https://doi.org/10.31449/inf.v45i5.3561 Informatica 45 (2021) 659–665 659 
 
An Optimized Deep Learning Based Technique for Grading and 
Extraction of Diabetic Retinopathy Severities 
Qiu-Ming Zhang  
School of Computer science, China University of Geosciences, Wuhan 430074, China 
E-mail: qiumingzhang98@outlook.com 
 
Jing Luo 
Practice training center, Wuhan Vocational College of Software and Engineering, Wuhan 430205, China 
E-mail: jingLuo12@hotmail.com 
 
Korhan Cengiz 
Department of Electrical - Electronics Engineering, Trakya University, 22030, Edirne, Turkey 
E-mail: korhancengiz@trakya.edu.tr  
 
Keywords: diabetic retinopathy, machine learning, deep learning, convolution neural network, severity grading 
 
Received: May 21, 2021 
The prognosis of Diabetic Retinopathy (DR) requires regular eye examinations, as ophthalmologists 
depends on fundus segmentation to treat DR pathologies. Automated approaches for detection, 
segmentation and classification have developed as an imperative area of research for the effective 
diagnosis of DR for the treatment of serious eye conditions that prevent visual impairment. Diagnosis of 
various DR lesions, as well as different severities, helping the ophthalmologists to analyze variations in 
fundus images and take the necessary measures before the disease progresses. Deep learning techniques 
have evolved as a recent advent to combat the issues of conventional machine leaning based methods. 
An optimized deep learning framework is proposed in this article for grading and extraction of diabetic 
retinopathy severities. This involves various steps like background segmentation, feature set extraction, 
feature optimization using Cuckoo search and Convolutional Neural Network (CNN) severity grade 
classification. The method was validated on two standard datasets MESSIDOR and IDRiD. The 
proposed method yields an accuracy value of 97.55%, cross entropy loss of 0.367 and time intricacy of 
20 mins and 15 secs for MESSIDOR and 98.02% cross entropy loss of 0.345 and time intricacy of 22 
mins and 21 secs for IDRiD dataset; respectively. The state-of-the-art comparison depicts that the 
proposed CNN based method provides a maximum accuracy improvement of 10.46% comparative to the 
existing methodology. The proposed framework yields better accuracy by procurement of the 
investigative outcomes acquired exhibits proficient DR determination.  
Povzetek:  Opisana je metoda ugotavljanja sladkorne bolezni iz slik očesnega ozadje na osnovi 
strojnega in globokega učenja. 
1 Introduction  
Prolonged diabetes leads to progressive microvascular 
complication, Diabetic Retinopathy (DR) which is the 
major source of serious vision impairments. DR cannot 
be recognized at the early stage but as the condition 
progresses, various lesions start appearing on retinal 
surface making early detection of disease, a crucial task. 
Various DR related lesions are depicted in Figure 1. 
DR is classified into two different forms: Non- 
Proliferative DR (NPDR) and Proliferative DR (PDR), 
however, in this paper NPDR stages are considered as 
they include the early signs of retinopathy [1-3].  
Different means are stated in the previous works for 
the identification of various DR lesion. A CLEAR-DR 
arrangement was anticipated by Devinder Kumar, et al. 
to enable the medical decision sustenance for DR [4]. 
The efficiency and usefulness of the anticipated system is 
established for the identification of DR which further 
improves the clinical practices. Ahmed ElTanboly, et al. 
developed a CAD based method for the grading of 
different DR stages [5]. A three-way Dr detection 
method was presented by Anam Tariq, et al. [6] for 
classification of different DR lesions. The system 
validation is done along with the assessment of various 
 
Figure 1: Different DR related Lesions. 

