VISVESVARAYA TECHNOLOGICAL UNIVERSITY

Belagavi, Karnataka, India

A Report on
“Detection of Diabetic Retinopathy using CNN Algorithm”

Submitted to RVITM Affiliated to Visvesvaraya Technological University (VTU),

Belagavi) in partial fulfillment of the requirements for the award of degree of

BACHELOR OF ENGINEERING
in
ELECTRONICS AND COMMUNICATION ENGINEERING
By
Project Team No. PT15

Akshat Gupta 1RF20EC00

Mohammed 3
Nadeem 1RF20EC02
Shivam Kumar Rai 8
Vikrant Rana 1RF20EC04
2
Under the guidance of
1RF20EC05
Dr. Kavitha N3
Professor Dept. of ECE, RVITM

R V Educational Institutions
R V Institute of Technology and Management, Bengaluru
Department of Electronics and Communication Engineering
2022-23
 ABSTRACT

Diabetes is a chronic end organ disease that occurs when the pancreas does not secrete enough insulin or
the body is unable to process it properly. Over time, diabetes affects the circulatory system, including
that of the retina. Diabetic retinopathy is a medical condition where the retina is damaged because fluid
leaks from blood vessels into the retina. Ophthalmologists recognize diabetic retinopathy based on
features, such as blood vessel area, exudes, hemorrhages, microaneurysms and texture.

Early detection and treatment can limit the potential for significant vision loss from diabetic
retinopathy. When DR is screened manually, it leads to extensive inconsistency among readers. If every
diagnosis is done manually then the demand for screening services will not be met as it would require a
lot of time and money.

CNN algorithms are widely used for the detection of diabetic retinopathy (DR), it helps in analyzing
fundus images in diagnosis . The proposed approach includes 3 stages: Preprocessing, Feature
extraction and segmentation and Classification. The proposed method is evaluated on a publicly
available dataset and compared with other state-of-the-art methods.

.
 MOTIVATION

According to the World Health Organization, diabetic retinopathy is the leading cause of blindness
among working-age adults, affecting up to 35% of people with diabetes worldwide.

In 2021, Approximately 537 million adults (20-79 years) are living with diabetes. The total number of
people living with diabetes is projected to rise to 643 million by 2030 and 783 million by 2045. 3 in 4
adults with diabetes live in low- and middle-income countries. One out of two people suffering from
diabetes has been diagnosed with some stage of DR. Detection of DR symptoms in time can avert the
vision impairment in majority of cases, however such revelation is difficult with present tools and
methods. There has been a need for comprehensive and automated DR detection tools and methods.

By developing an accurate and reliable method for detecting diabetic retinopathy using deep learning
techniques, One can potentially make a significant impact on patient outcomes and improve the quality
of life for many people. Generally this is what happens, suppose you have to show your eyes to
ophthalmologist , eyes are dilated and then there is consultation with the ophthalmologist and this entire
process takes a lot of time for diagnosis and the scarcity of ophthalmologist makes it harder and
expensive. By our technique we can simplify the process of detection and make the entire consultation
more efficient and cost effective.
 LITERATURE REVIEW
TITLE FEATURE LIMITATION
Pravin R. Kshirsagar et.al Contrary to typical machine learning Collecting enough data—is it
“Recognition of Diabetic techniques, deep learning ones learn difficult to know in advance
Retinopathy with Ground the proper features directly from the how much data will be needed
Truth Segmentation Using data instead of relying on the to train an algorithm. Ensuring
Fundus Images and Neural development of hand-crafted sufficient variability—the
Network Algorithm” features, a procedure that may be dataset must have enough data
(2022). very time-consuming and labor- about all relevant classes and
intensive. edge cases that must be
Under the preprocessing process, the handled by the algorithm.
following steps are performed: Color
conversion, Resizing, Filtering.
In this survey, GLCM is used to
isolate the components for feature
extraction
The proposed model has the
following performance metrics for
detecting diabetic retinopathy:
96.1538% sensitivity, 95.6522%
specificity, and 95.9184% accuracy.

