indicator as a good classifier.
The probability curve or the ROC curve is calculated and
shown in Figure 7. This graph shows the capability of our
model for distinguishing five classes. The performance of
classification at all classification thresholds is depicted in
the graph using the two-parameter i.e. true positive rate
and false-positive rate. The AUC is calculated by measur-
ing the entire two-dimensional area under the entire ROC
curve. It provides an aggregate measure of performance
for all possible thresholds in classification. AUC is scale-
invariant as it measures the ranking of prediction by ignor-
ing the absolute values. It is also a classification thresh-
old invariant, which means it measures the quality of the
models predictions without depending on the classification
threshold. The AUC for class 0, class 1, class 2, class 3,
and class 4 are 0.98, 0.96, 0.98, 0.98 and 0.99 respectively.
The maximum AUC achieved is of 0.99 by the PDR class
and the minimum achieved is 0.96 by the mild NPDR class.
A comparison between the models AUC and the AUC
from S. Qummar et al. [66] is shown in Table 8. Our
model outperforms in every class classification of the DR
by distinguishing each class at the highest rates. The maxi-
mum AUC of 0.99 shows the distinguishing capacity of the
proposed model is at its best.
The AUC values prove the model to be the best per-
former for the classification of diabetic retinopathy using
Figure 7: ROC Curve for the model.
fundus images. To give more detail about the model’s train-
ing capacity, the following curves are presented in Figure
8. The graph (a) in Figure 8 shows how the accuracy of
the model increases with increasing epochs. The graph (b)
of Figure 8 shows that the learning curves are in good fit.
The plot of the training loss decreases to a point of stability
which is good for the model loss.
From all evaluation parameters calculated and the graph
showing model accuracy, model loss with the ROC curve
and AUC value; it may be inferred that the model outper-
forms the existing model and can stand as a good classifier
for DR classification.
6.2 Novelty of the work
– Analysis of various parameters and using their best
values gives an outstanding performance in the pro-
posed 2-branch CNN model.
– The 2-branch classification model of CNN seems to
adopt the feature learning very well here.

Exploring the Parametric Impact on a Deep. . . Informatica 46 (2022) 205–221 217
Table 8: Comparison with state-of-the-art model’s output.
Various Class levels AUC from S. Qummar et al. [81] AUC of the proposed model
Class 0 0.97 0.98
Class 1 0.91 0.96
Class 2 0.78 0.98
Class 3 0.9 0.98
Class 4 0.94 0.99
Figure 8: Proposed model’s accuracy and loss.
– The detail learning of features i.e. the learning of lo-
cal and global features of fundus images helps a lot
to grasp the inter-class level differences of the dia-
betic retinopathy classification. The probability curve
shows it clearly.
– From the obtained result it can be concluded that, use
of VGG16 CNN model in this 2-branch classification
frame works well for diabetic retinopathy classifica-
tion.
– A good classification performance from a model can
be obtained through correct selection of parameters
and hyper-parameters for the model.
– The achieved AUC for each class levels summons the
proposed 2-branch model as a good classifier. The ac-
curacy and loss curve of the model also reciprocates it
as a good model.
6.3 Case study in IoT and Blockchain based
solution for smart healthcare
Blockchain technologies [67] in digital healthcare is very
important and showing its importance on routine basis.
Blockchain can be described as a powerful method to pro-
vide more secure and reliable data sharing between several
organizations. Secure and reliable communication between
several parties in a major challenge where blockchain tech-
nology can be very effective. One of the popular cases
uses of blockchain technology is patient-centric electronic
health records. In a report published by Johns Hopkins
University, it was mentioned that one of the causes of pa-
tient’s death was the medical errors in the patient records.
Therefore, blockchain technology can provide an effective
solution by introducing a blockchain based digital system
in order to provide a comprehensive and accurate record to
the healthcare provides about the patient by linking it with
the existing medical records of the patient. IoT in the other
hand can incorporate blockchain-based solutions into smart
devices that can be available in real time to the patient as
well as the health care provider. It will also help to monitor
the health of the people in real time basis. One framework
is proposed in Figure 9.
The framework consists of vital sign monitoring system,
an IoT server, a blockchain network, and a communica-
tion interface to collect the information from the healthcare
sensors from the patients. All the information is stored se-
curely and communicated to the medical staff for further
diagnosis and treatment. The approach once development
can be optimized to develop an IoT based smart device and
can be used by the medical staff efficiently.