660 Informatica 45 (2021) 659–665 Q. Zhang et al. 
 
parameters to justify the viability of the anticipated 
technique. Sunrita Poddar, et al. [7] further presented a 
DR and maculopathy grading mechanisms involving 
various sequential steps. The classification is done based 
on the shape and degree of the lesions identified. Another 
grading mechanism was anticipated by Alan D. Fleming, 
et al. [8] for the detection of microaneurysms (MAs) and 
its further differentiation from the noise like spots. This 
method provides satisfactory outcomes for the 
identification of MAs.  
Advances have been made in the field of medical 
imaging research and classification through the 
introduction of a new approach to bioimaging that relies 
on an ensemble of convolutional neural networks 
(CNNs). Li et al. [9] presented a DR classification 
approach using the pretrained networks tested on small 
datasets. Authors in [10] applied a CNN based transfer 
learning approach for DR recognition which yields much 
better sensitivity than the other conventional approaches. 
GoogleNet CNN platform was used by the authors in 
[11] in order to detect the symptoms of diabetic macular 
edema (DME) utilizing the benchmark MESSIDOR and 
the other is the E-Ophtha datasets. AlexNet based 
architecture was presented by the authors in [12] using 
the MESSIDOR dataset for diagnostic information 
extraction of DR severities. A zoon-in-Net based DR 
diagnostic platform was built by Wang, et al. [13] for 
retinopathy detection and severity grading. DR 
identification performance was improved by the authors 
in [14] using six different CNN architectures using the 
deep learning concept. A study was presented by the 
authors in [15] that compares the traditional DR 
classification methods to the deep learning-based 
approaches using various standard datasets. Authors in 
[16] developed a cross-disease review network for 
finding various specific characteristics of intrinsic 
relationship between diabetic macular edema (DME) and 
DR. Bhardwaj, et al. [17] presented a Quadrant based 
Inception-V3 network for DR grading using the two 
benchmark datasets (MESSIDOR and IDRiD). The 
methods proposed by then could be used for mass 
screening of DR and provide a second opinion to the 
ophthalmologists. Authors in [18] presented a significant 
DR grading method using IDRiD and MESSIDOR 
dataset considering five-fold validation of the CNN 
network. Various stages of pre-processing are used by 
the authors in order to transfer the architecture to various 
computational levels.  
This literature survey of various techniques suggests 
that the deep learning-based strategies are being focused 
on in the present scenario in order to provide better 
results for various image recognition and classification 
approaches [19-21]. The DR classification-based 
approaches have also shifted their concern towards the 
deep learning as the previously used traditional methods 
were labor intensive, require professional information 
[22-24]. However, despite of various advantages of DNN 
methods, still it possesses the challenge for healthcare 
application and its practical implementation. In order to 
address such limitations, this article aims at evaluating 
the appropriateness of CNN-enabled techniques for 
classifying DR into various DR severities and provide a 
significant contribution in medical imaging domain.  
 
This article focuses on developing an optimized deep 
learning-based technique for grading and extraction of 
diabetic retinopathy severities which can automatically 
detect lesions and can assess its severity grade as mild, 
moderate or severe NPDR stage. The stages involved are 
background segmentation, feature set extraction, feature 
optimization and severity grade classification. Cuckoo 
search (CS) optimization algorithm is applied on the 
extracted feature attributes to obtain the Optimized 
Feature Set which can be used for CNN based DR 
severity categorization whose efficacy is observed 
experimentally by the computation of various evaluation 
parameters. The method was validated on MESSIDOR 
[25] and IDRiD [26] dataset. The proposed deep learning 
framework exhibits the feasible ability for proficient DR 
determination.  
The remaining paper is systematized as follows: 
section 2 presents the methods and materials followed by 
the results and discussion of the proposed framework in 
section 3. Section 4 provides the concluding remarks of 
the article including the future considerations. 
2 Material and methods 
An optimized deep learning-based technique has been 
presented in this article for grading and extraction of 
diabetic retinopathy severities for providing a DR 
classification solution for retinal fundus images. The 
proposed framework which is divided into various 
successive steps signified in Figure 2. 
The undesired retinal components are masked using 
background segmentation and removal. These feature set 
is extracted from the final fully connected layers for the 
CNN architecture which are then are optimized using 
cuckoo search optimization algorithm and the optimized 
feature set is used for CNN based DR severity 
classification. The effectiveness of severity classification 
is experimentally observed using various evaluation 
 
Figure 2: Proposed Framework for optimized deep 
learning based DR grading technique. 

