Diabetic Retinopathy Classification Using Hybrid Deep Learning Approach

Abstract

During the recent years, diabetic retinopathy (DR) has been one of the most threatening complications of diabetes that leads to permanent blindness. Further, DR mutilates the retinal blood vessels of a patient having diabetes. Accordingly, various artificial intelligence techniques and deep learning have been proposed to automatically detect abnormalities in DR and its different stages from retina images. In this paper, we propose a hybrid deep learning approach using deep convolutional neural network (CNN) method and two VGG network models (VGG16 and VGG19) to diabetic retinopathy detection and classification according to the visual risk linked to the severity of retinal ischemia. Indeed, the classification of DR deals with understanding the images and their context with respect to the categories. The experimental results, performed on 5584 images, which are an ensemble of online datasets, yielded an accuracy of 90.60%, recall of 95% and F1 score of 94%. The main aim of this work is to develop a robust system for detecting and classifying DR automatically.

Full text

Vol.:(0123456789)
SN Computer Science (2022) 3:357
https://doi.org/10.1007/s42979-022-01240-8
SN Computer Science
ORIGINAL RESEARCH
Diabetic Retinopathy Classification Using Hybrid Deep Learning
Approach
BrahamiMenaouer1 · ZoulikhaDermane2· NourElHoudaKebir2· NadaMatta3
Received: 17 February 2022 / Accepted: 2 June 2022
© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd 2022
Abstract
During the recent years, diabetic retinopathy (DR) has been one of the most threatening complications of diabetes that leads
to permanent blindness. Further, DR mutilates the retinal blood vessels of a patient having diabetes. Accordingly, various
artificial intelligence techniques and deep learning have been proposed to automatically detect abnormalities in DR and its
different stages from retina images. In this paper, we propose a hybrid deep learning approach using deep convolutional
neural network (CNN) method and two VGG network models (VGG16 and VGG19) to diabetic retinopathy detection and
classification according to the visual risk linked to the severity of retinal ischemia. Indeed, the classification of DR deals
with understanding the images and their context with respect to the categories. The experimental results, performed on 5584
images, which are an ensemble of online datasets, yielded an accuracy of 90.60%, recall of 95% and F1 score of 94%. The
main aim of this work is to develop a robust system for detecting and classifying DR automatically.
Keywords Knowledge management· Deep learning· Convolutional neural networks (CNNs)· VGGNet· Diabetic
retinopathy· Image processing· Image classification· Healthcare decision support systems
Introduction
In most literature, diabetes is a chronic disease that occurs
either when the pancreas does not produce enough insu-
lin or when the body not producing enough insulin effec-
tively [5–8, 13, 37, 38]. According to Gurani etal. [22],
diabetes aimed at protracted time harms the blood vessels
of the retina, thereby affecting the vision of an individual in
addition to leading to diabetic retinopathy (DR). Accord-
ing to Gao etal. [21], the diagnosis of diabetic retinopathy
through eye fundus images traditionally performed by oph-
thalmologists for examining the presence and significance of
many subtle features is a cumbersome and time-consuming
process. Different complications arise due to diabetes, one
of which is diabetic retinopathy leading to blindness. Cur-
rently, diabetic retinopathy is an important disease leading
to blindness among elderly people and has become a global
medical problem over the last few decades. Likewise, dia-
betic retinopathy (DR), known as diabetic eye disease, is
a medical condition in which damage occurs to the retina
due to diabetes mellitus. According to Nair and Mishra [37,
38], diabetic retinopathy damages the blood vessels within
the retinal tissue causing them to leak fluid which distorts
vision. Besides, DR is a leading cause of vision loss, caused
by damage to the retina from complications of diabetes [18].
DR is a diabetes complication that affects eyes. According to
Wu etal. [57], DR is a progressive condition with microvas-
cular alterations that lead to retinal ischemia, retinal perme-
ability, retinal neovascularization and macular edema. Oth-
erwise, approximately one-third of 285 million people with
diabetes mellitus worldwide have signs of DR [31]. There
* Brahami Menaouer
mbrahami@gmail.com; brahami.menaouer@gmail.com;
menaouer.brahami@enp-oran.dz
Zoulikha Dermane
dermanekebir2020@gmail.com
Nour El Houda Kebir
houdakebir45@gmail.com
Nada Matta
nada.matta@utt.fr
1 Computer Science Department, LABAB Laboratory,
National Polytechnic School ofOran, BP: 1523 El M’naouer,
31000Oran, Algeria
2 National Polytechnic School ofOran, BP: 1523 El M’naouer,
31000Oran, Algeria
3 University ofTechnology ofTroyes, 12 Rue Marie Curie,
10300Troyes, France

SN Computer Science (2022) 3:357 357 Page 2 of 15
SN Computer Science
are several scientific and medical approaches to screen and
detect diabetic retinopathy in patients with medical imag-
ing. Comparing the results of several scientific and medical
approaches in prior studies, it does not only eliminate the
time and money costs, but also maintains a high accuracy
[52]. According to Benbassat and Polak and Pratt etal. [2,
43], the accuracy of care is of significant importance to both
the cost and effectiveness of treatment. If detected early
enough, effective treatment of DR is available.
In all the existing research works, innovations in deep
learning (DL) are tremendous and applications of DL tech-
niques are ever expanding and encompass a wide range
of services across many fields, namely feature extraction,
recognition, classification and prediction. According to Al
ayoubi etal. [1], DL has become one of the most common
techniques that has achieved better performance in many
areas, especially in medical image analysis and classifica-
tion. Recently, DL techniques have achieved tremendous
success in computer vision area. They can model high-level
abstractions in data relative to specific prediction task. This
very special potential of DL algorithms has made it a pre-
ferred tool for image analysis. Architectures of DL have
proven better at recognizing objects in pictures than human
detection and traditional image recognition. DL techniques
have a big advantage over machine learning techniques
because they learn high-level features from data in an incre-
mental manner, removing the need for domain expertise and
feature extraction. In the right context, convolutional neural
networks (CNNs) are one of the most widely adopted deep
neural network models for the current research literature. It
rapidly become a popular tool for medical image processing
and analysis. According to Suganthi etal. [51], CNNs are
a class of deep, feed-forward artificial neural networks that
has successfully been applied to analyzing visual imagery.
CNNs use relatively minimal pre-processing compared to
other image classification algorithms. During the recent
years, different methods of detecting abnormalities in DR
have been studied using deep learning, which also leads to
many solutions being provided. The classification of DR
involves the weighting of numerous features and the loca-
tion of such features [37, 38]. As per our objective, the new
techniques of deep learning are able to obtain much quicker
classifications to aid field experts (clinicians) in real-time
detection and classification. Therefore, this paper is focused
on applying end-to-end, accurate and computationally effi-
cient CNN and VGGNet models for automatic DR detec-
tion and classification. As well, the performance of the pro-
posed DR classification hybrid model is evaluated using the
APTOS 2019 dataset trained by different images of eyes that
have retinopathy and those which do not have retinopathy.
The rest of the paper is organized as follows: “Related
work” introduces the literature review on feature extrac-
tion, detection, and classification techniques proposed by
various researchers. “Proposed approach” presents the pro-
posed hybrid approach used for DR eye image classification,
and the experimental results and discussion are covered in
“Experimental and results”. Finally, “Conclusion” concludes
with the scope of the work combined with the challenges.
Related Work
A lot of researchers have conducted studies regarding
detection of diabetic retinopathy (DR) using various algo-
rithms, methodologies, techniques and procedures which
will be considered as part of the theoretical framework of
this research, enabling afterward the construction of the
conceptual framework of the study. For instance, Refs. [37,
38] have developed a network with CNN architecture and
data augmentation, which can identify the intricate fea-
tures involved in the classification task such as micro-
aneurysms and hemorrhages in the retina and consequently
provide a diagnosis automatically and without user input.
They achieved a sensitivity of 95% and an accuracy of 75%
on 5000 validation images. Gondal etal. [23] have pro-
posed a CNN model for referable diabetic retinopathy
(RDR) using two publicly available datasets. They per-
formed binary classification where normal and mild stages
are considered as non-referable DR and the rest of the
three stages are used as referable DR. The performance of
the CNN model is evaluated based on binary classification
resulting in sensitivity 93.6% and specificity 97.6% on
DiaretDB1. Wang etal. [54] have proposed a novel archi-
tecture that classifies the images as normal/abnormal, ref-
erable/non-referable DR. Their proposed method uses
three networks: main, attention and crop. The main net-
work uses the Inception model that is trained on ImageNet
where the attention network highlights different types of
lesions in the images and crop the network’s high attention
image. Seth and Agarwal [48] have proposed a hybrid deep
learning-based approach for detection of diabetic retinopa-
thy in fundus photographs. The authors used convolutional
neural network with linear support vector machine to train
the network on standard benchmark dataset EyePACS
dataset. Another research by Wan etal. [55] attempted
finding an automatic way to classify a given set of fundus
images. Coupled with transfer learning and hyperparam-
eter tuning, the authors adopt AlexNet, VggNet, Goog-
leNet, and ResNet, and analyze how well these models do
with the DR image classification. The best classification
accuracy is 95.68% and the results have demonstrated the
better accuracy of CNNs and transfer learning on DR
image classification. Wang and Yang [56] have proposed
a deep learning method for interpretable diabetic retinopa-
thy (DR) detection. The visual interpretable feature of the
proposed method is achieved by adding the regression

