A SEMINAR REPORT

ON

DEVELOPMENT OF A DEEP LEARNING ALGORIHM FOR

AUTOMATIC DIAGNOSIS OF DIABETIC RETINOPATHY

Submitted to
The faculty of Engineering and Technology of
Kakatiya University, Warangal
In Partial fulfilment of the requirements
For the award of
Bachelor of Technology
In
Information Technology
By
T.Shashank
B21IT083

Under the Supervison of

Mr.M.Kishore

Assistant Professor
Dept of IT

DEPARTMENT OF INFORMATION TECHNOLOGY

KAKATIYA INSTITUTE OF TECHNOLOGY & SCIENCE, WARANGAL-15
2023 - 2024
 CERTIFICATE

This is to certify that T.SHASHANK(B21IT083) of V- Semester B.Tech. Information

Technology has satisfactorily completed the Seminar entitled “DEVELOPMENT OF A

DEEP LEARNING ALGORITHM FOR AUTOMATIC DIAGNOSIS OF DIABETIC

RETINOPATHY” in the partial fulfillment of the requirement of B.Tech degree during this

academic year 2023-2024.

SUPERVISOR CHAIRMAN, DSEC

Mr.M.Kishore Dr.T.Senthil Murugan

AssistantProfessor Assoc Professor &
Head, Dept. of I.T.
Dept of IT
 ACKNOWLEDGEMENT

I wish to take this opportunity to express my deep gratitude to all the people who have extended
their cooperation in various ways during my Seminar. It is my pleasure to acknowledge the help
of all those individuals.

I thank Dr. K. Ashoka Reddy, Principle of Kakatiya Institute of Technology & Science,
Warangal, for his strong support.

I thank Dr. T.Senthil Murugan, Professor & Head, Department of Information Technology for
her constant support in bringing shape of this Seminar.

I would like to thank my supervisor, Mr.M.Kishore, Assistant Professor of Information

Technology for her guidance and help throughout the Seminar.

In completing this Seminar successfully all our faculty members have given an excellent
cooperation by guiding us in every aspect. All your guidance helped me a lot and I am very
grateful to you.

T.Shashank

B21IT083

iii
 ABSTRACT

In this , we concentrate on employing deep learning techniques for the

classification of diabetic retinopathy stages and the identification of eye
laterality through funduscopic images. Diabetic retinopathy poses a
significant threat to vision, characterized by progressive damage to
retinal blood vessels. The conventional diagnosis by ophthalmologists
through funduscopic screening often leads to delays in reporting and
intervention, accompanied by substantial financial costs and the risk of
blindness. Leveraging a convolutional neural network (CNN) approach,
our model was trained on a Kaggle dataset comprising around 35,000
images for diabetic retinopathy diagnosis. The network exhibited a
commendable sensitivity of 80.28% and specificity of 92.29% when
validated on a dataset of approximately 53,000 images. Additionally, a
separate aspect of the study involved training the CNN on 8,810 images
to detect eye laterality. The outcomes of this training phase are yet to be
detailed. This research contributes to the advancement of automated
diagnostic tools, addressing critical challenges in diabetic retinopathy
diagnosis and timely intervention.

CONTENTS

iv
1. INTRODUCTION
1.1 Background
1.2 Objectives
2. DIABETIC RETINOPATHY
2.1 Definition and Characteristics
2.2 Impact on Vision
3. DIAGNOSTIC CHALLENGE
3.1 Time Delays in Traditional Screening
3.2 Financial Costs and Risk of Blindness
4. METHODOLOGY
4.1 Deep Learning Approach
4.2 Neural Network (CNN)
4.3 Dataset Description Convolutional
5. DIABETIC RETINOPATHY DIAGNOSIS RESULTS
5.1 CNN Performance Metrics
5.2 Sensitivity and Specificity
5.3 Validation Set Results
6. EYE LATERALITY DETECTION
6.1 CNN Training for Laterality Detection
6.2 Dataset for Eye Laterality
7. SIGNIFICANCE AND CONTRIBUTION
7.1 Advancements in Automated Diagnostics
7.2 Addressing Critical Challenges
8.CONCLUSION
9.REFERENCES

LIST OF FIGURES

Figure No. Name of the Figure Page No.

