Diabetic Retinopathy Detection Using Deep Learning

Puneet Kumar1, Himanshu Ranjan2, Prabhav Singh 3,

Madhav Sharma4, Kakoli Banarjee5 and Harsha K.G6
1
JSS Academy of Technical Education/Computer Science and Engineering, Noida, India
Email: [email protected]
2-6
JSS Academy of Technical Education/Computer Science and Engineering, Noida, India
Email: {20cs126, 20cs182, 20cs191, Kakoli.banerjee, harshakg}@jssaten.ac.in

Abstract— A debilitating side effect of diabetes called diabetic retinopathy (DR) mostly affects
the blood vessels in the light-sensitive back tissue of the retina. This condition often appears
in cases of poorly managed diabetes and is the main cause of blindness in individuals in their
working years. The current methods of DR detection necessitate a manual examination of
patients by an ophthalmologist, which has led to the hunt for automated alternatives. The
Deep Learning-based approach for DR identification proposed in this article makes use of a
model built on 35,126 publicly accessible retinal pictures from eye PACS that are accessible
on the Kaggle platform. By utilizing GPU acceleration, the developed model attained an
exceptional accuracy of approximately 81%. Some of the main subjects of this research are
diabetic retinopathy, deep learning, and preprocessing.

Keywords— Diabetic Retinopathy, Convolutional Neural Network (CNN), Deep Learning

I. INTRODUCTION

The early detection of diseases significantly enhances the efficacy of treatment in healthcare. Diabetes, a
condition characterized by elevated blood glucose levels due to insufficient insulin production, affects
approximately 425 million individuals worldwide. Its impact extends beyond blood sugar regulation, impacting
organs such as the retina, heart, nerves, and kidneys.

Among the complications associated with diabetes is Diabetic Retinopathy (DR), a condition characterized by
the enlargement and leakage of blood and fluid from the blood vessels in the retina. If left untreated, DR can
progress to advanced stages, leading to visual impairment and contributing to 2.6% of global blindness cases.
Prolonged duration of diabetes increases the risk of developing DR, underscoring the importance of early
diagnosis and intervention through routine retinal screening.

The identification of various lesions in retinal imaging, including microaneurysms (MA), haemorrhages (HM),
and soft and hard exudates, is pivotal for DR detection. MA, the earliest indicator of DR, manifests as small,
round, red spots (<125 μm) due to vessel wall weakness. HM, categorized as flame (superficial) or blot
(deeper), appear as larger irregular spots (>125 μm). Hard exudates, distinguished by sharply margined,
brilliant yellow patches, result from plasma leakage in the outer retinal layers. Soft exudates, known as cotton
wool spots, present as round or oval white patches due to swollen nerve fibres[12]. While MA and HM lesions
exhibit red characteristics, soft and hard exudates appear bright.

DR is stratified into five stages—proliferative DR, mild DR, moderate DR, severe DR, and no DR—based on
the presence and severity of lesions.
To address the challenges associated with labour intensive human diagnostics, prone to errors and resource-
intensive, automated DR detection techniques have emerged as efficient alternatives[9][10]. This study
scrutinizes 33 recent research works employing automated deep learning algorithms for DR detection and
classification. The subsequent sections provide an overview of deep learning algorithms, describe various
fundus retina datasets, and present performance metrics. Additionally, automatic DR categorization techniques
are reviewed, followed by discussions and a comprehensive summary.

II. RELATED WORKS

Diabetic Retinopathy (DR), a prevalent complication of diabetes affecting millions worldwide, poses a
significant threat to vision health. Early detection of DR is imperative for timely intervention and treatment,
mitigating the risk of irreversible vision loss. In recent years, pioneering research has focused on developing
innovative algorithms and techniques for the early diagnosis of DR, leveraging cutting-edge technologies such
as deep learning, machine learning, and image processing.

Sumeet Dua proposes a novel blood vessel detection algorithm based on regional recursive hierarchical
decomposition using Quadtrees and post-filtration of edges. This technique effectively identifies blood vessels
in retinal images, significantly reducing false dismissals while maintaining computational efficiency [1].

In a similar vein, an interpretable Ensemble Deep Learning model for DR classification is introduced,
harnessing the power of multiple well-trained deep learning models integrated through the AdaBoost
algorithm. This ensemble approach demonstrates enhanced robustness and superior performance compared to
individual models, offering promising prospects for automated DR detection [2].