SN Computer Science (2022) 3:357 Page 3 of 15 357
SN Computer Science
activation map (RAM) after the global averaging pooling
layer of the convolutional networks (CNNs). The experi-
ments of this work were conducted on a large scale of
retina image dataset to achieve high performance on DR
detection compared with the state of the art, while achiev-
ing the merits of providing the RAM to highlight the sali-
ent regions of the input image. Besides, Pan etal. [41]
have proposed a novel and automatic diabetic retinopathy
(DR) detection method using deep convolutional neural
networks (DCNNs) to identify the region of interests
(ROIs). In this method, around 30,000 color retinal images
were used to train the proposed model and around 5000
images were collected to evaluate its classification perfor-
mance. Ni etal. [39] present a deep convolutional neural
network for DR stage classification, trained and evaluated
on a large dataset. The model uses high-resolution retinal
fundus images of both the left and right eyes as inputs to
take advantage of more detailed retinal lesion information
in images and strong correlation between both eyes.
Experiments show that the model proposed ouperforms
the fine-tuned Inception-v3 model by every measure,
achieving an accuracy of 87.2% and a Kappa score of
0.806 on the Kaggle dataset. Choudhury etal. [11] have
proposed a model for automatic detection of diabetic retin-
opathy using low complexity image processing technique
and modified convolutional neural network (CNN) with
better accuracy and precision to help the ophthalmologist
to detect changes in retina features. The model is used to
classify the fundus images into two categories, viz. healthy
and infected, and tested on Eye-PACS dataset, which
obtained a classification accuracy of 82% showing the
robustness of the system. Another study was conducted by
Jiang etal. [27] to propose an automatic retinal vessel
segmentation framework using deep fully convolutional
neural networks (FCN), which integrate novel methods of
data pre-processing, data augmentation, and full convolu-
tional neural networks. It is an end-to-end framework that
automatically and efficiently performs retinal vessel seg-
mentation. The framework was evaluated on three publicly
available standard datasets, achieving F1 score of 0.8321,
0.8531 and 0.8243, and an average accuracy of 0.9706,
0.9777, and 0.9773. Ni etal. [25] propose an alternative
hybrid solution method for diagnosing diabetic retinopathy
from retinal fundus images based on using both image
processing and deep learning for improved results. This
study validated 400 retinal fundus images within the
Messidor database and average values for different perfor-
mance evaluation parameters were obtained with accuracy
97%, sensitivity (recall) 94%, specificity 98%, precision
94%, F-score 94%, and GMean 95%. Moreover, various
computer vision-based techniques have been proposed by
Qummar etal. [44] to automatically detect DR and its dif-
ferent stages from retina images. In this work, the authors
used the publicly available Kaggle dataset of retina images
to train an ensemble of five deep convolution neural net-
work (CNN) models (Resnet50, Inceptionv3, Xception,
Dense121, Dense169) to encode the rich features and
improve the classification for different stages of DR.
Another research by Khalifa etal. [32] achieved automatic
diabetic retinopathy (DR) detection in retinal fundus pho-
tographs through the use of a deep transfer learning
approach using the Inception-v3 network. Zago etal. [58]
have designed a lesion localization model using a deep
network patch-based approach to reduce the complexity of
the model while improving its performance. For this, the
authors designed an efficient procedure (including two
convolutional neural network models) for selecting the
training patches, such that the challenging examples would
be given special attention during the training process.
Gadekallu etal. [19, 20] have developed an automatic DR
detection model with the aid of three main stages such as
(a) image pre-processing, (b) blood vessel segmentation
and (c) classification. This last part uses deep CNN, where
the improvement is exploited on the convolutional layer,
which is optimized by the same improved FP-CSO. A deep
neural network model was used in the study of Hemanth
etal. [24] in convergence with principal component analy-
sis (PCA) and firefly algorithm for the classification of the
diabetic retinopathy set. For this, the raw dataset is nor-
malized using the standard scalar technique and then prin-
cipal component analysis (PCA) is used to extract the most
significant features in the dataset. Another study by Castel-
lano etal. [9], has proposed a hybrid technique incorporat-
ing image processing and deep learning for detection and
classification of diabetic retinopathy. The model was vali-
dated using the retinal fundus dataset consisting of 400
images of the MESSIDOR database yielding good results.
A deep convolutional neural network [46] was used to train
a retinal image dataset consisting of 128,175 images. The
sensitivity and specificity scores in the study helped to
detect referable diabetic retinopathy (RDR) among dia-
betic patients using a deep neural network model. Bora
etal. [4] have created and validated two versions of a
deep-learning system to predict the development of dia-
betic retinopathy in patients with diabetes, who had teler-
etinal diabetic retinopathy screening in a primary care
setting. Supriya etal. [53] have proposed an approach for
the analysis of different DR stages using deep learning
technique. The approach trained a model called DenseNet
on an enormous dataset including around 3662 trained
images to automatically detect the DR stage and these are
classified into high-resolution fundus images. Another
study by Sampaul etal. [47], proposed a new methodology
based on convolutional neural networks (CNN) to diag-
nose and give a decision about the presence of retinopathy.
The CNN model is trained by different images of eyes that

