EYE DISEASE CLASSIFICATION - USING DEEP

LEARNING

Project report submitted in partial fulfilment of the requirement for

the degree of Bachelor of Technology

in

Computer Science and Engineering

By

HARSHITA SHARMA (191234)

Under the supervision of

DR. ABHILASHA SHARMA

to

Department of Computer Science & Engineering and Information

Technology

Jaypee University of Information Technology Waknaghat,

Solan-173234, Himachal Pradesh
 DECLARATION
I hereby declare that the work presented in this report entitled “Eye disease classification
using Deep Learning” in partial fulfilment of the requirements for the award of the degree
of Bachelor of Technology in Computer Science and Engineering submitted in the
department of Computer Science & Engineering and Information Technology, Jaypee
University of Information Technology Waknaghat is an authentic record of my own work
carried out over a period from March 2023 to May 2023 under the supervision of Dr.
Abhilasha Sharma, Assistant Professor (SG), Department of Computer Science &
Engineering and Information Technology.

The matter embodied in the report has not been submitted for the award of any other
degree or diploma.

Harshita Sharma

191234

This is to certify that the above statement made by the candidate is true to the best of my
knowledge.

Dr. Abhilasha Sharma

Assistant Professor (SG)

Department of Computer Science & Engineering and Information Technology.

Dated:

i
ii
 ACKNOWLEDGEMENT
First and foremost, I want to give God the highest praise for his heavenly grace, which
enabled me to successfully complete the project work. Supervisor Dr. Abhilasha Sharma,
Assistant Professor (SG), Department of CSE Jaypee University of Information
Technology, Wakhnaghat, I sincerely thank you and express my sincere gratitude. My
supervisor's extensive knowledge and genuine interest in the topic of “Deep learning"
made them the ideal choice for this project. This endeavour was made possible by her
never-ending tolerance, academic direction, constant encouragement, frequent and
energetic supervision, constructive criticism, insightful counsel, reading numerous subpar
draughts, and corrections at all levels.

I would like to extend my sincere gratitude to Dr. Abhilasha Sharma of the Department of
CSE for her gracious assistance in seeing my project through to completion. A warm
welcome would also be extended to everyone who has directly or indirectly assisted me
in making this project successful. In this particular circumstance, I would like to express
my gratitude to all of the staff members—teaching and non-teaching—who have provided
me with practical assistance and made my task easier.

Finally, I must acknowledge with due respect the constant support and patience of my
parents.

Dated:

iii
 TABLE OF CONTENTS

DECLARATION i
PLAGIARISM CERTIFICATE ii
ACKNOWLEDGEMENT iii
LIST OF ABBREVIATIONS v
LIST OF FIGURES vi
LIST OF TABLES viii
ABSTRACT ix

Chapter-1 INTRODUCTION 1
1.1 Introduction 1
1.2 Problem Statement 5
1.3 Objectives 5
1.4 Methodology 6
1.5 Language Used 6
1.6 Organisation 8

Chapter-2 LITERATURE SURVEY 9

Chapter-3: SYSTEM DEVELOPMENT 13

3.1 Feasibility Study 13
3.2 Model Development 13
3.3 SFD Diagram and DFD Diagram 14
3.4 Data Set Used in the Major Project 19
3.5 Algorithm/Pseudo code of the model 20

Chapter- 4: PERFORMANCE ANALYSIS 36

4.1 Selection of Best Performance Parameters 36

Chapter-5: CONCLUSIONS 46
5.1 Conclusion 46
5.2 Application of the Project 46
5.3 Limitation of the Major Project 46
5.4 Future Work 47

REFERENCES 48

iv
 LIST OF ABBREVIATIONS

ANN Artificial Neural Network

CNN Convolutional Neural Network
CCNN Coherent Convolution Neural Network
DL Deep learning
DFD Data Flow Diagrams
DT Decision Tree
FC Fully connected
FCNN Fully Connected Convolutional Neural Networks
KNN K-Nearest Neighbours
NB Naive Bayes
ML Machine Learning
MLP Multi layer Perceptron
OS Operating System
RAM Random Access Memory
RF Random Forest
SFD System Flow Diagrams
SVM Support Vector Machine
UI User Interface
UX User Experience

v
 LIST OF FIGURES

Fig. No. Title Page

No.

Fig. 1.1 Normal Eye vs Cataract Eye anatomy 2

Fig. 1.2 Normal Eye VS Diabetic Retinopathy Anatomy 3

Fig. 1.3 Normal Eye vs Glaucoma Anatomy 4

Fig. 3.1 System flow diagram 15

Fig. 3.2 DFD Level 0 17

Fig. 3.3 DFD Level 1 17

Fig. 3.4 DFD Level 2 18

Fig. 3.5 Importing libraries 20

Fig. 3.6 Importing dataset 20

Fig. 3.7 Data preprocessing 21

Fig. 3.8 Splitting the dataset into training and test sets 21

Fig. 3.9 Data augmentation 21

Fig. 3.10 Building the model and compilation 22

Fig. 3.11 Model training for 150 epochs on train set 22

Fig. 3.12 Model Summary 23

Fig. 3.13 Test Accuracy 23

Fig. 3.14 Structure of CNN model 24

Fig. 3.15 Human vs Artificial Neuron 25

Fig. 3.16 ANN 3 layer Architecture 26

vi
Fig. 3.17 CNN Architecture 30

Fig. 3.18 Pooling operation example 31

Fig. 3.19 Mathematical equation for cross entropy loss function 33

Fig. 4.1 A simple binary class classification confusion matrix 37

Fig. 4.2 Code for plotting confusion matrix 38

Fig. 4.3 Confusion Matrix 38

Fig. 4.4 Model final accuracy 40

Fig. 4.5 Classification Report for the model 44

vii
 LIST OF TABLES

Figure No. Title Page No.