R. K. Yadav et.al Resnet-50 architecture is used one major drawback is that a

“Automated Diagnosis of It has 48 convolution layers, 1 deeper network usually
Diabetic Retinopathy using max-pooling, and 1 average requires weeks for training.
ResNet-50 and Data pool layer. It has the detection of errors
Augmentation Techniques” approximately 23 million becomes difficult.
(2021). trainable parameters. It
reduces the training time.
Has high accuracy.
The ResNet-50 with KNN
classifier has a sensitivity,
specificity and accuracy of
95.99 %, 99.16% and 97.52%
respectively, which is superior
to the work reported.
TITLE FEATURE LIMITATION
S. M. Boudiaf et.al In VGG16 there are thirteen convolutional It is very slow to train(2-3
“Automated Detection of layers, five Max Pooling layers, and three weeks).The size of VGG-16
Diabetic Retinopathy Dense layers which sum up to 21 layers trained imageNet weights is
using Transfer Learning but it has only sixteen weight layers. 528 MB. So, it takes quite a
with VGG16 and instead of having a large number of hyper- lot of disk space and
Inception-v3 Models” parameters they focus on having bandwidth which makes it
(2021). convolution layers of 3x3 filter. inefficient. 138 million
The results using VGG16 show a parameters lead to exploding
performance measure of sensitivity. gradients problem.
92.75%, specificity 96.8%, and accuracy
94.1%

In the inception models instead of having

deep layers, we have parallel layers .it has
a total of 42 layers.
It has higher efficiency.
It has a deeper network compared to the
Inception V1 and V2 models, but its speed
isn't compromised.
It uses auxiliary Classifiers as regularizes.
it has an extremely low error rate
compared with its previous models and its
contemporaries.
The performance parameters of the model
attain an accuracy of 95.04%, a specificity
of 96.37%, and a sensitivity of 93.77%

Mahmut Karakaya et.al The smartphone-based retina imaging Overfitting: Transfer

“Smart phone based systems can be used as an learning algorithms may
retinal imaging system alternative to the direct ophthalmoscope. be prone to overfitting,
using Transfer However, the field of view of the which occurs when the
Learning” smartphone- based retina imaging model performs well on
(2020) systems plays an important role in the training data but
determining the automatic DR detection does not generalize well
accuracy. to new data. This can
lead to reduced
accuracy and poor
performance on new
Images.
TITLE FEATURES LIMITATION

J. A. Moreno et.al "Automated U-Net is a convolutional neural UNet has many parameters due to
Diagnosis of Diabetic network that was developed for the skip connections and the
Retinopathy using U-Net biomedical image additional layers in the expanding
Architecture“ segmentation.-net architecture path. This can make the model
(2020). is symmetric and consists of more prone to overfitting,
two major parts: the especially when working with
contracting path which small datasets.
constitutes the general
convolutional process and the
expansive path which
constitutes transposed 2D
convolutional layers.
Experiment evaluation states
that the proposed model
(UNET) outperforms
(accuracy: 0.97, sensitivity:
0.98, specificity: 0.98)
compared to other advanced
models
 PROBLEM FORMULATION

Diabetic retinopathy is a leading cause of vision loss and blindness among people with diabetes,
particularly in low- and middle-income countries. Early detection and timely treatment of the condition
are crucial for preventing vision loss and improving patient outcomes.

However, current methods for the detection and diagnosis of diabetic retinopathy, such as manual
grading by ophthalmologists and computer-aided diagnosis (CAD) systems, have limitations in terms of
accuracy, scalability, and cost-effectiveness.

There is a need for more effective and efficient methods for the detection and management of diabetic
retinopathy, particularly in resource-limited settings.

Deep learning-based approaches, such as convolutional neural network (CNN) algorithms, have shown
promise for the detection of diabetic retinopathy, but further research is needed to evaluate their clinical
utility and effectiveness, particularly in diverse patient populations.

Therefore, the problem statement is how to develop a more accurate, scalable, and cost-effective method
for the detection and management of diabetic retinopathy using deep learning-based approaches,
particularly in resource-limited settings.
 Scope and Objectives

With the advent of India’s own electronic health record system(EHR) . EHR will facilitate automatic
screening and diagnosis of diabetic retinopathy .This can assist medical professionals in identifying at-
risk individuals and initiating quick actions that improve patient outcomes.