SN Computer Science (2022) 3:357 357 Page 4 of 15
SN Computer Science
have retinopathy and those which do not have retinopathy.
The study presented by Lin etal. [33] combined the advan-
tages of graph convolution networks and self-supervised
learning for multilabel classification of fundus images. A
multiclass graph network contains seven layers that per-
form feature extraction and classification. The researchers
were able, not able to obtain significant results with this
methodology. Erciyas and Barışçı [17] have proposed a
deep learning-based technique for detecting diabetic retin-
opathy lesions automatically and independently of datasets
and then classifying the lesions found. A data pool is gen-
erated in the first stage of the proposed technique by gath-
ering diabetic retinopathy data from several datasets. Das
etal. [12] have described DR, its symptoms, features,
shape, size, and location of the features, and how DR
causes blindness. It also describes various ML and DL
techniques used for the detection of abnormal behavior of
RBVs and OD to identify DR lesions such as MAs, HEs,
EXs, CWS, FAZ, IRMA, and neovascularization in chron-
ological order. The purpose of this literature review in this
paper is to demonstrate and investigate the recent develop-
ment in automated scientific techniques (deep learning
methods) to detect DR based on computer-aided diagnosis
systems. Based on these related works, many researchers
have applied deep learning architectures for detecting DR.
Many of them consider deep convolutional neural
networks as an effective architecture as it provides feature
extraction without manual intervention.
Proposed Approach
As per our objective and motivations, this study is associated
with some background ideas and research efforts as shown
in Fig.1. Briefly, especially using deep learning for diag-
nosis of diabetic retinopathy and supporting it with image
processing have been remarkable ideas to follow. In gen-
eral, classification and diagnosis approach performed with
deep CNN method and two VGGNet models (VGG16 and
VGG19) have been widely followed for diabetic retinopathy
(see Fig.1).
In the context mentioned above, this study followed an
easy-to-design image pre-processing and hybrid deep learn-
ing approach for diagnosing diabetic retinopathy, by con-
sidering retinal fundus images as input data. In this respect,
Fig.2 represents the stages within the flow of the introduced
hybrid approach. For fundus photography enhancement
approach, a practical phase including OpenCV functions
was used accordingly. After the image pre-processing-
based enhancement, the classification was made by using
deep convolutional neural network (CNN) method and two
VGG NETWORK models (VGG16 and VGG19). In the next
Fig. 1 Ideas and research efforts in the background of this study

SN Computer Science (2022) 3:357 Page 5 of 15 357
SN Computer Science
stages, we evaluate our introduced hybrid approach by using
5584 retinal fundus images in the APTOS2019 dataset. The
image processing technique are important for a good image
enhancement, which will be effective for better detection
and classification at the end. The whole flow is a hybrid
approach applied to target image data, which is essential
for diagnosing from medical inputs in the form of visual
elements (see Fig.2).
Dataset (APTOS‑2019)
In our study, fundus images used in this study are publicly
available from Kaggle1 dataset. Images were provided by
the Asia Pacific Tele-Ophthalmology Society (APTOS)
as part of the 2019 Blindness Detection Competition [30,
40]. The Kaggle dataset is one of the widely used and well-
reported datasets for diabetic retinopathy. This dataset has
been used for analyzing the performance of algorithms used
for automated diagnosis of diabetic retinopathy. In addition,
the dataset “APTOS2019”, used for testing, contains almost
5584 high-resolution fundus images selected from the Kag-
gle dataset of 5597 images taken by different models. The
smallest native size among all of the datasets is 640 × 480.
The sample image from APTOS2019 is shown in Fig.3.
In this study, we rated each image for the severity of dia-
betic retinopathy on a scale from 0 to 4 taking into account
the clinician’s opinion where the numbers represent the
extent of the complication. The labels are provided by pro-
fessionals who rank the presence of DR in each image by a
scale of 0, 1, 2, 3, 4, which stand for “No DR”, “Mild DR”,
“Moderate DR”, “Severe DR”, “Proliferative DR” respec-
tively. For this, Table1 and Fig.4 show DR at different
stages.
As mentioned in the description of the dataset, the images
in the dataset come from different models and types of cam-
era, which can affect the visual appearance of left vs. right.
Fig. 2 Main steps of the proposed hybrid approach
Fig. 3 Sample of fundus image from the dataset
1 https:// www. kaggle. com/c/ aptos 2019- blind ness- detec tion.

SN Computer Science (2022) 3:357 357 Page 6 of 15
SN Computer Science
The data we use is usually split into training data, test data
and valid set. The training set was used to train a diagnostic
model of DR disease, and the test set (or subset) was used
to predict the accuracy of the disease diagnosis result. The
training set consists of 50% images and the rest is divided
over the test set (25% images) and valid set (25% images),
respectively (see Fig.5). During training, the validation set
is used to check and reduce overfitting.
Pre‑processing
In theory, image processing techniques are increasingly used
as a way of diagnosing diseases, including diseases of the
eye. According to Dutta etal. [14], image pre-processing
is a necessary step to remove the noise from images, to
enhance image features and to ensure the consistency of
Table 1 Class distribution in diabetic retinopathy dataset
Scale Severity Description of DR stages [10]
0 No DR The normal state (no anomalies)
1 Mild DR In this stage, microaneurysms occur; they are swellings which are ballon-like small areas in the retina’s tiny blood
vessels
2 Moderate DR In this stage, the blood vessels which provide nourishment to the retina are blocked
3 Severe DR In this stage, many blood vesselsare blocked, depriving blood supply to several areas of the retina. These areas of
retina send signals to the body for the growth of new blood vessels for nourishment
4 Proliferative DR This is the final stage of diabetic retinopathy, the advanced stage, where the signals which are sent by the retina for
nourishment trigger new vessels that are fragile and abnormal and grow with the retina and on the surface of the
clear vitreous gel which gets filled inside the eye
Fig. 4 Sample fundus images from APTOS dataset
Fig. 5 Splitting data folders into
training, validation, and testing
folders
Testing
Set
Dataset
Hybrid Model
Final Performance Evaluation
Evaluate the model based on various metrics
Validation
Set
Validate Models
Tune Hyper Parameters
Training
Set
Train
DL Algorithms
CNN VGGNet