An Optimized Deep Learning Based Technique for Grading and Extraction ... Informatica 45 (2021) 661–665 661 
 
 
Figure 4: Basic Convolutional Neural Network architecture. 
parameters and block-wise explanation of the proposed 
methodology is detailed in the following section. 
2.1 Background segmentation and removal 
The raw fundus images are obtained from standard 
dataset MESSIDOR [25] and latest IDRiD dataset [26] to 
assess the DR abnormalities. The original fundus images 
involves several image related artefacts, thus, some of 
the image pre-processing techniques are required to 
diminish these flaws preceding to Optical Disc (OD) 
localization. The image pre-processing stages which are 
utilized for this experimental set-up consists of 
conversion of RGB intensities, histogram equalization 
followed by the segmentation of optical disc and blood 
vessel retinal components are removed primarily due to 
their less relevance for anomaly identification and 
classification. Some of these segmented component, if 
not removed at the primary stage may lead to 
misclassification due to its similarity with the 
pathological indicators. The training data for neural 
networks is very essential as lack of data and small 
dataset problem may lead to class imbalance. For 
removing this redundancy, various data augmentation 
methods are used in this study like horizontal and vertical 
flips, random rotation, random scaling in order to 
increase the training samples. 
2.2 Feature optimization 
The wide range of feature set is extracted from the final 
fully connected layers for the CNN architecture which 
are then are optimized using cuckoo search optimization 
algorithm. The Cuckoo Search (CS) optimization 
algorithm to provide Optimized Feature Set (OFS) to the 
classification stage. This approach is based on the 
reproduction process of cuckoo birds. Cuckoo birds lay 
their eggs in the nest of other host species so that their 
off-springs can be raised by these proxy parents. If these 
proxy parents discover about the alien eggs in their nests, 
these eggs are either thrown away or the entire nest is 
uninhibited. The algorithm of CS optimization is 
restricted to certain constraints. 
Constraint 1: Individual cuckoo bird casually picks out 
a random nest and lays one egg in it. 
Constraint 2: The nest comprising the high quality of 
eggs is further approved for the next generation, 
however, others are discarded. 
Constraint 3:  The probability of finding the alien egg 
by the host species is 𝑝 𝑎 𝜖 [0,1] and these alien eggs are 
either thrown away or the nest can be uninhibited.  
Cuckoo search optimization algorithm is operative 
for solving global optimization problem and the major 
advantage of this algorithm is its simple implementation. 
It is less complex and efficient optimization approach as 
compared to other meta-heuristic approaches like particle 
swarm optimization and harmony search algorithms as it 
considers only single parameter [27]. 
The complete CS algorithm is given in Figure 3. 
 
 
2.3 CNN architecture for DR classification 
Convolutional deep neural networks are suited for image-
based retinopathy diagnostics. In this article, DR 
classification is accomplished using the CNN 
architecture. The basic CNN model depicted in Figure 4 
consisting of an input layer followed by a pile of 
convolutional layers, several pooling-merging layers, 
fully connected layers, and an output layer to derive four 
different DR severities [28]. 
The basic CNN layers are explained below: 
Convolution Layer: The convolution layer is designed to 
accept the input and then filter kernels are used followed 
by the activation unit to obtain the output features. While 
considering X i as the input image and y i as the output of 
convolutional operation. The function expressing the 
convolutional layer output is given by Eq. (1). 
𝑦 𝑖 = ∑ 𝑤 𝑖 ∗ 𝑋 𝑖 𝐽 𝑖 =0
+ 𝑏 
(1) 
where 𝑤 is weight vector, 𝑖 denoted the total nodes till 
the maximum number of node 𝐽 and ∗ denotes the 
convolutional operator.  
Rectified Linear Unit (ReLU) is utilized to compute the 
maximum of the output obtained using the function 
𝑓 (𝑥 ) = max (0, 𝑥 ). 
Pooling-merging Layer: It is common to append the 
convolutional layer with the pooling layer after the 
 
