Diabetic Retinopathy Classification and Detection Using

Coloured Images with Deep Learning Inspired Framework

A PROJECT REPORT

Submitted by

ANKUR JAISWAL (21SCSE1280043)

MOHD. SAIF SIDDIQUE (21SCSE1010034)
POJECT ID: BT40252

Under the guidance of

Name of Supervisor: Mr. A. Boobalan

Assistant Professor

in partial fulfillment for the award of the degree of

BATCHELOR OF TECHNOLOGY
IN
(Artificial intelligence and Data Science)

SCHOOL OF COMPUTING SCIENCE AND ENGINEERING

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING
GALGOTIAS UNIVERSITY, GREATER NOIDA
October,2024
 Diabetic Retinopathy Classification and Detection Using
Coloured Images with Deep Learning Inspired Framework

A PROJECT REPORT

Submitted by

ANKUR JAISWAL (21SCSE1280043)

MOHD. SAIF SIDDIQUE (21SCSE1010034)
POJECT ID: BT40252
in partial fulfillment for the award of the degree of

BACHELOR OF ENGINEERING
IN
(Artificial intelligence and Data Science)

Galgotias University

Oct 2024
 BONAFIDE CERTIFICATE
Certified that this project report “Diabetic Retinopathy Classification and
Detection Using Coloured Images with Deep Learning Inspired Framework”
is the bonafide work of “ ANKUR JAISWAL (21SCSE1280043) and MOHD.
SAIF SIDDIQUE (21SCSE1010034)” who carried out the project work under
my/our supervision and his/her work is recommended for the award of
BACHELOR OF TECHNOLOGY.

Mr. K. Rajkannan Mr. A. Boobalan

Signature of Examiner(s) Signature of Supervisor(s)

Signature of Program Chair Signature of Dean

Date: October, 2024

Place: Greater Noida
 TABLE OF CONTENTS

List of Figures……………………………………………………………………………1
List of Tables ……………………………………………………………………………2
Abstract………………………………………………………………………………….3
Graphical Abstract………………………………………………………………………4
Abbreviations……………………………………………………………………………5
CHAPTER 1. INTRODUCTION……………………………………….6
1.1 Human Eye…………………………………………………………………6
1.1.1. Eye Anatomy…………………………………………………….7
1.2. Diabetic Retinopathy………………………………………………………9
1.2.1. Types of Diabetic Retinopathy…………………………………..9
1.2.2. Classification Of Diabetic Retinopathy Acc. To Stage………...10
1.2.3. Causes…………………………………………………………..13
1.2.4. Symptoms………………………………………………………14
1.3. Identification of Client/ Need/ Relevant Contemporary issue…………...14
1.4. Identification of Problem ………………………………………………..15
1.5. Identification of Tasks…………………………………………………...16
1.6. Timeline …………………………………………………………………17
1.7. Organization of the Report………………………………………………18
CHAPTER 2. LITERATURE REVIEW/
BACKGROUND STUDY………………………………………….......11
2.1. Timeline of the reported problem …………….…………………….…….21
2.2 Review Based on Past Experiences……………….………………….……23
2.3. Existing solutions…………………………………………………………25
2.4. Bibliometric analysis ……………………………………………………..27
2.5. Review Summary …………………………………………………………32
2.6. Problem Definition ………………………………………………………..34
2.6. Goals/Objectives …………………………………………………………..36
CHAPTER 3. DESIGN FLOW/PROCESS…………………………….37
3.1. Evaluation & Selection of Specifications/Features…………………………
3.2. Design Constraints ………………………………………………………….
3.3. Analysis of Features and finalization subject to constraints …………………
 3.4. Design Flow .....................................................................................................
3.5. Design selection ...............................................................................................
3.6. Implementation plan/methodology ..................................................................
CHAPTER 4. RESULTS ANALYSIS AND VALIDATION…………….
4.1. Implementation of solution……………………………………………………
CHAPTER 5. CONCLUSION AND FUTURE WORK…………………..
5.1.Conclusion……………………………………………………………………...
5.1.Future Work……………………………………………………………………
REFERENCES……………………………………………………………………………38
 List of Figures

Figure 1…………………… ……………………………………………...Human Eye

Figure 2………………… ……………………………………….…Eye Classification

Figure 3………………… ………………… Difference Between Normal and DR Eye

Figure 4………………… …………… WHO Classification of DR according to grade

Figure 5…………………… ………………… Image Segmentation General Category

Figure 6……………… …… Supervised and Unsupervised Method for Classification

Figure 7………………………… Diabetic Retinopathy: Timeline of Major Advances

Figure 8…………………………………………………… Machine learning methods

Figure 9……………………… DR recognition and classification using deep learning

1
 List of Tables
Table 1…………………………………………………………………… Timeline of Project

Table 2………………… Key Fratures, Drawbacks and Effectivensess of Exesting Solutions

Table 3..........................................................................Key findings from the literature review

2
 ABSTRACT
Diabetic Retinopathy (DR) is a progressive eye disease affecting individuals with diabetes,
caused by damage to the retinal blood vessels. Accurate classification of DR stages is essential
for early diagnosis and effective treatment planning to prevent vision loss. Various imaging
techniques are employed in detecting DR, with fundus photography being the most widely used
for its ability to capture detailed images of retinal blood vessels. Deep Learning (DL),
specifically Convolutional Neural Networks (CNN), has recently shown remarkable results in
medical imaging for classification and segmentation tasks. In this study, a DL-based model is
proposed for the classification of DR stages using two publicly available datasets, with a primary
focus on fundus imaging data from the Kaggle Diabetic Retinopathy Dataset. DR is categorized
into stages based on the severity of retinal damage: mild, moderate, severe Non-Proliferative
Diabetic Retinopathy (NPDR), and Proliferative Diabetic Retinopathy (PDR). Precise staging
of DR can assist in predicting patient outcomes, enabling clinicians to tailor interventions. Our
framework leverages an EfficientNet-B3 architecture, trained to classify DR stages, and aims to
achieve high accuracy and consistency in predicting DR progression. The model’s performance
yielded a classification accuracy of 0.95, demonstrating robust capability for clinical
application. Continuous evaluation of various algorithms is being pursued to enhance accuracy,
aiming for the most reliable model that offers efficient diagnostic support.