7 Conclusion
Blockchain, AI and IoT based smart healthcare systems
needs some prior research that can be optimized further
and can be incorporated into portable smart devices. DR is
one of the complicated eye disease in all age group people
around the world and prior detection and proper treatment
is the only way to prevent the high percentage of visual
loss through surgical, pharmacological or laser treatment.
This is a pragmatic study about the reflected behaviour of a
CNN model due to various factors. In this paper, we have
investigated different CNN configurations by considering
several factors such as the type of pooling layer, activation

218 Informatica 46 (2022) 205–221 M. Jena et al.
Figure 9: A Typical IoT-Blockchain based smart healthcare
system.
used in architecture and the optimizer function, learning
rate and iterations for diabetic retinopathy classification.
To analyze the performance of different CNN variants, a
newly formed dataset is been used. The empirical evalua-
tion suggests that the max-min pooling with hyperbolic tan
activation when trained with Adam optimizer with 0.0001
learning rate for 1500 iteration gives the best accuracy of
86.50% among all the tested combination. This study will
help the future researchers to get ideas about the various
parameters and their effect on the model accordingly to
choose best possible combination for an architecture to be
designed. Here, to perform the experiment a two-branch
CNN model is proposed for the classification of fundus im-
age for diabetic retinopathy and the selected parameters are
used. For each class the average sensitivity, specificity, pre-
cision, recall, F1-score and AUC are achieved as 0.96, 0.99,
0.96, 0.99, 0.96 and 0.97 respectively from the proposed
approach which shows the outperformance of the model.
Data Availability
The fundus images used in this study are
available in the following data repository:
https://github.com/Manaswinijena/Fundus-
images/blob/master/REHLC Dataset.rar
Funding Statement
No funding was provided for this research.
Conflict of interest
The authors do not have any conflicts of interest.
Acknowledgment
We thank Dr. B.N.R. Subudhi (Retd. Professor and HOD
of Dept. of Opthalmology, M.K.C.G. Medical College,
Berhampur), Dr. Praveen Subudhi and Technician Bha-
bani Sankar Rath of Ruby Eye Hospital & Lasik Centre
(Berhampur) for their precious time to provide the fundus
images data for the experiment.
This work is supported by the Ministry of Science and
Higher Education of the Russian Federation (Government
Order FENU-2020-0022).
References
[1] Kumar S., Tiwari P. & Zymbler, M., “Internet of
Things is a revolutionary approach for future technol-
ogy enhancement: a review”, J Big Data, Vol. 6, pp.
111, 2019.
[2] Mansour R.F., El-Amraoui A., Nouaouri I., Díaz
V .G., Gupta D., Kumar S., "Artificial Intelligence and
Internet of Things Enabled Disease Diagnosis Model
for Smart Healthcare Systems",IEEE Access, vol. 9,
pp. 45137-45146, 2021.
[3] Sabanayagam C., Yip W., Ting D.S., Tan G., Wong
T.Y ., “Ten emerging trends in the epidemiology of
diabetic retinopathy”, Ophthalmic epidemiology, V ol.
23(4), pp.209-222, 2016.
[4] Piyasena M.M.P.N., Yip J.L., MacLeod D., Kim
M., Gudlavalleti V .S.M., “Diagnostic test accuracy of
diabetic retinopathy screening by physician graders
using a hand-held non-mydriatic retinal camera at
a tertiary level medical clinic”, BMC ophthalmol-
ogy, V ol.19(1), p.89, 2019.
[5] Paoletti M.E., Haut J.M., Plaza J., Plaza A., “A new
deep convolutional neural network for fast hyper-
spectral image classification”, ISPRS journal of pho-
togrammetry and remote sensing, V ol.145, pp.120-
147, 2018.
[6] Lee H., Kwon H., “Going deeper with contextual
CNN for hyperspectral image classification”, IEEE
Transactions on Image Processing, V ol.26 (10),
pp.4843-4855, 2017.
[7] W. Min, et al "Cross-Platform Multi-Modal Topic
Modeling for Personalized Inter-Platform Recom-
mendation," in IEEE Transactions on Multimedia,
vol. 17, no. 10, pp. 1787-1801, Oct. 2015
[8] Jena M., Mishra S. P., Mishra D., "A survey on appli-
cations of machine learning techniques for medical
imagesegmentation", International Journal of Engi-
neering & Technology, [S.l], vol.7 (4), pp.4489-4495,
nov. 2018. ISSN 2227-524X.

Exploring the Parametric Impact on a Deep. . . Informatica 46 (2022) 205–221 219
[9] He K., Zhang X., Ren S., Sun, J., “Deep resid-
ual learning for image recognition”, Proceedings of
the IEEE conference on computer vision and pattern
recognition, pp.770-778, 2016.