SN Computer Science (2022) 3:357 Page 7 of 15 357
SN Computer Science
images. Moreover, pre-processing methods are applied to the
images before actual processing to enhance the features of
the image. According to Boda-pati etal. [3], due to the way
APTOS2019 was collected, there are spurious correlations
between the disease stage and several image meta-features,
e.g., resolution, crop type, zoom level, or overall brightness.
All the images in the dataset are taken of different people,
using different clinical settings, and are of different sizes.
For this paper, we propose introducing some of the com-
monly used image processing techniques leveraging a very
popular computer vision library (OpenCV library in python
(cv2)) to adjust for the images and make more clearer images
so as to enable the model to learn features more effectively.
In short, we will read the images using OpenCV transfor-
mation functions and resize them to get the same height
and width (128 × 128 blocks of pixels). Thus, the image gets
divided into 8 rows and 8 columns, thus in total 64 blocks of
128 × 128 pixels are generated. The pixels are normalized to
improve the performance in the training of the CNNs mod-
els. After analyzing the data, we noticed that the data was
highly unbalanced among the diabetic retinopathy severity
image classes, which gave rise to the propensity of data aug-
mentation. For this, we applied augmentation on images in
real time to reduce overfitting. During each epoch, a random
augmentation of images that preserves the collinearity and
distance ratios was performed, to balance the data among the
diabetic retinopathy severity classes. Indeed, 5000 images
were obtained in each class after augmentation. The result-
ing images were presented as the input of our hybrid deep
learning techniques.
Convolutional Neural Networks (CNNs)
Deep Learning (DL) is a part of an artificial neural net-
work technique and a subclass of machine learning. More-
over, DL is part of a broader family of machine learning
methods based on learning data representations. In DL,
multiple layers are used for a higher level of feature from
the input dataset. DL technologies have rapidly improved
over the years, especially in the fields of engineering and
medical sciences [6–8]. In the fields of medical imaging
for the diagnosis of disease, DL techniques are very help-
ful for diabetic retinopathy detection due to their reliability
and accuracy. In this field, convolutional neural networks
(CNNs or ConvNets), a branch of deep learning, have an
impressive record for applications in image analysis and
interpretation, including medical imaging [6–8, 43]. Fur-
thermore, CNN is a class of deep, feed-forward artificial
neural networks that has successfully been applied to ana-
lyzing visual imagery. CNNs use relatively minimal pre-
processing compared to other image classification algo-
rithms [51]. According to earlier works, CNN is a type of
deep neural networks that learns features from the input
data and uses two-dimensional convolutional layers for the
processing of two-dimensional image data [28]. Besides,
CNN architecture is classified into many layers, such as
convolutional and pooling layers that are gathered into
modules. Moreover, one or more fully connected layers
follow these modules, same as in benchmarked feed-for-
ward neural network. Often, the modules are loaded on top
of each other to design a deep model [19, 20]. According
to Maeda-Gutiérrez etal. [35], the CNN settings usually
consist of a series of specific elements, which are the ones
that present the variations in the different architectures.
For Patel [42], CNN image classification takes an image
as input, processes it using hidden layers and classifies
it as an output. CNN uses convolution layers that extract
features from an image automatically. Deep convolutional
neural networks have been used to detect and distinguish
features related to diabetic retinopathy and macular edema
in colored images of the patient’s fund [19, 20]. Most of
the layers in CNN convert an input image to features, and
only the last few layers are used for classification. Finally,
Fig. 6 General architecture of a CNN

SN Computer Science (2022) 3:357 357 Page 8 of 15
SN Computer Science
Fig.6 graphically presents the general architecture of a
CNN, with its main elements.
Given the fact and conditions of CNN in the analysis of
medical image, set of m kernels are taken at each layer. Input
in the form of image is convoluted with the kernels. Let
us assume w as the kernel set, i.e,.
w={w1,w2, ..., wn}
. An
added bias set
B={b1,b2, ..., bm}
is taken for creating a new
feature map
Xm
. A nonlinear transformation is applied to
these newly generated features element-wise and it is iterated
for every convolutional layer l:
In this light, these advances are based on the CNN net-
work capability for extracting the features from input data
sources. In this study, the network's main focus is to detect
different stages of diabetic retinopathy.
VGG Network Architecture (VGG16
andVGG19).
The visual geometry group network (VGGNet) is a deep
neural network with a multilayered operation. The VGGNet
is based on the CNN model [36]. This deep learning method
is one of the first attempts at adding depth to improve clas-
sification accuracy. The major characteristic of this architec-
ture is instead of having a large number of hyperparameters,
they concentrated on simple 3 × 3 size kernels in convolu-
tional layers and 2 × 2 size in max pooling layers [15, 59].
(1)
Xl
m
=𝜎
(
W
l−1
m
×X
l−1
+b
l−1
k).
During testing, in VGGNet, the test image directly goes
through the VGGNet and obtains a class score map. This
class score map is spatially averaged to be a fixed-size vec-
tor [48]. For Setiawan and Damayanti [49], VGGNet created
the VGG16 [50] network architecture with 16 layers and
VGG19 [50] with 19 layers. According to Hieu and Hien
[26], VGG16 is a CNN architecture that was used to win
the ImageNet ILSVR competition 2014. It is as yet consid-
ered as one of the outstanding vision model architectures.
Moreover, VGG-19 is useful due to its simplicity, as 3 × 3
convolutional layers are mounted on the top to increase with
depth level [36, 46]. In Zhang etal. [60], VGG-19 model has
roughly 143 million parameters, where the parameters are
learned from the ImageNet dataset containing 1.2 million
general object images of 1000 different object categories
for training. As in Fig.7a and b, respectively, VGG16 and
Fig. 7 General architecture of VGG16 (a) and VGG19 (b)
Table 2 Comparison of VGG16 and VGG19 layers
Layer VGG16 VGG19
Size of layer 41 47
Image input size 244 × 244 pixel 224 × 244 pixel
Convolutional layer 13 16
Filter size 64 and 128 64,128,256, and 512
ReLU 5 18
Max pooling 5 5
FCL 3 3
Dropout 0.5 0.5
Softmax 1 1

SN Computer Science (2022) 3:357 Page 9 of 15 357
SN Computer Science
VGG19 each consists of five convolutional blocks (CBs),
followed by three dense layers.
The use of uniform and smaller filter sizes of VGG
can produce more complex features and lower computing
when compared to AlexNet. In summary, we present across
Table2 the difference between VGG16 and VGG19.
Experimental andResults
In the literature, most of the research works that apply deep
learning for DR detection and classification use hundreds
of thousands of images to train the model, meaning a huge
burden for the experts to label the images accordingly. In
this section, we present the performance results on the
APTOS2019 dataset and other datasets, namely Messidor-2
and Local public DR.
System Requirements
The experimental environment of this paper is Windows 10
system, Python 3.6.2, Tensorflow 1.11.0, and Jupyter. In the
hardware device section, the CPU is Intel Core i5 4300U
@ 1.90GHz 2.50GHz specifications, GPU 1060 6Gb D5
amp, Solid State Drive, Double Data Rate4 16Gb, and MSI
Z270 GAMING PRO CARBON Motherboard. The primary
software configuration included Python compiler, Spyder
4.0.1 editor, deep learning framework PyTorch, and uses the
neural network library Keras 2.2.4, Numpy 1.22.3, Pandas
1.4.2, SciPy 1.8.0, Scikit learn 1.0.2, tqdm, OpenCV 4.2.0,
and Matplotlib 3.1.3.
Evaluation Criteria
As per mostly authors in the literature, after extracting the
appropriate feature, the last step is to classify the attained
data and assign it to a specific class. The different classifica-
tion performance properties of the proposed hybrid approach
is evaluated based on the essential measures of true positive
(TP), true negative (TN), false positive (FP), and false nega-
tive (FN). With the help of these parameters, other essential
values, such as accuracy, precision, sensitivity, specificity
and F1 score are also computed. These popular parameters
are defined as follows:
The recall metric tells us how well a model finds all of the
true positives and is a ratio of true positives over all entities
in the testing set.
(2)
Sensitivity (recall)=TP∕TP +FN.
In general, sensitivity and specificity evaluate the effec-
tiveness of the algorithm on a single class, positive and
negative, respectively.
The precision metric will show the ratio of true posi-
tives over the total number of detected entities. In other
words, this metric will help us understand how well a
model returns only the true positives and not unrelated
entities.
Commonly, accuracy is the most used metric to evalu-
ate the classification performance. This metric calculates
the percentage of samples that are correctly classified. As
well, precision is how “precise” the model is out of those
predicted positive and how many of them are actually posi-
tive. A high value of the metric (F1 score) indicates that
the model performs better on the positive class. Thus, F1
score (also known as F-measure) might be a better meas-
ure when a balance between precision and recall is needed
with an uneven class distribution (large number of actual
negatives). This metric can be used to show the overall
performance of a tool.
Likewise, a confusion matrix is commonly used to visu-
alize the performance of a classification algorithm. Meas-
urement of TP, FP, TN and FN uses a confusion matrix of
a classification with n classes. When considering the class
k (
0≤k≤n
× n) as shown in Fig.8. Similarly, observa-
tions on correct and incorrect classifications are collected
into the confusion matrix C = cij, where cij represents the
frequency of class i being identified as class j.
In this study, an assessment of state-of-the-art pre-
trained models for the task of classification of DR disease
using images was done. The objective of this research was
(3)
Specificity =TP∕TP +FP.
(4)
Accuracy =TP +TN∕TP +TN +FN +FP,
(5)
Precision =TP∕TP +FP.
(6)
F
− score = 2×
precision × r ecall
precision + r ecall
.
Fig. 8 Confusion matrix for multiclass classification