Further, advancements in image processing techniques, such as morphological processes and SVM classifiers,
have paved the way for automatic detection and severity classification of DR lesions, particularly exudates.
These techniques not only facilitate early diagnosis but also provide insights into disease progression and
severity [3].

The utilization of deep learning methodologies, including bi channel convolutional neural networks (CNNs)
and multilayer perceptron neural networks (MLPNN), underscores a paradigm shift in DR detection. These
models exhibit remarkable accuracy in detecting retinal pathologies, thereby augmenting ophthalmologists'
diagnostic capabilities and enhancing the efficiency of automated retinal image analysis systems [4][5].

Moreover, Mohamed M. Abdelsalam and M. A. Zahran propose a multifractal geometry-based approach for
early DR detection, leveraging supervised machine learning algorithms to automate the diagnosis process with
high accuracy. This approach not only streamlines the diagnostic workflow but also holds promise for
classifying other retinal diseases affecting vascular distribution [6].

Lastly, efforts to develop AI-powered classification models for DR, leveraging machine learning algorithms
and transfer learning techniques, have garnered significant attention. These models exhibit promising results
in accurately classifying fundus images into normal and DR-affected categories, laying the groundwork for
efficient patient management and treatment [7][8].

In their study, Abràmoff et al. [12] conducted a comprehensive analysis of fundus images, with each eye
assessed by a retinal expert. Employing two distinct algorithms separately, the researchers achieved promising
results. The Eye Check algorithm yielded an AUC of 0.839, while the Challenge 2009 algorithm yielded an
AUC of 0.821, underscoring the efficacy of these methodologies in diagnosing diabetic retinopathy.

Gargey et al. [13] devised a novel device for the automatic detection of diabetic retinopathy, categorizing
images into healthy or afflicted with the condition. Their model underwent rigorous evaluation using publicly
available databases, MESSIDOR 2 and E-Ophtha, yielding impressive AUC values of 0.94 and 0.95
respectively, highlighting its robust performance in external validation.
In their endeavour, Wilfred Franklin and Edward Rajan [14] proposed an automated tool renowned for its high
accuracy in blood vessel detection. Leveraging an automatic segmentation algorithm on images sourced from
the DRIVE database, the authors achieved a notable accuracy rate of 95.03%, indicative of the efficacy of their
methodology.

Antal and Hajdu [15] adopted a multifaceted approach, incorporating image-level, lesion-specific, and
anatomical components in their analysis. Through meticulous classification efforts, their model, tested on the
publicly available Messidor dataset, yielded an impressive AUC of 0.989, underscoring the robustness of their
methodology in diabetic retinopathy detection.

Liskowski et al. [16] employed a supervised approach coupled with deep neural networks for image analysis.
Their method, which emphasized preprocessing techniques, demonstrated exceptional performance, with an
AUC exceeding 0.99 and accuracy surpassing 0.97. Notably, fine vessel sensitivity exceeded 0.87, highlighting
the efficacy of their structured prediction methodology.

Revathy et al. [17] pursued an SVM-based training approach, classifying retinal images into three distinct
categories of diabetic retinopathy severity. Leveraging various classification algorithms, the authors achieved
commendable accuracy, with results indicating an 82% accuracy rate, underscoring the viability of their
approach in disease stratification and diagnosis.
The early classification of digital retina fundus images relied on hand-engineered features and empirical
parameters. Cree et al. [18] demonstrated the viability of computer vision techniques for automating
microaneurysm detection, achieving results comparable to clinicians.

Recent advancements have seen methodologies evolve to encompass not only microaneurysm detection but
also the staging of diabetic retinopathy. Yun et al. [19] proposed a classification method for retina fundus
images, categorizing them into various severity stages of diabetic retinopathy, while Nayak et al. [20]
introduced a method incorporating image preprocessing and texture analysis for feature extraction.

Rosas et al. [21] proposed a microaneurysm recognition system utilizing computer vision techniques for image
preprocessing, demonstrating effectiveness in discerning the presence of microaneurysms in retinal images.

