IRJET -An Automatated Learning Approach for Detection of Diabetic Retinopathy using Deep Learning

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 02 | Feb 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 2826
AN AUTOMATED LEARNING APPROACH FOR DETECTION OF
DIABETIC RETINOPATHY USING DEEP LEARNING
S. Balaji1, Dr. R. Ramachandiran2, P. Karthikayan3, P. Udhayakumar4, V. Prathap5,
1,2Assistant Professor, Department of Information Technology, Sri Manakula Vinayagar Engineering College,
Puducherry
3Student, Department of Information Technology, Sri Manakula Vinayagar Engineering College, Puducherry
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Diabetic retinopathy (DR) is an across the board
issue for diabetic patient and it has been a primary
explanation behind visual deficiency in the dynamic populace.
A few troubles looked by diabetic patients in view of DRcan be
disposed of by appropriately keeping up the blood glucoseand
by auspicious treatment. As the DR accompanies various
stages and differing challenges, it is difficult to DR and
furthermore it is tedious. Right now, build up a computerized
division based order model for DR. At first, the Contrast
restricted versatile histogram evening out (CLAHE) is utilized
for portioning the pictures. Later, deepbelief network(DBN)is
employed for classifying the images into different grades of
DR. For exploratory investigation, the dataset is gotten from
Kaggle site which is open source stage that endeavors to
construct DR recognition model. The highest classifier
performance is attained by the presented model with the
maximum accuracy of 84.35 over compared models.
Key Words: Classification, DR, Segmentation, Deep
Learning, Histogram
1. INTRODUCTION
Diabetic retinopathy (DR) generally occur to patients who
acquires diabetes for long time and because of retinal
damage, it causes blindness[1]. By utilizing the strategy of
fundus imaging, the DR influenced retinal structure of eyes
may be recognized. By centering the eye, thefunduspictures
will be commonly caught through fundus camera. The
interior surface of eye is exhibited through fundus pictures
which involve fovea, retina, veins, optic circleandmacula[9].
A typical retina includes veins which conveys supplements
and blood required for eye. Ordinarily, the veins are
sensitive and in light of extra circulatory strain, they may
barge in diabetic patients. Through additional small blood
vessels count, the diabetic retinopathy progress because of
additional pressure might be found from retinal surface
through additional small blood vessels count, the diabetic
retinopathy progress because of additional pressure might
be found from retinal surface[6].
Fig -1: Stages of Diabetic Retinopathy
2. DIABETIC RETINOPATHY DETECTION
TECHNIQUES
In the section below, different fire detection techniques
are discussed in detail.
2.1 Classification of Diabetic Retinopathy using
ANN
To separate the exudates and veins, morphological
administrators are utilized by this strategy. In addition, an
accuracy of 96% is attained by the methods like genetic
algorithm (GA) andfuzzycmeans(FCM)[4].Thetechniqueof
multilayered thresholding is projected through for blood
vessels segmentation in retinopathy images. For retinal
structureanalysis, ridgelets,curveletandwavelettransforms
are employed additionally with fundus images [6].
Drawbacks:
• This is not a reliable method since the images is not
morphed those specific fundus areas.
• The loss of information in the interface between the
classifiers is detrimental to accurate classification.