SN Computer Science (2022) 3:357 357 Page 10 of 15
SN Computer Science
to compare all the models evaluating the accuracy, preci-
sion, sensitivity, specificity and F score.
Results andDiscussions
Diabetic retinopathy (DR) is one of the most severe dia-
betes complications, causing non-reversible damage to
retina blood vessels. In this work, the most widely used
APTOS2019 dataset has been chosen to verify the pro-
posed hybrid model using Python programming language
with Tensorflow framework. For this, we investigated the
classification of the cases of DR disease, using deeper and
dense networks. This method can perform diagnosis based
on the various status of the DR images (see Sect.1.2).
Multiple layered model has been designed for performing
convolution and feature extraction. Rectified linear unit
(ReLU) activation function is used to define the output of
internal layers. The graphical representation of training
loss vs validation loss and training accuracy vs validation
accuracy of the approach model is displayed in Fig.9. In
theory, losses are the errors that occurred in the process of
prediction while training of the model. The optimum train-
ing process always reduces the errors and increases the
accuracy. When consistent accuracy and loss are obtained,
training could be stopped. As per our model, lower the
loss better is the model and higher is the accuracy and
more satisfactory is the classification results. During the
training process, the hybrid model determines the graphs
for the model’s loss and accuracy for batch size, number
of epochs, verbose, learning rate to 32, 70, 1, and 0.0001,
respectively. Adam with β1 = 0.9 and β2 = 0.999 is used
for optimization, though there are many benefits to using
the Adam optimizer in terms of speed of training. As
well, Adam optimizer allows adjustment in the middle of
epochs, leading to great flexibility to improve the model
performance while training. Moreover, the extraction of
weights is done with the API provided by the Scikit Learn
library.
Upholding the stated objective of easy implementation,
good performance, and minimum cost resources, the fol-
lowing hyperparameter configurations are used:
1. Error metric: categorical cross-entropy, due to the type
of single-labeled multiple class problem, this metric is
the ideal metric for this project.
2. Performance metric: categoricalaccuracy, as it allows
finding the average number of hits regardless of the
class.
Fig. 9 Result of our hybrid model (represents the accuracy and loss
model)
Fig. 10 Results of the confusion matrices for APTOS2019 datasets,
using a trained deep and densely connected model. Actual and pre-
dicted labels are displayed on the y-axis and x-axis, respectively

SN Computer Science (2022) 3:357 Page 11 of 15 357
SN Computer Science
3. Number of epochs: 70, as it is the average where no
model exceeds the execution time allowed in Kaggle
dataset.
Besides, we employed weight decay to reduce the overfit-
ting of the models. A fully connected layer was trained with
two activation functions (ReLU and sigmoid).
The model evaluates the outcome on APTOS2019 dataset
images in the form of a confusion matrix. Furthermore, it
evaluates the performance table including precision, sensi-
tivity, specificity, and F1 score. This test is performed using
25% images base for performance evaluation. Figure10 rep-
resents the confusion matrices for APTOS2019 datasets. The
confusion matrices in Fig.10 show that for APTOS2019,
the model yields adequate results as it is a tiny and clean
dataset. We find that the best results are obtained in the two
classes ‘No RD’ and 'Moderate’ with an F1 score 0.94 and
0.67, respectively.
Moreover, the proposed approach can generate a report
table (see Table3) that delivers important information
regarding each class in the retinal datasets to support com-
prehensive analysis with an accuracy 90.60%. Precision, sen-
sitivity, and specificity are the key metrics for checking the
accuracy of a model. For this, it uses various mathematical
equations to evaluate the report on the input data (see “Sys-
tem requirements”). For our hybrid approach, we evaluate
the F1 score, which checks the accuracy of the test data
in the form of harmonic average specifically for imbalance
datasets.
Furthermore, based on the graphs, it can be seen that
after 70 epochs, the model loss and accuracy remain con-
stant, which can result in over fitting. Therefore, our hybrid
approach stores various weights at distinct learning rates
with a loss of 17.3% and 26.19% for the training and validat-
ing data, respectively. Conversely, the accuracy of the model
on APTOS2019, Messidor-2, and Local public dataset is
90.60%.
As per our objective, the proposed hybrid model for DR
detection offered consistency of interpretation on a specific
image. In summary, our hybrid model shows that:
1. The performance of the proposed hybrid model, trained
as a regressor for diabetic retinopathy detection and clas-
sification, allows it to improve multilevel classification
results when compared with the single deep learning
approach.
2. The performance of the proposed hybrid model was
yielded directly by the results of the training data, with
a human expert grading decisions, without the need
to focus on the underlying process of DR. In addition,
when performing a large-scale screening for DR, it was
critical to improving sensitivity and specificity for mini-
mizing misdiagnosed cases.
3. The most significant merit of our hybrid model was pos-
sibly the endeavor to simultaneously predict five levels
of DR with improved performance which was suitable
for more timely and reliable detection of DR.
4. The hybrid system developed in this study did not
require any specialized or advanced computer equip-
ment to classify fundus photographs, and it could be
deployed on standard low-cost computing equipment to
offer reproducible evaluation of DR images in patients
with suspected DR diseases.
5. The hybrid model produces comparable results to most
of the previous works without any feature-specific detec-
tion and using a much more general dataset.
Comparison Against State‑of‑the‑Art Methods
In the context mentioned above, we conducted a comparative
study of our hybrid model proposed with other existing diag-
nostic, detection, and classification models on grounds of
the approach used, a number of DR datasets (APTOS 2019
and other DDR) used in experimentation, methodology, fea-
tures used for extraction, and classification, and percentage
accuracy achieved. The results in the table further verify that
our proposed hybrid model achieves the best performance
among all the methods (see Table. 4).
In general, the results of the present study were evalu-
ated comparatively according to similar parameters, and we
found that the proposed hybrid model produced successful
results with respect to other studies. However, much better
results were produced in earlier studies as aforementioned,
and better results were obtained with slight differences with
the studies using similar formalization. Very good perfor-
mance metrics results were obtained according to studies
with similar DR dataset sizes. This shows that existing
detection and classification performance metrics values were
taken one step further with the proposed hybrid model in the
Table 3 A report on APTOS2019 with precision, sensitivity, specific-
ity and F1 score. The last column shows the images available for the
individual classes in the dataset
Class Precision Recall
(sensitiv-
ity)
F1 score Specificity Images
per
class
No DR 0.91 0.95 0.94 0.98 332
Mild DR 0.52 0.42 0.47 0.43 76
Moderate DR 0.58 0.79 0.67 0.80 187
Severe DR 0.36 0.14 0.20 0.15 37
Proliferative
DR
0.22 0.20 0.23 0.20 55

