A Convolutional Neural Network Approach for Diabetic Retinopathy Classification

Abstract

Diabetic Retinopathy (DR) is a kind of problem which affects diabetic patients, particularly those at their age of working, and can result in vision impairment and possibly irreversible blindness. For diagnosis and to prevent blindness or degeneration, early detection is critical. When ophthalmologists execute the diagnosis step of DR manually, it takes more time, effort, and money, and there are more possibility of misdiagnosis. The scientific community is focusing on developing a computer-aided recognition system for early identification and grading of DR severity. Ongoing AI research has highlighted the growth of the deep learning technique, which is better technique for doing medical image analysis and classification.

Full text

11th IEEE International Conference on Communication Systems and Network Technologies
978-1-6654-8038-3/22/$31.00 ©2022 IEEE 357
DOI: 10.1109/csnt.2022.63
A Convolutional Neural Network Approach for
Diabetic Retinopathy Classification
Nasmin Jiwani
Research Scientist
University of The Cumberlands
USA
nasminjiwani@gmail.com
Ketan Gupta
Research Scientist
University of The Cumberlands
USA
ketan1722@gmail.com
Neda Afreen
Department of Computer
Engineering
Jamia Millia Islamia
New Delhi, India
neda184441@st.jmi.ac.in
Abstract— Diabetic Retinopathy (DR) is a kind of problem
which affects diabetic patients, particularly those at their age
of working, and can result in vision impairment and possibly
irreversible blindness. For diagnosis and to prevent blindness
or degeneration, early detection is critical. When
ophthalmologists execute the diagnosis step of DR manually, it
takes more time, effort, and money, and there are more
possibility of misdiagnosis. The scientific community is
focusing on developing a computer-aided recognition system
for early identification and grading of DR severity. Ongoing AI
research has highlighted the growth of the deep learning
technique, which is better technique for doing medical image
analysis and classification.
Keywords— Convolutional Neural Network, Deep
Learning, Diabetic Retinopathy, IDRiD dataset
I. INTRODUCTION
In the world of medicine, disease therapy is more feasible
when diagnosed early. Diabetes develops when the body's
glucose levels rise owing to a lack of insulin. The retina,
kidneys, heart, and nerves are all affected by diabetes.
According to a WHO report, diabetes impacted 423 million
people globally in 2014, and the number expected to rise to
700 million by 2050.
Diabetic Retinopathy (DR) is a condition which affects
diabetic patients, particularly those in their working years.
DR is a complication of diabetes that causes the retina veins
to enlarge and leakage of blood and liquids. The problem of
vision is caused by DR. Patients with diabetes who have
been suffering from the disease for a long time are more
likely to develop DR. Patients should have their retinas
screened on a regular basis for analysis and DR treatment in
the early stages to reduce the risk of vision loss. The
presence of several types of signs on a retina image
distinguishes DR. Table 1 lists these symptoms as
Haemorrhages (HM), Microaneurysms (MA), Soft exudates
(EX) and Hard exudates (EX).
Early identification is critical for accurate diagnosis and
the prevention of blindness or degeneration[1]. Diagnosis
takes longer time and is more expensive when performed
manually. Automated detection, on the other hand, makes the
process simple[2]. This study investigated the Convolutional
neural network for DR to overcome this issue and maximize
the result. The main contributions of this work is to apply
deep learning to improve DR classification results on a
publically available dataset, as CNN uses less time and
computation than classical machine learning[3].
 Haemorrhages (HM) shows up as huge mark on
the retina and dimension range more than 125
µm with unpredictable edge. HM subdivided
into two types namely blot (deeper HM) and
flame (superficial HM).
 Hard exudates shows up as a bright yellow
stain on the retina brought about by spillage of
plasma. `
 Soft exudates named as cotton wool also shows
up like white markings on retina and it brought
about having a round or oval shape due to
irritation of the nerve fibre.
 Microaneurysms (MA), The first signs of DR
appear as small red round specks on the retina.
The length is less than 125m and there are sharp
edges present.