2.1 Comparison of various research papers 11

3.1 Relationship between human neuron and artificial neuron 26

viii
 ABSTRACT

Eye diseases play a significant role in causing visual impairments and even blindness
worldwide. To prevent eye illnesses, and such vision losses that are caused due to eye
diseases, also to improve the overall quality of life of patients, a method for early
detection is needed so that timely treatment of these diseases can be done. Currently,
diagnosis of eye diseases requires manual examination by a trained ophthalmologist.
However, this whole method can take up alot of time as well as can cost a lot of money,
also may not be readily available in certain regions.

In recent years, deep learning based techniques have proven tremendous promise in the
field of medicine and medical image analysis. This paper proposes a novel approach for
identifying various eye illnesses using deep learning based techniques such as CNN . The
proposed model consists of a convolutional neural network that can learn the features
from retinal images and classify the images into various eye illnesses.

The evaluation of the model is done on a large dataset of retinal images consisting of
various eye illnesses, including glaucoma, diabetic retinopathy, and cataracts. The model
proves to be better in terms of accuracy in prediction of these diseases and if trained on
with even larger dataset, the model can achieve even higher accuracies.

The results indicate that the supposed approach has the potential to revolutionise the
diagnosis of eye illnesses and improve access to eye care, particularly in remote regions
where trained ophthalmologists may not be readily available. The proposed approach
could also help minimise the workload of ophthalmologists and ameliorate the efficiency
of the healthcare system by enabling faster diagnosis and treatment of eye illnesses

ix
 Chapter-1 INTRODUCTION

The functionality of our eyes determines how well we can see the world around us. The
eyes are delicate organs that require routine care to be in good health and perform at their
peak. Our eyes need to be taken care of in a variety of ways, including protecting them
from harm, maintaining proper hygiene, and seeking expert help when necessary.

1.1 Introduction

There are various conditions that affect the eyes and may result in vision issues or even
blindness and such conditions are caused due to eye diseases. From infants to the elderly,
they can affect anyone, and they can be brought on by a number of things, such as
genetics, environmental influences, and way of life decisions. Refractive errors, cataracts,
glaucoma, dry eye, conjunctivitis, age-related macular degeneration (AMD), and
age-related macular edema are a few common eye conditions. Blurred or distorted vision,
light sensitivity, discomfort, redness, or discharge from the eyes are a few symptoms that
these disorders might produce. If a person notices any of these symptoms or changes in
their vision, it's crucial to consult a doctor right away so that they can get the necessary
diagnosis and therapy. Through proper eye care and routine eye exams, which can spot
early disease signs and help maintain their vision and general eye health, many eye
diseases can be avoided or managed. In order to correctly diagnose an eye disease, it is
important to analyse all the different range of symptoms.

Deep learning (DL), machine learning (ML), and artificial intelligence (AI) have the
potential to play significant roles in the prevention, diagnosis, and treatment of eye
diseases, according to studies and research done in recent years with the growth in
technology. These technologies have aided in the early diagnosis of diseases; AI and
machine learning (ML) can be used to analyse medical pictures like retinal scans to find
early indicators of eye ailments such diabetic retinopathy, glaucoma and age-related
macular degeneration. By being able to identify the characteristics and causes of eye
disorders, AI ML also plays a crucial role in individualised therapy, medicine
recommendations, and preventative care. This research studies and categorises and

1
assorts eye illnesses, including cataracts, glaucoma and diabetic retinopathy, using the
proposed model.

The following are the description of eye diseases that we are trying to classify using the
proposed CNN model:

(1) Cataract is a clouding of the natural lens in the eye, which can obstruct or haze
vision. It is a typical ailment that frequently develops as a normal part of ageing
but can also be brought on by other circumstances. As a person ages, protein
breakdown and clumping in the eye's lens can result in cloudiness and opaqueness
in the lens. The risk of getting cataracts can rise when people are exposed to UV
radiation from the sun, tanning beds, or other sources. Changes in blood sugar
levels in people with diabetes increase their chance of acquiring cataracts. The
risk of cataracts can be raised by several drugs, including corticosteroids. Despite
the fact that cataracts can arise for a number of reasons, surgery can be used to
treat them.

Fig 1.1 : Normal Eye vs Cataract Eye anatomy

2
(2) Diabetic retinopathy, is a disease which generally occurs to people with Type 1
and Type 2 diabetes, as the name suggests it is an impediment of diabetes, in
which blood vessels that are found in the light sensitive tissue at the back of the
eye are affected the most.

Although the exact cause of diabetic retinopathy is unknown, high blood sugar
levels over a longer time can harm the blood vessels in the retina. This harm can
result in blood vessels that leak or become blocked, which can impair vision or
even result in blindness. If a person doesn't care of his or her blood sugar levels or
does not properly manage or regulate their own blood sugar levels, or even high
blood pressure as well as raised cholesterol levels, and smoking can become other
major causes of diabetic retinopathy. Generally this aimenet is a greater risk for
persons with a longer time period of diabetes.