2.2 Modified Alexnet architecture forclassification
of diabetic retinopathy images
The present research aims to classifythefundusimageswith
high accuracy into various stages of diabetic retinopathy.
There is a massive growth in patients affected by diabetic
retinopathy. It is necessary to categorize the patients into
different stages of diabetic retinopathy in a swift manner.
Through the present research with the application of a
modified Alexnet architecture we have striven to increase
the classification accuracy in the study of DR images.
Among different CNN structures, Alexnet is one of the most
effective designs that are broadly utilized to address issues
in picture characterization [8]. The first step is to resize the
input fundus image to the size of 259 × 259 pixels
corresponding to the breadth and height and the three color
channels representing the depth of the input fundus image.
The output of neurons is computed as a scalar product of a
small portion of the image with theircorrespondingweights.
This process is repeated along the length and breadth. This
operation is performed in convolutional layer. In Rectified
Linear Unit (RELU) layer, an element-wise activation
function is employed. This layer replaces all the negative
activations with 0 by introducing nonlinearity to the system
and by applying the function –f (k) = max (0, k). In pooling
layer, the samples are reduced along the spatial coordinates.
This process is known as decimation. Fully Connected (FC)
layer computes the Class scores for each image andgivesthe
prediction. The probability score for each of the prediction
class is computed and the class that is scoring maximum
probability score is chosen as the predicted class.
Drawbacks:
• The results can be improved by increasing the size
of the dataset.
• There is lack of augmentation and normalization in
the images.
2.3 Classification of DR using CNN
We bring convolutional neural systems (CNNs) [5]
capacity to DR recognition, which incorporates 3
significant troublesome difficulties: arrangement,
division and detection. Coupledwithmovelearningand
hyper-parameter tuning, we embrace AlexNet, VggNet,
GoogleNet, ResNet[2], and examine how well these
models do with the DR picture classification. We utilize
openly accessible Kaggle stage for preparing these
models. The best classification accuracy is 80.62% and
the results have demonstrated the better accuracy of
CNNs and transfer learning on DR image classification.
Studies in this domain involve feature segmentation
and blood vessels [4].
Drawbacks:
• The segmentation of blood vessel is not accurate
• the precision of the strategy can't be guaranteed as
the dataset highlights are inferred exactly and
physically
• the datasets are of low quality and comprise small
size some fundus images with single collection
environment relatively offer complexities to
compare the algorithm's performance for
experimental purposes
2.4 Classification and Detection of Retinal Changes
Due to Red lesions in Longitudinal Fundus Images
This project have presented a robust and flexible multistage
approach for tracking retinal changes due to small red DR
lesions such as microaneurysms and dot hemorrhages in
longitudinal fundus images. The system was applied to both
small and large retinal fields of 81 diabetic eyes. Strength to
intra and between picture enlightenment varieties was
accomplished by misusing fundus pictures that are
standardized for glow and differentiation over the whole
field of view. The improvement in the visibility and contrast
of especially small retinal features in the normalized fundus
images enabled our approach to track subtleretinal changes,
including those that are visually difficult to detect on the
colour fundus images [2].
A simple and effective criterion for blobness (BM) was
defined fordetectingspatiotemporal retinal changelocations
from longitudinal normalized fundus images[3].TheBMcan
likewise be handily adjusted to other related issues for the
identification and following of little round articles in a
progression of enrolled longitudinal pictures. The proposed
approach was assessed with regards to an ordinary diabetic
retinopathy screening programincludingsubjectsextending
from solid (no retinal sore) to direct (with clinically
pertinent retinal lesions) DR levels. Evaluation was done on
both a large field-of-view fundusmosaics,whichconsistedof
the macula, optic nerve, temporal, and superior fields, and a
small field-of-view of the retina consistingonly ofthemacula
centered fields[2]. The resultsshowthatthesystemwasable
to detect retinal changes due to small DR lesions with a
sensitivity of 80% from large field fundus mosaics and small
field fundus images at an average false positive rate of 2:5
and 1, respectively.
Drawbacks:
• The detection of red DR lesions from single time
point images can be very difficult due to the subtle
nature of most of the lesions and limited number of
lesion pixels