India's very own telemedicine programme eSanjeevani will play a major role in our project as. The
created project may be connected with telemedicine platforms, allowing for remote diabetic retinopathy
screening and diagnosis. For patients in rural and underdeveloped areas in particular, this may boost
access to affordable healthcare.

By enhancing the model, it is possible to diagnose diabetic retinopathy in real-time, which would
increase screening effectiveness and reduce the burden on healthcare providers.

The project can be further extended to be able to directly be applied to a fundus camera where after
taking an image of the retina of the patient , the algorithm can easily detect to which stage of diabetic
retinopathy( DR) it is. Thereby reducing the time spent on diagnosis of a patient and result in more
efficiency and accuracy.
 METHODOLOGY

The images are classified into 5 categories namely: No DR, Mild DR, Moderate DR, Severe DR and

PDR.

Preprocessing
A fundus camera provides an upright, magnified view of the fundus. Image pre-processing is the steps
taken to format images before they are used by model training and inference. Image preprocessing may
also decrease model training time and increase model inference speed.
Preprocessing is an essential step to clean image data. Fully connected layers in convolutional neural
networks, require that all images are the same sized arrays. If your images are not in the same size, your
model may not perform as expected.

Color conversion
The fundus images are obtained in the RGB format and since RGB has higher complexity, converting
it to grayscale or extracting the green component increases the intensity, and as a result,
efficiency is also increased.

Green Component
The retinal image is taken in the RGB form by fundus camera. The green channel of the RGB space is
extracted and chosen for detection of exudates because exudates appear most contrasted in this channel.
So, the first step is to separate this channel to a new image.
Grayscale Conversion
The acquired image resolution is 1280 X 1024 in 24-bit JPEG format. The color image of an eye is
taken as the input image and is converted to a grayscale image.

Image resizing
If input images are particularly large, reducing the size of these images will dramatically
improve model training time without significantly reducing model performance. For example,
the standard size of images on iPhone 11 are 3024 × 4032. Image resizing refers to the
scaling of pixels wherein lowering the range of pixels from a picture is facilitated. It also helps
to zoom in on pixels. In case of resizing the picture, it does both reduction and scale-up to
satisfy the dimension requirements.

Filtering
Most of the images are influenced by some kind of unwanted noises causing disturbance in image
quality and resolution. , Analyzing the images are usually processed right after removing these
noises from the images. Filtering provides fine results comparing with other available methods

Adaptive Histogram Equalization

Adaptive histogram equalization, which is said to improve contrast in image, is applied to the grayscale
converted eye image. As a result of this adaptive histogram equalization, the dark area in the input eye
image that was badly illuminated becomes brighter in the output eye image while the side that was
highly illuminated remains or reduces so that the whole illumination of the eye image is same.

Median filtering
The median filter approach is utilized for noise removal from images and signals. Median filter is very
crucial in the image processing field as it is well known for the preservation of edges during noise
removal. The main idea of the median filter is to run through the signal entry by entry, replacing each
entry with the median of neighboring entries. The pattern of neighbors is called the window which
slides, entry by entry, over the entire signal.
Segmentation
Image segmentation is a commonly used technique in digital image processing and analysis to partition
an image into multiple parts or regions, often based on the characteristics of the pixels in the image.
Image segmentation could involve separating foreground from background, or clustering regions of
pixels based on similarities in color or shape.

Threshold Segmentation
Threshold segmentation is the simplest method of image segmentation and also one of the most
common parallel segmentation methods. It is a common segmentation algorithm which directly divides
the image gray scale information processing based on the gray value of different targets. Threshold
segmentation can be divided into local threshold method and global threshold method. The global
threshold method divides the image into two regions of the target and the background by a single
threshold. The local threshold method needs to select multiple segmentation thresholds and divides the
image into multiple target regions and backgrounds by multiple thresholds. The advantage of the
threshold method is that when the target and the background have high contrast, the segmentation effect
can be obtained.