Fig 1.2: Normal Eye VS Diabetic Retinopathy Anatomy

(3) Glaucoma, it is the disease in which the optic nerve is affected. Optic nerve is the
nerve which connects the eye and the brain of a person. Although it can happen
with normal or low intraocular pressure (IOP), it is frequently related to elevated
IOP. Globally, glaucoma is one of the main causes of permanent blindness.
Although the precise cause of glaucoma is unknown, it is believed to be a result of

3
 a combination of genetic and environmental factors. The additional causes of
glaucoma are age, genetics, ocular injuries or surgeries, and long-term
corticosteroid use. Regular eye exams are essential for detecting and tracking the
progression of glaucoma since early diagnosis and treatment can reduce the risk of
vision loss and improve results.

Fig 1.3: Normal Eye vs Glaucoma Anatomy

Traditional methods for predicting eye diseases can be time-consuming, subjective, and
subject to human error, such as manual examination and picture interpretation. Inaccurate
or inconsistent outcomes might result from this, as well as delays in diagnosis and
treatment. On the contray, artificial intelligence (AI), deep learning-based techniques, can
offer several benefits over traditional techniques, as they are quicker and more efficient,
show improved accuracy, aid in early detection and individualised treatments, and AI
algorithms can provide consistent and objective analysis of medical images, lowering the
risk of human error and variability. Therefore, in this study, creation of a deep learning
based CNN model is presented, that will aid in the classification of eye diseases.

4
1.2 Problem Statement

People frequently neglect their eye irritations or changes until they become urgent when it
comes to the human eye, unless there is a vision issue. The most probable cause for this
issue worldwide is - absence of such an application which is explicitly designed for
identifying prior or premature signs of disorders that occurs in the eye, if such an
application must have existed, then the patients could have got appropriate information
regarding such ailments and they could have better assessed treatment for the same. This
research’s methodology addresses the above problem by enlightening the end user to
examine his own eye to educate them more about their eye ailment and its current state.
Conventional procedures for diagnosing eye problems are expensive, time-consuming,
and call for highly specialised knowledge. CNNs have demonstrated promising results in
a variety of image-based medical applications, including the diagnosis of eye disorders,
thanks to breakthroughs in deep learning techniques. However creating a CNN-based
system for diagnosing eye disorders that is accurate and effective involves overcoming a
number of obstacles, including choosing the right architecture, correcting class
imbalances, coping with noisy and low-quality images, and training the model on a huge
and varied dataset. Consequently, the goal of this research is to create a solid and
trustworthy CNN-based system for the rapid and widespread screening and monitoring of
patients in clinics and hospitals, as well as the early detection and diagnosis of a variety
of eye disorders.

1.3 Objectives

The three main objectives of this project are:

● The goal of the project "Eye diseases classification using deep learning" is to
develop automated and precise methods for the early detection and diagnosis of
different eye disorders utilising medical images such as eye retinal images.

● The objective is to lessen ophthalmologists' workload, enhance the precision and

effectiveness of screening and diagnosis, and eventually enhance patient
outcomes.

● Using Convolutional neural networks (CNNs), which is a type of deep learning

technique, this research will be demonstrating three different types of eye diseases
and inform patients beforehand so they can take required treatment at the right

5
 time.

● Using deep learning techniques, to result in more effective patient management

and treatment plans.

1.4 Methodology

The main methodology of this project revolves around using CNN for successful
identification of four different categories of eye diseases.

six main steps that aides in this are :

(1) Collecting the database.

(2) Preprocessing the dataset.

(3) Splitting the dataset into training and testing.

(4) Making of the CNN model.

(5) Model evaluation and results.

1.5 Language Used

Python 3 - The main reasons for the use of this programming language for the
implementation of the project are stated below:

Simple and dependable:

Python generates readable and condensed code. Python's simplicity enables programmers
to create solid solutions because machine learning, deep learning and artificial
intelligence are based on complex algorithms and adaptable workflows. Developers can
concentrate solely on resolving an DL, ML issue rather than on the language's finer
technical details.

Extensive selection of libraries and frameworks:

Many modules and frameworks are available for Python. Programmers use these
frameworks and libraries to speed up development. Here are a few examples:

Deep learning frameworks include Keras, TensorFlow, and Scikit-learn.

● For scientific computing and data-analysis we can use numpy.

6
● Matplotlib is used for creating static, animated, and interactive visualisations.

● Pandas is a data analysis tool that can be used for a variety of purposes.

● Keras is used for adding layers and neurons in our CNN model.

● Tensorflow is used for training and inference of our deep neural network model.

Independency of platform

A programming language or framework that is "platform independent" enables

developers to write things on one system and then use them with little (or no)
modification on another. The popularity of Python is due to the fact that it is a
platform-neutral language. It can be used with a number of operating systems, including
Windows, Linux, and macOS. Python software can be disseminated and utilised without
the requirement for a Python interpreter by using Python code to create standalone
executable programmes for the majority of popular operating systems.

This study has employed Python as a programming language for this project. Python
contains a large number of deep learning libraries, which expedites method development
and makes it easier. These libraries include a range of deep learning techniques as well as
GPU support to reduce training times, such as Keras, TensorFlow, etc.

The proposed model is built on the Jupyter notebook. Jupyter notebook provides a
server-client programme that enables users to edit and run all their notebook papers
through a web browser. This Jupyter notebook can be made to run locally on the desktop
without any internet connection and it can be deployed on a remote server as well that can
be made accessible via the internet.

