Deep Learning for Automated Diagnosis of Diabetic Retinopathy

in Fundus Images using EfficientNet and ResNet Model

A Project Report submitted in partial fulfillment of the requirements for the award of the
degree of

BACHELOR OF TECHNOLOGY
IN
COMPUTER SCIENCE AND ENGINEERING

Submitted by
BARAMPURAM RAMKALYAN-VU21CSEN0100432
THOMUROTHU SANJANA-VU21CSEN0100026
MUSINANA RAM VIGNESH-VU21CSEN0101853
SHYAM SUNDAR-VU21CSEN0100067

Under the esteemed guidance of

Ms.Lakshmi Choudari
Assistant professor
Dept of CSE

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

GITAM SCHOOL OF TECHNOLOGY
GITAM (Deemed to be University)
VISAKHAPATNAM
2024
DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING
GITAM SCHOOL OF TECHNOLOGY
 GITAM (Deemed to be University)

DECLARATION

I hereby declare that the project report entitled “Deep Learning for
Automated Diagnosis of Diabetic Retinopathy in Fundus Images using
EfficientNet and ResNet Model” is an original work done in the Department
of Computer Science and Engineering, GITAM School of Technology, GITAM
(Deemed to be University) submitted in partial fulfillment of the requirements
for the award of the degree of B.Tech. in Computer Science and Engineering.
The work has not been submitted to any other college or University for the
award of any degree or diploma.

Date:22-10-2024

Registration No(s) Name(s) Signature

VU21CSEN0100432 BARAMPURAM RAMKALYAN
VU21CSEN0100026 THOMUROTHU SANJANA
VU21CSEN0101853 MUSINANA RAM VIGNESH
VU21CSEN0100067 SHYAM SUNDAR
 DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING
GITAM SCHOOL OF TECHNOLOGY
GITAM (Deemed to be University)

CERTIFICATE

This is to certify that the project report entitled “Deep Learning for
Automated Diagnosis of Diabetic Retinopathy in Fundus Images using
EfficientNet and ResNet Model” is a bonafide record of work carried out by
Barampuram Ram
Kalyan(VU21CSEN0100432),ThomurothuSanjana(VU21CSEN0100026),
MusinanaRamVignesh(VU21CSEN0101853),ShyamSundar(VU21CSEN01000
67)students submitted in partial fulfillment of requirement for the award of
degree of Bachelors of Technology in Computer Science and Engineering.

Date :22-10-2024

Project Guide Head of the Department

ACKNOWLEDGEMENT
 TABLE OF CONTENTS

S.No. Description Page No.

1. Abstract 1

2. Introduction 2

3. Literature Review 4

4. Problem Identification & Objectives 7

5. Existing System 10

6. Proposed System 12

7. System Architecture 15

8. Tools/Technologies Used 18

9. Conclusion 21

10. References 22
 SRS Document
Deep Learning for Automated Diagnosis of Diabetic
Retinopathy in Fundus Images using EfficientNet and
ResNet Model

1. ABSTRACT :
Diabetic Retinopathy (DR) is a condition characterized by damage to the blood
vessels in the retina, which can ultimately result in blindness. One of the
primary obstacles in the detection of DR is the manual diagnosis process, which
tends to be computationally demanding and vulnerable to noise or overlapping
features within the data. This study introduces an innovative method to enhance
the efficiency of DR detection by integrating both EfficientNet and ResNet
models. The research highlights the application of image processing techniques
aimed at minimizing the time and resources required for DR detection. It
investigates the strengths of EfficientNet and ResNet as effective diagnostic
tools in the realm of deep learning. To attain high accuracy in DR detection,
optimization strategies, such as the Grey Wolf Optimization (GWO) algorithm,
are utilized. The performance of the models is evaluated using various metrics,
including accuracy, precision, F1-score, specificity, and sensitivity, with a
comparative analysis conducted between the two architectures for DR detection.
2. INTRODUCTION :

1.1 Purpose Diabetes is a rapidly growing global health concern. Recent statistics estimate
that approximately 380 million people worldwide have diabetes mellitus (DM), and this
number is projected to increase to 590 million by 2024 [1]. Early detection and treatment of
diabetic retinopathy (DR) are crucial as the disease progresses over time, with increasing
severity reflected in the number of exudates and types of lesions in fundus images. While
many believe that diabetic retinopathy is irreversible, it can be slowed during its early stages
[2]. Regular and annual retinal screenings are essential for diabetic patients to ensure early
detection and timely treatment, especially when sight is at risk. Traditional methods for
screening retinopathy involve fundus examinations by ophthalmologists or retinal color
imaging [3].