No DR, mild DR, moderate DR, severe DR, and
proliferative DR are the five stages of DR based on the
occurrence, which are briefly listed in Table 1 and we put
mild, moderate, severe and proliferative DR in one class and
no DR in aother class for creating binary class. Figure 1 and
2 shows a retina image with certain anomalies and its
grading into distinct stages. DR detection via automated
approaches saves time and money, and it is more accurate
than manual analysis. When performed manually, the risk of
misdiagnosis is higher[4].
(i) (ii) (iii) (iv) (v)
Fig. 1. Retina images with grading levels (i) normal, (ii) mild, (iii) moderate,
(iv) severe and (v) proliferative DR
2022 IEEE 11th International Conference on Communication Systems and Network Technologies (CSNT) | 978-1-6654-8038-3/22/$31.00 ©2022 IEEE | DOI: 10.1109/CSNT54456.2022.9787577
Authorized licensed use limited to: University of the Cumberlands. Downloaded on June 08,2022 at 23:47:22 UTC from IEEE Xplore. Restrictions apply.

358
Fig. 2. Normal retinal image and DR related anomalies
Table 1
DR grading on the basis of severity and signs
DR severity grading
Signs
No DR
There are no anomalies
lesion ranging between MA
and severe DR
With DR
It may be any of them:
 Intraretinal HM
number for every 4
quadrants is more
than 20.
 Venous beeding in
more than 2 quadrant.
 More than one
quadrant has a visible
intraretinal
microvascular
problem
 Pre-retinal or
Vitreous HM
The following is how the paper is laid out: The second
section introduces related literature. Section 3 exhibits the
working model process in detail, Section 4 displays our
dataset together with results, and Section 5 concludes the
paper.
II. RELATED WORK
Early DR research relied on measurements of optic circle
and veins manually, as well as the presence of flaws such as
microaneurysms, haemorrhages soft exudates and hard
exudates. Then, utilization of several machine learning
approaches such as k-nearest neighbour and support vector
machines (SVM), grading was done by hand built features
(KNN). With the use of support vector machine and K-
nearest neighbour classifier, Jaykumar et al. [5] proposed a
procedure for microaneurysms detection and exudates from
the retina. GLCM feature extraction is used for further
classification after preprocessing. The SVM classifier
outperforms the KNN classifier. Patton et al. [6] investigated
and established standards for retinal imaging evaluation, as
well as the approach for recognizing retinal marker and
indications associated with DR. Jordon et al. [7] provides a
brief introduction to quantifiable methodologies for
reviewing fundus photographs, with a focus on identifying
retinal symptoms and examine retinal illness using automatic
algorithms.
Deep learning reduces the requirement for human
intervention in feature engineering by immediately learning
data depiction at a low level with high level parameters.
Recep et al. [8] proposed a strategy on the basis of AlexNet,
GoogleNet and CNN to improve the outcome of DR
identification using mobile phone and traditional fundus
camera retina photos. The result of emloying photos from
diverse groups is examined by retraining these frameworks
on datasets such as EyePACS, Messidor and IDRiD. On
independent test datasets, these approaches exhibit great
accuracy. Alyoubi et al. [9] revealed state-of-the-art
solutions for DR colour fundus picture localization and
categorization using deep learning process. Furthermore, the
colour fundus retina DR datasets were examined. They also
take on contrast testing difficulties that require greater
investigation. Using a cross disease attention network,
Xiameng et al. [10] described a new approach for rating DR
and DME together. With only picture level inspection, it
analyses the inner links between the diseases. They created
two independent attention modules for disease dependent
and disease specified aspects learning, then combined them
for grading DR and DME to improve grading outcomes. For
testing, they use the Messidor and IDRiD datasets.
III. PROPOSED METHODOLOGY
This part presents the organization of the used deep learning.
Deep learning is a part of Artificial Intelligence that gets its
inspiration from the human brain structure[11]. Different
layers of hierarchy exist in DL, each of which includes
indiscriminate processing steps for pattern classification and
unsupervised feature learning. Segmentation, classification
and image registration are just a few of the uses of DL in
medical image analysis. DL performs features extraction
from the system using training set photos to learn the
structure. Although this learning ability eliminates the need
for creating particular features, the strategy is based on
comprehensive end-to-end DL training[12].
