Volume 5, Issue 6, June – 2020 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Diabetic Eye Disease Detection Using

Machine Learning Techniques
Prof. Jayshree Aher Araadhya Sharma, Sashank Vemulapalli,
Department of Computer Engineering Pragat Singh, Meet Shah
MIT College of Engineering, Pune Department of Computer Engineering
Pune, Maharashtra, India MIT College of Engineering, Pune
Pune, Maharashtra, India

Abstract:- Diabetic Retinopathy (DR) is an eye disease very accurate at identifying the disease, it is extremely
that affects people that suffer from diabetes over inefficient since it requires a lot of manual work and no
prolonged periods of time. If not detected and diagnosed automation currently exist for it. Also, many times the
at the right time, it often leads to weakening of vision medical tools as well as equipment and the expertise
and can even lead to absolute loss of vision. The disease required for carrying out the procedure is lacking in regions
generally affects people who are aged between 35 to 50 of the world where a large number of people have diabetes
years, but recent cases involving teenagers have also and where the need of such a system for detection is most
been reported widely. The process for diagnosing needed. Since the current lifestyle of people around the
Diabetic Retinopathy is often difficult since very few world would only lead to a greater number of people across
visible symptoms appear in patients until it is too late the world with diabetes, there is an ever increasing need for
for treatment and the point of no return is met. Current automation in this procedure and the current system will
techniques that exist for detecting Diabetic Retinopathy soon not be able to cope up with the increasing demands in
are extremely time consuming and require a manual the respective field.
procedure to be carried out by lab technicians which
involves inserting medical tools into the patient’s eye. The process consists of recognizing very minute
The proposed methodology is to utilize the neoteric details, such as microaneurysms, to some broader and more
branch of computer science i.e. Machine Learning visible details, such as exudates. Also, the position of these
techniques to assist in identifying and diagnosing the physically observable deformities are important to carry out
disease by analysing the images of the eye. As per the the procedure for detecting the disease.
research study, the images will be preprocessed, and
converted to the Gray Scale following which the Microaneurysms are the first symptoms that are
extraction of relevant features using appropriate observable and are visible in the images as red dots that
supervised learning techniques are carried out to obtain typically vary from 25 to 100 micrometres in diameter [3].
the final trained model. Exudate on the other hand is a fluid that constitutes of
proteins, cells and other solid entities and leaks out of from
Keywords:- Diabetic Retinopathy, Diabetic eye disease, blood vessels into the surrounding tissues [4]. Exudate
Microaneurysms, Exudates, Machine learning, Supervised seeps from areas of inflammation, cuts and infections and is
learning Introduction. also commonly referred to as pus. In people suffering with
diabetic retinopathy, exudates form in the retina of the eye
I. INTRODUCTION as shown below [5].

The Department of Public Health in the United States

estimates that about 29 million people suffer from diabetes
in the United States [1]. Diabetic Retinopathy is an eye
disease that affects people suffering from diabetes over
prolonged periods. Weakening and loss of vision due to
Diabetic Retinopathy can be avoided if it is detected at the
right time, however this is a very challenging task since the
disease displays very few symptoms before it is too late for
diagnosis and providing treatment. Current techniques that
exist for detecting Diabetic Retinopathy are extremely time
consuming and require a manual procedure to be carried out
by the lab technicians which involves examining the eye of
the patient with the help of lab tests which involve inserting
medical tools into the eye.

Symptoms that help identify and detect Diabetic

Fig 1:- Sample Image (Thin Arrows: Exudates; Triangle:
Retinopathy are lacerations and other deformations that are
Microaneurysms; Thick arrow: Retinal haemorrhage)
caused by the disease [2]. Even though this procedure is

IJISRT20JUN435 www.ijisrt.com 752

Volume 5, Issue 6, June – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
II. LITERATURE SURVEY III. PROPOSED SYSTEM