Figure 3: Cuckoo Search Algorithm. 
Objective function is defined as 𝑓 (𝑥 ) where 𝑥 =
 (𝑥 1
, 𝑥 2
, … … , 𝑥 𝑖 ) 
Create a initial population 𝑥 𝑖 consisting of 𝑛 host nests 
where 𝑖 = 1,2, … … , 𝑛 
WHILE (t < Maximum Generation) or (stopping 
criteria) 
Get any 𝑖 𝑡 ℎ
 cuckoo arbitrarily by Levy distribution, 
Assess its fitness function : 𝐹 𝑖 
Among the initial population of host nests, choose any 
𝑗 𝑡 ℎ
 nest randomly  
Evaluate its fitness function : 𝐹 𝑗 
IF  (𝐹 𝑖 > 𝐹 𝑗 ) 
Replace 𝑗 with a new solution 
END 
A segment (𝑝 𝑎 ) of inferior nests are uninhibited and the 
new nests are constructed  
Keep the best solutions/ nest 
The best solutions are ranked in an order to find the 
existing best solution 
Permit the existing best solution to the next generation 
END WHILE 
 

662 Informatica 45 (2021) 659–665 Q. Zhang et al. 
 
convolution operation. This is done so as to obtain the 
down sampled output. The most commonly used pooling 
operator is max pooling that computes the local 
maximum of input features for intermediate 
dimensionality reduction.  
Fully connected Layer: A series of fully connected 
layers are stacked subsequently after adjacent 
convolution and pool-merging layers. This particular 
stack of layer is responsible for facilitating the flattening 
of output into single vector for classification purpose. 
This layer uses the softmax activation function for 
computing the output and it is expressed by Eq. (2). 
𝑆𝑜𝑓𝑡𝑚𝑎𝑥 (𝑋 ) =
𝑒 𝑚𝑛
∑𝑒 𝑚𝑛
𝑛 
(2) 
where 𝑒 𝑚𝑛
 is exponential operator to obtain the output 
vector. 
The CNN architecture used in this work for DR 
recognition and analysis consist of three subsequent 
convolutional and pooling layers. These layers consist of 
convolutional filters of size 5×5 and the pooling filters of 
size 2×2. The number of filters keeps in increasing from 
to increase the network depth. The parameter 
considerations of the network layers are tabulated in 
Table 1. 
2.4 Performance evaluation parameters  
The performance of the presented CNN model for DR 
recognition and classification is measured for the metrics 
like accuracy, time intricacy, and cross entropy cost 
function are formulated in the following equations.  
i. Accuracy: It defines the percentage of exactly 
corrected outcomes obtained out of the total number 
of events. The formula for accuracy is expressed by 
Eq. (3). 
𝐴𝐶𝐶 =
𝑇𝑃 + 𝑇𝑁
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁
 
(3) 
ii. Sensitivity: The percentage of accurately indicated 
chronic retinal diseased class cases which actually 
possess chronic DR symptoms are indicated by 
Sensitivity. It is expressed by Eq. (4). 
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 =
𝑇𝑃
𝑇𝑃 + 𝐹𝑁
 
(4) 
 
iii. Specificity: Specificity provides the higher 
probability of truly indicated non-chronic DR class 
without giving the false positive outcomes. It is 
expressed by Eq. (5).  
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =
𝑇𝑁
𝑇𝑁 + 𝐹𝑃
 
(5) 
 
iv. Precision: This value indicated the correctly or 
precisely obtained accurate results from the total 
number of positive predictions. Precision is expressed 
by Eq. (6). 
 
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃
𝑇𝑃 + 𝐹𝑃
 
(6) 
 
v. Cross-Entropy: The cross-entropy loss is able to 
provide the correct classification capabilities of a 
neural network. The expression is given in Eq. (7).  
𝐶𝑟𝑜𝑠𝑠 𝐸𝑛𝑡𝑟𝑜𝑝𝑦 = 
1
∑ 𝑂𝑢𝑡𝑝𝑢𝑡 𝑜 ,𝑡 log(𝑝 𝑜 , 𝑡 )
𝑁 𝑡 =1
  
 
 