SN Computer Science (2022) 3:357 357 Page 12 of 15
SN Computer Science
present study. It suggests that the proposed hybrid model
has improved specificity compared with the previous works.
In summary, classifying the diabetic retinopathy remains
a major challenge for researchers and domain experts who
need more research to clarify the problem. The current work
opens the way to building a complete automated monitoring
system for diabetic retinopathy (DR) which is a long-term
underlying disease.
The monitoring of disease will prevent blindness in
patients and limit vision impairment. In our future works,
YOLOv5 may be used to detect all diabetic retinopathy (DR)
to obtain their benefits, such as accuracy and speed.
Conclusion
In the world that we are living today, there is a great demand
for automated diagnostic system for DR. It is always accept-
able to have devices that directly diagnose the disease from
the fundus image without much clinical intervention. Today,
computer-assisted detection of medical images is a recently
emerging application of artificial intelligence, machine and
deep learning that can save time and manpower. In medi-
cal image processing, the image processing techniques
are important for good image enhancement, which will be
effective for better diagnosis and classification at the end.
Recently, deep learning (DL) techniques have achieved
superior performance in classification and segmentation.
Currently, DL techniques are applied to handle compli-
cated anomalies to improve the accuracy of DR due to its
efficiency in feature learning. The present paper is devoted
to the early detection and classification of DR in retinal
images using a hybrid deep learning approach based on
fine-tuned versions of (CNN, VGG16, and VGG19). Our
hybrid approach is fully automated with an end-to-end struc-
ture without the need for manual feature extraction. The
numerical experiments were conducted on the Asia Pacific
Tele-Ophthalmology Society (APTOS) 2019 dataset. Our
developed hybrid approach is able to perform multiclass
tasks with an accuracy of 90.60%. The performance of the
developed hybrid approach is assessed by expert DR clini-
cians and is ready to be tested with a larger database.
To conclude, the potential benefit of using our trained
hybrid approach (CNN, VGG16, and VGG19) is that it can
classify thousands of images every minute allowing it to
be used in real time whenever a new image is acquired.
The experiment also shows that our classification hybrid
approach can assist the oculist in diagnosing DR accurately
with more speed and could potentially boost DR patients’
screening rate. Overall, the networks have the potential to
be incredibly useful to DR specialists (clinicians) in the
future, as the networks and the datasets continue improving
and they will offer real-time classifications. Otherwise, we
demonstrate that the integration of DL techniques is highly
Table 4 Comparison between the hybrid model and the state-of-the-art models on the DDR dataset
Authors Number of
classes
Images/dataset Methodology Performance metrics and results
[57] 5 Kaggle Dataset CNN models (AlexNet,
VGG16, and Inception-
Net V3)
63.23% accuracy
[29] 5 APTOS2019 Modified xception 83.09% accuracy
88.24% sensitivity
87.00% specificity
[34] 5 EyePACS, APTOS2019, and DeepDR CNNs models 85.44% accuracy
98.48% sensitivity
71.82% specificity
90.27% precision
93.62% F1 score
[16] 5 EyePACS, APTOS2019, and DeepDR Transfer learning VGG16 73.7% accuracy
67.82% recall
66.85% Precision
64.28% F1 score
[45] 5 EyePACS Set Of CNN architectures 75% accuracy
0.7588 kappa coefficient
Alyoubi etal. (2021) 5 DDR and APTOS2019 CNN512 and YOLOv 89% accuracy
89% sensitivity
97.3 specificity
Proposed model 5 APTOS2019
Messidor-2
Local public DR
Hybrid models (CNN,
VGG16 and VGG19)
90.60% accuracy
95.00% recall
94.66% precision
94.00% F1 score

SN Computer Science (2022) 3:357 Page 13 of 15 357
SN Computer Science
feasible in applications with small datasets, taking advantage
of the theoretical robustness and the representational power
of DL methods.
A limitation of the study is the use of a limited num-
ber of our current datasets images. For this, the quality
and balance of the datasets used to build a DR screening
system are very critical. In the future, we aim to combine
several similar datasets to achieve the balance of the data-
set and to increase the number of images in each class.
Currently, we intend to make our model more robust and
accurate by using more such images from our local hos-
pitals. The results were generated from the experiments
in several numbers of epochs. It is concluded that more
number of epochs are needed to improve the accuracy
level. Therefore, one of our future works is to develop
deeper collaborative relations with hospitals and clin-
ics to acquire more data. With more data, we believe the
classification accuracy will be further increased. Another
limitation is the requirement of large processing power. As
the DR database is complex, training a DL model requires
high computational resources such as high RAM and core
processors. Quantum computation appears very conveni-
ent for decreasing processing time and intricacy required
by DL.
As part of the future study, we have plans to collect
a much cleaner dataset from real Algeria screening set-
tings. The performance of the proposed hybrid approach
therefore motivates to conduct similar studies in various
other domains having high-dimensional and heterogene-
ous data. At the same time, the ongoing developments in
CNNs allow much deeper networks, namely fuzzy CNN
and YOLOv5, which could learn better the intricate fea-
tures that this network struggled to learn. In contrast, we
assumes the augmentation of the deep techniques with
pre-processing to reveal clini–-pathological features and
performance upgrades. As clinical challenges, the medi-
cal validation and real-time implementation of DL meth-
ods in clinical practice remain the important challenge,
as these depend on the understanding of the patients to
entrust medical concerns to machines. Moreover, research
may emphasize advancing innovative schemes toward
conquering the shortcomings of current state-of-the-art
technology.
Acknowledgements This project is registered in the context of
National Program of Research, launched in collaboration between the
healthcare sector, our research team of the System Engineering Depart-
ment, and the laboratory Tech-CICO in University of Technology of
Troyes (UTT). The authors also acknowledge the service team health-
care for her assistance ining this project. Also, the dedication of our
team members and enthusiasm helped us to move forward.
Declarations
Conflict of interest The authors declare that they have no conflict of
interestrelated to the content of this manuscript.
Ethical approval This study does not contain any studies with human
participants or animals performed by any of the authors.
References
1. Al Ayoubi W, Shalash WM, Abulkhair MF. Diabetic retinopathy
detection through deep learning techniques: a review. Inf Med
Unlocked. 2020;20:1–11.
2. Benbassat J, Polak BC. Reliability of screening methods for
diabetic retinopathy. Diabet Med. 2009;26(8):783–90.
3. Bodapati JD, Naralasetti V, Shareef SN, Hakak S, Bilal M,
Maddikunta PKR, Jo O. Blended multi-modal deep ConvNet
features for diabetic retinopathy severity prediction. Electronics.
2020;9(6):1–16. https:// doi. org/ 10. 3390/ elect ronic s9060 914.
4. Bora A, Balasubramanian S, Babenko B, Virmani S, Venugo-
palan S, Mitani A, Bavishi P. Predicting the risk of developing
diabetic retinopathy using deep learning. Lancet Digit Health.
2020;3(1):10–9. https:// doi. org/ 10. 1016/ s2589- 7500(20)
30250-8.
5. Brahami M, Sabri M, Matta N. Towards a model to improve
Boolean knowledge mapping by using text mining and its
applications: case study in healthcare. Int J Inf Retr Res.
2020;10(3):40–58.
6. Brahami M, Dermane Z, Kebir N-H, Sabri M, Matta N. Coronavi-
rus pneumonia classification using X-ray and CT scan images with
deep convolutional neural networks models. J Inf Technol Res
(JITR). 2022;15(3):1–23. https:// doi. org/ 10. 4018/ JITR. 299391.
7. Brahami M, Kebir N-H, Dermane Z, Sabri M, Matta N. Detec-
tion and classification of brain tumors from MRI images using a
deep convolutional neural network approach. Int J Softw Innov
(IJSI). 2022;10(1):1–25. https:// doi. org/ 10. 4018/ IJSI. 293269.
8. Brahami M, Abdeldjouad FZ, Sabri M. Multi-class sentiment
classification for healthcare tweets using supervised learning
techniques. Int J Serv Sci Manag Eng Technol. 2022;13(1):1–
23. https:// doi. org/ 10. 4018/ IJSSM ET. 298669.
9. Castellano G, Castiello C, Mencar C, Vessio G (2020) Crowd
detection for drone safe landing through fully-convolutional
neural networks. In: Proceedings of the international confer-
ence on current trends in theory and practice of informatics,
Dortmund, Germany, 17–21 February, p. 301–312
10. Chandore V, Asati S. Automatic detection of diabetic retin-
opathy using deep convolutional neural network. Int J Adv Res
Ideas Innov Technol. 2017;3(4):633–41.
11. Choudhury AR, Bhattacharya D, Debnath A, Biswas A (2019)
An integrated image processing and deep learning approach
for diabetic retinopathy classification. In: Saha A, Kar N, Deb
S, editors. The proceeding of second international conference
ICCISIoT 2019, Agartala, India, December 13–14, Communica-
tions in Computer and Information Science, Springer, p. 3–15
12. Das D, Biswas SK, Bandyopadhyay S. A critical review on diag-
nosis of diabetic retinopathy using machine learning and deep
learning. Multimedia Tools Appl. 2022. https:// doi. org/ 10. 1007/
s11042- 022- 12642-4.
13. Doshi D, Shenoy A, Sidhpura D, Gharpure P. Diabetic retin-
opathy detection using deep convolutional neural networks. In:
Proceeding of the international conference on computing, ana-
lytics and security trends, Qingdao, China, Published in IEEE;
2016. pp. 261–266.