[10] Pandey S.K., Sharma, V ., “World diabetes day 2018:
battling the emerging epidemic of diabetic retinopa-
thy”, Indian journal of ophthalmology, V ol. 66(11),
p.1652, 2018.
[11] Xu K., Feng D., Mi H., “Deep convolutional neural
network-based early automated detection of diabetic
retinopathy using fundus image”, Molecules, V ol.
22(12), p.2054, 2017.
[12] Otálora S., Perdomo O., González F., Müller H.,
“Training deep convolutional neural networks with
active learning for exudate classification in eye fun-
dus images”, Intravascular Imaging and Computer
Assisted Stenting and Large-Scale Annotation of
Biomedical Data and Expert Label Synthesis, pp.146-
154, 2017.
[13] Quellec G., Charrière K., Boudi Y ., Cochener
B.,Lamard M., “Deep image mining for dia-
betic retinopathy screening”, Medical image analy-
sis, V ol.39, pp.178-193, 2017.
[14] Porwal P., Pachade S., Kamble R., Kokare M.,
Deshmukh G., Sahasrabuddhe V . Meriaudeau F.,
“Indian diabetic retinopathy image dataset (idrid):
A database for diabetic retinopathy screening re-
search”. Data, V ol.3 (3), p.25, 2018.
[15] Jena M., Mishra S. P., Mishra D. "Detection of Di-
abetic Retinopathy Images Using a Fully Convo-
lutional Neural Network", 2018 2nd International
Conference on Data Science and Business Analytics
(ICDSBA), pp. 523-527, September 2018.
[16] Jena M., Mishra S. P., Mishra D., “A Fully
Convolutional Neural Network for Recognition
of Diabetic Retinopathy in Fundus Images”,
Recent Patents on Computer Science, V ol.12 (1),
https://doi.org/10.2174/2213275912666190628124008.,
[17] Liskowski P., Krzysztof K., "Segmenting Reti-
nal Blood Vessels with Deep Neural Networks",
IEEE transactions on medical imaging, V ol.35 (11),
pp.2369-2380, 2016.
[18] Picek S., Samiotis I.P., Kim J., Heuser A., Bhasin
S., Legay, A., “December. On the performance of
convolutional neural networks for side-channel anal-
ysis”, International Conference on Security, Privacy,
and Applied Cryptography Engineering, pp. 157-176,
2018.
[19] Yang, X. et al. , "Deep Relative Attributes," IEEE
Transactions on Multimedia, vol. 18, no. 9, pp. 1832-
1842, Sept. 2016.
[20] Basirat M., Roth P.M., “The quest for the golden ac-
tivation function”. arXiv preprint arXiv:1808.00783,
2018.
[21] Pedamonti D., “Comparison of non-linear activation
functions for deep neural networks on MNIST clas-
sification task”, arXiv preprint arXiv: 1804.02763,
2018.
[22] Nwankpa C., Ijomah W., Gachagan A., Marshall S.,
“Activation functions: Comparison of trends in prac-
tice and research for deep learning”, arXiv preprint
arXiv: 1811.03378, 2018.
[23] Riegler G., Osman Ulusoy A., Geiger A., “Learning
deep 3d representations at high resolutions”, Proceed-
ings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 3577-3586, 2017.
[24] Fu H., Gong M., Wang C., Batmanghelich K., Tao
D., “Deep ordinal regression network for monocular
depth estimation”, Proceedings of the IEEE Confer-
ence on Computer Vision and Pattern Recognition,
pp. 2002-2011, 2018.
[25] Jena M., Mishra S.P., Mishra D., “Empirical Anal-
ysis of Activation Functions and Pooling Layers in
CNN for Classification of Diabetic Retinopathy”, In-
ternational Conference on Applied Machine Learn-
ing, pp.34-39, 2019.
[26] Smith L.N., “A disciplined approach to neural net-
work hyper-parameters: Part 1--learning rate, batch
size, momentum, and weight decay”, arXiv preprint
arXiv:1803.09820, 2018.
[27] Dauphin Y ., De Vries H., Bengio Y ., “Equilibrated
adaptive learning rates for non-convex optimization”,
Advances in neural information processing systems,
pp. 1504-1512, 2015.
[28] Ranganath R., Wang C., David B., Xing E., “An adap-
tive learning rate for stochastic variational inference”,
International Conference on Machine Learning, pp.
298-306, 2013.
[29] Zhao H., Fuxian L., Han Z., Zhibing L., "Research
on a learning rate with energy index in deep learn-
ing", Neural Networks, Vol.110, pp.225-231, 2019.