• In contrast to the small fields, the higher falsealarm
rate gets occurred in the large field fundus mosaics
images is mainly caused by the lower image quality
• The presence of significant illumination artefacts
such as white spots
2.5 Diagnosis of DR using Machine Learning
Classification Algorithm
In this they Proposed a method which we train individual
classifier algorithm and not the ensemble of that. The
features extracted and will be used to train the classifiers
and the best individual classifiers are used to identify DR or
non-DR categories by the help of Support Vector Machine,
logistic Regression and Neural Network[6]. This paper
proposed methods todevelopanautomatedsystemtodetect
the case of diabetic retinopathy among the diabetic patients
and is aimed at helping ophthalmologists to detect early
symptoms of diabetic retinopathy with ease. The ideas
proposed for the intelligent system can be understood by
this paper[4]. Also this paper highlightsvarioustechnologies
used for diagnosis and detection of diabetic eye disease.
Drawbacks:
• Here by using the machine learning algorithm the
extraction of featureandclassificationareoccurred
in separate phases and output will be generated.
• It does not work with multiple hiddenlayerofunits.
3. Proposed Method
Keeping the constraints of the current models, right now,
present a productive division based characterizationmodel.
To approve the exhibited model on the DR groupingprocess,
a benchmark dataset from Kaggle site is utilized. Initiallythe
fundus images from the dataset will get preprocessed bythe
method of conversion from RGB to grayscale due to
detection of the neurons we increase the contrast in the
Green fundus area of the images and the images will get
preprocessed . The preprocessed image will undergo
segmentation process. When the picture is sectioned
utilizing CLAHE strategy, the marking of classes happens.
Next, DBN based classification model will be built by proper
training phase. Once the model is created using DBN, test
input images can be provided to attain proper output. The
proposed work involves in the process of three stages:
a) Preprocessing b) Segmentation c) Classification of stages.
Fig -2: Work Flow of Proposed Method
4. CONCLUSIONS
The motivation of this project work is to implement an
automatic diagnosisofDRusingfundusimagesclassification.
Extreme vision misfortune in diabetic patients can be
maintained a strategic distance from by identifying and
treating diabetic retinopathy at a beginning period. The
method proposed in this paper aims at providing an optimal
solution for the classification ofdiabeticretinopathypatients
according to the severity of the disorder. Deep learning is
one of the state of the art techniquestoaddressclassification
problems and it provides better accuracy. Efficient Deep
Belief Network architecture to detect andclassifythefundus
images will be helpful for the ophthalmologist to a greater
extent in eradicating the vision loss due to diabetic
retinopathy. The testing and training the evaluation of
proposed architecture is done using Kaggle dataset.
Classifying the images collected from the Kaggle datasetinto
Healthy retina, Normal,Mild, Moderate,Severe,Proliferateof
stages using the proposed work.
REFERENCES
[1] C. P. Wilkinson, F. L. Ferris, R. E. Klein, P. P. Lee, C. D.
Agardh, M. Davis, D. Dills, A. Kampik, R.
Pararajasegaram, and J. T. Verdaguer, "Proposed
international clinical diabetic retinopathy and
diabetic macular edema disease severity scales,"
Opthalmology, vol. 110, no. 9, pp. 1677–1682, Sep.
2003
[2] M. M. Fraz, P. Remagnino,A.Hoppe,B.Uyyanonvara,
A. R. Rudnicka, C. G. Owen, and S. A. Barman, "Blood
vessel segmentation methodologies in retinal

images -A Survey," Comput. Meth. Prog. Bio., vol.
108, pp. 407–433, Mar. 2012.
[3] J. Nayak, P. S. Bhat, U. R. Acharya, C. M. Lim, and M.
Kagathi, "Automated identification of diabetic
retinopathy stages using digital fundus images," J.
Med. Syst., vol. 32, pp. 107–115, 2008.
[4] R. Pires, S. Avila, H. F. Jelinek, J. Wainer, E. Valle, and
A.Rocha,"Beyondlesion-baseddiabetic retinopathy:
a direct approach for retinal," IEEE J. Biomed,
Health Inform., vol. 21, no. 1, pp. 193–200, Jan.
2017.
[5] Doshi D, Shenoy A , Sidhpura D , Gharpure P .
Diabetic retinopathy detection using deep
convolutional neural networks. 2016 international
conference on computing, analytics and security
trends (CAST), December; 2016 .
[6] Bhatkar AP, Kharat GU. Detection of diabetic
retinopathy in retinal images using MLP classifier.
2015 IEEE international symposium on nano
electronic and information systems, December;
2015.
[7] Haloi M, Dandapat S, Sinha R (2015) A Gaussian
scale space approach for exudates detection,
classification and severity prediction. ArXiv. 2015.
pp 1–7.
[8] T. Shanthi , R.S. Sabeenian, “ Modified Alexnet
architecture for classification of diabetic
retinopathy images”, Dec. 2018.
[9] S. Balaji, “A New Perception Of Grey Wolf
Optimization in Cloud Classification. A Cloud Based
Application for Solving Medicine Oriented Real
World Complex Problems”, International journal of
Pure andAppliedMathematics(IJPAM),Volume119,
Issue 14, 2018.
[10] Kedir M. Adal_, Peter G. van Etten,JoseP.Martinez,
Kenneth W. Rouwen, Koenraad A. Vermeer,
Member, IEEE and Lucas J. van Vliet, Member,
IEEE, “An Automated System fortheDetectionand
Classification of Retinal Changes Due to Red
Lesions in Longitudinal Fundus Images, ”IEEE
Transcations on Biomedical Engineering, 2017.

IRJET -An Automatated Learning Approach for Detection of Diabetic Retinopathy using Deep Learning

More Related Content

What's hot (19)

Similar to IRJET -An Automatated Learning Approach for Detection of Diabetic Retinopathy using Deep Learning (19)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET -An Automatated Learning Approach for Detection of Diabetic Retinopathy using Deep Learning