Automated Detection of Diabetic

Retinopathy Using Deep Learning

A Project Work Synopsis

Submitted in the partial fulfillment for the award of the degree of

BACHELOR OF ENGINEERING
IN

COMPUTER SCIENCE WITH SPECIALIZATION IN

ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING

Submitted by:
UID NAME
21BCS9622 HARSH HARSHIT
21BCS9479 RAHUL YADAV
21BCS8816 BRIJESH KISHORE PUROHIT
21BCS9076 MRINAL BHATT

Under the Supervision of: Rosevir Singh

CHANDIGARH UNIVERSITY, GHARUAN, MOHALI - 140413,

PUNJAB
September, 2024

1
 Abstract

Diabetic Retinopathy (DR) is a leading cause of vision impairment and blindness, particularly in
individuals with diabetes. Early detection and timely treatment are critical to prevent severe visual loss.
Traditional diagnostic methods, which rely on manual interpretation of retinal images by
ophthalmologists, are time-consuming, subject to human error, and often inaccessible to underserved
populations. To address these challenges, this paper presents an automated system for detecting diabetic
retinopathy using deep learning techniques. The proposed system leverages convolutional neural
networks (CNNs) trained on large datasets of labeled retinal fundus images to accurately identify various
stages of DR, from mild to proliferative. The deep learning model automatically classifies the presence
and severity of the disease, offering high sensitivity and specificity. This automation promises to
enhance screening efficiency, reduce diagnostic costs, and increase access to early-stage DR detection in
remote and resource-limited areas. Experimental results demonstrate that the proposed deep learning
approach achieves competitive performance compared to expert ophthalmologists, highlighting its
potential for clinical deployment in real-world healthcare settings.
Diabetic Retinopathy (DR) is a leading cause of vision impairment and blindness, particularly in
individuals with diabetes. Early detection and timely treatment are critical to prevent severe visual loss.
Traditional diagnostic methods, which rely on manual interpretation of retinal images by
ophthalmologists, are time-consuming, subject to human error, and often inaccessible to underserved
populations. To address these challenges, this paper presents an automated system for detecting diabetic
retinopathy using deep learning techniques.

Keywords:
Automated Detection of Diabetic Retinopathy Using Deep Learning, Convolutional Neural Networks
(CNNs)

2
 Table of Contents

Title Page i

Abstract ii

1. Introduction

1.1 Problem Definition

1.2 Project Overview

1.3 Hardware Specification

1.4 Software Specification

2. Literature Survey

2.1 Existing System

2.2 Proposed System

2.3 Literature Review Summary

3. Problem Formulation

4. Research Objective

5. Methodologies

6. Experimental Setup

7. Conclusion

8. Tentative Chapter Plan for the proposed work

9. Reference

3
1. INTRODUCTION

Diabetic Retinopathy (DR) is a serious complication of diabetes that affects the blood
vessels of the retina, leading to progressive vision loss and, in severe cases, blindness. It is
one of the most prevalent causes of blindness worldwide, particularly in working-age adults.
With the global rise in diabetes cases, the need for effective and accessible screening
methods for early detection of DR has become increasingly urgent. Early diagnosis is
critical, as timely treatment can significantly reduce the risk of vision loss. However, current
screening practices, which rely on manual analysis of retinal fundus images by
ophthalmologists, are resource-intensive, prone to subjectivity, and often unavailable in rural
or underdeveloped regions.

To address these limitations, this project aims to develop an automated detection system for
diabetic retinopathy using deep learning techniques. By leveraging convolutional neural
networks (CNNs), a type of deep learning model, the system is capable of learning complex
patterns from retinal images and accurately identifying signs of DR. Deep learning models,
particularly CNNs, have demonstrated remarkable success in image recognition tasks,
making them well-suited for medical image analysis. This project will focus on designing,
training, and validating a deep learning model to classify DR stages based on large datasets
of retinal fundus images.