CNNs are more commonly employed for analysis in
clinical image than other approaches, and are quite efficient.
Convolution layers, pooling layers, and fully connected
(FC) layers are the three primary layers in the CNN
architecture[13]. The CNN's number of layers, size, and
filters are all determined by the vision of author. Every layer
in the CNN architecture has a distinct function. Various
filters performs convolution of an image in the CONV layers
Authorized licensed use limited to: University of the Cumberlands. Downloaded on June 08,2022 at 23:47:22 UTC from IEEE Xplore. Restrictions apply.

359
to extract the features[14]. To minimize the feature maps
size, after CONV layer pooling layer is applied. Various
pooling techniques are there but the most popular are
average pooling and maximum pooling. FC layers are used
to characterize the input image entirely. The frequently
utilized classification function is the SoftMax activation
function[15]. The model summary of CNN is illustrated in
figure 3.
Fig 3. CNN model summary
In general, dataset collection and the necessary pre-
processing to enhance and upgrade the images is the prior
step in the process of detection and classification of DR
using DL. Then, DL technique is applied, which performs
feature extraction and classification of the images as shown
in figure 3.
.
Fig 4. Classification of DR using DL method
IV. EXPERIMENTAL RESULT
Using the IDRiD dataset, we compare and evaluate the
model's performance in this section.
A. Dataset
The IDRiD dataset was created using real clinical data from
an eye hospital in Maharashtra (Nanded), India. The whole
dataset contains 516 photos with different DR classes which
we have taken as binary class[16]. DR is graded into classes
based on disease severity and here taken as one without DR
and other with DR. There are 413 photos for training and
103 images for testing. Table 2 shows the statistics labelling
inside the IDRiD dataset. The IDRiD dataset is related with
three types of ground truth, which are listed below:
 Annotation at the pixel size- This type of notation is
used to locate specific lesions within a photograph, as
well as to split and pinpoint region of interest in data.
A total of eighty one colour fundus pictures with DR
signs and 164 with no sign of DR are included in the
dataset. Color fundus images are available in the
format of .jpg, with binary masks for each lesion type
in .tif format, as well as a specific optic disc mask
(OD) for entire eighty one colour fundus images[17].
These annotations are significant in study because
they allow researchers to examine how lesion
segmentation is computed inside an image.
 Diseases Grading of DR- It consists of data that
indicates the risk factor associated with the entire
image. The 516 pictures with varying pathological
stages of DR and DME were rated by a clinical
specialist. The CSV record allows to grade all photos
on a DR severity level. By retaining the desired ratio
of disease stratification, the train and test data consists
of 413 and 103 photos, respectively.
 Optic disc and Fovea centre co-ordinates- For the
entire 516 image data, the fovea centre and OD co-
ordinates are stated, and is presented as a CSV record
B. Evaluation Metric
There are various performance standard measurements for
assessing Deep Learning algorithms categorization
performance. The area under the curve (AUC), sensitivity,
accuracy and specificity are the most commonly used
standard measurements[18]. The number of correctly
categorized photos, expressed as a percentage, is used to
determine accuracy. Specificity is the percentage of normal
images that are classified as normal, whereas sensitivity
indicates the percentage of abnormal images which are
labelled as aberrant by the classifier[19]. AUC is a graph
that shows the relationship between specificity and
sensitivity.
Input images
Preprocessing
DL Method
Classification
Authorized licensed use limited to: University of the Cumberlands. Downloaded on June 08,2022 at 23:47:22 UTC from IEEE Xplore. Restrictions apply.

360
Where, TP stands for True Positive which means correctly
classified diseased images. True Negative (TN) refers to
occurrences of non-diseased images that are categorized as
such. False Positive (FP) images are non-diseased images
that are classified as sick, whereas False Negative (FN)
images are diseased images that are classified as normal[20].