The Jupyter notebook provides a computational engine that runs the code in a notebook
document and this engine is known as kernel. Python code is generally executed by the
ipython kernel, which is mentioned in the detailed documentation, the link for the which
is provided in the references. Official kernels are available for a wide variety of other
languages.

Whenever a person opens anaconda and Jupyter notebook, then the notebook document’s
related kernel is automatically launched. To generate results, the person has to execute
run the notebook, which can be done cell by cell or through the menu Cell-> Run All

7
option and this will execute the file. Depending on the type of computations, the kernel
may use lot of CPU and RAM, and the RAM is not released until the kernel is terminated.

1.6 Organisation

The organisation of the report is as following:

In chapter 1, the introduction of eye diseases and factors causing these diseases has been
discussed. The chapter presented the importance of eye disease prediction and
identification and research methodology for eye diseases prediction

In chapter 2,the literature survey has been discussed. This chapter gives a brief
introduction and an abstract of various researches which were previously carried out by
other researchers around the globe for the same, and brief about their techniques and
approaches to their limitations. Also a comparison table is made for the same purpose.

In chapter 3,the system development has been discussed. This chapter discusses the DL
algorithm used and a brief explanation about artificial neurons, forward and backward
propagation and CNN. How the data set is first imported and then split into training and
testing sets and then the CNN model is trained on them for the purpose of testing new
dataset and then results are predicted. Using an accuracy metric, performance analysis is
depicted.

In chapter 4, this chapter has discussed the performance of the model for which an
accurate use of accuracy metric is done, this chapter also depicts here the losses this
model made while training and how it achieved the final accuracy after several rounds of
forward and backward propagation.

In chapter 5, in this chapter the conclusion of this project with the discussion on the
algorithm used to build this project and proposed solutions with the brief of existing
limitations of this project and future scope of the project, are being discussed.

8
 Chapter-2 LITERATURE SURVEY

Mohammad Monirujjaman Khan et al. (2022) the study presented in this research puts
forward a fanatical ocular detection method which is mainly based on deep learning. For
this study, they picked a cutting-edge image classification algorithm like VGG-19 and
trained the model on the ODIR dataset, which contains 5000 photos. The photos consist
of eight different classes of the fundus images. Many varieties of ocular illnesses are
constitued by these classes. Nevertheless the dataset for these classes they picked was
very imbalanced. And to solve this problem they sought to turn this multiclass
classification problem into a binary classification problem and take the same number of
images for both categories in order to solve this challenge.The accuracy of the VGG-19
model was 98.13% for the normal (N) versus pathological (M) myopia class, 94.03% for
the normal (N) against cataract (C), and 90.94% for the normal (N) versus glaucoma
classes (G). [1]

Sushma K Sattigeri et al. (2022) their study provides a novel approach to provide an
automated eye illness identification model utilising visually discernible symptoms. It uses
deep learning techniques like convolution neural networks and digital image processing
techniques like segmentation and morphology. Using the suggested procedure, four eye
diseases—crossed eyes, bulging eyes, cataracts, uveitis, and conjunctivitis—are examined
and grouped.The dataset for the purpose was collected partially from kaggle and partially
from local optometrist's assistance. Two different models are used to train for single-eye
and dual-eye pictures. Using two-eye pictures, one model forecasts disorders including
crossed eyes and bulging eyes. The eye configuration of single eye achieved accuracy of
96% and that of two eye achieved 92.31%.[2]

Nouf Badha, et al.(2022). This study focuses on identifying eye infections of Glaucoma
disease and for this purpose the authors of this paper used a variety of machine learning
(ML) classifiers, including Support Vector Machine (SVM), K-Nearest Neighbors
(KNN), Naive Bayes (NB), Multi-layer perceptron (MLP), Decision Tree (DT), and
Random Forest (RF), as well as Deep Learning (DL) models, such as Convolutional
Neural Network (CNN) based on Resnet152 model.. On the dataset for Ocular Disease
Intelligent Recognition, the proposed technique is evaluated. The collected findings

9
demonstrated that, in comparison to the other ML classifiers, the RF and MLP classifiers
had the best accuracy of 77%. For the identical task and dataset, the deep learning model
(CNN model: Resnet152) offers an even higher accuracy of 84%. [3]

Grzegorz Meller,et al.(2020). The main objective of this project is to delineate a

convolutional neural network-based advanced classification model that can recognise eye
disorders from images of eye-fundus. This model will be able to distinguish various eye
diseases. It made use of the "Ocular Disease Intelligent Recognition (ODIR)" dataset, a
structured ophthalmology database of 5,000 individuals with age, colour fundus pictures
of the left and right eyes, and doctors' diagnostic keywords. Patient labels include normal
(N), diabetes (D), glaucoma (G), cataract (C), AMD (A), hypertension (H), myopia (M),
and various diseases/abnormalities (O).Using a relatively simple network and only
images labelled as N (normal) or C (cataract), a simple CNN model was trained to
determine whether an eye had a normal fundus or a cataract. After 12 epochs , the model
achieved 93% accuracy, and when the experiment was run on the entire ODIR dataset,
50% validation accuracy was attained. [4]