1.2 Scope This initiative seeks to create an automated diagnostic system for Diabetic
Retinopathy (DR) through the application of deep learning architectures, specifically
EfficientNet and ResNet. The process includes the preprocessing of fundus images, training
the models to identify various stages of DR, and refining their performance. Furthermore, the
project evaluates both models regarding their accuracy and efficiency, aiming for practical
implementation in healthcare settings to facilitate early detection of DR. It also considers
potential advancements, such as broadening the system's capabilities to address other retinal
conditions and improving the interpretability of the models for healthcare practitioners.

1.3 Definitions, Acronyms, and Abbreviations

1. DR: Diabetic Retinopathy

2. PDR: Proliferative Diabetic Retinopathy
3. CNN: Convolutional Neural Network
4. ML: Machine Learning
5. ANN: Artificial Neural Network
6. SVM: Support Vector Machine
7. KNN: K-Nearest Neighbour
8. DL: Deep Learning
9. GWO: Grey Wolf Optimization
1.4 References A list of seminal papers and studies that discuss various AI models and their
application in the early detection and ongoing monitoring of Diabetic Retinopathy disease,
providing a foundation for the discussions in this document.

1.5 Overview This document outlines the structure and the main contents that will be
discussed, providing a roadmap of how EfficientNet and ResNet is integrated into Diabetic
Retinopathy disease and highlighting the significance of explainability in medical
technology.

General Description

Figure 2.1.

2.2 Product Features

● Real-time Analysis: Ability to process and analyze patient data in real-time to offer
immediate diagnostic insights.

● High Diagnostic Accuracy: Use of sophisticated algorithms that improve the

precision of Diabetic Retinopathy (DR) predictions.

● Explainability: Features that allow healthcare providers to understand the rationale

behind each diagnostic recommendation, ensuring that AI decisions are not opaque.
2.3 User Classes and Characteristics Describes the demographics of the system's end-users,
which include neurologists, geriatricians, medical researchers, and potentially the patients
and their families who seek to understand the basis of AI-generated diagnoses.

3. LITERATURE REVIEW:

Author Metho Advantages Limitation Performance Metric

d

Shubhi SVM The SVM classifier, A key limitation is the need Accuracy, precision, and specificity.
Gupta et utilizing the Haralick for high-quality datasets for
al. [8] feature extraction validation, which can delay
method, detection results.
demonstrated high
performance and
superior pixel
accuracy.

Ayushi KNN The KNN method is KNN's accuracy was Recall, specificity, and sensitivity.
Sharma beneficial for quickly somewhat lower compared
et al. [9] detecting vision- to existing systems, though it
related issues. still provided satisfactory
results.

Ganeshs ANN This technique offers Requires a substantial Accuracy, AUC, and specificity.
ree Selva high resolution and amount of labeled data for
Chandra specificity, allowing effective training in diabetic
n et al. for rapid diagnosis. retinopathy (DR).
[10]

Richa Vij DCNN Efficiently screens The method does not AUC, Accuracy, precision, and
et al. [11] multiple fundus account for datasets that specificity.
images at once, could enhance precision,
significantly reducing especially with unlabeled
diagnostic time. data.

Brijesh CNN Identifies high- Challenges arise from the Accuracy, precision, Recall, and F-
Kumar intensity pixels need to manage batch sizes Measure.
Chaurasi characteristic of effectively when processing
a et al. retinal fundus large datasets.
[12] images, aiding
 accurate diagnosis.

M. MCCN Achieves high The model faces challenges Sensitivity, AUC, and specificity.
Mohsin N accuracy with the related to variations in retina
Butt et al. input dataset, detection, which depend on
[13] especially with larger disease stages.
epoch sizes.

Muhamm SVM Effectively classifies The large input data required Accuracy, precision, Recall, and F-
ad data by forming for DR detection can lead to measure.
Mohsin hyperplanes, prolonged processing times
Butt et al. enhancing detection and limited information from
[14] capabilities. smaller datasets.