[30] Sharma A., "Guided stochastic gradient descent algo-
rithm for inconsistent datasets", Applied Soft Comput-
ing, Vol.73, pp.1068-1080, 2018.
[31] Kingma D. P., Jimmy B., "Adam: A method
for stochastic optimization", arXiv preprint arXiv:
1412.6980, 2014.
[32] Zeiler M. D., "ADADELTA: an adaptive learning rate
method", arXiv preprint arXiv: 1212.5701,2012.

220 Informatica 46 (2022) 205–221 M. Jena et al.
[33] Lam C., Yi D., Guo M., Lindsey T., “Automated
detection of diabetic retinopathy using deep learn-
ing”, AMIA Summits on Translational Science Pro-
ceedings, pp.147-155, 2018.
[34] Xu K., Feng D., Mi H., “Deep convolutional neural
network-based early automated detection of diabetic
retinopathy using fundus image”, Molecules, V ol.
22(12), pp.2054, 2017.
[35] Gharaibeh N., Al-Hazaimeh O.M., Al-Naami B., Na-
har K.M., “An effective image processing method for
detection of diabetic retinopathy diseases from reti-
nal fundus images”, International Journal of Signal
and Imaging Systems Engineering, V ol.11(4), pp.206-
216, 2018.
[36] Lin G.M., Chen M.J., Yeh C.H., Lin Y .Y ., Kuo H.Y .,
Lin M.H., Chen M.C., Lin S.D., Gao Y ., Ran A., Che-
ung C.Y ., “Transforming retinal photographs to en-
tropy images in deep learning to improve automated
detection for diabetic retinopathy”, Journal of oph-
thalmology, 2018.
[37] Adem K., “Exudate detection for diabetic retinopa-
thy with circular Hough transformation and convolu-
tional neural networks”, Expert Systems with Appli-
cations, V ol.114, pp.289-295, 2018.
[38] Li Y .H., Ye N.N., Chen, S.J., Chung, Y .C., Computer-
assisted diagnosis for diabetic retinopathy based on
fundus images using deep convolutional neural net-
work. Mobile Information Systems, 2019.
[39] Darshit D., Shenoy A., Sidhpura D., Ghapure P., "Di-
abetic retinopathy detection using deep convolutional
neural networks", 2016 International Conference on
Computing, Analytics and Security Trends (CAST),
pp.261-266, 2016.
[40] Nikhil M., Rose A., “Diabetic retinopathy stage clas-
sification using CNN”, International Research Jour-
nal of Engineering and Technology (IRJET), V ol.6,
p.5969, 2019.
[41] Jaiswal A.K., Tiwari P., Kumar S., Al-Rakhami M.S.,
Alrashoud M., Ghoneim A., “Deep Learning-Based
Smart IoT Health System for Blindness Detection Us-
ing Retina Images”, IEEE Access, Vol. 9, pp. 70606-
70615, 2021.
[42] L. Hu, et al, "Software defined healthcare networks,"
IEEE Wireless Communications, V ol. 22(6), pp. 67-
75, December 2015.
[43] Pratt H., Coenen F., Broadbent D. M., Harding S.P.,
Zheng Y ., “Convolutional neural networks for dia-
betic retinopathy”, Procedia Computer Science, V ol.
90, pp. 200-205, 2016.
[44] MM Althobaiti, KPM Kumar, D Gupta, S Ku-
mar, RF Mansour., An Intelligent Cognitive Comput-
ing based Intrusion Detection for Industrial Cyber-
Physical Systems,Measurement 186, 110145, 2021.
[45] A Tandon, A Dhir, A.K.M. Najmul Islam, M Män-
tymäki, Blockchain in healthcare: A systematic lit-
erature review, synthesizing framework and future
research agenda, Computers in Industry, V olume
122,103290, 2020
[46] MI Ali, S Kaur, A Khamparia, D Gupta, S Kumar,
A Khanna, F Al-Turjman, "Security challenges and
cyber forensic ecosystem in IOT driven BYOD envi-
ronment", IEEE Access 8, 172770-172782, 2020.
[47] Wang, H., “IoT based clinical sensor data man-
agement and transfer using blockchain technology”,
Journal of ISMAC, V ol. 2(03), pp.154-159, 2020.
[48] Poux, F. and Billen, R., 2019. V oxel-based 3D point
cloud semantic segmentation: unsupervised geomet-
ric and relationship featuring vs deep learning meth-
ods. ISPRS International Journal of Geo-Information,
8(5), p.213.