Hao Gu et al. (2020) In this paper demonstrates a design of a novel approach that is based
on hierarchical deep learning network is proposed. This network aids in identification of
corneal diseases and is composed of a family of multi-task and multi-label learning
classifiers which represents different levels of eye diseases which are indeed derived from
a predefined hierarchical eye disease taxonomy.Then, in order to understand the
fine-grained heterogeneity of eye illness traits, they presented a multi-level eye
disease-guided loss function. Using a retrospective dataset of 5,325 ocular surface
pictures, the proposed algorithm was trained directly on an end-to-end basis. In a
prospective, a dataset which is majorly clinic-based consists of 510 outpatients who are
freshly enrolled with diseases of infectious like keratitis, non-infectious keratitis, corneal
dystrophy or degeneration, and corneal neoplasm, the algorithm's performance was lastly
evaluated against 10 ophthalmologists. For each form of corneal disease, the algorithm's
area under the ROC curve was over 0.910, and generally speaking, its sensitivity and
specificity were on par with or better than the average values of all ophthalmologists.[5]

Pawan Kumar Upadhyay et al.(2022).The automatic detection of retinal disorders using

OCT pictures is one such deep learning application that is covered in this work. The new
method for detecting retinal diseases has been proposed, and it effectively handles the
four-class problem by differentiating between the images of choroidal neovascularization

10
(CNV), diabetic macular edema (DME), DRUSEN, and NORMAL class. The pre-trained
VGG-16 model was pruned in order to build more precise neuron connections, develop
more optimised layers, and perform important tasks for occult disease detection in the
proposed network (CCNN). The model achieved 97.16% accuracy.

Author(s)/Referen Journal/ Methodology Disadvantage

ce
conference,
year

Mohammad 2022 A deep learning based The dataset with the

Monirujjaman approach- VGG-19 is used classes was highly
Khan at el/1 to classify ODIR dataset unbalanced.
into multiple classes.
Dataset has 5000 images of
fundus

Sushma K 2022 A deep learning based The neural classifier model

approach -CNN is used to was fed images for
Sattigeri et al/2
classify a dataset picked training with different
from kaggle of 5 types of informational indexes
eye diseases. from images.

Nouf Badha et al/3 2022 ML based algorithms such The machine learning
as KNN, SVM, NB, DT, RF models picked for the
are used to detect human eye classification of ODIR
infections of Glaucoma images performed poorly
disease. Later DL based and cannot be relied on for
CNN Resnet152 model is real medical diagnosis..
used for better accuracy.

11
Grzegorz Meller 2020 Deep learning based CNN Many image variants or
model is used for the classes could not have
et al/4
purpose of classifying ODIR their representation in the
images dataset. validation set of the model.

Hao Gu et al/5 2020 A novel hierarchical deep The algorithm is trained

learning network - CNN is exclusively to detect the
used for the purpose of ocular surface diseases
classifying ocular surface listed in the methods
images consisting of corneal
section, so it may miss
diseases
eye diseases with a
normal ocular surface.

Pawan Kumar 2022 A Coherent Convolution There are cases where

Neural Network (CCNN) is models fail to detect the
Upadhyay/ 6
used for the purpose of retinal disease accurately
classifying four different as the model accuracy
classes of eye diseases. relies on tuning of
hyperparameters.

Table 2.1 : Comparison of various Research Papers

12
 Chapter-3: SYSTEM DEVELOPMENT

3.1 Feasibility Study

The purpose of the feasibility study is to assess if the project is feasible—that is, whether
it can be executed effectively and conveniently—from an operational, technical,
economic, and organisational perspective.

1. Economic Feasibility:

The project was created exclusively with open-source software and libraries. The creators
of these libraries have all made them freely accessible online. As a result, there are no
costs associated with project development.

2. Operational Feasibility:

This is a feasible operating concept that could help prediction of eye diseases.

1.5.1 Software Requirements:

Python 3 was used to build the project for the reasons outlined above. Kaggle was used to
collect the dataset. GitHub and StackOverflow were consulted as resources in the case
that programming grammar errors occurred. An open-source, high-GPU Jupyter
Notebook or an interface similar to that - called Google Colaboratory. A completely
cloud-based, setup-free Jupyter notebook environment, Google Colab is available for
free. With Colab, users may access advanced computational resources from the browser,
write and run code, store and share analysis, and more. Python data analysis scripts can
be iterated and written using Jupyter Notebook. Code lines can be written and executed
one at a time rather than having to write and rewrite an entire programme.

3.2 Model Development

The process of creating an DL model includes gathering of data from various trustworthy
sources and processing it by an DL model to make it suitable for modelling, designing the
model, developing the model's algorithm, and calculating performance metrics, selecting

13
the best performance parameters for the model evaluation. Once the model is put into use,
model maintenance becomes extremely important.

3.3 SFD Diagram and DFD Diagram

3.3.2 System Flow Diagram

System flow diagram is one of the very famous ways of graphical depictions in
which data flow in any system is designed and it is based on software engineering
practices being followed by developers and researchers. The diagram shows a
series of steps that describe where the input and output for the system come from
and where they go to as part of the diagram. The graphic makes it possible to
manage how the system decides which events to process and how data enters the
system. The system flow diagram, which leaves out the minor components and
shows the important components of the system in order of importance, is
essentially a visual depiction of data flow.

With its visual representations and sequential overall process, it offers a lot of
advantages to both the user and the organisations.

● Different engineers and programmers utilise system flow diagrams to

visualise decisions and potential results from the system, assisting system
designers in designing in accordance with it.
● System flow diagrams employ symbols that are very simple as well as
effortless for end users to understand, showing how the process works and
where choices are made.
● The diagrams assist in visualising the key inputs for a system that
considers and generates the key output utilising key steps, which is
advantageous for the organisation for their system requirements.
● Many well-known organisations utilise system flow diagrams for business
objectives since they are relatively simple to grasp thanks to the use of
symbols, which removes the difficulty of utilising complicated processes
or codes that are typically seen using a sophisticated algorithm or in
pseudo-codes.