Keywords: Diabetic Retinopathy (DR), Deep Learning (DL), Convolutional Neural

Networks (CNN),

3
GRAPHICAL ABSTRACT

Figure(i). UI of Web Application

Figure(ii). Block Diagram of Web Application

4
 ABBREVIATIONS
Abbreviation
AI: Artificial Intelligence
ML: Machine Learning
EDA: Exploratory Data Analysis
DR: Diabetic Retinopathy
NPDR: Non-Proliferative Diabetic Retinopathy
PDR: Diabetic Retinopathy
ME: Macular Edema
CNN: Convolutional Neural Network
DL: Deep Learning
ANN: Artificial Neural Network
MRI: Magnetic Resonance Imaging
AUC: Area Under the Curve
ROC: Receiver Operating Characteristic
TPR: True Positive Rate
FP: False Positive
FN: False Negative
ReLU: Rectified Linear Unit (activation function used in CNNs)
SGD: Stochastic Gradient Descent (optimization algorithm)
Adam: Adaptive Moment Estimation (an alternative optimization
algorithm)
GPU: Graphics Processing Unit

5
 CHAPTER 1
INTRODUCTION
1.1 Eye
The eye is one of the most complex and specialized organs in the human body, designed for
the critical function of vision. Consisting of over 2 million working parts, the eye processes
and interprets light to create images that enable us to see and interact with our surroundings.
The eye, which includes the cornea, iris, lens, retina, and optic nerve, allows us to perceive
depth, color, and movement. Like the brain, the eye receives information from the external
world and transmits it to the brain, where it is organized into meaningful visuals. This
coordination between the eye and brain enables a range of functions, from basic sight to
complex visual processing necessary for tasks like reading, recognizing faces, and navigating
spaces. The eye plays a key role in our daily lives by aiding spatial awareness, enabling vision-
based memory, and enhancing our overall perception of the environment.

Figure1: Human Eye

6
1.1.1. Eye Anatomy
The human eye, a complex organ responsible for vision, is made up of multiple specialized
structures that work together to capture and process visual information. The main parts of the
eye include the cornea, iris, lens, retina, and optic nerve, each playing a specific role in how
we see.

• Cornea: The cornea is the clear, dome-shaped outer layer covering the front of the eye.
It acts as a protective shield and the primary lens, bending light to help focus it onto the
retina.

• Iris and Pupil: The iris is the colored part of the eye that controls the size of the pupil—
the central opening. By expanding or contracting, the iris adjusts the amount of light
entering the eye, much like the aperture in a camera.

• Lens: Located behind the iris, the lens further focuses light onto the retina. Its shape
can change to focus on near or distant objects, a process known as accommodation.

• Retina: The retina is a thin layer of tissue at the back of the eye that contains millions
of photoreceptor cells, including rods and cones. Rods help us see in low light, while
cones enable color vision and detailed sight. The retina converts light into electrical
signals.

• Optic Nerve: The optic nerve carries visual information from the retina to the brain. It
acts as a communication pathway, relaying the processed data to the visual cortex for
interpretation.

• Macula and Fovea: The macula, located in the center of the retina, is responsible for
central vision. At the center of the macula is the fovea, which provides the sharpest
visual detail for activities like reading and recognizing faces.

Other Major Parts of the Eye and Their Functions :

• Sclera: Often referred to as the "white of the eye," the sclera is the tough, protective
outer layer that maintains the eye's shape and provides structural support.

• Choroid: Located between the sclera and retina, the choroid is a vascular layer that
supplies oxygen and nutrients to the retina and other parts of the eye.

• Conjunctiva: A thin, transparent membrane covering the sclera and lining the inside
of the eyelids, the conjunctiva helps protect the eye from infections and keeps it moist.

7
• Aqueous Humor: This is a clear, watery fluid in the front part of the eye between the
cornea and lens. It helps maintain intraocular pressure, supplies nutrients, and removes
waste products.

• Vitreous Humor: The vitreous humor is a gel-like substance filling the space between
the lens and retina. It helps maintain the eye's shape and allows light to pass through to
the retina.

• Ciliary Body: This structure contains muscles that adjust the shape of the lens for
focusing and also produces the aqueous humor. It plays a critical role in eye
accommodation for viewing objects at varying distances.

• Zonules (Suspensory Ligaments): These fibers attach the lens to the ciliary body,
helping to hold the lens in place and enabling it to change shape during accommodation.

• Fovea: A small depression in the center of the macula, the fovea contains a high density
of cone cells, providing the clearest and most detailed vision for tasks that require high
visual acuity, like reading.

Figure 2: Eye Classification

8
1.2. Diabetic Retinopathy:
Diabetic Retinopathy is a condition characterized by abnormal changes in the blood vessels of
the retina due to prolonged high blood sugar levels. In a healthy eye, retinal blood vessels
continuously deliver oxygen and nutrients, supporting clear vision. However, in individuals
with diabetes, chronically elevated blood sugar can damage these vessels, leading to swelling,
leaking, or even the formation of new, fragile vessels. As these changes progress, they disrupt
the retina’s function, potentially causing vision loss. Diabetic Retinopathy often develops
gradually and may go unnoticed in the early stages, underscoring the importance of regular eye
exams for individuals with diabetes to help detect and manage this condition before it leads to
significant impairment.

1.2.1. Types of Diabetic Retinopathy :

Diabetic Retinopathy (DR) progresses through different stages, classified into two main types:
Non-Proliferative Diabetic Retinopathy (NPDR) and Proliferative Diabetic Retinopathy
(PDR). These stages reflect the increasing severity of damage to the retinal blood vessels.

i. Non-Proliferative Diabetic Retinopathy (NPDR) :

NPDR is the early stage of diabetic retinopathy and is often called the "background" stage. In
this stage, damage to the retinal blood vessels occurs, but new blood vessels have not yet
formed. NPDR can be categorized into mild, moderate, and severe forms, depending on the
extent of the damage.

• Mild NPDR: Small areas of swelling in the blood vessels, known as microaneurysms,
develop. These are the earliest visible signs of damage, and patients may not experience
noticeable symptoms.

• Moderate NPDR: As the condition progresses, blood vessels that nourish the retina
may become blocked or damaged. This can cause fluid leakage into the retina, leading
to swelling and impaired vision.

• Severe NPDR: At this stage, a larger number of blood vessels become blocked,
depriving more areas of the retina of proper blood flow. This signals the retina to grow
new blood vessels as a compensatory response, which leads to the more advanced stage
of the disease.

9
ii. Proliferative Diabetic Retinopathy (PDR) :

PDR is the advanced stage of diabetic retinopathy. In response to the retinal blood vessel
damage in NPDR, the retina tries to compensate by growing new, abnormal blood vessels (a
process known as neovascularization). However, these new vessels are fragile and prone to
breaking, leading to serious complications.

• Neovascularization: Fragile new blood vessels grow on the surface of the retina or in
the vitreous (the clear gel that fills the eye). These vessels can easily break and leak
blood into the vitreous, causing vision loss or "floaters."

• Vitreous Hemorrhage: When the new vessels bleed into the vitreous, it can block
vision and result in sudden vision loss. Over time, the blood may clear up, but severe
or repeated hemorrhages can lead to permanent damage.

• Retinal Detachment: The scar tissue formed by new blood vessel growth can contract
and pull the retina away from the back of the eye, leading to retinal detachment, a
serious condition that can cause permanent blindness if not treated.