Advancements in image processing have led to automated feature acquisition using convolutional neural
networks (CNNs). Gargeya et al. [22] automated diabetic retinopathy screening, achieving notable results
through CNN-based methodologies, while Dutta et al. [23] utilized statistical features extracted from images
to outperform CNNs in classification tasks.

Mansour [24] introduced a computer-aided diagnosis system, employing background subtraction and feature
extraction through the AlexNet model, exhibiting high validation accuracy in diabetic retinopathy diagnosis.

Qummar et al. [25] proposed an ensemble approach utilizing multiple deep learning models, demonstrating
robust performance in handling unbalanced data, while Gadekallu et al. [26] optimized a deep learning model
using intelligent computing techniques and achieved high accuracy in classification tasks.

Despite advancements, challenges persist in differentiating between diabetic retinopathy stages, underscoring
the need for further research and standardized evaluation methodologies.
III. METHODOLOGIES
A. Problem Formulation
The project was initiated with the comprehensive study of scientific literature on diabetic retinopathy and
integration of deep learning methodologies to analyse retinal images. This initial investigation has given rise
to a precise definition of the research objective, which is timely identification of diabetic retinopathy by
novel application of deep learning algorithms in visual image data.
B. Data Collection and Annotation
A carefully selected set of retinal images has been collected from various credible sources, including
established repositories and clinical databases like IEEE’s IDRiD and Kaggle. In order to carefully label each
image, to capture subtle indicators of diabetic retinopathy severity and to provide a solid basis for model
training and evaluation, machine annotation has been performed.
C. Model Architecture Design
Recognizing the intricate patterns inherent in retinal images and the subtle features indicative of Diabetic
Retinopathy, a bespoke convolutional neural network (CNN) architecture was meticulously crafted. The
specific architecture was designed to extract and interpret the essential characteristics which are decisive for
accurate disease detection.
D. Model Training and Validation
CNN has undergone rigorous training using an annotated dataset, with a particular focus on optimising key
performance indicators relevant to the diagnosis of diabetic retinopathy. In order to ensure the model's ability
to generalize effectively to unseen data and real world clinical scenarios, training iterations were
accompanied by rigorous verification procedures.
E. Data Augmentation and Preprocessing
A series of data augmentation techniques have been carefully applied to strengthen the model against
overestimation and increase its adaptability to different retinal image changes. In addition, to standardize and
enhance the image quality of input images, preprocessing steps have been carefully carried out including
normalization and noise reduction.

F. Transfer Learning and Fine Tuning

The principles of transfer learning have been exploited, building on the foundations of pre-established CNN
architectures that have been already trained on diverse image datasets. These pretrained models were invaluable
tools for the start, undergoing meticulous fine tuning to seamlessly adapt them to the subtleties of Diabetic
Retinopathy detection so as to speed up convergence and enhance performance.

G. Interpretability and Explainability

Acknowledging the imperative of interpretability in medical AI applications, efforts were directed towards
elucidating the model's decision-making processes. Techniques such as class activation maps (CAM) and
attention mechanisms were thoughtfully employed, offering clinicians invaluable visual insights into the
model's predictive rationale and facilitating informed clinical decision-making.

IV. IMPLEMENTATION
A. Model Architecture:

• We defined a custom CNN model architecture specifically tailored for our project.

• The model consists of two main parts: feature extraction layers and classifier layers.

• The feature extraction layers include convolutional and max pooling layers to extract meaningful features
from input images. We used two sets of convolutional layers followed by max pooling to progressively reduce
the spatial dimensions of the input.
 • The classifier layers include fully connected layers (also known as dense layers) to perform classification
based on the extracted features.

Fig 1: CNN Architecture

B. Layer Configurations:

• We configured the convolutional layers to have appropriate kernel sizes, padding, and output channels to
capture relevant patterns in the input images.

• ReLU activation functions were used after each convolutional layer to introduce non-linearity and enable
the network to learn complex relationships in the data.

• Max pooling layers were inserted to down sample the feature maps and reduce computational complexity
while preserving important features.

Fig2: Confusion Matrix

C. . Model Parameters:

• We defined the number of output classes (num_classes) based on the severity levels of diabetic retinopathy
in our dataset.
 • The number of output channels in the convolutional layers and the size of the fully connected layers were
chosen based on experimentation and architectural considerations.