SN Computer Science (2022) 3:357 357 Page 14 of 15
SN Computer Science
14. Dutta S, Manideep BC, Basha SM, Caytiles RD, Iyengar NCSN.
Classification of diabetic retinopathy images by using deep
learning models. Int J Grid Distrib Comput. 2018;11(1):89–106.
https:// doi. org/ 10. 14257/ ijgdc. 2018. 11.1. 09.
15. El Asnaoui K, Chawki Y. Using X-ray Images and Deep Learn-
ing for Automated Detection of Coronavirus Disease. J Biomol
Struct Dyn. 2020;38:1–22.
16. El Houby MF. Using transfer learning for diabetic retinopathy
stage classification. Appl Comput Inf. 2021;17(1):1–11. https://
doi. org/ 10. 1108/ ACI- 07- 2021- 0191.
17. Erciyas A, Barışçı N. An effective method for detecting and
classifying diabetic retinopathy lesions based on deep learning.
Comput Math Methods Med. 2021;34(2021):1–13. https:// doi.
org/ 10. 1155/ 2021/ 99288 99.
18. Fleming AD, Goatman KA, Philip S, Prescott GJ, Sharp PF.
Automated grading for diabetic retinopathy: a large-scale
audit using arbitration by clinical experts. Br J Ophthalmol.
2010;94(12):1606–10.
19. Gadekallu TR, Khare N, Bhattacharya S, Singh S, Maddikunta
PKR, Srivastava G. Deep neural networks to predict diabetic
retinopathy. J Ambient Intell Human Comput. 2020;11(3):1–14.
20. Gadekallu TR, Khare N, Bhattacharya S, Singh S, Reddy Mad-
dikunta PK, Ra IH, Alazab M. Early detection of diabetic retin-
opathy using PCA-firefly based deep learning model. Electron-
ics. 2020;9(2):3–16. https:// doi. org/ 10. 3390/ elect ronic s9020
274.
21. Gao J, Leung C, Miao C. Diabetic retinopathy classification using
an efficient convolutional neural network. In: Proceeding of the
IEEE international conference on agents (ICA), October 18–21,
Jinan, China, 2019. p. 80–85
22. Gurani VK, Ranjan A, Chowdhary CL. Diabetic retinopathy
detection using neural network. Int J Innov Technol Explor Eng.
2019;8(10):2936–40.
23. Gondal WM, Köhler JM, Grzeszick R, Fink GA, Hirsch M.
Weakly-supervised localization of diabetic retinopathy lesions in
retinal fundus images. In: Proceeding of the IEEE international
conference image process (ICIP'2017), September 17–20, 2017.
p. 2069–2073.
24. Hemanth DJ, Deperlioglu O, Kose U. An enhanced diabetic retin-
opathy detection and classification approach using deep convo-
lutional neural network. Neural Comput Appl. 2020;32:707–21.
https:// doi. org/ 10. 1007/ s00521- 018- 03974-0.
25. Hemanth DJ, Deperlioglu O, Kose U. An enhanced diabetic retin-
opathy detection and classification approach using deep convolu-
tional neural network. Neural Comput Appl. 2019;31:1–15.
26. Hieu NV, Hien NLH. Recognition of plant species using
deep convolutional feature extraction. Int J Emerg Technol.
2020;11(3):904–10.
27. Jiang Y, Zhang H, Tan N, Chen L. Automatic retinal blood vessel
segmentation based on fully convolutional neural networks. Sym-
metry. 2019;11(1112):1–22.
28. Jyotiyana M, Kesswani N. Deep learning and the future of bio-
medical image analysis. In: Dash S etal, editors. Deep learning
techniques for biomedical and health informatics, Studies in big
data. vol 68. Springer; 2020. p. 329–345
29. Kassani SH, Kassani PH, Khazaeinezhad R, Wesolowski MJ, Sch-
neider KA, Deters R. Diabetic retinopathy classification using a
modified xception architecture. In: Proceedings of the 2019 IEEE
international symposium on signal processing and information
technology (ISSPIT'19), Ajman, United Arab Emirates, 10–12
December, 2019. p 1–6
30. Khalifa NEM, Loey M, Taha MHN, Mohamed HNET. Deep trans-
fer learning models for medical diabetic retinopathy detection.
Acta Inf Med. 2019;27(5):327–32.
31. Kumar SPN, Deepak RU, Satharb A, Sahasranamam V, Kumar
RR. Automated detection system for diabetic retinopathy using
two field fundus photography. In: Proceeding of the 6th interna-
tional conference on advances in computing & communications
(ICACC'16), vol. 93. 6–8 September, Cochin, India, published in
Procedia Computer Science; 2016. p. 486–494
32. Li F, Liu Z, Chen H, Jiang M, Zhang X, Wu Z. Automatic detec-
tion of diabetic retinopathy in retinal fundus photographs based on
deep learning algorithm. Transl Vis Sci Technol. 2019;8(6):1–13.
https:// doi. org/ 10. 1167/ tvst.8. 6.4.
33. Lin J, Cai Q, Lin M. Multi-label classification of fundus images
with graph convolutional network and self-supervised learning.
IEEE Signal Process Lett. 2021;28:454–8. https:// doi. org/ 10.
1109/ LSP. 2021. 30575 48.
34. Liu H, Yue K, Cheng S, Pan C, Sun J, Li W. Hybrid model
structure for diabetic retinopathy classification. J Healthc Eng.
2020;2020:1–9. https:// doi. org/ 10. 1155/ 2020/ 88401 74.
35. Maeda-Gutiérrez V, Galván-Tejada CE, Zanella-Calzada LA,
Celaya-Padilla JM, Galván-Tejada JI, Gamboa-Rosales H, Luna-
García H, Magallanes-Quintanar R, Guerrero Méndez CA,
Olvera-Olvera CA. Comparison of convolutional neural network
architectures for classification of tomato plant diseases. Appl Sci.
2020;10(1245):1–15.
36. Mateen M, Wen J, Nasrullah SS, Huang Z. Fundus image classifi-
cation using VGG-19 architecture with PCA and SVD. Symmetry.
2019;11(1):1–12.
37. Nair M, Mishra DS. Categorization of diabetic retinopathy sever-
ity levels of transformed images using clustering approach. Int J
Comput Sci Eng. 2019;7(1):642–8.
38. Nair M, Mishra D. Classification of diabetic retinopathy sever-
ity levels of transformed images using K-means and thresholding
method. Int J Eng Adv Technol. 2019;8(4):51–9.
39. Ni J, Chen Q, Liu C, Wang H, Cao Y, Liu B. An effective CNN
approach for diabetic retinopathy stage classification with dual
inputs and selective data sampling. In: Proceeding of the 18th
IEEE international conference on machine learning and applica-
tions (ICMLA'2019), 2019. p. 1578–1584.
40. Pak A, Ziyaden A, Tukeshev K, Jaxylykova A, Abdullina D. Com-
parative analysis of deep learning methods of detection of diabetic
retinopathy. Cogent Eng. 2020;7(1):1–9.
41. Pan J, Yong Z, Sui D, Qin H. Diabetic retinopathy detection based
on deep convolutional neural networks for localization of dis-
criminative regions. In: Proceeding of the 8th international confer-
ence on virtual reality and visualization (ICVRV), October 22–24,
Qingdao, China, Published in IEEE; 2018. p. 46–52
42. Patel S. A comprehensive analysis of convolutional neural network
models. Int J Adv Sci Technol. 2020;29(4):771–7.
43. Pratt H, Coenen F, Broadbent DM, Harding SP, Zheng CY. Con-
volutional neural networks for diabetic retinopathy. In: Proceeding
of the international conference on medical imaging understanding
and analysis (MIUA'2016), vol. 90. 6–8 July, Loughborough, UK,
Published in Procedia Computer Science; 2016. p. 200–2005.
44. Qummar S, Khan FG, Shah S, Khan A, Shamshirband S, Rehman
ZU, Khan IA, Jadoon AW. A deep learning ensemble approach for
diabetic retinopathy detection. IEEE Access. 2019;7:150530–40.
45. Rodriguez-Leon C, Arevalo W, Banos O, Villalonga C. Deep
learning for diabetic retinopathy prediction. In: The proceeding
of the international work-conference on artificial neural networks,
vol. 12861. June 16–18, Virtual Event, Springer, LNCS 2021. p.
537–546
46. Roshini T, Ravi RV, Reema Mathew A, Kadan AB, Subbian PS.
Automatic diagnosis of diabetic retinopathy with the aid of adap-
tive average filtering with optimized deep convolutional neural
network. Int J Imaging Syst Technol. 2020;30(1):1–21.
47. Sampaul TGA, Robinson YH, Julie EG, Shanmuganathan V, Nam
Y, Rho S. Diabetic retinopathy diagnostics from retinal images
based on deep convolutional networks. Preprints. 2020;2020:1–21.