14
● A keyboard, mouse, or other types of sensors serve as a flow diagram's
data inputs, while a monitor or printer serves as the output device through
which we see the results.
● System flow diagrams are applied in many real-world applications, such as
automatic washing machines and airline reservation systems, to efficiently
handle all system processes.

The following is the SFD of this study:

Fig 3.1: System flow diagram

15
3.3.3 Data Flow Diagrams

The Data flow diagram (DFD) is a prototypical visual representation which tracks
and represents the movement and overall flow of data through a proposed system
of a study or research or a new method. An orderly and comprehensible DFD can
graphically represent the pertinent quantity of the system demand. It can be drawn
manually, or automatically with the advent of AI based softwares available online,
or both.
Data flow diagrams denotes the flow of information, by depicting the entrance of
the the data to the system, as well as its exit from the system, it also denotes the
system or the method which is generally a process that modifies that data, as well
as depicts the place where the information is kept.
One of the major objective of DFD is to outline the boundaries and scope of the
system as a whole. It can be best utilised in cases of system redesign or in cases of
understanding the foundation of the system from the user point of view, as it can
act as a communication tool between the user that is any participant and a system
analyst or architect as it will serve them information in the sequence required.
Many a times a DFD is also coined as a bubble chart or data flow graph.
Each level of abstraction for a system or piece of software can be performed using
the DFD. In reality, DFDs can be divided into stages that signify a growth in the
flow of information and functional granularity. In DFD, levels are denoted by the
numbers 0, 1, or 2. The data flow diagram in this study has made use of three
main levels: 0-level DFD, 1-level DFD, and 2-level DFD.

3.3.3.1 Data Flow Diagram - Level 0

The complete programme need is shown as a single bubble in the context

diagram, sometimes referred to as the core system model, with input and output
data marked by incoming and exiting arrows. The system is then broken down and
represented as a DFD with several bubbles. The system components that each of
these bubbles represents are then broken down and documented as
ever-more-detailed DFDs. To ensure that the programme at hand is thoroughly
grasped, this step may be performed as many times as necessary. Maintaining the

16
amount of inputs and outputs between levels is crucial; this idea as levelling was
given by DeMacro.

The following image depicts the Level 0 DFD for this study:

Fig 3.2: DFD Level 0

3.3.3.2 Data Flow Diagram - Level 1

The key processes identified in the level 0 DFD are divided into sub-processes at
this level to give a more in-depth picture of the system. On the level 1 DFD, each
subprocess is represented as a distinct process. Additionally, each sub-process’s
data flows and data stores are displayed.

The level 1 DFD for this study is as shown:

Fig 3.3: DFD Level 1

17
 3.3.3.3 Data Flow Diagram - Level 2

By further subdividing the sub-processes mentioned in the level 1 DFD, this level
offers an even more in-depth look at the system. On the level 2 DFD, each
subprocess is represented as a distinct process. Each sub-process's data flows and
data stores are also displayed.

The level 2 DFD for this study is as shown:

Fig 3.4: DFD Level 2

3.4 Data Set Used in the Major Project

For the purpose of this study, the dataset used for model training is picked up from
kaggle. The dataset consists of four categories of retinal images out of these, three
categories are of different eye diseases - cataract, glaucoma and diabetic retinopathy and
the fourth category is of normal retinal eye image. These images are collected from
various sources like IDRiD, Oculus recognition, HRF etc. There are a total of 4217
images, out of which 1038 images are of cataract, 1098 images of diabetic retinopathy,
1007 images of glaucoma, 1074 of normal. The retinal images consist of left and right eye
images in almost equal proportions. The total size of the dataset is 772.22 MB

18
The Kaggle.com site's dataset was gathered there. Kaggle is a branch of Google and is a
website community for data scientists and ML practitioners. Customers can browse and
post data sets on Kaggle, explore and build models, interact with other data scientists, and
compete to solve any data science challenges. Because kaggle is a reputable website with
over a million users globally, one can be sure that the dataset is unique and authentic.

3.5 Algorithm/Pseudo code of the model

In order to analyse and classify the eye diseases DL based CNN model has been
employed, successfully built using tensorflow aided keras library. The following is the
proposed model’s pseudocode :

3.5.1 Pseudocode of the proposed method:

i. Importing the libraries

ii. Importing dataset

iii. Data Preprocessing all the images are reshaped and recolored into one uniform size
of 180px X 180px and RGB colour respectively.

iv. Splitting the dataset into training, validation and test sets The retinal images are
split into 80% training and 20% test sets precisely 3374 retinal images into training set
and 843 retinal images into test set.

v. Data augmentation: Random rotation, contrast and zooming is done

vi. Training the model: Keras sequential method is used for the purpose of adding
layers.

vii. Evaluation of model

19
The following is the screenshot that depicts importing of all libraries:

Fig 3.5: Importing Libraries

The following is the screenshot that depicts importing of dataset :

Fig 3.6: Data Preparation and importing the dataset

20
The following is the screenshot that depicts preprocessing of data:

Fig 3.7: Data Preprocessing

The following figure depicts the split in dataset into training and testing sets

Fig 3.8: Splitting the data set into training and testing sets

The following is the depiction of data augmentation:

Fig 3.9: Data Augmentation

21
The following figure shows model training and compilation:

Fig 3.10: Building the model, and compilation

The following shows the model is trained on training data on 150 epochs.