Fig3: ROC Curve

D. Optimization:

• We aimed to optimize the model's performance by experimenting with different architectures, activation
functions, and hyperparameters.

• We used ReLU activation functions for their effectiveness in preventing the vanishing gradient problem and
accelerating convergence during training.

• The choice of kernel sizes, padding, and stride in convolutional layers, as well as the number of output
channels, were optimized to balance model complexity and expressiveness.
 Fig4: Validation Prediction Samples

Fig5: Error Analysis

Overall, by customizing the architecture and parameters of the CNN model, we aimed to design a model
capable of accurately classifying diabetic retinopathy severity levels while minimizing overfitting and
maximizing generalization performance.

E. Technologies Used
1. PyTorch:

•PyTorch is an open-source deep learning framework developed by Facebook's AI Research lab (FAIR).
• It provides a dynamic computational graph, allowing for flexible and efficient model construction and
training.
• PyTorch offers a rich ecosystem of libraries and modules for tasks such as data loading, model optimization,
and visualization.
• Key features include automatic differentiation, GPU acceleration, and support for distributed computing.
• We leveraged PyTorch's nn.Module API to define our custom CNN model architecture, specifying the layers,
activations, and connections.
• Additionally, PyTorch's torch.optim module enabled us to choose and configure optimization algorithms
such as SGD, Adam, or RMSprop, controlling parameters like learning rate and momentum.

2. NumPy
 •NumPy is a fundamental package for numerical computing in Python, providing support for multi-
dimensional arrays and mathematical operations.
• It offers efficient data structures and functions for array manipulation, linear algebra, and random number
generation.
• We utilized NumPy arrays to represent image data and labels, enabling efficient computation and
manipulation during preprocessing, training, and evaluation.
• NumPy's broadcasting and vectorized operations facilitated fast and concise code implementation for tasks
such as data normalization, array reshaping, and statistical analysis.

3. Scikit-learn (sklearn)

•scikit-learn is a versatile machine learning library in Python, offering a wide range of algorithms and utilities
for data mining and analysis.
• It provides tools for classification, regression, clustering, dimensionality reduction, and model evaluation.
• We may have used scikit-learn for tasks such as data preprocessing (e.g., scaling, feature selection), model
selection (e.g., cross-validation), and performance evaluation (e.g., computing metrics like accuracy, precision,
recall).
• scikit-learn's consistent API and extensive documentation made it easy to experiment with different machine
learning techniques and integrate them into our pipeline seamlessly.

4. Matplotlib and Seaborn

•Matplotlib and Seaborn are Python libraries for creating static, animated, and interactive visualizations.
• Matplotlib provides a MATLAB-like interface for plotting, while Seaborn offers high-level functions for
statistical data visualization.
• We employed these libraries to generate various types of plots and charts, including line plots, bar plots,
scatter plots, histograms, and heatmaps.
• By visualizing data distributions, model performance metrics, and training/validation curves, we gained
insights into the dataset characteristics, model behaviour, and optimization process

5. PIL (Python Imaging Library) / Pillow

• PIL (Python Imaging Library) is a library for opening, manipulating, and saving many different image file
formats.
• Pillow is a maintained fork of PIL, offering additional functionality and improvements.
• We used PIL/Pillow for image preprocessing tasks such as resizing, normalization, and format conversion.
• Its comprehensive image processing capabilities allowed us to prepare the input data efficiently and ensure
compatibility with the neural network model.

6. Optimizers:

• Optimization algorithms play a crucial role in training deep learning models by updating model parameters
to minimize the loss function.
• Common optimizers include Stochastic Gradient Descent (SGD), Adam, RMSprop, and AdaGrad, each with
its own advantages and hyperparameters.
• We experimented with different optimizers and configurations to find the most suitable strategy for our task,
considering factors such as convergence speed, generalization performance, and robustness to noise.
• Hyperparameters such as learning rate, momentum, and weight decay were fine-tuned to achieve optimal
training dynamics and model performance.

7. Dataset Specifications:

• The dataset used in our project consisted of retinal fundus images annotated with severity levels of diabetic
retinopathy.
• Each image was associated with a severity label indicating the degree of diabetic retinopathy, ranging from
0 to 4, representing different stages of the disease progression.
• The dataset may have been sourced from reputable medical databases or research institutions, ensuring
quality, diversity, and representativeness.
• Key statistics of the dataset, including the number of samples, class distribution, image resolution, and
annotation quality, were carefully analyzed and documented to guide model development and evaluation.