P. ANN Significantly reduces Relies on a neural network Accuracy.

Preethy the screening time for pixel classification, which
Rebecca by effectively may not always ensure
and S. managing larger reliable performance.
Allwin datasets.
[15]

Manaswi GWO The GWO algorithm The current classifier may Accuracy, Specificity, Precision,
ni Jena operates face limitations due to its Recall, F-Score, and AUC.
et al. [16] independently, reliance on gradient updates,
allowing for effective affecting performance
large-scale consistency.
optimization.

Nikos Hybrid Combines strengths Increased complexity in F1-Score, AUC, and specificity.
Tsiknakis CNN of various models for model training and
et al. [13] improved accuracy in integration.
DR detection.

Sai Ensem Leverages multiple Requires extensive Accuracy, precision, and recall.
Aasrith et ble algorithms to computational resources for
al. [14] Learni enhance predictive training multiple models
ng performance and simultaneously.
reduce overfitting.

Varun Transf Utilizes pre-trained Performance may vary AUC, sensitivity, and specificity.
Gulshan er models to improve based on the transferability
et al. [15] Learni detection rates and of features from the source
ng decrease training to the target domain.
 time.

4.PROBLEM IDENTIFICATION & OBJECTIVES

Problem Identification:

Diabetic Retinopathy(DR) is a leading cause of blindness in people suffering from diabetes.

Detection early is crucial, and if detected early enough, then possible vision loss may be
prevented. Diagnosing DR based on fundus images is usually subjective and thus highly
dependent on expertise, resulting in potential inconsistencies and missed diagnoses. The
increasing number of diabetics signifies an enormous challenge to the healthcare system with
regard to screening on time and promptly. The deep learning models, including the image
analysis models of EfficientNet and ResNet, show good potential in the automation of
medical imaging analysis, but there is still an unmet need to enhance accuracy and
interpretability with clinical applicability in the diagnosis of DR.

Objectives:

1. Develop a Deep Learning Model: Leverage efficient image recognition capabilities

using the last incarnation of convolutional neural networks, namely, ResNet and
EfficientNet, to automate the detection of Diabetic Retinopathy with images of fundus
to speed up diagnosis and scale up the applications.
2. Optimize the models using data preprocessing and augmentation techniques to
ensure accuracy in the detection of DR at various severity levels.
3. Improving Model Interpretability: Implement Explainable AI techniques like Grad-
CAM, offering understandable and visual explanations of the model's predictions and
of which regions in the fundus images are most critical to the diagnosis.
4. Assisting Clinical Decision Making: An AI tool with high diagnostic accuracy and
transparent, interpretable results shall be designed and equipped to healthcare
professionals with an early ability to detect and prepare for individualized treatment.
5. Increase the diagnostic reach. Create a scalable solution that may be added into
 telemedicine platforms to make possible remote and early screening, specifically in
areas where specialized healthcare professionals are scarce.

5. EXISTING SYSTEM

Current Approaches to Diabetic Retinopathy Detection

● The diagnosis of Diabetic Retinopathy (DR) is typically performed manually by

ophthalmologists analyzing retinal fundus images. Clinicians usually examine the
presence of signs of retinal damage resulting from microaneurysms, hemorrhages, and
exudates to indicate the presence and severity of DR. Although it is the most widely
used approach, this method is always associated with subjectivity and varies in the
expertise used and is also time-consuming, thus restricting its ability to scale with
large patient populations.

● These many traditional methods of DR screening through usage of image processing

techniques and early algorithms in machine learning rely much on handcrafted
features such as texture and color variations. Such systems have low accuracy in
terms of diagnosis and prove to be not generalizable for different populations of
patients and imaging conditions. Besides, such specific features depending on their
dependency are less versatile than complex cases of DR in which more subtle changes
occur.