The primary objective of this project is to create a cost-effective, efficient, and scalable
solution for DR detection, enabling early diagnosis in both clinical settings and remote areas
with limited access to specialist care. By automating the screening process, the system is
expected to enhance diagnostic accuracy, reduce the burden on healthcare providers, and
ultimately improve patient outcomes.

Early diagnosis is critical, as timely treatment can significantly reduce the risk of vision loss.
However, current screening practices, which rely on manual analysis of retinal fundus
images by ophthalmologists, are resource-intensive, prone to subjectivity, and often
unavailable in rural or underdeveloped regions.

4
 To address these limitations, this project aims to develop an automated detection system for
diabetic retinopathy using deep learning techniques. By leveraging convolutional neural
networks (CNNs), a type of deep learning model, the system is capable of learning complex
patterns from retinal images and accurately identifying signs of DR. Deep learning models,
particularly CNNs, have demonstrated remarkable success in image recognition tasks,
making them well-suited for medical image analysis. This project will focus on designing,
training, and validating a deep learning model to classify DR stages based on large datasets
of retinal fundus images.

1.1 Problem Definition

Handwritten Digit Recognition poses a formidable challenge in the realm of computer

vision, as the diversity in individual writing styles introduces complexities in accurately
interpreting handwritten numerical characters. This project addresses the need for a robust
solution that can effectively discern and classify handwritten digits, thereby enabling
applications such as automated document processing and optical character recognition.

The variability in writing styles leads to challenges in generalization across different

datasets, making it imperative to develop a solution that can adapt to diverse handwriting
patterns. Traditional methods often struggle to handle these variations, necessitating the
exploration of advanced machine learning techniques. By utilizing Convolutional Neural
Networks (CNNs), Support Vector Machines (SVMs), and a mix of datasets, including the
foundational MNIST Dataset, this project aims to provide a comprehensive solution to the
Handwritten Digit Recognition problem.

5
1.2 Problem Overview

Handwritten Digit Recognition involves the development of a system capable of accurately

Diabetic Retinopathy (DR) is a common and potentially blinding complication of diabetes
that occurs when high blood sugar levels damage the retinal blood vessels. If left
undiagnosed or untreated, DR can lead to irreversible vision loss. Early detection and
treatment are critical to prevent severe visual impairment, but the current methods of
diagnosis face several challenges:
1. Manual Diagnosis and Resource Constraints: Traditional DR screening involves the
manual examination of retinal fundus images by ophthalmologists. This process is time-
consuming, requires highly trained specialists, and is subject to human error and variability in
diagnosis. Additionally, there is a shortage of trained professionals, particularly in remote and
underdeveloped regions, leading to delays in diagnosis and treatment.
2. High Prevalence of Diabetes: With the increasing prevalence of diabetes worldwide, there is
a corresponding rise in the number of DR cases. Healthcare systems, especially in low-
resource settings, are struggling to meet the growing demand for regular screenings and early-
stage detection of DR. This leads to many patients not receiving timely care, increasing their
risk of vision loss.

Given these challenges, the problem this project seeks to address is the development of an
automated system for the early detection and classification of diabetic retinopathy using deep
learning. The system must be capable of accurately identifying DR from retinal images, be
scalable for mass screening, and provide an affordable solution that can be implemented in
both urban and rural healthcare settings.

1.3 Hardware Specification

6
The project is designed to be implemented on a standard laptop or desktop computer.
However, due to the computational demands associated with machine learning tasks,
especially during the training and testing phases, it is advisable to consider more powerful
hardware configurations. For optimal performance, it is recommended to use a computer

equipped with a dedicated Graphics Processing Unit (GPU), higher Random Access
Memory

7
1.4 Software Specification

The software environment is crucial for developing, training, and deploying machine
learning models for Automated Detection of Diabetic Retinopathy Using Deep Learning.
The following software specifications are recommended:

1. Operating System: Linux-based operating systems like Ubuntu or CentOS are

commonly used for machine learning tasks. Windows or macOS can also be suitable,
depending on user preferences.
2. Python and Package Management: Python 3.x is the preferred programming language
for machine learning. Use package managers like Anaconda or pip to install and
manage Python libraries.
3. Deep Learning Frameworks: TensorFlow and PyTorch are popular frameworks for
developing neural networks. Keras, a high-level API, can be integrated with
TensorFlow for ease of use.