(7) 
where 𝑁 = classes, true classification = 𝑡 for observation 
𝑜 , prediction probability 𝑝 . 
3 Results and discussion 
The experimentation done for DR severity grading is 
done utilizing the optimized framework and the network 
performance is accessed using several evaluation 
parameters and the cost analysis is done in terms of cross 
entropy as well as time intricacy. The images are taken 
from both the datasets: MESSIDOR and IDRiD which 
are initially background segmented followed by the 
optimized feature set reduction using the Cuckoo search 
methodology. The outcomes of background segmentation 
and removal are depicted in Figure 5. 
The proposed framework the CNN classification 
network and the performance assessment in terms of 
various performance parameters are shown in Figure 6.   
The comparison done in Figure 6 indicates that the 
favorable outcomes are achieved with the performance 
parameters as; 97.55% accuracy, 98.04% sensitivity, 
97.06% specificity and 97.09% precision value for 
MESSIDOR dataset and 98.02% accuracy, 98.04% 
sensitivity, 98.00% specificity and 98% precision value 
for IDRiD dataset. 
The analysis of cost effectiveness is done in regards 
of cross entropy and time intricacy. For a reliable 
performance of a CNN network, these values should be 
Layer Kernel Size Stride 
Size 
Number of 
Filters 
Conv layer 
1 
5×5×1 1 25 
Pool layer 
1 
2×2×1 2 25 
Conv layer 
2 
5×5×25 1 50 
Pool layer 
2 
2×2×1 2 50 
Conv layer 
2 
5×5×50 1 75 
Pool layer 
2 
2×2×1 2 75 
Table 1: Parameter considerations for the CNN 
architecture used in the proposed system. 
 
Figure 5: (a) Original image, (b) OD segmented part (c) 
Retinal image after segmentation. 

An Optimized Deep Learning Based Technique for Grading and Extraction ... Informatica 45 (2021) 663–665 663 
 
minimum. The graphical presentation of these parameters 
in done in Figure 7.  
The cross-entropy loss of 0.367 is obtained for time 
intricacy of 20 minutes and 15 seconds for MESSIDOR 
dataset and 0.345 cross entropy loss with time intricacy 
of 22 minutes and 21 seconds for IDRiD dataset. As the 
cross entropy reduces for the proposed method, time 
intricacy increases, thereby, maintaining a tradeoff.  
3.1 Comparison with other methods 
The comparative analysis is done two way in this article. 
Firstly, the pretrained CNN models are compared with 
the proposed model in order to validate its flexibility and 
performance. The various pretrained CNN models used 
are AlexNet [29], GoogleNet [30], ResNet [31] and 
VggNet [32, 33]. The comparison is drawn in terms of 
time intricacy and accuracy which are shown in Figure 8. 
This comparative analysis reveals that the proposed 
method withstands among all the pretrained model 
providing the maximum performance for DR severity 
grading in terms of accuracy. However, the time 
intricacy comparison reveals that proposed framework 
maintains a trade-off in this case and is a bit complex 
comparative to the other CNN pretrained architectures.  
In this article, a comparative analysis with the state-
of-the-art methods indicated in literature is also done in 
terms of percentage accuracy improvement. The 
comparison is drawn in Figure 9.  
The competency of the proposed optimized approach 
for DR severity grading is assessed in terms of 
percentage accuracy improvement. It provides reliable 
solution with maximum percentage improvement of 
10.46% from the state-of-the-art literary work. 
4 Conclusions 
This article proposes an optimized deep learning-based 
technique for grading and extraction of diabetic 
retinopathy severities. The qualitative and quantitative 
analysis is based on various performance parameters as 
well as cost analysis. The two benchmark datasets 
evaluated for the study are MESSIDOR and IDRiD. The 
background segmentation, feature set extraction, feature 
optimization and CNN based severity grade classification 
stages are involved in this work for improving the 
network performance. The proposed method yields an 
accuracy value of 97.55% for MESSIDOR and 98.02% 
for IDRiD dataset maintaining a cross entropy loss of 
0.367 and 0.345 respectively. The time intricacy of 20 
mins and 15 secs is observed for MESSIDOR whereas 22 
mins and 21 secs are seen for IDRiD dataset which 
maintains a tradeoff with the network performance. The 
state-of-the-art comparison reveals the feasibility of the 
proposed classification method which provides a 
maximum accuracy improvement of 10.46% with the 
 
Figure 6: Performance parameters obtained for proposed 
framework. 
 