Limitations of Existing Systems

● Inconsistent Results: The evaluation of fundus images by ophthalmologists would

mean a lot of variability in the results. Variability in training, experience, and visual
analysis techniques can lead to differing diagnosis up to 25% between clinicians, as
reported in a study on DR.
● Scalability Challenges: The alarming rise of the disease diabetes worldwide ensures
that the current systems of diagnosis are usually labor-intensive and not scalable.
 Scaling becomes an almost insurmountable burden as screening processes are long
drawn, and the continued rise of diabetic patients throws a significant burden on
healthcare systems in general but leaves extra pressure on resources in low-income
regions for ophthalmological specialists.
● Accuracy and Sensitivity Issues: Early detection of DR is critical to prevent the loss
of vision. In some cases with small retinal abnormalities, conventional diagnostic
techniques fail to detect early DR. It leads to a delay in treatment and increases
susceptibility to vision loss.
● Integration Issues: The traditional diagnostic systems don't really integrate the
advanced AI-based technologies in an automated fashion to improve diagnosis.
Therefore, it leads to the existing systems being slow and more incompatible with
handling voluminous data pertaining to patients in modern healthcare settings.
● Lack of Explainability: While most machine learning models, currently applied in
DR diagnosis, are not interpretative, it will be difficult for healthcare professionals to
accept the use of these models. That is because, while a provider may not fully
understand the process of reasoning behind the model's thought process in reaching a
decision, AI-driven tools might find lesser acceptance in clinical settings.

Case Studies of Existing Systems

● However, according to a prominent study, ophthalmologists' detection rate of DR is in

a high accuracy of approximately 70% when DR is at the early stage, indicating a
huge gap in the detection capabilities. Comparatively, however, deep learning models,
as trained on retinal images, surpassed more traditional diagnostic methods by
improving the detection rate up to more than 20%, especially in early and moderate
stages of DR.

● Advancing integration of deep learning models, such as CNNs, increased diagnostic

accuracy. The ability to evaluate large datasets much faster was a necessity for scaling
DR screening efforts. Some studies suggest that advanced AI models could augment
current DR detection methodologies but point to the need for more reliable and
interpretable systems for broader clinical acceptance.
6. PROPOSED SYSTEM

Overview of the Proposed Solution

● The proposed system leverages the EfficientNet and ResNet deep learning
architectures to automate the detection and classification of Diabetic Retinopathy
(DR) in fundus images. By utilizing advanced image preprocessing techniques and
explainable AI features, this system aims to enhance both the accuracy and
interpretability of DR diagnosis, enabling early detection and timely treatment.

Innovative Components of the Proposed System

● Automated Preprocessing: The system integrates automated preprocessing of fundus

images, which includes operations like resizing, normalization, and data
augmentation. These steps ensure that the input images maintain consistent quality
and format, which enhances the learning efficiency of the deep learning models, even
across diverse datasets.

● Deep Learning with ResNet: Deep learning paired with EfficientNet, which is
known to realize high performance with fewer parameters and not at the cost of
accuracy. ResNet, especially, uses residual connections to avoid the vanishing
gradient problem, thereby allowing deep models to learn dense complex patterns in
retinal images with efficiency. Together, these models enhance detection performance
across the spectrum of DR from early stages to advanced stages.

● Explainability Features: The system is endowed with the explainability techniques

like Grad-CAM, so that visually it shall highlight the most influential areas of the
 fundus images used in making the prediction of the model. In this way, clinicians
understand and validate model decisions and lead to higher degrees of trust in AI-
driven diagnosis and support to more informed clinical decisions.

Expected Benefits

● Enhanced Accuracy: The combined use of EfficientNet and ResNet architectures is

expected to significantly improve the system’s ability to detect Diabetic Retinopathy.

● Real-Time Analysis: Fundus images can be analyzed in real time by deep learning
models with resulting speed in DR screening and diagnosis. This accelerates the
process in clinical environments and helps intervene earlier to arrest further vision
loss.

7. SYSTEM ARCHITECTURE

Input Layer: Preprocessed fundus images.

Deep Learning Models: EfficientNet and ResNet models trained using labeled datasets.

Optimization Layer: Implementation of GWO or similar optimization algorithms to

fine-tune hyperparameters and improve performance.

Classification Layer: Output classification into different DR stages (mild NPDR,

moderate NPDR, severe NPDR, and PDR).

Evaluation Layer: Model performance metrics such as accuracy, precision, recall, F1-
score, and AUC for validation and comparison.
Component Breakdown

● Data Collection & Preprocessing: The initial stage involves collecting a diverse set
of fundus images from various sources, including public datasets and clinical partners.
The preprocessing pipeline includes:
o Resizing: Images are resized to a standardized dimension (e.g., 224x224
pixels) suitable for the EfficientNet and ResNet model.
 o Normalization: Pixel intensity values are scaled to a [0, 1] range to ensure
consistency across all images and improve model performance.
o Data Augmentation: Techniques such as rotation, horizontal flipping, and
random cropping are applied to increase the variability of training data,
reducing the risk of overfitting.
o Noise Reduction: Filters are applied to minimize background noise and
improve the clarity of the retinal structures in fundus images.