4. Machine Learning Libraries: Scikit-learn for traditional machine learning algorithms.

OpenCV for image preprocessing tasks. NumPy and Pandas for data manipulation and
analysis.
5. Development Environment: Integrated Development Environment (IDE) such as Jupyter
Notebooks or Visual Studio Code for code development and experimentation.
6. Version Control: Git for version control and collaboration with team members.
7. Additional Tools: Docker for containerization to ensure consistent environments across
different systems. Virtual environments for isolating project dependencies.

8
LITERATURE SURVEY

1.4 Existing System

The landscape of Handwritten Digit Recognition has witnessed significant advancements,
Diabetic Retinopathy (DR) has become a significant focus in medical image analysis
research, driven by its increasing prevalence and the need for early detection to prevent
vision loss. Several traditional methods, as well as recent developments using artificial
intelligence (AI), particularly deep learning, have emerged for the automated detection of
DR. The literature survey below summarizes the evolution of these systems and highlights
the transition from traditional techniques to modern deep learning approaches.

Existing System

1. Manual Diagnosis by Ophthalmologists

Traditionally, DR screening and diagnosis are performed manually by ophthalmologists who

visually inspect retinal fundus images for signs of DR, such as microaneurysms,
hemorrhages, and neovascularization. The key challenges associated with manual screening
include:

 Time-Consuming: Each image must be examined carefully, making it impractical for large-scale
screenings.

 Subjectivity: Diagnosis can vary between specialists due to differences in experience and
judgment.

 Limited Accessibility: In many remote and underdeveloped regions, the availability of trained
ophthalmologists is limited, leading to delayed diagnoses and treatment.

2. Image Processing and Machine Learning-Based Approaches

Prior to the rise of deep learning, several automated systems relied on classical image

9
 processing and machine learning techniques. These systems often involved:

 Feature Extraction: Specific features such as texture, color, and shape were manually extracted
from retinal images using image processing techniques.

 Classification: Traditional machine learning classifiers such as Support Vector Machines

(SVM), k-Nearest Neighbors (k-NN), or Random Forests were used to classify images based on
the extracted features.

Limitations of these early approaches:

 Feature Engineering: These methods heavily relied on handcrafted features, which required
domain expertise and often missed subtle patterns in the data.

 Limited Accuracy: The performance of these systems was often suboptimal compared to human
experts, with relatively high false-positive and false-negative rates.

 Scalability: These approaches struggled to scale to large datasets and were sensitive to variations
in image quality and lighting conditions.

3. Shallow Neural Networks

Before the dominance of deep learning, shallow neural networks were used for DR
detection. While these systems showed some promise, they were limited by their inability to
automatically learn complex hierarchical features from images. They required careful
feature selection and tuning, which restricted their flexibility.

Examples of Existing Systems:

 Messidor Database (2004): An early dataset for DR detection that was used to test various
machine learning algorithms, focusing on feature extraction methods.

10
 DiaretDB1 Database: Used in studies to develop automated systems based on image processing
and machine learning, achieving moderate success but limited generalization to real-world
settings.