By leveraging these advanced technologies and tools, we designed a robust and efficient pipeline for diabetic
retinopathy classification, integrating state-of-the-art methodologies and best practices from the fields of deep
learning, machine learning, and image processing.

F. Result
Throughout the development of our diabetic retinopathy detection project, we immersed ourselves in a
plethora of research papers and embraced established methodologies to create an effective solution. Our
objective was to craft a model not confined to clinical settings but accessible for routine use, offering precise
identification of diabetic retinopathy (DR) through everyday technology.

Our achievement lies in the implementation of a state-of-the-art convolutional neural network (CNN) model
using Python, TensorFlow, and Keras. By harnessing the power of machine learning, specifically CNNs, we
tailored a solution adept at discerning subtle DR patterns from retinal images with remarkable accuracy.

In lieu of frontend frameworks like React.js, our focus remained steadfast on refining the core machine
learning model to ensure optimal performance. We dedicated extensive effort to dataset preparation and
exploration, employing Python libraries such as Pandas and scikit-learn for comprehensive data analysis and
model training.

The culmination of our efforts manifests in a robust CNN model capable of precise DR detection, a testament
to the efficacy of machine learning in healthcare applications. Our project underscores the potential for
leveraging advanced technologies to democratize medical diagnosis, ultimately enhancing patient care and
outcomes in the realm of diabetic retinopathy management.

G. Discussion
Our project, centred on the detection of diabetic retinopathy through deep learning methodologies, serves as
a significant advancement in proactive healthcare. Much like the drive for early detection and intervention in
medical conditions, our application aims to provide users with the tools to monitor and manage their ocular
health effectively.

While existing models for diabetic retinopathy detection often cater to clinical or industrial settings, our focus
is on creating a solution accessible to everyday users. We recognize the importance of personalized
healthcare solutions, particularly for conditions like diabetic retinopathy, which require regular monitoring
and early intervention.

The rising prevalence of diabetic retinopathy underscores the critical need for accurate and accessible
detection methods. By harnessing deep learning algorithms, our model delivers precise assessments of retinal
health, empowering individuals to take proactive steps in managing their condition and potentially preventing
vision loss.

Critics may question the accuracy of our model's estimations, but our rigorous approach to data analysis and
model training ensures reliable results. Our calculations are based on robust mathematical formulas,
providing users with valuable insights into their ocular health status.
The outcomes of our project reflect our commitment to improving healthcare accessibility and outcomes. Our
application offers a user-friendly interface, allowing individuals to effortlessly track and monitor their retinal
health. These insights enable informed decision-making and early intervention, ultimately enhancing the
quality of life for individuals affected by diabetic retinopathy.

In summary, our project represents a significant step forward in democratizing access to diabetic retinopathy
detection. By providing individuals with the tools to monitor their ocular health proactively, we aim to make
a meaningful impact on healthcare outcomes and improve the lives of those affected by this condition.

TABLE I. COMPARISON BETWEEN DIFFERENT DEEP LEARNING TECHNIQUES

S. No Model/Algorithm Strengths Weaknesses

1 Convolutional Captures spatial hierarchies, Limited sequential data handling

Neural Networks effective for image
(CNNs) recognition

2 Handles sequential data, Vulnerable to

Recurrent Neural retains context and vanishing/exploding gradients,
Networks (RNNs) memory slower training on long
sequences

3 Long Short- Overcomes vanishing Complex architecture, higher

Term Memory gradient problem, captures computational cost
(LSTM) long-term dependencies
Networks
4 Generative Produces realistic synthetic Training instability, mode
Adversarial data, high-quality samples collapse
Networks
(GANs)
5 Learns meaningful Limited by input data quality,
Autoencoders representations, prone to overfitting
reconstructs input data

V. CONCLUSION
Automated screening systems represent a pivotal advancement in the realm of medical diagnostics, offering
expedited assessments and substantial cost savings for healthcare providers while ensuring timely
interventions for patients. Within this context, the application of automated systems is particularly
transformative in the early detection of Diabetic Retinopathy (DR), a leading cause of blindness globally.
The distinctive stages of DR are characterized by discernible lesions on the retina, making it amenable to
automated analysis.