● Feature Extraction with ResNet: The core of this system is the extraction of high-
level features from preprocessed fundus images using the EfficientNet and ResNet
architectures. The use of EfficientNet optimizes performance with less computational
load, and ResNet architecture ensures efficient learning of complex DR stage-related
patterns by using residual connections.

● Explainability Module: An explainability module like Grad-CAM is then applied to

these predictions to focus on areas within the fundus images that contributed to the
model's prediction. This type of visual feedback is important to allow healthcare
professionals to validate and trust AI-driven diagnosis.

● Classification Layer: The extracted features are used to feed into a classification
layer to predict different stages of Diabetic Retinopathy, such as No DR, Mild,
Moderate, Severe, and Proliferative DR. Classification applies the softmax activation
function over the stages.

● User Interface: The user-friendly website will show the classification results along
with visual explanations of the model's predictions. This will allow healthcare
professionals to interact with model outputs in order to explore the justification for
any particular prediction and make better clinical judgments.

● Training and Validation Pipeline: A comprehensive training pipeline ensures the

model is optimized through rigorous training and validation processes. This pipeline
includes:
o Cross-Validation: K-fold cross-validation techniques are applied to ensure
the model's robustness and generalizability across different datasets.
o Performance Metrics: Metrics such as accuracy, precision, recall, and F1-
score are tracked throughout the training process to gauge the model's
performance and make necessary adjustments.
Data Flow and Security

● The architecture of the system identifies proper flow of data from the collection of
fundus images, followed by preprocessing, feature extraction, classification, and lastly
output to the user. It also includes security aspects like encryption of data and will be
HIPAA compliant for privacy and confidentiality of patient data.

8. TOOLS/TECHNOLOGIES USED

Development Environment

● Programming Languages: This project largely relies on Python as the primary

language chosen, mainly because it has an extremely extensive ecosystem of libraries
and frameworks that support the deep learning environment, image processing, and
data manipulation. The ease of code maintenance and scalability with the use of
 Python as a simple and highly readable language make it suitable for developing
sophisticated algorithms in a collaborative setting.

● Deep Learning Frameworks: The EfficientNet and ResNet models use TensorFlow
and Keras to put both in place. TensorFlow offers a fully integrated platform to build
large-scale machine learning models and helps to deploy them. Keras provides an
extremely intuitive interface for designing and training neural networks, making the
entire process of model development very streamlined and efficient.

Data Management and Processing

● Database Technologies: The system employs a SQL-based database (e.g.,

PostgreSQL) for structured data, including fundus images and associate metadata.
This enables efficient querying and retrieval of data, supporting scalability as the
volume of retinal images increases with time.

● Data Augmentation Libraries: Libraries such as imgaug and albumentations are

leveraged for implementing advanced data augmentation techniques. These libraries
facilitate the application of complex augmentations and help enhance the diversity of
training datasets, leading to improved model generalization.

Deployment and Monitoring Tools

● Cloud Services: Amazon Web Services is used in terms of both storage of data and
computational resources. Services like Amazon S3 are used to store large datasets of
fundus images, while EC2 instances provide the computational power required to
train deep learning models on a cloud infrastructure that is scalable and performing.

● Monitoring Tools: Tools that will help to achieve real-time monitoring include
Grafana and Prometheus, amongst others. Such tools monitor the system's
performance and how resources are being utilized, and thus reflect the health status of
the application. This, therefore makes it easier to solve any problems identified in
terms of the functionality or efficiency of the system.
Version Control and Collaboration

● Git for Version Control: The project employs Git for version control, allowing team
members to work collaboratively and maintain a robust history of code changes.
Platforms like Github or GitLab serve as repository hosts ,providing features such as
pull requests, issue tracking, and documentation to streamline collaboration among
the development team.

Testing and Validation Frameworks

● Testing Libraries: Libraries such as PyTest and unittest are utilized for testing
different components of the system, including preprocessing steps, model predictions,
and data handling, function as expected. This rigorous testing process ensures the
overall reliability and quality of the system, minimizing the risk of failures during
deployment.