The challenge with detecting and diagnosing Diabetic The system that is proposed in this research paper
Retinopathy is that the patient is unaware that they are employs supervised machine learning techniques in order to
affected by the disease until the disease has developed to an classify an images of the eye that is given as input into two
extent that treatment will prove to be ineffective. categories: “Positive” and “Negative”. The positive category
Automation of this procedure for detecting and identifying indicates the presence of the disease and the negative category
the disease has great potential in reducing the cost of the indicates its absence.
procedure as well as saving time and reducing the amount
of manual labor required. Techniques that involve
analysing the images with the help of computer systems can
help to a great extent in identifying the symptoms precisely
as well as in determining the damage that has already been
caused by the disease. Correct diagnosis of the disease on
time reduces the probability for total loss of vision by 50%
in the patient [6]. Since there are almost millions of patients
each year that undergo the procedure for determining the
presence of the disease, doctors and lab technicians are
finding it extremely difficult to cope up with the ever
increasing demand for their skill and expertise. This is
especially the case in rural areas where majority of people
do not have access to good healthcare facilities as well as
the shortage of doctors and other experts in these regions
[7]. Therefore, automating the process for detecting the
disease can assist the doctors and other experts in the field
and can also help to limit the damage that is caused by this
disease.

Processing the images of the eye with the help of

computer systems help in quick detection of the disease as
well as provides convenience to the patient since the need
for inserting medical tools into the eye does not exist
anymore.
Fig 3:- Proposed Methodology
The primary goal of the automated system in detecting
the disease is to accurately identify patients who have The system is first trained on the images from the
progressed to threshold disease in a reliable manner so that training dataset which are labelled with their respective
quick and effective treatment can be provided to them. As categories. Prior to training, the required preprocessing
shown in fig. 2 [8], current techniques for detecting the steps such as RGB to Gray conversion are carried out. After
disease are manual and time consuming. achieving accuracy above a specific threshold, the system is
used to classify new incoming images of the patient’s eye.

A. Image Pre-processing
In this step, the image is converted from the RGB to
Gray scale is subsequently used as an input for extracting
the features. The Gray scale involves different shades
where the darkest shade that is possible on the scale is
black whereas the lightest possible shade is white. A black
shade on the scale signifies the absolute absence of any
kind of light whereas the white shade signifies the absolute
reflection or transmission of light across all the visible
wavelengths [9]. The RGB to gray scale conversion is
shown in fig.4 and fig.5 below [10].

Fig 2:- Manual Screening of the Eye

IJISRT20JUN435 www.ijisrt.com 753

Volume 5, Issue 6, June – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
for each neuron in the second layer. The convolutional
neural network is able to tackle this problem by reducing
the number of parameters that are considered for further
processing thereby allowing to construct a deeper network
with fewer parameters.

Fig 6:- CNN

Fig 4:- Colored image
As shown in fig.6 [11], CNN consists of multiple
convolutional and pooling layers. The connected layers of
the network will serve as a classifier on top of the extracted
features and will assign probabilities for the objects in the
image being what the algorithm predicts them to be.

 Support Vector Machine (SVM)

SVM is used to determine the optimal hyperplane that
is able to linearly separate the data points into two
components while maximizing the distance and margins of
the hyperplane from data points in each category. In two
dimension, the hyperplane is represented in the form of a
straight line whereas it is represented as a plane in the three
dimension. SVM is widely used to classify data that is
linearly separable since it cannot accurately classify non-
linear data due to the lack of flexibility in SVM to bend the
hyperplane. Once the training data is proved to be
sufficiently linear, SVM can be used to implement the
Fig 5:- Gray Scale Image system for classification into the “Positive” and “Negative”
categories.
The blood vessels that are present in the original
images are mostly unclear, and the size of each image in
the dataset is different from each other, therefore it is
essential to preprocess images so that they are brought to
the same size and the visibility of the blood vessels and
various other features that assist in detecting the disease are
improved. Before we proceed to extract the features, the
images are first rescaled to the same size.

B. Pooling
Pooling is a technique that is used to convert the input
matrix Min × Nin into a smaller Mout × Nout output matrix.