The Fuzzy C-Means Segmentation

Fuzzy c-means (FCM) segmentation is a method of clustering that allows one piece of data to belong to
two or more clusters. Here, it is used to segment the input eye image and detect the blood vessels.
Information about blood vessels can be used in grading disease severity or as part of the process of
automated diagnosis of diseases with ocular manifestations. It is based on minimization of certain
objective function.
Feature extraction
Feature extraction is a part of the dimensionality reduction process, in which, an initial set of the raw
data is divided and reduced to more manageable groups. It helps to get the best feature from those big
data sets by selecting and combining variables into features, thus, effectively reducing the amount of
data. These features are easy to process, but still able to describe the actual data set with accuracy and
originality.

Feature extraction and classification

Feature Extraction and Classification (DRFEC) employs DL CNN models such as VGG-16, VGG-19,
Xception, InceptionV3, InceptionResNetV2, MobileNet, MobileNetV2, EfficientNets B0-B7,
DenseNet121, DenseNet169, DenseNet201, ResNet50, ResNet50V2, ResNet101, ResNet101V2,
ResNet152, ResNet152V2, NASNetLarge and NASNetMobile. Amongst all the models, DenseNet201
has achieved the highest training accuracy whereas VGG-16 has achieved the lowest training accuracy.
Again, VGG-16 has achieved lowest validation accuracy whereas EfficientNetB4 has achieved highest
validation accuracy. ResNet50 has shown highest overfitting whereas InceptionV3 has shown the lowest
overfitting. It has been identified that EfficientNetB4, InceptionResNetV2, NasNetLarge and
DenseNet169 are the top-4 best performing models on Kaggle’s EyePACS DR detection dataset.

EfficientNets
EfficientNets rely on AutoML and compound scaling to achieve superior performance without
compromising resource efficiency. The EfficientNetB4 model has a total of19,466,823 of which
19,341,616 parameters are trainable for a targe
t size of (224 × 224). It has a training accuracy
of 99.37% and a validation accuracy of 79.11% .
INCEPTIONRES-NETV2

Inception-ResNet-v2 is a convolutional neural network that is trained on more than a million images
from the ImageNet database. The network is 164 layers deep and can classify images into 1000
object categories, such as keyboard, mouse, pencil, and many animals. As a result, the network has
learned rich feature representations for a wide range of images. The network has an image input size
of 299-by-299. It has a training accuracy of 99.36% and a validation accuracy of 79.05%.It is observed
that the training accuracy and validation loss of the model are higher than validation accuracy and
training loss, respectively which has led to the overfitting of the model.
 Flowchart

IMAGES

PREPROCESSING

FEATURE
EXTRACTION AND
SEGMENTATION

CLASSIFICATION

NO DR PROLIFIC DR

MILD DR SEVERE DR

MODERATE
 Expected Outcome of the project

• By using different DL model we can classify the stages of diabetic retinopathy into 5 different types
• The model would be trained on three parameters training, validation and testing.
• The expected accuracy of the proposed model would be in the range of 80-98%.

Facilities required for proposed work

Software :
Google collab/Tensor flow lite(IDE)

Programming languages :
Python(backend)
HTML(frontend)
CSS(frontend)
Java script(frontend)

Datasets :
Kaggle dataset for diabetic retinopathy(96,500 images)

Processor :
I5 12th Gen with Nvidia graphics card(for training data)
 Bibliography/Reference

Shuang Yu, Di Xiao and Yogesan Kanagasingam, “Exudate Detection for Diabetic Retinopathy With
Convolutional Neural Networks”.

Oliver Faust & Rajendra Acharya U. & E. Y. K. Ng & Kwan-Hoong Ng & Jasjit S. Suri, “Algorithms
for the Automated Detection of Diabetic Retinopathy Using Digital Fundus Images: A Review”.

Yashal Shakti Kanungo, Bhargav Srinivasan, Dr. Savita Choudhary, “Detecting Diabetic Retinopathy
using Deep Learning”.

Mahmut Karakaya* and Recep E. Hacisoftaoglu, “Comparison of smartphone-based retinal imaging

systems for diabetic retinopathy detection using deep learning”.