SN Computer Science (2022) 3:357 Page 15 of 15 357
SN Computer Science
48. Seth S, Agarwal B. A hybrid deep learning model for detecting
diabetic retinopathy. J Stat Manag Syst. 2018;21(4):569–74.
49. Setiawan W, Damayanti F. Layers modification of convolu-
tional neural network for pneumonia detection. J Phys Conf Ser.
2020;1477:1–10.
50. Simonyan K, Zisserman A. Very deep convolutional networks for
large-scale image recognition. In: Proceeding of the 3rd IAPR
Asian conference on pattern recognition (ACPR'2015), November
3–6, Kuala Lumpur, Malaysia; 2015. p. 730–734.
51. Suganthi SRL, Sneha UK, Shwetha S. Diabetic retinopathy clas-
sification using machine learning techniques. Int J Eng Trends
Technol. 2020;68(1):51–6.
52. Sun Y, Zhang D. Diagnosis and analysis of diabetic retinopa-
thy based on electronic health records. IEEE Access Spec Sect
Healthc Inf Technol Extreme Remote Environ. 2019;7:86115–20.
53. Supriya M, Seema H, Zia S. Diabetic retinopathy detection using
deep learning. In: Proceeding of the international conference
on smart technologies in computing, electrical and electronics
(ICSTCEE). 9–10 Oct. 2020, Bengaluru, India, 2020. p. 515–520.
https:// doi. org/ 10. 1109/ ICSTC EE496 37. 2020. 92775 06
54. Wang Z, Yin Y, Shi J, Fang W, Li H, Wang X. Zoom-in-net: deep
mining lesions for diabetic retinopathy detection. In: International
conference on medical image computing and computer-assisted
intervention (MICCAI'2017), Canada, September 11–13, Pub-
lished in Lecture Notes in Computer Science book series, vol.
10433, 2017. p. 267–275.
55. Wan S, Liang Y, Zhang Y. Deep convolutional neural networks
for diabetic retinopathy detection by image classification. Comput
Electr Eng. 2018;72:274–82.
56. Wang Z, Yang J. Diabetic retinopathy detection via deep convolu-
tional networks for discriminative localization and visual explana-
tion. In: Proceeding of the workshops of the thirty-second AAAI
conference on artificial intelligence (AAAI-18), February 2–7,
New Orleans, Louisiana, USA, 2018. p. 514–522.
57. Wang X, Lu Y, Wang Y, Chen WB. Diabetic retinopathy stage
classification using convolutional neural networks. In: Proceeding
of the international conference on information reuse and integra-
tion for data science, Salt Lake City, USA, July 7–9, 2018. p.
465–471.
58. Zago GT, Andreão RV, Dorizzi B, Teatini Salles EO. Diabetic
retinopathy detection using red lesion localization and convolu-
tional neural networks. Comput Biol Med. 2020;116:1–12.
59. Zhang X, Zou J, He K, Sun J. Accelerating very deep convolu-
tional networks for classification and detection. IEEE Trans Pat-
tern Anal Mach Intell. 2016;38:1943–55.
60. Zhang Q, Wang H, Yoon SW, Won D, Srihari K. Lung nodule
diagnosis on 3D computed tomography images using deep con-
volutional neural networks. Procedia Manuf. 2019;39:363–70.
Publisher's Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.