Figure 7: Cost analysis outcomes for cross entropy and 
time intricacy. 
96,50%
97,00%
97,50%
98,00%
98,50%
Accuracy Sensitivity Specificity Precision
Performance (%)
Parameters
Performance Comparison
For MESSIDOR For IDRiD
19:40
20:09
20:38
21:07
21:36
22:04
22:33
0,33
0,34
0,35
0,36
0,37
Time Intricacy (mins:secs)
Cross Entropy loss
Cross Entropy vs Time Intricacy
Cross Entropy loss Time Intricacy
 
Figure 8: Comparison with other CNN pretrained 
models. 
 
Figure 9: Comparison of existing state-of-the-art 
methods with the proposed framework. 
 
00:00
04:48
09:36
14:24
19:12
00:00
86%
88%
90%
92%
94%
96%
98%
100%
AlexNet
GoogleNet
ResNet
VggNet
Proposed…
AlexNet
GoogleNet
ResNet
VggNet
Proposed…
MESSIDOR dataset IDRiD dataset
Time Intricacy (mins:secs)
Accuracy (%)
Accuracy Time Intricacy (mins:secs)
0,00%
2,00%
4,00%
6,00%
8,00%
10,00%
12,00%
82%
84%
86%
88%
90%
92%
94%
96%
98%
100%
PErcentage Improvement
Accuracy
Comparison with sate of the art methods
Accuracy Percentage Improvement