C. Classification algorithms
 Convolutional Neural Network (CNN)
Each neuron that is a part of the neural network is
affected by only its receptive field. It is impractical to use
feed forward neural networks to classify images since an
extremely large number of neurons would be required due
to the large input size of the images where each and every
pixel is represented as a different variable that is used for
processing. For example, a neural network that processes a Fig 7:- SVM
small image of size 50 x 50 has 2,500 corresponding values

IJISRT20JUN435 www.ijisrt.com 754

Volume 5, Issue 6, June – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
From the above figure [12], SVM problem can be [4]. Salman Sayed, Dr. Vandana Inamdar, Sangram Kapre,
formulated as: “Detection of Diabetic Retinopathy using Image
Processing and Machine Learning” ( IJIR, Vol. 6,
Issue 1, January 2017).
[5]. Sagar Honnungar, Sanyam Mehra, Samuel Joseph,
“Diabetic Retinopathy Identification and Severity
Classification”, Stanford University, 2016
[6]. Harry Pratt, Frans Coenen, Deborah M. Broadbent,
Simon P. Harding, Yalin Zheng, “Convolutional
Here yi =+1 represents the positive class (having eye Neural Networks for Diabetic Retinopathy”
(International Conference on Medical Imaging
disease) and yi = -1 represents the negative class (not
Understanding and Analysis, July 2016).
having eye disease).
[7]. Yung-Hui Li, Nai- Ning Yeh, Shih-Jen Chen and Yu-
D. Performance Analysis Chien Chung, “Computer-Assisted Diagnosis for
Diabetic Retinopathy Based on Fundus Images Using
 Accuracy: Accuracy is a metric that is used to represent Deep Convolutional Neural Network” (National
Central University, Taiwan 2019).
the number of predictions that are correct out of the
[8]. Fenner BJ, Wong RLM, Lam WC, “Advances in
total number of predictions that are made.
retinal imaging and applications in diabetic
TP + TN retinopathy screening: a review” (Ophthalmol Ther,
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = Vol. 6, Issue 3, 2018)
TP + TN + FN + FP [9]. Julius A, Hopper W. “A non-invasive, multi-target
approach to treat diabetic retinopathy”, Biomed
 Sensitivity: Sensitivity is used to measure the accuracy Pharmacother 2019; 109:708–15.
of the predictions of diabetic diseased eye that are made [10]. Krause, J. et al. Grader Variability and the Importance
by the model. of Reference Standards for Evaluating Machine
Learning Models for Diabetic
TP Retinopathy. Ophthalmology. 125(8), 1264–1272
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 =
TP + FN (2018)
[11]. Ali Karsaz, M. Roshan, "A comparative analysis of
 Specificity: Specificity is used to measure the accuracy classification algorithms in diabetic retinopathy
of the predictions of normal eye that are made by the screening", Computer and Knowledge Engineering
model. (ICCKE) 7th International Conference on, pp. 84-89,
2017.
TN [12]. Gurmeen Kaur, Shveta, “Detection of Diabetic
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =
TN + FP Retinopathy using SVM and MDA”, International
Journal of Computer Applications, Vol. 117-No 19,
Here, TP- True Positives; FP- False Positives; TN- May 2018.
True Negatives; FN- False Negatives.

IV. CONCLUSION

In the proposed work, a non-invasive procedure has

been presented to identify and diagnose diabetic eye
disease with the help of supervised Machine Learning
techniques.

REFERENCES

[1]. Zhentao Gao, Jie Li, Jixiang Guo, Yuanyuan Chen,

Zhang Yi, and Jie Zhong, “Diagnosis of Diabetic
Retinopathy Using Deep Neural Networks” (Institute
of Electrical and Electronics Engineers, January 2019)
[2]. Yogesh Kumaran, Chandrashekhar M. Patil, “A Brief
Review of the Detection of Diabetic Retinopathy in
Human Eyes Using Pre-Processing and Segmentation
Techniques” ( IJRTE, Volume-7, December 2018)
[3]. Carson Lam, Darvin Yi, Margaret Guo, Tony Lindsey,
“Automated Detection of Diabetic Retinopathy using
Deep Learning”, Stanford School of Medicine, 2018

IJISRT20JUN435 www.ijisrt.com 755