[49] Bengio Y ., Courville A.C., Vincent P., “Represen-
tation Learning: A Review and New Perspectives”,
IEEE Trans.Pattern Anal. Mach. Intell., V ol. 35,
pp.1798–1828, 2013.
[50] Li B., Lu, H., Zhang H., Tan S., Ji Z., “A
multi-branch convolutional neural network for de-
tecting double JPEG compression”, arXiv preprint
arXiv:1710.05477, 2017.
[51] Feng J., Wu X., Shang R., Sui C., Li J., Jiao L., Zhang
X., “Attention multibranch convolutional neural net-
work for hyperspectral image classification based on
adaptive region search”, IEEE Transactions on Geo-
science and Remote Sensing, 2020.
[52] Gao H., Yang Y ., Lei S., Li C., Zhou H., Qu X.,
“Multi-branch fusion network for hyperspectral im-
age classification”, Knowledge-Based Systems, V ol.
167, pp.11-25, 2019.
[53] Ge Z., Cao G., Li X., Fu, P., “Hyperspectral Image
Classification Method Based on 2D–3D CNN and
Multibranch Feature Fusion”, IEEE Journal of Se-
lected Topics in Applied Earth Observations and Re-
mote Sensing, V ol.13, pp.5776-5788, 2020.
[54] Sahu S., Singh A.K., Ghrera S.P., Elhoseny M., “An
approach for de-noising and contrast enhancement of
retinal fundus image using CLAHE”, Optics & Laser
Technology, V ol.110, pp.87-98, 2019.
[55] Ruby Eye Hospital & Lasik Centre (REHLC Dataset).
URL: https://github.com/Manaswinijena/Fundus-
images/blob/master/REHLC Dataset.rar

Exploring the Parametric Impact on a Deep. . . Informatica 46 (2022) 205–221 221
[56] Raman R., Srinivasan S., Virmani S., Sivaprasad
S., Rao C., Rajalakshmi R., “Fundus photograph-
based deep learning algorithms in detecting diabetic
retinopathy. Eye, V ol. 33(1), pp.97-109, 2019.
[57] Hemanth D.J., Deperlioglu O., Kose, U., “An en-
hanced diabetic retinopathy detection and classifica-
tion approach using deep convolutional neural net-
work”, Neural Computing and Applications, V ol.32
(3), pp.707-721, 2020.
[58] Pawara, P., Okafor, E., Schomaker, L., & Wiering, M.
Data augmentation for plant classification. In Interna-
tional Conference on Advanced Concepts for Intelli-
gent Vision Systems (2017, September) 615-626.
[59] Salamon J., Bello, J. P., “Deep convolutional neu-
ral networks and data augmentation for environmen-
tal sound classification”, IEEE Signal Processing Let-
ters, V ol. 24(3), pp. 279-283, 2017.
[60] Wong H.B., Lim G.H., “Measures of diagnostic ac-
curacy: sensitivity, specificity, PPV and NPV”, Pro-
ceedings of Singapore healthcare, V ol. 20(4), pp.316-
318, 2011.
[61] Zhang D., Wang J., Zhao X., “September. Estimating
the uncertainty of average F1 scores”, Proceedings of
the 2015 International Conference on The Theory of
Information Retrieval, pp. 317-320, 2015.
[62] Hamel P., Lemieux S., Bengio Y ., Eck D., “October.
Temporal Pooling and Multiscale Learning for Auto-
matic Annotation and Ranking of Music Audio”, IS-
MIR, pp. 729-734, 2011.
[63] Simonyan K., Zisserman A., “Very deep convolu-
tional networks for large-scale image recognition”,
arXiv preprint arXiv:1409.1556, 2014.
[64] J. A. Hanley, B. J. McNeil, “The meaning and use
of the area under a receiver operating characteris-
tic (ROC) curve”, Radiology, vol. 143(1), pp. 29-36,
1982.
[65] A. Dalyac M. Shanahan, J. Kelly, “Tackling Class Im-
balance With Deep Convolutional Neural Networks”,
pp. 30-35, 2014.
[66] Qummar S., Khan F.G., Shah S., Khan A., Shamshir-
band S., Rehman Z.U., Khan I.A., Jadoon W., “A deep
learning ensemble approach for diabetic retinopathy
detection”, IEEE Access, V ol. 7, pp.150530-150539,
2019.
[67] M. A. Rahman et al., "Blockchain-Based Mobile
Edge Computing Framework for Secure Therapy Ap-
plications," IEEE Access, vol. 6, pp. 72469-72478,
2018.

222 Informatica 46 (2022) 205–221 M. Jena et al.