Fig 3.11: Model training for 150 epochs on train set

The following figure shows model summary:

22
 Fig 3.12: Model Summary

The following depicts the test accuracy:

Fig 3.13: Test Accuracy

23
The following figure shows the overall architecture of training of model’s subsequent
layers:

Fig 3.14: Structure of CNN model

24
3.5.2 Data Augmentation and its importance

Data augmentation is a very impressive and efficient technique which makes use of a
method that creates additional data points from the existing data; this approach helps in
enhancing the overall amount of data.For the purpose of creating new data points in the
deep learning models, this approach makes some minor changes in the section already
existing imported dataset. The process of artificially deriving fresh data from previously
collected training data is known as data augmentation. Techniques include cropping,
cushioning, flipping, rotating, and resizing.This method helps in strengthening the overall
performance of the model as well as problems of overfitting and missing values of data
are avoided very efficiently. For the purpose of training the proposed model data
augmentation is performed, using the keras libraries experimental library - preprocessing
method. And random rotation, contrast and zooming techniques are employed.

3.5.3 Artificial Neural Network

The phrase "artificial neural network" refers to a branch of artificial intelligence that was
inspired by biology and is based on the brain. Artificial Neural networks is a man made or
as the name suggests artificial computational network which is based on the biological
neural network of the humans, which are the basic building blocks of structure of human
brain. This neural network works in parallel design constructs as that of the human brain,
as an ANN also make use of neurons that are further linked to each other in different
layers of a network. In AI,these neurons are termed as nodes. The following image
depicts the human neuron and the artificial neuron:

Fig 3.15: Human Vs Artificial Neuron

25
The following table shows the relationship between biological neuron and an artificial
neuron:

Human Neuron Artificial Neuron

Cell nucleus Nodes

Dendrites Inputs

Synapse Weights

Axon Output

Table 3.1: Relationship between human neuron and artificial neuron

A neural network can be visualised as a very large number of artificial neurons, also
known as nodes or units are made to stack in a form of hierarchy of layers. There are
several different kinds of layers that can be found in an artificial neural network and can
be better understood as the following diagram.

Fig 3.16 : ANN 3 layers architecture

26
Input layer : As the name of the layer itself implies it is the place where the user given
inputs or the programmer provides the input data for further processing and the input can
be of different formats.

Hidden layer : It is the place where all the computations are performed, in order to make
all the necessary processing and feature extraction, to understand the hidden and buried
patterns in the input data. This layer lies between the input layer and output layer of the
ANN.

Output layer : The function of this layer is to receive the various outputs which are
generated by the hidden layer by several processing and then to further communicate
these output nodes to some activation function and then as a result. The artificial neural
network computes some weighted total of all the inputs and then also incorporates a bias,
whenever ann is fed with inputs. Further there is presence of a transfer function which is
used to visualise all this processing and this computation for the purpose of producing the
output, it then passes the weighted total as an input to an activation function. The purpose
of the activation function is to decide upon the activation of a node, and the output layer
only receives those nodes which are activated by the activation function. There are
various different kinds of activation functions and they are used according to the type of
classification or regression problem they are used in.

3.5.4 Forward and Backward Propagation

Forward propagation - as the name implies, forward propagation feeds the network with
input data moving forward.All the inputs moves in forward direction, from input layer the
hidden layer receives the input nodes and then process them in accordance with the
activation function and further forwards these nodes or transfers these nodes to the
upcoming immediate next layer..

In ANN emphasis is laid on supplying all the input nodes only in forward direction
because it is the only way the output will get produced in such an architecture. If the data
starts flowing in some other direction then this might create a cycle or a loop and will
hinder the formation of output nodes. Therefore to prevent this the data must flow only in
forward direction during the process of output generation. This is what is implied by

27
forward propagation which is also referred to as feed forward propagation, as it helps in
spread of information.

Processing takes place at each neuron in a hidden or output layer in two steps:

Preactivation: It is the linear transformation of weights with respect to the inputs

available. It is a weighted sum of inputs. In preactivation the neurons decide upon the
activation of input nodes in intermediate layers of ANN architecture. And the function is
the same, to decide whether or not a node should be activated and transmitted to layers
ahead which is based on the aggregated total and the type of activation function used.

Activation: The activation function receives the estimated weighted sum of the inputs. It
is basically a mathematical function which gives the network more nonlinearity. The
sigmoid, hyperbolic tangent(tanh), ReLU, and Softmax activation functions are the four
most widely used and well-liked activation functions.

One of the major foundations and back support of neural networks is backpropagation.
In this method all the weights of the neural network are adjusted or tweaked again on the
basis of the error rate which is record at the end of preceding epoch or iteration during
training, and by properly tuning and tweaking of the weights, the error rate recorded can
be lowered to much levels and overall model’s reliability is enhanced. This also aids in
ultimate broadening of the model’s applicability.

In neural networks, backpropagation is the abbreviation for "backward propagation of

errors." It is the most common technique used for developing and training the whole
neural network. This method helps in calculation of the gradient of loss function with
respect to each weight in the neural network.

The following are the steps for backpropagation:

Step 1: Firstly all the input nodes must enter through the preconnected path.

Step 2:With the help of the weights all the input nodes are modelled. Generally the
weights initially are selected at random.