R.BHARGAVI (316126510107), R. GEETHA NALINI (316126510106), K.POORNA SUBHASH

(316126510089), K. SAINADH, “DIAGNOSIS OF DIABETIC RETINOPATHY USING
TRANSFER LEARNING”

D. Xiao, A. Bhuiyan, S. Frost, J. Vignarajan, M. L. T. Kearney, and Y. Kanagasingam, “Major

automatic diabetic retinopathy screening systems and related core algorithms: a review,” Machine
Vision and Applications, vol. 30, no. 3, pp. 423–446, 2019. B. Harangi and A. Hajdu, “Automatic
exudate detection by fusing multiple active contours and regionwise classification,” Computers in
Biology and Medicine, vol. 54, pp. 156–171, 2019.

B. Harangi, I. Lazar, and A. Hajdu, “Automatic exudate detection using active contour model and
regionwise classification,” in 2019 Annual International Conference of the IEEE Engineering in
Medicine and Biology, pp. 5951–5954, IEEE, San Diego, CA, USA, August 2019
M. Mateen, J. Wen, Nasrullah, S. Song, and Z. Huang, “Fundus image classification using VGG-19
architecture with PCA and SVD,” Symmetry, vol. 11, no. 1, 2019.

D. Y. Carson Lam, “Automated detection of diabetic retinopathy using deep learning,” AMIA Summits on
Translational Science Proceedings, vol. 2018, 147 pages, 2020

F. Zabihollahy, A. Lochbihler, and E. Ukwatta, “Deep learning based approach for fully automated
detection and segmentation of hard exudate from retinal images,” in Proceedings of the Medical Imaging
2019: Biomedical Applications in Molecular, Structural, and Functional Imaging, Springer, San Diego, CA,
USA, January 2020

D. S. W. Ting, G. C. M. Cheung, and T. Y. Wong, “Diabetic retinopathy: global prevalence, major risk
factors, screening practices, and public health challenges: a review,” Clinical & Experimental
Ophthalmology, vol. 44, no. 4, pp. 260–277, 2020

J. H. Tan, H. Fujita, S. Sivaprasad et al., “Automated segmentation of exudates, haemorrhages,

microaneurysms using single convolutional neural network,” Information Sciences, vol. 420, pp. 66–76,
2021

O. Perdomo, S. Otalora, F. Rodríguez, J. Arevalo, and F. A. González, “A novel machine learning model
based on exudate localization to detect diabetic macular edema,” in Proceedings of the Ophthalmic Medical
Image Analysis Third International Workshop, Athens, Greece, October 2021

B. Harangi and A. Hajdu, “Detection of exudates in fundus images using a Markovian segmentation
model,” in 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology
Society, pp. 130–133, IEEE, Chicago, IL, USA, November 2022
T. Kauppi, V. Kalesnykiene, and J.-K. Kamarainen, “The diaretdb1 diabetic retinopathy database and
evaluation protocol,” in Procedings of the British Machine Vision Conference 2007, pp. 1–10, British
Machine Vision Association, University of Warwick, UK, September 2020

C. Stauffer and W. E. L. Grimson, “Adaptive background mixture models for real-time tracking,” in
Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
pp. 246–252, IEEE, Collins, Colorado, June 2021
 Additional Information

Is the project proposed relevant to the Industry / Society or Institution?

Yes, The identification of diabetic retinopathy using the CNN algorithm is a proposed project that is
very important to both business and society.

The creation of a reliable and effective CNN algorithm for the diagnosis of diabetic retinopathy can
have important commercial implications. Such algorithms can be used by medical imaging firms and
healthcare professionals to give quicker and more precise diagnosis of diabetic retinopathy, improving
patient outcomes and lowering costs.

The suggested project can also help society because it could significantly affect how diabetic
retinopathy is prevented and treated. The proposed project can help in the early identification of diabetic
retinopathy, allowing for prompt interventions and therapies to stop the disease's progression and the
onset of blindness. It will do this by creating a dependable and effective instrument for the detection of
diabetic retinopathy.

All things considered, the suggested project on the diagnosis of diabetic retinopathy using CNN
algorithm has huge ramifications for both business and society, making it extremely relevant and
significant.

Can the product or process developed in the project be taken up for filing a Patent?
No