4.3.1 Sample Fundus Eye Images 4

5.3.2 Validation datasets 6

v
5.3.3 Confusion Matrix 7

6.2.4 Non-inverted,inverted image,Noisy Fundus Images 8

6.2.5 Laterality detection network 10

vi
 CHAPTER 1
INTRODUCTION

1.1 Background

The emergence of deep learning has revolutionized medical diagnostics,

particularly in addressing challenges posed by conditions like diabetic retinopathy.
This progressive disease affecting retinal blood vessels requires swift and accurate
diagnosis. Conventional methods, though crucial, suffer from delays and financial
burdens. Our study employs convolutional neural networks (CNNs) on a Kaggle
dataset of 35,000 images to offer a more efficient and cost-effective approach to
diabetic retinopathy diagnosis, aiming to contribute significantly to automated
medical diagnostics.

1.2 Objectives

our primary objectives are , to explore the application of deep learning,

specifically convolutional neural networks (CNNs), in classifying the stages of
diabetic retinopathy; and secondly, to investigate the CNN's efficacy in detecting
the laterality of the eye using funduscopic images. The imperative nature of timely
and accurate diagnosis for diabetic retinopathy, coupled with the limitations of
traditional screening methods, underscores the need for innovative and
technologically advanced approaches. By focusing on these objectives, we aim to
contribute valuable insights and potential solutions to enhance the efficiency and
effectiveness of diabetic retinopathy diagnosis, ultimately paving the way for
improved patient outcomes.

1
 CHAPTER 2

DIABETIC RETINOPATHY

2.1 Definition and Characteristics

Diabetic retinopathy is a progressive condition affecting the blood vessels in the

retina, primarily arising from chronic diabetes. It poses a significant threat to vision
and is characterized by the gradual deterioration of retinal vessels. As the disease
advances, it can lead to severe complications, emphasizing the critical need for
early and accurate diagnosis to mitigate potential sight-threatening consequences.

2.2 Impact on Vision

The impact of diabetic retinopathy on vision is profound, representing a serious

threat to visual health. As the condition progresses, it can lead to vision impairment
and, in severe cases, blindness. The damage to the retinal blood vessels disrupts the
normal functioning of the eye, underscoring the urgency for effective diagnostic
measures. Timely intervention is crucial to mitigate the risk of irreversible vision
loss associated with diabetic retinopathy.

CHAPTER 3

DIAGNOSTI CHALLENGES

3.1 Time Delays in Traditional Screening

Traditional screening methods for diabetic retinopathy are encumbered by inherent time
delays in reporting and subsequent intervention. The reliance on manual examination by

2
ophthalmologists, coupled with the growing demand for screenings, often results in delays that
can impact patient outcomes. The integration of automated and technologically advanced
diagnostic tools, as explored in this study, seeks to address and minimize these time delays,
enabling more timely and effective interventions for individuals with diabetic retinopathy.

3.2 Financial Costs and Risk of Blindness

The traditional methods of diagnosing diabetic retinopathy not only entail time delays but
are also associated with significant financial costs and heightened risks of blindness. The
financial burden arises from the extensive resources required for manual screenings by
ophthalmologists and subsequent medical interventions. Moreover, the inherent risk of blindness
underscores the urgent need for cost-effective and efficient diagnostic approaches. This study
delves into leveraging deep learning techniques to mitigate these financial costs and reduce the
peril of blindness through timely and accurate detection of diabetic retinopathy.

\ CHAPTER 4
METHODOLOGY

4.1 Deep Learning Approach

The methodology begins by elucidating the choice of a deep learning approach, specifically
Convolutional Neural Networks (CNNs). The discussion highlights the inherent capability of
CNNs to automatically learn hierarchical features from images, making them well-suited for
complex visual recognition tasks. Emphasis is placed on the ability of CNNs to capture spatial
dependencies in retinal images, crucial for diabetic retinopathy diagnosis.

3
4.2 Convolutional Neural Network (CNN)

This section provides a more detailed overview of the CNN architecture employed in the study. It
covers the fundamental building blocks, such as convolutional layers, pooling layers, and fully
connected layers. The discussion emphasizes how the architecture's depth and layer-specific
functions contribute to learning features from the funduscopic images.

Detailed information on the CNN's training parameters is included, covering aspects like learning
rate, batch size, and optimization algorithms. The rationale behind the chosen values is explained,
considering the trade-off between model convergence speed and avoidance of overfitting.

4.3 Dataset Description

FIG 4.3.1

EyePacs Dataset:

Expanding on the dataset, this section provides an in-depth description of the EyePacs dataset
from Kaggle. Details include the number of images, distribution across different stages of diabetic
retinopathy, and any preprocessing steps applied to enhance the quality of the images. Challenges
encountered in the dataset, such as variations in resolution, lighting conditions, and the presence
of artifacts, are discussed.