1.5 Proposed System

Recent advancements in Handwritten Digit Recognition underscore a paradigm shift

towards Deep learning, particularly Convolutional Neural Networks (CNNs), has
revolutionized medical image analysis, offering superior performance in automated DR
detection. The proposed system aims to overcome the limitations of traditional methods by
leveraging deep learning for end-to-end detection and classification of DR from retinal
images.
1. Convolutional Neural Networks (CNNs) for DR Detection
CNNs are a class of deep learning models specifically designed for image data. They have
shown exceptional performance in identifying complex patterns and features from images
without the need for manual feature extraction. In the context of DR detection, CNNs
automatically learn to identify key indicators of the disease, such as microaneurysms,
hemorrhages, and exudates, directly from raw retinal images.
Advantages of CNN-based systems:
 Automatic Feature Extraction: CNNs learn hierarchical features directly from data, eliminating
the need for manual feature engineering.
 High Accuracy: Deep learning models have outperformed traditional machine learning methods
and sometimes even human experts in identifying early signs of DR.
 Scalability: CNN-based systems can scale to large datasets and adapt to variations in image
quality, lighting, and patient demographics.
 Generalization: Well-trained CNNs generalize well across diverse populations, making them
suitable for global deployment.
2. Transfer Learning
Given the availability of large pre-trained models like ResNet, VGG, and Inception,
11
 transfer learning can be used to fine-tune models specifically for DR detection. This
significantly reduces the training time and improves performance, especially when the
labeled dataset is relatively small.
3. Multi-Stage Classification
The proposed system will classify DR into multiple stages:
 No DR: No visible signs of diabetic retinopathy.
 Mild DR: Early signs such as microaneurysms.
 Moderate DR: Increased damage, including hemorrhages.
 Severe DR: Significant damage with risk of progression.
 Proliferative DR: Advanced stage with neovascularization, posing a high risk of vision loss.

1.6 Literature Review Summary

.
Year Citation (Article/Author) Technique Source Evaluation
Parameter

2004 Messidor Database Traditional Machine Public Dataset Sensitivity,

Learning (SVM, K-NN) Specificity

2010 DiaretDB1 Database Image Processing + Feature Public Dataset Accuracy,

Extraction Specificity

2015 Kaggle Diabetic Convolutional Neural Kaggle Dataset Accuracy, AUC, F1-
Retinopathy Challenge Networks (CNN) score

2016 Google AI's DR Screening Deep Learning (CNN) Internal Dataset Sensitivity,
System Specificity

2018 Gulshan et al. CNN with Transfer Learning Google AI/Retinal AUC, Sensitivity,
Fundus Images Specificity

2019 Lam et al. Hybrid Deep Learning + Private Dataset Accuracy, Precision
Image Processing

2020 Zago et al. CNN with Data Public Dataset (IDRiD) Sensitivity, F1-
Augmentation score

2021 Nayak et al. Ensemble Deep Learning APTOS Dataset AUC, Sensitivity,
Models Accuracy

12
 2022 Wang et al. Attention-based CNN Public Dataset Precision, Recall
Models (Messidor)

2023 Raj et al. Deep Learning with Retinal Fundus Image Accuracy,
Preprocessing Dataset Specificity

This table summarizes the progression of techniques from traditional machine learning approaches
to deep learning models, particularly convolutional neural networks (CNNs), highlighting
improvements in accuracy and scalability over time. Each entry provides a snapshot of the
techniques, datasets, and evaluation metrics used in the research.

13
 2. PROBLEM FORMULATION

Diabetic Retinopathy (DR) is a leading cause of vision loss among individuals with diabetes
and, if undiagnosed or untreated, can lead to permanent blindness. The timely detection of DR is
essential for preventing its progression, but traditional screening methods are costly, time-
consuming, and depend on the availability of trained ophthalmologists. The increasing global
prevalence of diabetes, particularly in low-resource and rural areas, presents a critical challenge in
diagnosing DR early enough for effective treatment. There is an urgent need for an automated,
scalable, and accurate method for DR screening that can be deployed widely and effectively.

Problem Statement

The goal of this project is to design and implement an automated detection system for
Diabetic Retinopathy using deep learning, specifically Convolutional Neural Networks
(CNNs). The system must be able to analyze retinal fundus images, detect the presence of
DR, and classify the disease into various stages. The model must achieve high accuracy,
sensitivity, and specificity, and it should be adaptable to diverse datasets to ensure robust
real-world performance.