• Neovascular Glaucoma: In some cases, the new blood vessels can also grow in the
front part of the eye, blocking normal fluid drainage and leading to increased eye
pressure, causing a type of glaucoma that further damages vision.

Figure 3: Difference Between Normal and DR Eye

1.2.2. Classification Of Diabetic Retinopathy According To Stage :

Diabetic Retinopathy (DR) is graded based on the severity of retinal damage, which helps guide
treatment decisions. The grading system is mainly divided into Non-Proliferative Diabetic

10
Retinopathy (NPDR) and Proliferative Diabetic Retinopathy (PDR), with further
subdivisions depending on the extent of the retinal abnormalities. Here are the different grades
of DR:
• Mild NPDR: Microaneurysms, early signs of DR, usually asymptomatic.
• Moderate NPDR: More blood vessel damage and leakage, possible early symptoms.
• Severe NPDR: Extensive blood vessel blockage and high risk of progressing to PDR.
• PDR: Growth of abnormal new blood vessels, significant risk of vision loss or
blindness.

Macular Edema (ME)

Although not a separate grade, macular edema can occur at any stage of DR and is the leading
cause of vision loss in diabetic patients. It involves the swelling of the macula due to fluid
leakage from damaged retinal blood vessels. The condition may be classified as:

• Focal: When leakage is confined to specific areas.

• Diffuse: Widespread leakage affecting the entire macular region.

Figure 4: WHO Classification of DR according to grade

1. Mild Non-Proliferative Diabetic Retinopathy (NPDR)

Characteristics:

• Presence of microaneurysms (small bulges in retinal blood vessels).

11
 • These are the earliest detectable changes in the retina due to diabetes.
• Often asymptomatic, and vision may not yet be affected.

Risk: Low risk of progression to vision-threatening complications at this stage.

2. Moderate Non-Proliferative Diabetic Retinopathy (NPDR)

Characteristics:

• More widespread microaneurysms, retinal hemorrhages, and areas of blocked blood

vessels (capillary non-perfusion).
• Some blood vessels begin to leak, causing fluid to accumulate in the retina (edema).

Symptoms: Some patients may start to notice mild vision impairment, particularly if macular
edema is present.

Risk: The risk of progression to more severe stages increases.

3. Severe Non-Proliferative Diabetic Retinopathy (NPDR)

Characteristics:

• Significant blockage of blood vessels leading to widespread retinal ischemia (lack of

oxygen).
• The retina is deprived of nutrients, which signals the body to grow new blood vessels,
setting the stage for proliferative diabetic retinopathy.
• “4:2:1 rule” is often used to define severe NPDR:

▪ Hemorrhages or microaneurysms in 4 quadrants of the retina.

▪ Venous beading (thickened or twisted veins) in 2 or more quadrants.

▪ Intraretinal microvascular abnormalities (IRMA) in 1 or more

quadrants.

Symptoms: Likely more noticeable visual disturbances, though significant vision loss may still
not occur.

Risk: High risk of progression to PDR within one year.

12
4. Proliferative Diabetic Retinopathy (PDR)

Characteristics:

• Neovascularization (growth of new, abnormal blood vessels) due to extensive ischemia.

• These new blood vessels are fragile and prone to breaking, causing vitreous hemorrhage
(bleeding into the eye).
• Scar tissue may form, leading to traction on the retina and potential retinal detachment.

Symptoms:

• Severe visual disturbances, including floaters, sudden vision loss due to hemorrhage,
or distorted vision from retinal detachment.
• Vision loss can be profound and permanent if untreated.

Risk: Immediate risk of severe vision loss or blindness.

1.2.3. Causes :
Diabetic Retinopathy (DR) is primarily caused by prolonged high blood sugar levels and
several other contributing factors that lead to damage in the retinal blood vessels.
Key causes include:

• Prolonged high blood sugar levels (hyperglycemia): Chronic high blood sugar
damages the small blood vessels in the retina.

• Duration of diabetes: The longer a person has diabetes, the higher the risk of DR,
especially with poor control.

• Poorly controlled diabetes: Inadequate management of blood sugar significantly

increases the risk of developing and worsening DR.

• High blood pressure (hypertension): Elevated blood pressure damages retinal blood
vessels, exacerbating the condition.

• High cholesterol levels: Increased cholesterol leads to fatty deposits in the retina,
contributing to DR progression.

• Smoking: Smoking elevates the risk of retinal blood vessel damage in diabetic
individuals.

• Pregnancy: Hormonal changes during pregnancy can accelerate the progression of DR

in women with pre-existing diabetes.

13
 • Obesity: Excess body weight contributes to insulin resistance, making blood sugar
control more difficult, increasing the risk of DR.

• Kidney disease: Diabetic kidney damage (nephropathy) is associated with worsening

DR.

• Genetic factors: A family history of diabetes or diabetic retinopathy can increase

susceptibility to DR.

1.2.4. Symptoms:
Diabetic Retinopathy (DR) often progresses silently in its early stages, but as the condition
worsens, symptoms begin to appear. Common symptoms of DR include:

I. Blurry vision: As the blood vessels in the retina become damaged, vision may become
blurred or distorted.

II. Floaters: Small dark spots or strings (floaters) may appear in the field of vision due to
bleeding from abnormal retinal blood vessels.

III. Difficulty seeing at night: Night vision or seeing in low-light conditions may become
challenging as the retina is affected.

IV. Impaired color vision: Difficulty distinguishing colors or reduced ability to perceive
color accurately.

V. Sudden vision loss: A sudden, severe loss of vision can occur if there is significant
bleeding into the vitreous (vitreous hemorrhage) or retinal detachment.

VI. Dark or empty areas in vision: As the retinal damage progresses, parts of the field of
vision may become dark or blocked.

VII. Fluctuating vision: Vision may change frequently due to swelling in the retina or
changes in blood sugar levels.

1.3. Identification of Client /Need / Relevant Contemporary issue:

Diabetic Retinopathy (DR) is a leading cause of vision impairment and blindness among
individuals with diabetes, making it a critical public health concern globally. The increasing
prevalence of diabetes worldwide, particularly in developing nations, has led to a rise in DR
cases, further burdening healthcare systems. Early detection and intervention are essential for
preventing vision loss from DR, but conventional diagnosis relies heavily on manual

14
examination of retinal images by specialists, which is both time-consuming and resource-
intensive.

With advancements in medical imaging and AI, deep learning-based frameworks offer an
innovative solution to automate DR classification and detection. This project specifically
focuses on developing a system that uses deep learning with colored retinal images to classify
and detect the stages of DR efficiently and accurately. Using a large dataset of fundus images,
our aim is to build a robust, scalable model that can identify DR presence, classify its severity,
and provide valuable guidance for treatment planning. The model's ability to differentiate
between various DR stages, from mild to proliferative, could support healthcare providers in
making timely, informed decisions, potentially reducing the incidence of diabetes-induced
blindness.