4
 CHAPTER 5

DIABETIC RETINOPATHY DIAGNOSIS RESULTS

5.1 CNN performance metrics

Sensitivity and specificity are pivotal metrics in assessing the efficacy of the Convolutional
Neural Network (CNN) for diabetic retinopathy diagnosis. These metrics quantify the model's
ability to correctly identify positive and negative cases, respectively.

5.2 sensitivity and specificity

Sensitivity (True Positive Rate):

Sensitivity, also known as recall, represents the proportion of actual positive cases correctly
identified by the CNN. It is calculated as the ratio of true positives to the sum of true positives and
false negatives.

Sensitivity= TruePositives+FalseNegatives/TruePositives
​

A high sensitivity value indicates the model's proficiency in capturing instances of diabetic
retinopathy, minimizing false negatives.

Specificity (True Negative Rate):

Specificity gauges the model's accuracy in identifying true negative cases. It is computed as the
ratio of true negatives to the sum of true negatives and false positives.

Specificity= TrueNegatives+FalsePositives/TrueNegatives
​

A high specificity value signifies the model's competence in correctly classifying non-diabetic
retinopathy cases, mitigating false positives.

5
 5.3 Validation Set Results

FIG 5.3.2

The validation set, comprising 53,126 images with diverse characteristics, serves as a robust
benchmark for evaluating the CNN's overall performance.

Confusion Matrix:
The confusion matrix delineates the model's predictions against the ground truth, providing a
granular view of classification outcomes.
Predicted0
TrueNegatives
FalseNegatives
​

Predicted1
FalsePositives
TruePositives
​

Predicted2
TruePositives
​

Predicted3
TruePositives
​

Predicted4
TruePositives
​

Overall Accuracy:
The overall accuracy assesses the CNN's ability to correctly classify images across all categories.
It is computed as the ratio of correctly classified images to the total number of images.

Accuracy= TruePositives+TrueNegatives/TotalImages

6
​ FIG 5.3.3

Incorporating these metrics provides a comprehensive evaluation of the CNN's diagnostic

performance for diabetic retinopathy.

and strict capital controls. It's been suggested that during the2012–2013 Cypriot financial
crisis bitcoin purchases rose due to fears that savings accounts would be confiscated or taxed.

CHAPTER 6

EYE LATERALITY DETECTION

6.1 CNN Training for Laterality Detection

Convolutional Neural Network Architecture:

The CNN architecture is specifically designed for laterality detection, incorporating

features aimed at distinguishing left and right eyes. Larger filter sizes, such as 7x7, are utilized
to capture directional patterns like blood vessel orientation.

Training Set Composition:

The training dataset is crucial, comprising 8,810 labeled retinal images with an equal
distribution of left and right eye samples to prevent bias during training.

Data Preprocessing:

7
Similar to diabetic retinopathy diagnosis, preprocessing involves cropping, resizing, and
normalization. However, during augmentation, images are not flipped to preserve the
directionality of anatomical features.

6.2 Dataset for Eye Laterality

FIG 6.2.4

Diverse Image Characteristics:

The dataset utilized for Eye Laterality Detection is meticulously curated to encompass a wide
array of retinal images, ensuring the model's robustness across diverse scenarios. Images are
sourced from EyePacs, a telemedicine platform dedicated to preventing vision impairment from
diabetic retinopathy.

Inclusion of Various Resolutions and Cameras:

Retinal images are captured using different variants of fundus cameras, each employing a unique
recording method. This leads to variations in image resolutions, ranging up to 4500 x 3500 pixels,
and diverse camera characteristics. The inclusion of images from various cameras allows the
model to adapt to the intricacies of different imaging devices.

Varied Angles and Lighting Conditions:

8
To simulate real-world conditions, the dataset includes images captured from different angles
and under various lighting conditions. This variation helps the model learn to identify laterality
features regardless of the orientation or lighting during image acquisition.

Noise and Artifact Levels:

Real retinal images often contain noise and artifacts that can complicate the laterality detection
task. The dataset intentionally incorporates images with varying levels of noise and artifacts,
challenging the model to discern relevant patterns amidst potential distractions.

Challenges in Generalization:

Given the diversity in the dataset, the model faces the challenge of generalizing across different
cases, including variations in camera types, resolutions, lighting, and the presence of noise. This
necessitates robust training strategies to avoid overfitting to specific characteristics of the
training data.