C. Result
CNN framework is employed for DR classification. The
system is trained using the ADAM optimizer with a 1e-4
learning rate and a batch size of 10 for 30 epochs, a dropout
layer with a dropout rate of 0.5 is applied after that sparse
categorical loss is used as a loss function[21][22]. Accuracy
of CNN model for DR classification is shown in Table 2.
Here train accuracy is 73% and test accuracy is 68%.
Table 2. Train and test accuracy for DR classification
Grading
Train Accuracy
Test Accuracy
DR
73%
68%
Accuracy and loss in training and validation of DR
grading is given in Fig.5.
(a)
(b)
Fig 5. (a) Accuracy in training and validation (b) Loss in Training and
validation
V. CONCLUSION
Deep learning is used for DR classification based on its
binary grading to increase study in the medical domain,
particularly for the diagnosis of diabetic retinopathy.
Diabetic Retinopathy is a condition that affects diabetic
patients, particularly those in their working years, and causes
vision impairment and, in some cases, permanent blindness.
For diagnosis and to prevent blindness or degeneration, early
detection is critical. Ongoing AI research has highlighted the
growth of the deep learning technique, which is the best
technique for doing medical picture analysis and
classification. And here, the CNN model classifies DR based
on binary class on the public benchmark dataset IDRiD,
achieving higher training and testing accuracy.
REFERENCES
[1] M. Sameer and B. Gupta, “Beta Band as a Biomarker for
Classification between Interictal and Ictal States of Epileptical
Patients,” in 2020 7th International Conference on Signal
Processing and Integrated Networks (SPIN), 2020, pp. 567–570,
doi: 10.1109/SPIN48934.2020.9071343.
[2] S. K. B. Sangeetha, N. Afreen, and G. Ahmad, “A Combined
Image Segmentation and Classification Approach for COVID-19
Infected Lungs,” J. homepage http//iieta. org/journals/rces, vol.
8, no. 3, pp. 71–76, 2021.
[3] M. Sameer, A. K. Gupta, C. Chakraborty, and B. Gupta,
“Epileptical Seizure Detection: Performance analysis of gamma
band in EEG signal Using Short-Time Fourier Transform,” in
2019 22nd International Symposium on Wireless Personal
Multimedia Communications (WPMC), 2019, pp. 1–6, doi:
10.1109/WPMC48795.2019.9096119.
[4] A. Mahajan, K. Somaraj, and M. Sameer, “Adopting Artificial
Intelligence Powered ConvNet To Detect Epileptic Seizures,” in
2020 IEEE-EMBS Conference on Biomedical Engineering and
Sciences (IECBES), 2021, pp. 427–432, doi:
10.1109/IECBES48179.2021.9398832.
[5] J. Lachure, A. V Deorankar, S. Lachure, S. Gupta, and R. Jadhav,
“Diabetic Retinopathy using morphological operations and
machine learning,” in 2015 IEEE International Advance
Computing Conference (IACC), 2015, pp. 617–622, doi:
10.1109/IADCC.2015.7154781.
[6] N. Patton et al., “Retinal image analysis: concepts, applications
and potential,” Prog. Retin. Eye Res., vol. 25, no. 1, pp. 99–127,
Authorized licensed use limited to: University of the Cumberlands. Downloaded on June 08,2022 at 23:47:22 UTC from IEEE Xplore. Restrictions apply.

361
2006.
[7] K. Jordan, M. Menolotto, N. Bolster, I. Livingstone, and M.
Giardini, “A Review of Feature-based Retinal Image Analysis,”
Expert Rev. Ophthalmol., vol. 12, Mar. 2017, doi:
10.1080/17469899.2017.1307105.
[8] R. E. Hacisoftaoglu, M. Karakaya, and A. B. Sallam, “Deep
learning frameworks for diabetic retinopathy detection with
smartphone-based retinal imaging systems,” Pattern Recognit.
Lett., vol. 135, pp. 409–417, 2020, doi:
https://doi.org/10.1016/j.patrec.2020.04.009.