664 Informatica 45 (2021) 659–665 Q. Zhang et al. 
 
currently prevailing methodology. The proposed 
technique can be used in the future for providing a 
reliable solution enabling the deep learning concept for 
providing the real time DR severity grading services to 
the ophthalmologists. 
5 References  
[1] Akram, M. U., Khalid, S., Tariq, A., Khan, S. A., & 
Azam, F. (2014). Detection and classification of 
retinal lesions for grading of diabetic retinopathy. 
Computers in biology and medicine, 45, 161-171. 
https://doi.org/10.1016/j.compbiomed.2013.11.014. 
[2] Fleming, A. D., Philip, S., Goatman, K. A., Olson, J. 
A., & Sharp, P. F. (2006). Automated 
microaneurysm detection using local contrast 
normalization and local vessel detection. IEEE 
transactions on medical imaging, 25(9), 1223-1232. 
https://doi.org/10.1109/tmi.2006.879953. 
[3] ElTanboly, A., Ghazal, M., Khalil, A., Shalaby, A., 
Mahmoud, A., Switala, A., ... & El-Baz, A. (2018, 
April). An integrated framework for automatic 
clinical assessment of diabetic retinopathy grade 
using spectral domain OCT images. In 2018 IEEE 
15th International Symposium on Biomedical 
Imaging (ISBI 2018), pp. 1431-1435. IEEE. 
https://doi.org/10.1109/ISBI.2018.8363841. 
[4] Kumar, D., Taylor, G. W., & Wong, A. (2019). 
Discovery radiomics with CLEAR-DR: interpretable 
computer aided diagnosis of diabetic retinopathy. 
IEEE Access, 7, 25891-25896.  
https://doi.org/10.1109/ACCESS.2019.2893635. 
[5] ElTanboly, A., Ismail, M., Shalaby, A., Switala, A., 
El‐Baz, A., Schaal, S., ... & El‐Azab, M. (2017). A 
computer‐aided diagnostic system for detecting 
diabetic retinopathy in optical coherence 
tomography images. Medical physics, 44(3), 914-
923. https://doi.org/10.1002/mp.12071. 
[6] Tariq, A., Akram, M. U., Shaukat, A., & Khan, S. A. 
(2013). Automated detection and grading of diabetic 
maculopathy in digital retinal images. Journal of 
digital imaging, 26(4), 803-812.  
https://doi.org/10.1007/s10278-012-9549-4. 
[7] Poddar, S., Jha, B. K., & Chakraborty, C. (2011, 
November). Quantitative clinical marker extraction 
from colour fundus images for non-proliferative 
diabetic retinopathy grading. In 2011 International 
Conference on Image Information Processing, pp. 1-
6. IEEE.  
https://doi.org/10.1109/ICIIP.2011.6108956. 
[8] Fleming, A. D., Goatman, K. A., Philip, S., 
Williams, G. J., Prescott, G. J., Scotland, G. S., ... & 
Olson, J. A. (2010). The role of haemorrhage and 
exudate detection in automated grading of diabetic 
retinopathy. British Journal of Ophthalmology, 
94(6), 706-711.  
https://doi.org/10.1136/bjo.2008.149807. 
[9] Li, X., Pang, T., Xiong, B., Liu, W., Liang, P., & 
Wang, T. (2017, October). Convolutional neural 
networks based transfer learning for diabetic 
retinopathy fundus image classification. In 2017 
10th international congress on image and signal 
processing, biomedical engineering and informatics 
(CISP-BMEI) (pp. 1-11). IEEE.  
https://doi.org/10.1109/CISP-BMEI.2017.8301998. 
[10] García-Zurdo, Rubén. "Creation of facial composites 
from user selections using image gradient." 
Informatica 44, no. 1 (2020).   
https://doi.org/10.31449/inf.v44i1.2340 
[11] Perdomo, O., Otalora, S., Rodríguez, F., Arevalo, J., 
& González, F. A. (2016). A novel machine learning 
model based on exudate localization to detect 
diabetic macular edema.  
https://doi.org/110.17077/omia.1057. 
[12] Johari, M. H., Hassan, H. A., Yassin, A. I. M., Tahir, 
N. M., Zabidi, A., Rizman, Z. I., ... & Wahab, N. A. 
(2018). Early detection of diabetic retinopathy by 
using deep learning neural network. International 
Journal of Engineering and Technology (UAE), 7(4), 
198-201. 
https://doi.org/10.14419/ijet.v7i4.11.20804. 
[13] Shijina, V., Adithya Unni, and Sunil Jacob John. 
"Similarity measure of multiple sets and its 
application to pattern recognition." Informatica 44, 
no. 3 (2020).  
DOI: https://doi.org/10.31449/inf.v44i3.2872 
[14] Chen, Y. W., Wu, T. Y., Wong, W. H., & Lee, C. Y. 
(2018, April). Diabetic retinopathy detection based 
on deep convolutional neural networks. In 2018 
IEEE international conference on acoustics, speech 
and signal processing (ICASSP), pp. 1030-1034. 
IEEE. 
https://doi.org/10.1109/ICASSP.2018.8461427. 
[15] Gonçalves, J., Conceiçao, T., & Soares, F. (2019). 
Inter-observer Reliability in Computer-aided 
Diagnosis of Diabetic Retinopathy. In HEALTHINF 
(pp. 481-491).  
https://doi.org/10.5220/0007580904810491. 
[16] Li, X., Hu, X., Yu, L., Zhu, L., Fu, C. W., & Heng, 
P. A. (2019). CANet: cross-disease attention 
network for joint diabetic retinopathy and diabetic 
macular edema grading. IEEE transactions on 
medical imaging, 39(5), 1483-1493.  
https://doi.org/10.1109/TMI.2019.2951844. 
[17] Bhardwaj, C., Jain, S., & Sood, M. (2021). 
Hierarchical severity grade classification of non-
proliferative diabetic retinopathy. Journal of 
Ambient Intelligence and Humanized Computing, 
12(2), 2649-2670.  
https://doi.org/10.1007/s12652-020-02426-9. 
[18] Saranya, P., & Prabakaran, S. (2020). Automatic 
detection of non-proliferative diabetic retinopathy in 
retinal fundus images using convolution neural 
network. Journal of Ambient Intelligence and 
Humanized Computing, 1-10.  
https://doi.org/10.1007/s12652-020-02518-6. 
[19] Dhiman, G., Oliva, D., Kaur, A., Singh, K. K., 
Vimal, S., Sharma, A., & Cengiz, K. (2021). BEPO: 
A novel binary emperor penguin optimizer for 
automatic feature selection. Knowledge-Based 
Systems, 211, 106560.  
https://doi.org/10.1016/j.knosys.2020.106560. 