Key Objectives

1. Automated Detection: To develop a system capable of automatically identifying the

presence of Diabetic Retinopathy from retinal fundus images without human
intervention.
2. Stage Classification: To classify DR into various stages (No DR, Mild, Moderate,
Severe, and Proliferative DR) based on the severity of retinal damage.
3. Performance Optimization: To ensure that the deep learning model achieves high
accuracy, sensitivity, and specificity, comparable to human experts.
4. Scalability: To design a scalable solution that can be deployed in large-scale
screening programs, including in low-resource environments.
5. Real-Time Deployment: To integrate the model into telemedicine platforms,
enabling real-time screening and diagnosis in remote areas with limited access to
ophthalmologists.

Challenges

1. Data Quality and Diversity: Retinal fundus images vary in quality, lighting, and
resolution, which can impact the model’s ability to generalize across datasets.
2. Class Imbalance: The prevalence of advanced DR stages (Severe and Proliferative)
is lower compared to earlier stages, leading to a class imbalance in the dataset.
3. Interpretability: Medical professionals need interpretable models to understand how
predictions are made, particularly for high-stakes decisions like disease diagnosis.
 4. Deployment in Resource-Constrained Areas: Hardware limitations in remote or
rural areas may require the system to be optimized for deployment on edge devices or
cloud platforms.

Expected Outcomes

1. Accurate DR Detection: The system should detect DR with a high level of precision
and correctly classify its stages.
2. Improved Accessibility: The automated system should make DR screening
accessible in remote and underserved regions where manual screening is impractical.
3. Scalable Solution: The system should be scalable to accommodate large populations
and integrate into existing healthcare infrastructure.
4. Reduced Workload on Healthcare Professionals: By automating the initial
screening process, the system should reduce the burden on ophthalmologists,
allowing them to focus on cases requiring detailed review.
 3. OBJECTIVES

Develop a Robust Handwritten Digit Recognition Model: Implement a machine:

 Develop an Automated Detection Model: Create a deep learning model, specifically a

Convolutional Neural Network (CNN), capable of accurately detecting the presence of Diabetic
Retinopathy in retinal fundus images.

 Classify DR Stages: Implement a multi-class classification system to categorize Diabetic

Retinopathy into distinct stages: No DR, Mild DR, Moderate DR, Severe DR, and Proliferative
DR.

 Optimize Model Performance: Achieve high accuracy, sensitivity, and specificity in the
model’s predictions, ensuring that it performs comparably to human ophthalmologists in
diagnosing DR.

 Enhance Scalability and Accessibility: Design the system to be scalable for large-scale
screening programs, making it accessible for use in diverse healthcare settings, including rural and
underserved areas.

 Integrate with Telemedicine Platforms: Develop a framework for integrating the automated
detection system into telemedicine applications, enabling real-time analysis and diagnosis of retinal
images in remote locations.

1.
 4. METHODOLOGY

Data Preprocessing
• Image Resizing Resize all images to a consistent resolution (e.g., 512x512 pixels) to ensure
uniformity and compatibility with the neural network.
• Grayscale Conversion: Convert the images to grayscale to reduce computational complexity while
retaining critical diagnostic features.
• Noise Reduction: Apply filters such as Gaussian blur or median filters to remove noise and
improve image clarity.

Labeling
• Assign appropriate labels to the retinal images to indicate the severity of diabetic retinopathy.
Common classes include: Healthy retina without signs of the disease.
• Retinas showing varying levels of disease severity, based on lesion type and number. Advanced
stage with abnormal blood vessel growth.
• Ensure that the labeling is consistent with clinical guidelines to maintain accuracy in the training
and evaluation phases.
 5. EXPERIMENTAL SETUP