This article provides a comprehensive overview of recent advancements in automated systems tailored
specifically for the detection and classification of diabetic retinopathy, with a focus on leveraging deep
learning methodologies. Notably, deep learning techniques have emerged as a cornerstone in this domain,
owing to their capacity to autonomously learn intricate patterns and features from vast datasets.
A key aspect of this review is the elucidation of publicly available fundus DR datasets, which serve as
indispensable resources for training and validating automated systems. These datasets encompass a wide
spectrum of retinal images, annotated with clinically relevant labels, facilitating robust model development
and evaluation.

Central to the discussion are deep learning techniques, with Convolutional Neural Networks (CNNs)
emerging as a predominant choice among researchers. CNNs are adept at extracting hierarchical features
from images, enabling efficient classification and detection of DR-related abnormalities. The inherent
architecture of CNNs, characterized by convolutional and pooling layers, facilitates the automatic
identification of subtle lesions indicative of DR progression.

Moreover, this review delves into the nuanced methodologies and strategies within the realm of deep learning
specifically tailored for detecting and classifying DR. Techniques such as transfer learning, data
augmentation, and ensemble learning are elucidated for their efficacy in enhancing model performance and
generalization capabilities.

In summation, this article encapsulates the dynamic landscape of automated systems for diabetic retinopathy
detection, underscored by the transformative potential of deep learning methodologies. By synthesizing
recent advancements and elucidating key techniques, this review serves as a valuable resource for researchers
and practitioners alike, driving further innovation in the field of automated medical diagnostics.

ACKNOWLEDGEMENT
The creation of this research paper has been a collaborative endeavor, made possible by the unwavering
support, guidance, and encouragement of a remarkable group of individuals. Their contributions have been
essential, and we express our sincere gratitude to each one of them.

We extend our deepest gratitude to our esteemed project mentor: Mr. Harsha KG, whose expertise and guidance
played a pivotal role in shaping the proposed methodology. Dr. Kakoli Banerjee, our revered Head of the
Department (HOD) and project coordinator, also deserves special recognition for providing invaluable
comments that significantly enhanced the overall quality of the manuscript. Appreciation is extended to J.S.S
ACADEMY OF TECHNICAL EDUCATION, NOIDA, for providing us with a conducive environment and
the opportunity to pursue this research. Their support has been instrumental in the success of this endeavor.

Finally, we express our gratitude to our esteemed project team members collectively. Their collaborative spirit,
valuable suggestions, and guidance played a crucial role at various stages of completing this research paper.
Their unwavering assistance has been nothing short of invaluable, and we are truly grateful for their
contributions. This acknowledgment reflects the collective effort and synergy that has fueled the successful
completion of our review paper. Each individual mentioned here has made a unique and vital contribution,
contributing to the depth and quality of our work

REFERENCES
[1] S. Dua, N. Kandiraju, H. Thompson, “Design and implementation of a unique blood vessel Detection
Algorithm towards Early diagnosis of Diabetic Retinopathy”, 2005

[2] H Jiang ,K Yang, M Gao,D Zhang “An interpretable Ensemble Deep learning model for Diabetic
Retinopathy disease classification”,2019

[3] M Gandhi, R Dhanasekaran “Diagnosis of Diabetic Retinopathy Using Morphological Process and SVM
Classifier”,2013

[4] Shu-I Pao, Hong-Zin Lin, Ke-Hung Chein, Ming-Chena Tai, “Detection of Diabetic Retinopathy Using
Bichannel Convolutional Neural Network”,2020
[5] Deepthi K Prasad, Vibha L, and Venugopal K R, “Early Detection of Diabetic Retinopathy from Digital
Retinal Fundus images”, December 2015

[6] Amol Prataprao Bhatkar , “Detection of Diabetic Retinopathy in Retinal images using MLP Classifier”
2015

[7] Mahendran Gandhi, “Diagnosis of Diabetic Retinopathy Using Morphological Process and SVM
Classifier”, 2013

[8] Mayush Shelar, Sonali Gaitonde, Amudha Senthilkumar, Mradul Mundra and Anurag Sarang “Detection
of Diabetic Retinopathy and its Classification from the Fundus images” 2021

[9] Yadav, N., Banerjee, K., & Bali, V. (2020). A survey on fatigue detection of workers using machine
learning. International Journal of E-Health and Medical Communications (IJEHMC), 11(3), 1-8.