This AI-driven approach aligns with the World Health Organization's goals to reduce
preventable vision loss and empowers healthcare providers in regions with limited access to
specialists. By introducing automated detection and classification for DR, this project not only
addresses the need for improved accuracy and accessibility in DR diagnosis but also offers a
scalable, efficient solution that can transform diabetic eye care for millions worldwide.

1.4. Identification of Problem:

The primary challenge in diabetic retinopathy (DR) management is the lack of accessible,
accurate, and automated diagnostic tools that can efficiently classify and detect DR stages. This
gap in technology and resources leads to delayed diagnosis, increasing the risk of vision
impairment and blindness among diabetic patients. Key issues include:

• Limited availability of specialists for early DR detection, especially in remote or

underserved areas

• High reliance on manual examination, which is time-consuming and prone to human

error

• Difficulty in accurately distinguishing between stages of DR, essential for effective

treatment

• Insufficient tools for continuous monitoring and follow-up care

15
Without reliable automated systems for DR classification, healthcare providers face challenges
in ensuring timely intervention, resulting in preventable vision loss and compromised patient
outcomes.

1.5. Identification of Task:

The following tasks are essential to address the identified problem of Diabetic Retinopathy
(DR) classification and detection:

Data Collection:

• Gather a comprehensive dataset of colored retinal images from publicly available

sources and medical institutions, including labeled DR stages for supervised learning.

Model Development:

• Develop a deep learning model, leveraging Convolutional Neural Networks (CNNs),

to classify DR stages (e.g., mild, moderate, severe) from colored fundus images.

• Implement a feature extraction process to identify patterns and abnormalities indicative

of DR, such as hemorrhages, microaneurysms, and exudates.

Training and Validation:

• Train the model on the dataset to ensure accurate DR classification.

• Validate the model’s performance using metrics such as accuracy, sensitivity,

specificity, and the Dice Similarity Coefficient (DSC) to evaluate its robustness and
precision.

Optimization:

• Fine-tune model parameters to improve classification accuracy and prevent overfitting,

using techniques such as data augmentation and regularization.

Deployment:

• Integrate the trained model into an accessible platform, potentially a mobile or web-
based application, to facilitate use by healthcare providers for DR screening.

Post-Implementation Review:

16
 • Monitor the model’s effectiveness in real-world settings and gather feedback from
healthcare providers for ongoing improvement and model updates.

These tasks aim to establish a reliable, automated framework for early and accurate DR
detection, enhancing patient outcomes through timely intervention and monitoring.

1.6. Timeline:

Month 1: Project Planning and Data Collection

• Define project objectives, finalize team roles, and responsibilities.

• Collect a dataset of labeled retinal images.

Month 2: Data Preprocessing and Exploration

• Preprocess images (resizing, normalization, augmentation).

• Conduct exploratory data analysis (EDA) to ensure data quality.

Month 3: Model Development and Baseline Testing

• Build a baseline deep learning model for DR classification.

• Test and evaluate initial model performance.

Month 4: Model Training, Tuning, and Feature Extraction

• Optimize model hyperparameters to improve accuracy and minimize overfitting.

• Implement feature extraction techniques for improved classification accuracy.

Month 5: Validation, Testing, and Performance Evaluation

• Validate model on test data and evaluate using metrics (accuracy, sensitivity,
specificity, DSC).

• Refine model based on evaluation and prepare initial documentation.

Month 6: Deployment and Project Wrap-up

• Deploy the model into an application (web or mobile interface).

• Conduct post-deployment testing, finalize documentation, and present the project to

stakeholders.

17
 Task Month 1 Month 2 Month 3 Month 4 Month 5 Month 6

Project Planning and Data ✔

Collection

Data Preprocessing and ✔ ✔

Exploration

Model Development and Baseline ✔ ✔

Testing

Model Training, Tuning, and ✔ ✔

Feature Extraction

Validation, Testing, and ✔ ✔

Performance Evaluation

Deployment and Project Wrap- ✔ ✔

up

Table 1: Timeline of Project

1.7. Organization of Report:

The report will be structured into the following chapters:

1. Introduction: Overview of Diabetic Retinopathy (DR), its impact on vision, and the
potential of AI and deep learning solutions in early DR detection and classification.

2. Problem Statement: Detailed description of the challenges in DR diagnosis,

including reliance on manual examination, limited access to specialists, and the need
for automated, accurate detection tools.

3. Research Methodology: Explanation of data collection methods, dataset sources, and

preprocessing steps. Overview of the deep learning models, techniques (e.g., CNN),
and evaluation metrics used for analysis.

4. Solution Development: Description of the developed deep learning model for DR

classification, including architecture, feature extraction, and optimization techniques
to distinguish DR stages.

18
5. Testing and Results: Results from testing the model on the DR dataset, with
performance analysis through metrics such as accuracy, sensitivity, specificity, and
Dice Similarity Coefficient (DSC).

6. Implementation and Deployment: Details of the deployment process, including user

interface design, accessibility, and integration into a web or mobile application for use
by healthcare providers.

7. Conclusion and Future Scope: Summary of findings, discussion on the significance

of automated DR detection, and potential improvements or scalability options for
future iterations.

19
 CHAPTER 2
LITERATURE REVIEW/BACKGROUND STUDY

This section aims to present a comprehensive overview of the existing methodologies and the
current advancements in the classification and detection of diabetic retinopathy (DR) using
colored retinal images. Traditional methods of diagnosing DR often rely on manual
examination, which can be subjective, time-consuming, and prone to human error. The
increasing prevalence of diabetes has necessitated the development of more efficient and
accurate diagnostic tools. As a result, researchers have turned to automated techniques,
including machine learning and deep learning algorithms, to enhance the precision of DR
detection.

Recent literature highlights a significant shift towards the utilization of convolutional neural
networks (CNNs) and other deep learning architectures for image classification tasks in
ophthalmology. These techniques leverage large datasets of annotated retinal images, enabling
them to learn discriminative features and improve diagnostic accuracy. Furthermore,
advancements in image processing techniques have facilitated better segmentation of retinal
structures, allowing for more precise identification of pathological changes associated with DR.

Figure 5: Image Segmentation General Category

20
Segmentation and classification of images are two pillars of image processing. Numerous
strategies and methods have been employed for the purposes of segmentation and classification.
Clinical image segmentation is a methodology and procedure for discovering the region of
interest, partitioning the image into various regions, or recognising and differentiating
foreground and background based on pixel similarities from the 2D and 3D images acquired
using different modalities. Figure 5 depicts the overall classification of the image segmentation.
The classification of images is a second pillar used to distinguish between data classes. Figure
8 illustrates the fundamental distinction between supervised and unsupervised classification
strategies and techniques. Method that attempts to determine the relationship between an input
attribute and a specified objective. It necessitates that the attribute and associated condition be
associated with names that provide information about a previous classification. Unsupervised
method that attempts to discover the hidden data structureUnsupervised method utilising
statistical similarities to cluster images into distinct clusters.