The experimental setup involves several key components, including data acquisition, preprocessing, model
development, training, evaluation, and deployment. Below is a detailed outline of each stage in the
experimental setup:
1. Data Acquisition
 Dataset Selection: Choose relevant public datasets, such as:
o Kaggle Diabetic Retinopathy Detection Dataset: Contains thousands of labeled retinal
images classified into various stages of DR.
o Messidor Database: A widely used dataset for DR research.
o IDRiD Dataset: Includes images with various stages of DR and corresponding annotations.
 Data Diversity: Ensure the dataset includes diverse images with variations in resolution, lighting
conditions, and patient demographics to improve model generalization.
2. Data Preprocessing
 Image Resizing: Resize images to a consistent input size suitable for the CNN (e.g., 224x224
pixels).
 Normalization: Normalize pixel values (e.g., scale to [0, 1] range) to enhance model convergence
during training.
 Data Augmentation: Apply techniques such as rotation, flipping, zooming, and brightness
adjustment to artificially expand the dataset and reduce overfitting.
 Train-Test Split: Divide the dataset into training (e.g., 70%), validation (e.g., 15%), and test sets
(e.g., 15%) to ensure proper evaluation of the model.
3. Model Development
 Model Architecture: Select or design a CNN architecture suitable for image classification tasks.
Options may include:
o Pre-trained Models: Utilize transfer learning with models like VGG16, ResNet50, or
EfficientNet to leverage features learned on large image datasets.
o Custom CNN: Build a custom CNN model tailored to the specific features of retinal
images.
4. Model Training
 Loss Function: Choose an appropriate loss function (e.g., categorical cross-entropy) to measure the
performance during training.
  Optimizer: Select an optimizer (e.g., Adam, SGD) to adjust the learning rate and minimize the loss
function.
 Training Configuration: Set hyperparameters such as batch size, learning rate, and number of
epochs. Use techniques like early stopping to prevent overfitting.
 Training Process: Train the model using the training dataset while validating it with the validation
set to monitor performance.
5. Model Evaluation
 Metrics: Evaluate the trained model using metrics such as:
o Accuracy: Overall correct predictions.
o Sensitivity (Recall): True positive rate for detecting DR.
o Specificity: True negative rate for identifying healthy eyes.
o F1-Score: Harmonic mean of precision and recall.
o Area Under the ROC Curve (AUC): Measure of the model's ability to distinguish between
classes.
 Test Set Evaluation: Assess model performance on the unseen test set to determine generalization
capabilities.
6. Model Deployment
 Integration: Integrate the trained model into a telemedicine platform or a standalone application for
real-time analysis.
 User Interface: Develop a user-friendly interface for healthcare professionals to upload images and
receive diagnostic results.
 Cloud Deployment: Consider deploying the model on cloud infrastructure (e.g., AWS, Google
Cloud) to ensure scalability and accessibility.
7. Validation and Feedback
 Clinical Validation: Collaborate with ophthalmologists to validate the model's predictions against
expert assessments in real-world scenarios.
 User Feedback: Gather feedback from healthcare providers to refine the system and improve
usability.
8. Continuous Improvement
 Model Retraining: Implement a mechanism for continuous learning by retraining the model with
new data over time to enhance performance.
 Monitoring: Establish a monitoring system to evaluate the model's performance in the field and
make necessary adjustments.
CONCLUSION
The development of an automated system for detecting Diabetic Retinopathy (DR)
using deep learning offers a transformative solution to the challenges posed by
traditional screening methods. Leveraging the power of Convolutional Neural
Networks (CNNs), the proposed system can effectively and accurately identify DR and
classify its severity from retinal fundus images, potentially reducing the dependency on
manual diagnosis by ophthalmologists.

This deep learning-based approach provides significant advantages, including

enhanced accuracy, scalability, and the ability to operate in low-resource environments
where access to healthcare is limited. By integrating the system into telemedicine
platforms, real-time screening and diagnosis become accessible to populations in
remote and underserved areas, addressing the global need for early DR detection and
timely treatment.