[10] Banerjee, K., & Bali, V. (2020). Design and Development of Bioinformatics Feature Based DNA
Sequence Data Compression Algorithm. EAI Endorsed Trans. Pervasive Health Technol., 5(20), e5.

[11] N, Yadav, K. Banerjee, & Bali, V. (2020). A survey on fatigue detection of workers using machine
learning. International Journal of E-Health and Medical Communications (IJEHMC), 11(3), 1-8.

[12] Abràmoff MD, Reinhardt JM, Russell SR, Folk JC, Mahajan VB, Niemeijer M, Quellec G (2010)
Automated early detection of diabetic retinopathy. Ophthalmology 177(6)

[13] Gargeya R, Leng T (2010) Automate identification of diabetic retinopathy using deep learning.
Ophthalmology 124(7)

[14] Wilfred Franklin S, Edward Rajan S (2014) Computerized screening of diabetic retinopathy employing
blood vessel segmentation in retinal images. Biocybern Biomed Eng 34(2)

[15] Antal B, Hajdu A (2014) An ensemble-based system for automatic screening of diabetic retinopathy.
Knowl-Based Syst 60

[16] Abràmoff MD (2013) Automated analysis of retinal images for detection of referable diabetic
retinopathy. JAMA Ophthalmol 131

[17] Revathy R, Nithya BS, Reshma JJ, Ragendhu SS, Sumithra MD (2020) Diabetic retinopathy detection
using machine learning. IJERT 9(6)

[18] Cree, M.J.; Olson, J.A.; McHardy, K.C.; Sharp, P.F.; Forrester, J.V. A fully automated comparative
microaneurysm digital detection system. Eye 1997, 11, 622–628. [CrossRef]

[19] Yun, W.L.; Rajendra Acharya, U.; Venkatesh, Y.; Chee, C.; Min, L.C.; Ng, E. Identification of different
stages of diabetic retinopathy using retinal optical images. Inf. Sci. 2008, 178, 106–121. [CrossRef]

[20] Nayak, J.; Bhat, P.S.; Acharya, U.R.; Lim, C.M.; Kagathi, M. Automated Identification of Diabetic
Retinopathy Stages Using Digital Fundus Images. J. Med. Syst. 2008, 32, 107–115. [CrossRef] [PubMed]

[21] Rosas-Romero, R.; Martínez-Carballido, J.; Hernández-Capistrán, J.; Uribe-Valencia, L.J. A method to
assist in the diagnosis of early diabetic retinopathy: Image processing applied to detection of microaneurysms
in fundus images. Comput. Med. Imaging Graph. 2015, 44, 41–53. [CrossRef] [PubMed]
[22] Gargeya, R.; Leng, T. Automated identification of diabetic retinopathy using deep learning.
Ophthalmology 2017, 124, 962–969. [CrossRef] [PubMed]

[23] Dutta, S.; Manideep, B.; Basha, S.M.; Caytiles, R.D.; Iyengar, N. Classification of diabetic retinopathy
images by using deep learning models. Int. J. Grid Distrib. Comput. 2018, 11, 89–106. [CrossRef]

[24] . Mansour, R.F. Deep-learning-based automatic computer-aided diagnosis system for diabetic
retinopathy. Biomed. Eng. Lett. 2018, 8, 41–57. [CrossRef] [PubMed]

[25] Qummar, S.; Khan, F.G.; Shah, S.; Khan, A.; Shamshirband, S.; Rehman, Z.U.; Khan, I.A.; Jadoon, W.
A deep learning ensemble approach for diabetic retinopathy detection. IEEE Access 2019, 7, 150530–
150539. [CrossRef]

[26] Gadekallu, T.R.; Khare, N.; Bhattacharya, S.; Singh, S.; Maddikunta, P.K.R.; Srivastava, G. Deep neural
networks to predict diabetic retinopathy. J. Ambient. Intell. Humaniz. Comput. 2020, 1–14. [CrossRef]