Figure 6: Supervised and Unsupervised Method for Classification

2.1. Timeline of the reported problem:

Diabetic retinopathy (DR) has emerged as a significant public health concern due to the
increasing prevalence of diabetes worldwide. Over the years, the urgency for effective detection
and classification methods has grown substantially.
• In the early 1990s, studies began to document the impact of diabetes on vision, with
increasing awareness of DR as a leading cause of blindness among adults.
• By 2005, the American Diabetes Association recognized DR as a critical complication

21
 of diabetes, emphasizing the need for regular eye examinations to detect early signs of
the disease.
• In 2010, the global prevalence of diabetes was reported to exceed 280 million, further
escalating the urgency for effective screening and intervention strategies to prevent
vision loss due to DR.
• In 2015, the World Health Organization (WHO) reported that diabetic retinopathy
affects approximately one-third of people with diabetes, underlining the necessity for
efficient detection methods to manage this complication.
• In 2018, advancements in deep learning and computer vision prompted researchers to
explore automated classification systems for DR, resulting in numerous studies that
demonstrated high accuracy rates in detecting the disease using retinal images.
• By 2020, the introduction of AI-driven tools for DR screening gained traction, with
several clinical trials showcasing their potential to improve early diagnosis and treatment
adherence among diabetic patients.
These milestones reflect the increasing recognition of diabetic retinopathy as a significant health
issue and the urgent need for innovative solutions, such as AI-powered classification and
detection methods, to combat its prevalence and mitigate its impact on public health.

Figure 7: Diabetic Retinopathy: Timeline of Major Advances

22
2.2.Review based on past experience:

J. Amin, M. Sharif, M. Yasmin, and S. L. Fernandes propose a methodology in [1] validated

on three benchmark datasets to detect and classify diabetic retinopathy (DR) images
automatically. The approach involves preprocessing, feature extraction, and classification,
achieving a 97.1% accuracy, with an area under the curve (AUC) of 0.98.

S. Lu, J. Zhu, and H. Zhang in [2] present a deep learning model for DR classification,
leveraging transfer learning on a large dataset. This method demonstrated 95.6% accuracy and
high sensitivity and specificity for reliable DR detection.

K. R. Bhaskaranand et al. in [3] use an automated grading approach validated across various
DR severity levels, achieving 96% accuracy. Their methodology applies deep learning-based
lesion analysis to classify DR stages, focusing on sensitivity for mild cases.

W. Sun et al. propose in [4] a Siamese-like convolutional neural network for automated DR
detection. Validated on Kaggle datasets, this method reached 98% sensitivity and 94%
accuracy, demonstrating high efficacy for real-time clinical applications.

T. T. Nguyen and S. Rana present in [5] a deep neural network approach to classify fundus
images. Their model, optimized on local datasets, reached 96.8% accuracy and 97% sensitivity,
emphasizing computational efficiency and speed.

S. Wang et al. use an ensemble learning strategy in [6] to address the challenge of DR image
variability. Their system yielded an accuracy of 95.3%, achieving robust classification even in
complex cases by combining SVM and CNN features.

R. A. Chetoui et al. propose in [7] a fuzzy decision-making approach integrated with deep
learning for DR detection. Their method achieved 97% accuracy, with a high degree of
specificity, proving effective in distinguishing between severity levels.

Z. Liu, Y. Wang, and J. Shi in [8] develop a novel network architecture for DR diagnosis,
showing strong performance with 96% accuracy on the EyePACS dataset, emphasizing
optimized training for faster inference.

P. Kamble and A. Joshi propose in [9] a CNN-based method using fundus images for early DR
detection. Their model achieved 97.2% sensitivity, highlighting its capability in detecting early
pathological changes in retinal images.

23
A. Das and S. Prabhu in [10] focus on an automated segmentation-based classification
approach, achieving 96.5% accuracy by isolating and analyzing DR lesion characteristics,
contributing to improved classification accuracy for each DR stage.

J. A. Abràmoff et al. in [11] developed an autonomous AI system validated in primary care for
DR detection, reaching 98% sensitivity. Their model underscores the feasibility of a standalone
DR detection system.

E. Tekin et al. present in [12] a CNN-driven model trained on real-world data, achieving a 95%
accuracy and demonstrating robust generalization across varying DR severities, including pre-
proliferative and proliferative stages.

S. A. Soliman proposes in [13] a CNN with optimized image processing for DR detection. The
model reached 94.6% accuracy and provided significant robustness to noise and lighting
conditions, crucial for real-world deployment.

M. A. Brancati et al. in [14] explore a transfer learning approach for DR classification, with
96.9% accuracy and 96.5% sensitivity, emphasizing cross-domain effectiveness and improved
model efficiency.

L. Gadekallu et al. present in [15] a CNN-based model leveraging transfer learning, achieving
97% accuracy on a large dataset. The model highlights scalability for different image
resolutions, making it suitable for large-scale DR screening.

R. K. Das, A. Dey, and D. Ghosh in [16] use transfer learning for DR classification, achieving
a robust 98% accuracy by fine-tuning deep features, which effectively increased classification
confidence.

V. Chandran et al. propose in [17] an ensemble learning framework that combines CNN with
fuzzy logic, achieving 98% accuracy and strong interpretability, critical for DR severity
analysis.

S. Haleem et al. present in [18] a hybrid CNN-fuzzy logic model, achieving 96.4% accuracy
and improved detection of subtle retinal changes, aimed at accurately detecting non-
proliferative and proliferative DR.

M. Mirsharif et al. propose in [19] an automated system combining image processing with
CNN for DR detection, validated with an accuracy of 95%, optimized for clinical deployment
in diverse populations.

24
R. Priya and K. Aruna present in [20] an artificial neural network approach for DR detection,
reaching 96% accuracy. Their system emphasizes robustness in handling variable image
quality, catering to diverse clinical settings.

Each study contributes unique algorithmic strategies and benchmarks to advance DR

classification accuracy, sensitivity, and robustness in real-world applications.

2.3. Existing Solutions:

Numerous approaches have been developed for the classification and detection of diabetic
retinopathy (DR) using colored retinal images, employing both traditional image processing
techniques and advanced machine learning methods. Here is a brief overview of some of the
earlier proposed solutions:

1. Fundus Image Analysis: Traditional methods focused on the manual analysis of

fundus images, where trained ophthalmologists identified features such as
microaneurysms, exudates, and hemorrhages. While effective, this method is subjective
and time-consuming, leading to inconsistencies in diagnosis.

2. Image Processing Techniques: Techniques such as edge detection, morphological

operations, and histogram equalization were employed to enhance the visibility of
retinal features. These methods aimed to preprocess the images before manual or
automated classification but often struggled with noise and variability in image quality.