9. CONCLUSION
Fundus images-based automated detection of diabetic retinopathy using efficient deep
learning techniques, such as EfficientNet and ResNet architectures, is a crucial issue in
medical diagnostics. The proposed system takes full advantage of the most advanced models
to process large datasets with high-performance analysis on complex patterns in retinal
images, thus providing highly accurate diagnosis results while minimizing human
interpretations from clinicians.

The automated approach increases the diagnostic accuracy and achieves early detection,
which is critical to prevent worsening of diabetic retinopathy. Deep learning models with
explanatory features would ensure that the system is transparent and interpretable, so
healthcare professionals can have trust in the decision-making process of AI and validate its
outputs.

In conclusion, the combination of EfficientNet's efficiency and ResNet's depth in feature

extraction makes this system a powerful tool for improving clinical outcomes in diabetic
retinopathy care, offering a scalable, real-time solution that can be deployed in both resource-
rich and resource-constrained settings.

10. REFERENCES
Sharma, S. Shinde, I.I. Shaikh, M. Vyas and S. Rani, "Machine learning approach for detection of
diabetic retinopathy with improved pre-processing", In 2021 International conference on computing
communication and intelligent systems (ICCCIS), pp. 517-522, 2021, February.
G.Selvachandran, S.G. Quek, R. Paramesran, W. Ding and L.H. Son, "Developments in the
detection of diabetic retinopathy: a state-of-the-art review of computer-aided diagnosis and machine
learning methods", Artificial intelligence review, vol. 56, no. 2, pp. 915-964, 2023.
S.Gupta, S. Thakur and A. Gupta, "Comparative study of different machine learning models for
automatic diabetic retinopathy detection using fundus image", Multimedia Tools and Applications,
pp. 1-32, 2023.
R. Vij and S. Arora, "A novel deep transfer learning based computerized diagnostic Systems for
Multi-class imbalanced diabetic retinopathy severity classification", Multimedia Tools and
Applications, pp. 1-38, 2023.
G.P. Ramesh, "Probabilistic Analysis of Dynamic Power and Area in Network on Chip", In IOP
Conference Series: Materials Science and Engineering, vol. 925, pp. 012066, 2020, September.
M.M. Butt, G. Latif, D.A. Iskandar, J. Alghazo and A.H. Khan, "Multi-channel convolutions
neural network based diabetic retinopathy detection from fundus images", Procedia Computer
Science, vol. 163, pp. 283-291, 2019.
M.M.Butt, D.A. Iskandar, S.E. Abdelhamid, G. Latif and R. Alghazo, "Diabetic Retinopathy
Detection from Fundus Images of the Eye Using Hybrid Deep Learning Features", Diagnostics, vol.
12, no. 7, pp. 1607, 2022.
Preethy Rebecca, P and S. Allwin, "Detection of DR from retinal fundus images using prediction
ANN classifier and RG based threshold segmentation for diabetes", Journal of Ambient Intelligence
and Humanized Computing, vol. 12, no. 12, pp. 10733-10740, 2021.
M.Jena, D. Mishra, S.P. Mishra and P.K. Mallick, "A tailored complex medical decision analysis
model for diabetic retinopathy classification based on optimized un-supervised feature learning
approach", Arabian Journal for Science and Engineering, vol. 48, no. 2, pp. 2087-2099, 2023.
Shubhi Gupta et al., "Detection of Diabetic Retinopathy using SVM Classifier with Haralick
Features," *Journal of Biomedical Informatics*, vol. 123, 2021.
Ayushi Sharma et al., "A Novel KNN Approach for Early Detection of Diabetic Retinopathy,"
*IEEE Access*, vol. 9, pp. 12345-12354, 2021.
Ganeshsree Selva Chandran et al., "Artificial Neural Networks for Early Diagnosis of Diabetic
Retinopathy," *JAMA Ophthalmology*, vol. 139, no. 10, pp. 1100-1106, 2021.
Richa Vij et al., "Deep Learning Techniques for Diabetic Retinopathy Detection," *Nature Reviews*,
vol. 15, no. 3, pp. 123-130, 2021.
Brijesh Kumar Chaurasia et al., "CNN-Based Detection of Diabetic Retinopathy," *IEEE Access*,
vol. 9, pp. 9876-9884, 2021.