[9] W. Alyoubi, W. Shalash, and M. Abulkhair, “Diabetic
retinopathy detection through deep learning techniques: A
review,” Informatics Med. Unlocked, vol. 20, p. 100377, Jun.
2020, doi: 10.1016/j.imu.2020.100377.
[10] X. Li, X. Hu, L. Yu, L. Zhu, C.-W. Fu, and P.-A. Heng, “CANet:
cross-disease attention network for joint diabetic retinopathy and
diabetic macular edema grading,” IEEE Trans. Med. Imaging,
vol. 39, no. 5, pp. 1483–1493, 2019.
[11] N. Nasir, N. Afreen, R. Patel, S. Kaur, and M. Sameer, “A
Transfer Learning Approach for Diabetic Retinopathy and
Diabetic Macular Edema Severity Grading,” Rev. d’Intelligence
Artif., vol. 35, pp. 497–502, Dec. 2021, doi: 10.18280/ria.350608.
[12] M. Sameer and B. Gupta, “ROC Analysis of EEG Subbands for
Epileptic Seizure Detection using Naive Bayes Classifier,” J.
Mob. Multimed., pp. 299–310, 2021.
[13] M. Sameer and B. Gupta, “Time–Frequency Statistical Features
of Delta Band for Detection of Epileptic Seizures,” Wirel. Pers.
Commun., 2021, doi: 10.1007/s11277-021-08909-y.
[14] S. M. Beeraka, A. Kumar, M. Sameer, S. Ghosh, and B. Gupta,
“Accuracy Enhancement of Epileptic Seizure Detection: A Deep
Learning Approach with Hardware Realization of STFT,”
Circuits, Syst. Signal Process., 2021, doi: 10.1007/s00034-021-
01789-4.
[15] S. Gupta, M. Sameer, and N. Mohan, “Detection of Epileptic
Seizures using Convolutional Neural Network,” in 2021
International Conference on Emerging Smart Computing and
Informatics (ESCI), 2021, pp. 786–790, doi:
10.1109/ESCI50559.2021.9396983.
[16] P. Porwal et al., “Indian Diabetic Retinopathy Image Dataset
(IDRiD): A Database for Diabetic Retinopathy Screening
Research,” Data , vol. 3, no. 3. 2018, doi: 10.3390/data3030025.
[17] M. Sameer and P. Agarwal, “Coplanar waveguide microwave
sensor for label-free real-time glucose detection,”
Radioengineering, vol. 28, no. 2, p. 491, 2019.
[18] M. Sameer and B. Gupta, “Detection of epileptical seizures based
on alpha band statistical features,” Wirel. Pers. Commun., vol.
115, no. 2, pp. 909–925, 2020, doi: 10.1007/s11277-020-07542-
5.
[19] M. Sameer, A. K. Gupta, C. Chakraborty, and B. Gupta, “ROC
Analysis for detection of Epileptical Seizures using Haralick
features of Gamma band,” in 2020 National Conference on
Communications (NCC), 2020, pp. 1–5, doi:
10.1109/NCC48643.2020.9056027.
[20] N. Afreen, R. Patel, M. Ahmed, and M. Sameer, “A Novel
Machine Learning Approach Using Boosting Algorithm for Liver
Disease Classification,” in 2021 5th International Conference on
Information Systems and Computer Networks (ISCON), 2021, pp.
1–5.
[21] N. Jiwani, K. Gupta, and P. Whig, “Novel HealthCare
Framework for Cardiac Arrest With the Application of AI Using
ANN,” in 2021 5th International Conference on Information
Systems and Computer Networks (ISCON), 2021, pp. 1–5, doi:
10.1109/ISCON52037.2021.9702493.
[22] M. Sameer and B. Gupta, “CNN based framework for detection
of epileptic seizures,” Multimed. Tools Appl., 2022, doi:
10.1007/s11042-022-12702-9.
Authorized licensed use limited to: University of the Cumberlands. Downloaded on June 08,2022 at 23:47:22 UTC from IEEE Xplore. Restrictions apply.