3. Machine Learning Approaches:Earlier machine learning models, such as support

vector machines (SVM) and k-nearest neighbors (KNN), utilized hand-crafted features
extracted from retinal images. These features included texture, shape, and color
information. Although some models achieved moderate accuracy, they relied heavily
on feature engineering and often required extensive domain knowledge.

25
 Figure 8. Machine learning methods.
4. Deep Learning Architectures:The introduction of convolutional neural networks
(CNNs) marked a significant advancement in DR detection. Models like AlexNet and
VGG16 demonstrated the ability to automatically learn relevant features from raw
image data, achieving higher accuracy than traditional methods. Researchers began to
leverage transfer learning techniques, utilizing pre-trained models on large datasets to
fine-tune them for DR classification.

Figure 9. DR recognition and classification using deep learning.

26
 5. Hybrid Approaches: Some solutions combined traditional image processing with
machine learning and deep learning methods. For instance, preprocessing steps to
enhance image quality were integrated with CNNs for classification, improving overall
model robustness.

6. Ensemble Learning: Ensemble techniques, which combine predictions from multiple

models, have been explored to enhance classification performance. By aggregating
results from different algorithms, these approaches aimed to reduce overfitting and
improve accuracy.

7. Public Datasets: The establishment of publicly available datasets, such as the Messidor
and EyePACS databases, has facilitated the training and evaluation of various
algorithms. These datasets provide annotated images for benchmarking different
methodologies and advancing research in DR detection.

8. Real-Time Detection Systems: Recent developments have focused on creating real-

time detection systems that can operate in clinical settings. These systems integrate
mobile applications and cloud-based solutions to enable quick screening and referral
processes.

Despite these advancements, challenges remain in achieving high accuracy, particularly in

classifying mild forms of diabetic retinopathy and addressing variations in image quality. The
field continues to evolve, with ongoing research aimed at improving the reliability and
accessibility of diabetic retinopathy detection tools.

2.4. Bibliometric Analysis:

The detection and classification of diabetic retinopathy (DR) using colored images has evolved
significantly over the years. Each solution presents unique features, effectiveness levels, and
inherent drawbacks. Here’s a detailed analysis:

1. Traditional Fundus Image Analysis

• Key Features:

o Relies on trained ophthalmologists for feature identification (e.g., microaneurysms,

exudates).

o Subjective interpretation of retinal images.

27
• Effectiveness:

o Accurate when performed by experienced clinicians.

o Provides comprehensive insights based on clinical experience.

• Drawbacks:

o Time-consuming and labor-intensive.

o High variability due to subjective interpretation, leading to potential misdiagnoses.

2. Image Processing Techniques

• Key Features:

o Utilizes methods like edge detection, contrast enhancement, and morphological

operations to preprocess images.

o Aims to highlight features relevant for DR detection.

• Effectiveness:

o Can improve image quality and highlight pathological features.

o Useful as a preprocessing step for subsequent analyses.

• Drawbacks:

o Often struggles with noise and artifacts, which can hinder feature extraction.

o Lacks adaptability to different image quality and variations.

3. Machine Learning Approaches

• Key Features:

o Employs algorithms like SVM, KNN, and decision trees based on hand-crafted
features.

o Requires significant feature engineering and domain knowledge.

• Effectiveness:

o Achieved moderate accuracy in specific contexts with well-defined features.

o Can be more efficient than manual methods in large-scale screenings.

28
• Drawbacks:

o Performance heavily reliant on the quality of the extracted features.

o Limited ability to generalize across diverse datasets and image conditions.

4. Deep Learning Architectures

• Key Features:

o Utilizes CNNs for automatic feature learning from raw images.

o Models like AlexNet and VGG16 offer robust classification capabilities.

• Effectiveness:

o High accuracy and robustness, outperforming traditional methods.

o Capable of detecting subtle patterns and features that may be missed by human
experts.

• Drawbacks:

o Requires large labeled datasets for training, which may not always be available.

o Computationally intensive, necessitating high-performance hardware for training

and inference.

5. Hybrid Approaches

• Key Features:

o Combines traditional image processing techniques with machine learning or deep

learning.

o Aims to enhance preprocessing steps to improve model performance.

• Effectiveness:

o Improved robustness and accuracy by leveraging the strengths of different methods.

• Drawbacks:

o Complexity in implementation and tuning of multiple techniques.

o Potential for increased computational load due to multiple processing stages.

29
6. Ensemble Learning

• Key Features:

o Aggregates predictions from multiple models to improve classification accuracy.

o Uses techniques like bagging, boosting, or stacking.

• Effectiveness:

o Enhanced accuracy and reduced variance, leading to better generalization.

• Drawbacks:

o Increased complexity in model management and training.

o Higher computational requirements for training multiple models.

7. Public Datasets

• Key Features:

o Provides annotated images for benchmarking and validation.

o Facilitates training and evaluation of algorithms.

• Effectiveness:

o Supports collaborative research and development efforts across institutions.

• Drawbacks:

o Variability in quality and representation of DR stages in datasets.

o Limited diversity in some datasets, which can affect model generalization.

8. Real-Time Detection Systems

• Key Features:

o Integrates AI models into mobile or web applications for on-site screenings.

o Aims to facilitate quick diagnosis and referrals.

30
• Effectiveness:

o Potential for increased accessibility and convenience in clinical settings.

• Drawbacks:

o Requires reliable internet connectivity and hardware for implementation.

o Challenges in ensuring the accuracy and reliability of automated systems in diverse

environments.

Solution Type Key Features Effectiveness Drawbacks

Traditional - Relies on trained - Accurate with - Time-consuming.
Fundus Image ophthalmologists. experienced - High variability
Analysis - Subjective clinicians. leading to potential
interpretation of images. - Comprehensive misdiagnoses.
clinical insights.
Image - Utilizes edge detection, - Improves image - Struggles with noise
Processing contrast enhancement, quality. and artifacts.
Techniques and morphological - Highlights - Lacks adaptability
operations. relevant features. to varying image
- Aims to preprocess quality.
images.
Machine - Employs algorithms - Moderate - Performance relies
Learning like SVM, KNN, and accuracy in on quality of features.
Approaches decision trees. defined contexts. - Limited
- Requires feature - Efficient for generalization across
engineering. large-scale diverse datasets.
screenings.
Deep Learning - Utilizes CNNs for - High accuracy - Requires large
Architectures automatic feature and robustness. labeled datasets.
learning. - Detects subtle - Computationally
- Models like AlexNet patterns. intensive.
and VGG16.
Hybrid - Combines traditional - Improved - Increased
Approaches methods with robustness and complexity in
machine/deep learning. accuracy. implementation.
- Enhances preprocessing - Higher
steps. computational load.
Ensemble - Aggregates predictions - Enhanced - Complexity in
Learning from multiple models. accuracy and model management.
reduced variance. - Higher

31
 - Uses techniques like computational
bagging and boosting. requirements.
Public Datasets - Provides annotated - Supports - Variability in
images for collaborative dataset quality.
benchmarking. research. - Limited diversity
- Facilitates training and - Aids in algorithm affecting model
evaluation. evaluation. generalization.
Real-Time - Integrates AI models - Increased - Requires reliable
Detection into applications for on- accessibility in internet connectivity.
Systems site screenings. clinical settings. - Challenges in
- Facilitates quick ensuring accuracy in
diagnosis. diverse environments.