9
FIG 6.2.5

10
 CHAPTER 7

SIGNIFICANCE AND CONTRIBUTION

7.1 Advancements in Automated Diagnostics

Innovative Deep Learning Approach:

This study significantly contributes to the field of automated diagnostics by pioneering a deep
learning approach, specifically leveraging Convolutional Neural Networks (CNNs). The use of
CNNs, a cutting-edge subset of artificial intelligence, demonstrates a departure from traditional
methods, emphasizing the potential of neural networks in image-based diagnostics.

Real-Time Diabetic Retinopathy Diagnosis:

The application of CNNs enables the real-time diagnosis of diabetic retinopathy (DR) based on
funduscopic retinal images. By harnessing the power of deep learning algorithms, the system
achieves high sensitivity and specificity in detecting various stages of DR. This not only expedites
the diagnostic process but also offers a more efficient and accurate alternative to traditional
screening methods.

7.2 Addressing Critical Challenges

Timely Diagnosis to Prevent Visual Loss:

The study directly addresses the critical challenge of time delays in traditional screening for
diabetic retinopathy. By automating the diagnostic process through CNNs, the model facilitates
early detection of DR, a pivotal factor in preventing visual loss. The expedited diagnosis
contributes to timely intervention, potentially mitigating the risk of vision-threatening
complications.

Financial Costs and Risk Reduction:

11
Beyond the clinical aspects, the automated diagnostic system introduced in this study holds the
promise of reducing financial costs associated with traditional screening. The implementation of
an AI-driven approach minimizes the need for extensive human resources and streamlines the
diagnostic workflow. Additionally, by enabling early intervention, the system contributes to a
reduction in the long-term financial burden and the risk of blindness associated with DR.

Detection of Eye Laterality:

A noteworthy contribution lies in the incorporation of a second CNN for detecting the laterality of
the eye. This novel aspect sets the system apart, offering a comprehensive diagnostic solution.
The ability to discern between left and right eyes enhances the diagnostic report, providing a
holistic evaluation of a patient's ocular health.

CONCLUSION

In summary, the application of Convolutional Neural Networks (CNNs) in diabetic retinopathy

(DR) diagnosis represents a significant leap forward in addressing the drawbacks of traditional
screening methods. By achieving a delicate balance between sensitivity and specificity, the CNN-
based system provides a quick and accurate classification of DR stages, overcoming time delays
and financial constraints associated with manual evaluation.The incorporation of a second CNN
for eye laterality detection adds a unique dimension to the diagnostic process. This innovative
feature enhances the precision of the diagnostic report, contributing to a more comprehensive
understanding of a patient's ocular health.

By tackling critical challenges in DR screening, this research aligns with the broader trend of
integrating artificial intelligence into healthcare. The success of this deep learning model suggests
a transformative potential for automated diagnostics, promising efficiency and cost-effectiveness.

Looking ahead, the system's integration into clinical practice could redefine DR screening
standards. Ongoing research could explore additional parameters for improved accuracy, ensuring
the model's reliability in diverse clinical settings.

In conclusion, this research signifies a significant step towards accessible, timely, and accurate
DR screening, leveraging CNNs to revolutionize medical diagnostics. The combination of

12
technological innovation and a holistic approach positions this study as a promising advancement
in improving patient outcomes and alleviating the societal burden of diabetic retinopathy.

REFERENCES

 G.G. Gardner, D. Keating, T.H. Williamson, and A.T. Elliott, Automatic

detection of diabetic retinopathy using an artificial neural network: a
screening tool, British Journal of Ophthalmology 80 (1996), 940-944.
 D. Usher, M. Dumskyj, M. Himaga, T.H. Williamson, S. Nussey, and J.
Boyce, Automated detection of diabetic retinopathy in digital retinal
images: a tool for diabetic retinopathy screening, Diabetic Medicine 21
(2004), 84-90.
 H. Pratt, F. Coenen, D.M. Broadbent, S.P. Harding, and Y. Zheng,
Convolutional Neural Networks for Diabetic Retinopathy, Procedia
Computer Science 90 (2016), 200-205.

 J.W. Yau, S.L. Rogers, R. Kawasaki, E.L. Lamoureux, J.W. Kowalski, T. Bek, S.J.
Chen, J.M. Dekker, A. Fletcher, and J. Grauslund, Global prevalence and major risk
factors of diabetic retinopathy, Diabetes Care 35 (2012), 556-564.

13
14