An Optimized Deep Learning Based Technique for Grading and Extraction ... Informatica 45 (2021) 665–665 665 
 
[20] Dhiman, G., Singh, K. K., Soni, M., Nagar, A., 
Dehghani, M., Slowik, A., ... & Cengiz, K. (2021). 
MOSOA: a new multi-objective seagull optimization 
algorithm. Expert Systems with Applications, 167, 
114150. https://doi.org/10.1016/j.eswa.2020.114150. 
[21] Bhardwaj, C., Jain, S., & Sood, M. (2021). 
Hierarchical severity grade classification of non-
proliferative diabetic retinopathy. Journal of 
Ambient Intelligence and Humanized Computing, 
12(2), 2649-2670.  
https://doi.org/10.1007/s12652-020-02426-9. 
[22] Bhardwaj, C., Jain, S., & Sood, M. (2020). Diabetic 
retinopathy lesion discriminative diagnostic system 
for retinal fundus images. Advanced Biomedical 
Engineering, 9, 71-82.  
https://doi.org/10.14326/abe.9.71. 
[23] Rathee, G., Sharma, A., Kumar, R., Ahmad, F., & 
Iqbal, R. (2020). A trust management scheme to 
secure mobile information centric networks. 
Computer Communications, 151, 66-75. 
https://doi.org/10.1016/j.comcom.2019.12.024. 
[24] Yuvaraj, N., Srihari, K., Dhiman, G., 
Somasundaram, K., Sharma, A., Rajeskannan, S., ... 
& Masud, M. (2021). Nature-Inspired-Based 
Approach for Automated Cyberbullying 
Classification on Multimedia Social Networking. 
Mathematical Problems in Engineering, 2021. 
https://doi.org/10.1155/2021/6644652.  
[25] Sharma, A. and Kumar, N., 2021. Third Eye: An 
Intelligent and Secure Route Planning Scheme for 
Critical Services Provisions in Internet of Vehicles 
Environment. IEEE Systems Journal.  
DOI: https://doi.org/10.1109/JSYST.2021.3052072 
[26] Porwal, P., Pachade, S., Kamble, R., Kokare, M., 
Deshmukh, G., Sahasrabuddhe, V., & Meriaudeau, 
F. (2018). Indian diabetic retinopathy image dataset 
(IDRiD): a database for diabetic retinopathy 
screening research. Data, 3(3), 25.  
https://doi.org/10.3390/data3030025. 
[27] Mareli, M., & Twala, B. (2018). An adaptive 
Cuckoo search algorithm for optimisation. Applied 
computing and informatics, 14(2), 107-115. 
https://doi.org/ 10.1016/j.aci.2017.09.001. 
[28] Gao, Z., Li, J., Guo, J., Chen, Y., Yi, Z., & Zhong, J. 
(2018). Diagnosis of diabetic retinopathy using deep 
neural networks. IEEE Access, 7, 3360-3370. 
https://doi.org/10.1109/ACCESS.2018.2888639. 
[29] Li, G., Liu, F., Sharma, A., Khalaf, O.I., Alotaibi, 
Y., Alsufyani, A. and Alghamdi, S., 2021. Research 
on the Natural Language Recognition Method Based 
on Cluster Analysis Using Neural 
Network. Mathematical Problems in Engineering, 
2021.  
DOI: https://doi.org/10.1155/2021/9982305  
[30] Keshavarz-Ghorabaee, M., Amiri, M., Zavadskas, E. 
K., Turskis, Z., & Antucheviciene, J. (2018). 
Simultaneous evaluation of criteria and alternatives 
(SECA) for multi-criteria decision-making. 
Informatica, 29(2), 265-280,  
DOI: https://doi.org/10.15388/Informatica.2018.167 
[31] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep 
residual learning for image recognition. In 
Proceedings of the IEEE conference on computer 
vision and pattern recognition (pp. 770-778). 
https://doi.org/10.1109/CVPR.2016.90. 
[32] Sayago-Heredia, J., Pérez-Castillo, R., & Piattini, M. 
(2021). A Systematic Mapping Study on Analysis of 
Code Repositories. Informatica, 1-42,  
DOI: https://doi.org/10.15388/21-INFOR454 
[33] Liu, L., Wang, L., Xu, D., Zhang, H., Sharma, A., 
Tiwari, S., ... & Shah, M. A. (2021). CT Image 
Segmentation Method of Liver Tumor Based on 
Artificial Intelligence Enabled Medical Imaging. 
Mathematical Problems in Engineering, 2021. 
https://doi.org/10.1155/2021/9919507 
  

666 Informatica 45 (2021) 659–665 Q. Zhang et al.