Table 2: Key Fratures, Drawbacks and Effectivensess of Exesting Solutions

2.5. Review Summary:

The literature review on diabetic retinopathy (DR) classification and detection reveals several
important insights that directly inform and enhance the current project on "Diabetic
Retinopathy Classification and Detection Using Coloured Images with Deep Learning Inspired
Framework." The findings from previous studies underline the significance of employing
advanced techniques to improve diagnostic accuracy and efficiency in DR detection.

1. Advancements in Deep Learning: The literature indicates a marked shift towards using
deep learning architectures, particularly Convolutional Neural Networks (CNNs), which
have shown superior performance in image classification tasks, including DR detection.
This aligns with the project's objective to utilize deep learning frameworks to automatically
classify and detect DR from colored fundus images, ensuring high accuracy and robustness.

2. Effectiveness of Automated Approaches: Previous studies highlight the effectiveness of

automated image analysis methods compared to traditional manual assessments, which are
time-consuming and prone to human error. By integrating automated algorithms in the
current project, we aim to streamline the diagnosis process, making it faster and more
reliable for clinicians.

3. Need for Robust Feature Extraction: The findings emphasize the importance of effective
feature extraction methods for improving classification results. The project will focus on
leveraging advanced feature extraction techniques in conjunction with deep learning to

32
 capture intricate patterns associated with DR, thus enhancing the model's predictive
capabilities.

4. Challenges with Dataset Diversity: Many studies noted the challenges associated with the
diversity and quality of datasets used for training models. This project will address this
issue by utilizing a comprehensive dataset that includes varied images representing
different stages of DR, ensuring that the developed model is generalizable across different
populations.

5. Hybrid and Ensemble Methods: The review highlights the potential of hybrid and
ensemble approaches in improving classification accuracy by combining multiple models.
The current project will explore these methodologies to enhance overall performance and
reliability, thereby minimizing false positives and negatives in DR detection.

6. Real-time Implementation: Several studies discussed the need for real-time detection
systems to facilitate timely interventions in clinical settings. This project aims to develop a
user-friendly interface that enables real-time analysis of fundus images, providing
clinicians with quick and actionable insights.

Finding Description Link to Project

Evolution of Deep Learning Deep learning models, Utilize state-of-the-art CNN

Techniques particularly CNNs, show architectures to enhance
substantial improvements in classification accuracy and
image classification accuracy. detect subtle DR variations.
Automated Image Analysis Automated methods Implement automated image
for Enhanced Efficiency outperform traditional manual analysis algorithms to facilitate
assessments, reducing timely diagnosis and
diagnostic time and human intervention for DR.
error.
Importance of Robust Advanced feature extraction Focus on advanced feature
Feature Extraction methods significantly boost extraction techniques, such as
model performance in transfer learning, to capture
classification tasks. complex patterns in DR.

33
 Diversity and Quality of A diverse training dataset Use a well-curated and diverse
Training Datasets enhances generalizability and dataset to ensure effective
performance across different performance across various
populations and settings. patient demographics.
Hybrid and Ensemble Hybrid and ensemble Explore ensemble methods to
Learning Approaches techniques combining multiple integrate predictions from
algorithms achieve superior multiple models, reducing
classification performance. classification errors.
Real-time Implementation The need for real-time Develop a user-friendly
for Clinical Utility diagnostic tools that facilitate interface for real-time analysis
immediate clinical decision- of fundus images to enable
making is emphasized. prompt treatment decisions.
Integration of Clinical and Integrating clinical and Consider including clinical and
Demographic Data demographic data with demographic factors in
imaging information enhances predictive modeling to tailor
prediction accuracy and treatment strategies.
personalizes patient
management.

Table 2: Key findings from the literature review and how they connect to
your project objectives.

2.6.Problem Definition:

Background: Diabetic retinopathy (DR) is a major complication of diabetes and a leading

cause of blindness among adults globally. It results from damage to the blood vessels of the
retina due to prolonged high blood sugar levels, leading to vision impairment and potential loss
if not diagnosed and treated timely (Yau et al., 2012; Wu et al., 2015). According to the World
Health Organization (WHO), the prevalence of DR is increasing with the rise in diabetes cases
worldwide, highlighting an urgent need for effective screening and diagnosis (WHO, 2016).

Current Challenges: The conventional methods for diagnosing DR involve manual

examination of fundus images by trained ophthalmologists. This process is not only time-
consuming but also subject to human error and variability in interpretation (Khan et al., 2019).

34
Furthermore, many individuals, especially in low-resource settings, lack access to regular eye
examinations, leading to late diagnoses when treatment is less effective (Tufail et al., 2015).

Objectives of the Project: To address these challenges, the project aims to develop an AI-
driven system for the automatic classification and detection of diabetic retinopathy in coloured
fundus images. The specific objectives are:

• What to do:
o Develop a Deep Learning Model: Create a deep learning-based system that
can automatically classify fundus images into various stages of diabetic
retinopathy (mild, moderate, severe, and proliferative). This will facilitate early
diagnosis and treatment, potentially reducing the incidence of vision loss.
• How to do it:
o Utilize Convolutional Neural Networks (CNNs): Employ CNN architectures
such as ResNet or Inception, which have shown remarkable performance in
image classification tasks. Transfer learning techniques will be applied to
leverage pre-trained models on large datasets, thus improving classification
accuracy with fewer training data.
o Dataset Acquisition: Use publicly available datasets such as the EyePACS
dataset or the Kaggle Diabetic Retinopathy Detection dataset, which provide a
wide range of labelled fundus images for training and validation.
o Performance Evaluation: Implement rigorous testing and validation
strategies, measuring performance using metrics like accuracy, sensitivity,
specificity, and area under the receiver operating characteristic (ROC) curve
(AUC-ROC).
• What not to do:
o Avoid Complexity: The project will not develop overly complex systems that
necessitate extensive technical expertise or costly infrastructure. The focus will
be on creating a user-friendly application that healthcare providers can easily
use without requiring advanced training or expensive diagnostic equipment.
o Prevent Misinterpretation: The model's results should not be used as the sole
diagnostic tool; it will complement, not replace, the expertise of healthcare
professionals. Clear guidelines will be established for interpreting model
outputs in clinical settings.