Overall, the automated detection system not only promises to reduce the workload on
healthcare professionals but also ensures earlier intervention, helping to prevent vision
loss for millions of diabetic patients worldwide. Continuous improvements, including
model retraining and clinical validation, will further strengthen the system's
performance and real-world impact.
 TENTATIVE CHAPTER PLAN FOR THE PROPOSED
WORK

CHAPTER 1: INTRODUCTION

 Introduce the topic of Automated Detection of Diabetic Retinopathy Using Deep

Learning using machine learning.
 Provide background information on the importance of the problem.
 Clearly state the objectives and scope of your proposed work.
 Outline the significance and potential impact of the research.

CHAPTER 2: LITERATURE REVIEW

 Summarize existing research on Automated Detection of Diabetic Retinopathy Using

Deep Learning
 Analyze different methodologies, models, and datasets used in related works.
 Highlight gaps or limitations in current approaches.
 Establish a foundation for your proposed work based on insights gained from the
literature.

CHAPTER 3: OBJECTIVES

 Clearly define the specific objectives of your project.

 Break down the objectives into actionable steps, such as data collection, pre-
processing, model selection, and evaluation.
 Ensure alignment with the broader goals outlined in the introduction.

CHAPTER 4: METHODOLOGIES
  Detail the methodologies you plan to employ in your project.
 Discuss the step-by-step process, including data collection, pre-processing, feature
engineering, model selection, and validation.
 Justify the choice of methodologies and explain how they address the identified
gaps in the literature.

CHAPTER 5: EXPERIMENTAL SETUP

 Describe the hardware and software components of your experimental setup.

 Discuss the selection of datasets, models, and tools for the Handwritten Digit
Recognition project.
 Provide insights into how the experimental setup contributes to the reliability and
validity of your results.

CHAPTER 6: CONCLUSION AND FUTURE SCOPE

 Summarize key findings and contributions of your work.

 Discuss the implications of your results in the context of Handwritten Digit
Recognition.
 Present possible avenues for future research and enhancements to the proposed
system.
 Conclude with a reflection on the significance of your work in advancing the field.
 REFERENCES

 Gulshan, V., Peng, L., Coram, M., Stumpe, M. C., Wu, D., Narayanaswamy, A., ... &
Webster, D. R. (2016).
Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy
in Retinal Fundus Photographs.
JAMA, 316(22), 2402-2410.
[DOI: 10.1001/jama.2016.17216]

 Kaggle Diabetic Retinopathy Detection Dataset (2015).

Available from: https://www.kaggle.com/c/diabetic-retinopathy-detection

 Abràmoff, M. D., Lou, Y., Erginay, A., Clarida, W., Amelon, R., Folk, J. C., & Niemeijer,
M. (2016).
Improved Automated Detection of Diabetic Retinopathy on a Publicly Available Dataset through
Integration of Deep Learning.
Investigative Ophthalmology & Visual Science, 57(13), 5200-5206.
[DOI: 10.1167/iovs.16-19964]

 Lam, C., Yu, C., Huang, L., & Rubin, D. (2018).

Retinal Lesion Detection with Deep Learning Using Image Patches.
Investigative Ophthalmology & Visual Science, 59(1), 590-596.
[DOI: 10.1167/iovs.17-22329]

 Messidor Database (2004).

Available from: http://www.adcis.net/en/Download-Third-Party/Messidor.html

 LeCun, Y., Bengio, Y., & Hinton, G. (2015).

Deep Learning.
Nature, 521(7553), 436-444.
[DOI: 10.1038/nature14539]

 Li, H., Jiang, H., Sun, X., Tang, X., & Sze, K. Y. (2019).
Automated Diabetic Retinopathy Detection Using Deep Learning Algorithms.
Journal of Healthcare Engineering, 2019, Article ID 6529538.
[DOI: 10.1155/2019/6529538]

 Wang, L., Pedersen, P., & Hafner, L. (2020).

Automated Diagnosis of Diabetic Retinopathy Based on Machine Learning Algorithms and Retinal
Fundus Images: A Review.
IEEE Access, 8, 138734-138749.
[DOI: 10.1109/ACCESS.2020.3012941]