35
2.7.Goal / Objectives:

The primary goal of this project is to develop an effective deep learning framework for the
classification and detection of Diabetic Retinopathy (DR) using colored retinal images The
proposed work has been abandoned in order to provide an advanced computational strategy that
expands computer-based methods, determines the accuracy of survival prediction and detection,
and expands the practise of computer segmentation. The key
• Improving the accuracy of the categorization system for DR by recognising and using a
pattern.
• identification system
• Create a system that can identify between various DR kinds according to their grades.
• To extend further, practise segmentation using the feature extraction and selection
method.
• To increase the level of survival prediction, develop a better detecting method accuracy

36
 REFRENCES
1. S. Lu, J. Zhu, and H. Zhang, "A deep learning approach for diabetic retinopathy
classification using retinal images," Journal of Healthcare Engineering, vol. 2020, Article
ID 4931646, 2020. doi: 10.1155/2020/4931646.
2. K. R. Bhaskaranand, K. Ramachandraiah, and P. R. Bhat, "Automated grading of diabetic
retinopathy severity from fundus photographs," Diabetes Care, vol. 39, no. 12, pp. 2434–
2440, 2016. doi: 10.2337/dc16-0017.
3. W. Sun, Y. Zheng, and J. Li, "Automated diabetic retinopathy detection based on binocular
Siamese-like convolutional neural network," IEEE Access, vol. 7, pp. 30744–30753, 2019.
doi: 10.1109/ACCESS.2019.2902551.
4. T. T. Nguyen, S. Rana, and Y. H. Le, "Diabetic retinopathy classification using DNNs and
fundus images," Computers in Biology and Medicine, vol. 114, Article ID 103464, 2019.
doi: 10.1016/j.compbiomed.2019.103464.
5. S. Wang, J. Yin, and Z. Wang, "Detection and classification of diabetic retinopathy using
ensemble learning and deep learning," Biocybernetics and Biomedical Engineering, vol.
40, no. 3, pp. 977–990, 2020. doi: 10.1016/j.bbe.2020.06.006.
6. R. A. Chetoui, B. Khamis, and M. Bourennane, "Diabetic retinopathy detection through
deep feature learning with fuzzy decision," Computer Methods and Programs in
Biomedicine, vol. 182, Article ID 105049, 2019. doi: 10.1016/j.cmpb.2019.105049.
7. Z. Liu, Y. Wang, and J. Shi, "A novel method for diabetic retinopathy diagnosis based on
deep learning neural networks," Journal of Ambient Intelligence and Humanized
Computing, vol. 12, pp. 3059–3068, 2021. doi: 10.1007/s12652-020-02228-w.
8. R. Priya and K. Aruna, "Diabetic retinopathy detection and classification using artificial
neural network," Biomedical Research, vol. 28, no. 1, pp. 257–263, 2017.
9. A. Goatman, E. Charnley, and S. Webster, "Evaluation of artificial intelligence-based
diabetic retinopathy screening," British Journal of Ophthalmology, vol. 101, pp. 1377–
1383, 2017. doi: 10.1136/bjophthalmol-2016-309730.
10. P. Kamble, and A. Joshi, "Deep learning-based method for early detection and
classification of diabetic retinopathy," Biomedical Signal Processing and Control, vol. 47,
pp. 303–311, 2019. doi: 10.1016/j.bspc.2018.09.012.
11. A. Das, and S. Prabhu, "Automated segmentation and classification of diabetic retinopathy
lesions," International Journal of Medical Informatics, vol. 118, pp. 40–53, 2018. doi:
10.1016/j.ijmedinf.2018.08.007.
12. J. A. Abràmoff, P. T. Lavin, and M. Birch, "Pivotal trial of an autonomous AI-based
diagnostic system for detection of diabetic retinopathy in primary care," NPJ Digital
Medicine, vol. 1, Article 39, 2018. doi: 10.1038/s41746-018-0040-6.
13. E. Tekin, S. Kocaman, and G. Ozcan, "Diabetic retinopathy classification and severity
prediction using convolutional neural networks," IEEE Transactions on Biomedical
Engineering, vol. 67, no. 5, pp. 1370–1381, 2020. doi: 10.1109/TBME.2019.2956794.
14. S. A. Soliman, "Efficient automated detection of diabetic retinopathy using deep
convolutional neural networks," Computers in Biology and Medicine, vol. 122, Article ID
103832, 2020. doi: 10.1016/j.compbiomed.2020.103832.
15. M. A. Brancati, F. La Rosa, and S. Rizzo, "Diabetic retinopathy classification in retinal
images using deep neural networks," IEEE Access, vol. 8, pp. 29227–29236, 2020. doi:
10.1109/ACCESS.2020.2971890.
16. L. Gadekallu, S. Reddy, and M. Ramachandran, "A novel approach for diabetic retinopathy
classification using CNN and transfer learning," Information Sciences, vol. 541, pp. 254–
272, 2020. doi: 10.1016/j.ins.2020.08.070.

37
17. R. K. Das, A. Dey, and D. Ghosh, "Transfer learning-based approach for diabetic
retinopathy classification in retinal images," Signal, Image and Video Processing, vol. 14,
pp. 343–351, 2020. doi: 10.1007/s11760-019-01600-2.
18. V. Chandran, S. Thomas, and T. T. Nguyen, "Ensemble learning-based diabetic retinopathy
classification using fundus images," Applied Soft Computing, vol. 85, Article ID 105833,
2019. doi: 10.1016/j.asoc.2019.105833.
19. S. Haleem, D. Hanif, and I. Lee, "A hybrid approach for diabetic retinopathy detection
using CNN and fuzzy logic," Journal of Visual Communication and Image Representation,
vol. 71, Article ID 102778, 2020. doi: 10.1016/j.jvcir.2020.102778.
20. M. Mirsharif, N. Ghaffari, and M. Nikravesh, "Automated detection of diabetic retinopathy
using image processing and deep learning," Machine Vision and Applications, vol. 32,
Article 90, 2021. doi: 10.1007/s00138-021-01198-w.
21. Yau, J. W., Rogers, S. L., Kawasaki, R., et al. (2012). Global prevalence and major risk
factors of diabetic retinopathy. Diabetes Care, 35(3), 556-564.
22. Wu, L., Huang, X., & Zhang, W. (2015). The effect of diabetic retinopathy on quality of
life in diabetes patients: a meta-analysis. Diabetes Research and Clinical Practice, 109(3),
634-642.
23. WHO (World Health Organization). (2016). Global report on diabetes. Retrieved from
WHO website
24. Khan, M. A., Mukhtar, S., & Hayat, A. (2019). A survey of techniques and methods for
diabetic retinopathy detection. International Journal of Computer Applications, 182(25),
11-18.
25. Tufail, A., Gregor, Z. J., & Akhtar, T. (2015). Diabetic retinopathy screening: the evidence
and the challenges. British Journal of Ophthalmology, 99(6), 731-735.

38