28
Step 3: Determine each neuron's output from the input layer through the hidden layer and
out to the output layer.

Step 4: Determine the output error

Backpropagation Error = Actual Output - Desired Output

Step 5: Return from the hidden layer to the output layer and change the weights to lower
the error.

Step 6: Continue the procedure until the intended result is obtained.

3.5.5 Convolutional Neural Networks

Convolutional neural networks (CNNs) are a subset of deep learning neural networks that
are frequently employed in pattern recognition, computer vision, and image
processing.Since CNN is based on image processing, one must understand what exactly
an image is. A digital image is a binary representation of any visual data which is made
up of an arrangement most similar to a grid. Image can be thought of as a grid of pixels,
and each and every value of the pixel indicates some information like the colour it is and
how bright it is. Convolutional, pooling, and fully linked layers are some of the layers
that make up CNNs. In mathematics, a procedure called convolution on two functions
creates a third function that expresses how the shapes of the first two are changed. CNN
operates on similar concepts. The role of CNN or ConvNet is to transform the images into
a format that is simpler to interpret without sacrificing details that are essential for
making accurate predictions.

When humans first see an image, their brains process a huge amount of information. They
as a whole system encompass practically the whole vision field, like the human brain,
each neuron has a specific receptor and is linked to other nodes in order to transmit the
information gathered. Analogue to how each neuron in the biological vision system reacts
to stimuli only in the restricted region of the visual field known as the receptive field,
each neuron in a CNN analyses data only in its receptive field. The layers first pick up on
lines, curves, and other simpler patterns before moving on to more complex patterns like
faces and objects. By using a CNN, one can enable human like sight for computers.

29
A CNN typically has three layers: starting from convolutional layer then next subsequent
laye is a pooling layer and there can be repetitions of such layers and finally a fully
connected layer

Fig 3.17: CNN Architecture

3.5.6 ReLu in Hidden layers

A frequent form of activation function in deep neural networks, such as Convolutional

Neural Networks (CNNs), is the ReLU (Rectified Linear Unit) layer. The element-wise
rectification process, which is applied by the ReLU function, simply outputs the input
value if it is positive and 0 otherwise. In comparison to sigmoid, ReLU boasts of training
sophisticated models and neural networks with constant values. Relu shows greater
effectiveness: ReLU has an advantage over other sigmoid functions in that it runs
significantly faster and with a lot lower run time max(0,f). This is due to the fact that it
does not suffer from the vanishing gradients problem.

30
3.5.7 Pooling

The pooling layer acts as a stand-in for the output of the network at particular locations
by computing a combined statistic from the nearby outputs. This helps to reduce the
spatial size of the representation, which minimises the volume of processing and weights
required.

The pooling procedure is applied to each layer of the representation independently.One of

the pooling operations is in addition to the average of the rectangular neighbouring
values in matrix and the L2 norm of the rectangular neighbourhood, is based on the
distance that exists from the central pixel. The most used approach is max pooling, which
reports the largest output from the neighbourhood.

Fig 3.18: Pooling Operation Example

3.5.8 Flattening

A Convolutional Neural Network (CNN) technique known as flattening converts the

output of the convolutional and pooling layers into a one-dimensional vector that is sent
to a fully connected layer for additional processing. The output of a convolutional or

31
pooling layer is typically a three-dimensional tensor with dimensions (height, width,
depth), which is further turned into a one-dimensional vector in order to send it to a fully
connected layer. By simply stacking the values of each channel on top of one another, the
flattening operation converts the tensor into a long vector.

3.5.9 Full Connection

In typical Fully Connected Convolutional Neural Networks (FCNN), all neurons in this
layer are fully connected to all neurons in the layer before it and the layer after it. This
connection enables the computation of the units of this layer using the standard matrix
multiplication and bias effect method. The representation that exists across the input and
the output is mapped using the FC layer.

3.5.10 Activation Function For Final Layer

A Convolutional Neural Network (CNN)'s final layer's activation function is determined

by the type of problem being solved. The final layer of a CNN commonly employs a
sigmoid activation function for binary classification issues. The network's output is
translated by the sigmoid function into a probability value between 0 and 1, which
represents the likelihood that the input belongs to the positive class. The final layer of a
CNN commonly uses a softmax activation function for multi-class classification issues.
The network's output is normalised by the softmax function so that the total of the
probabilities for each class equals 1. As a result, the network may generate a probability
distribution for each potential class and select the class with the highest probability as the
output. The final layer of a CNN commonly employs a linear activation function for
regression problems, which yields a continuous output value. Depending on the particular
requirements of the problem, alternative activation functions, such as tanh or ReLU, may
occasionally also be utilised in the final layer of a CNN.

32
3.5.10 Sparse Categorical Cross-Entropy

When working on a deep learning or machine learning problem, the model is optimised
during training utilising loss/cost functions. Almost always, the objective is to decrease
the loss function. The lower the loss, the better the model performs. Cross-entropy loss is
one of the most significant cost functions. It aids in the optimisation of classification
models.

Logits are transformed into probabilities by Softmax. The Cross-Entropy is used to

calculate the deviation from the truth values by taking the output probabilities (P).

Making the model output as close as feasible to the desired output (truth values) is the
goal. In order to reduce the Cross-Entropy loss, the model weights are iteratively
modified as necessary during model training. Model training is the process of changing
the weights, and as the model continues to train and the loss is reduced, we say the model
is learning.

Cross-Entropy Loss Function is also known as logistic, logarithmic, or logarithmic loss.

Loss of Cross-Entropy Logarithmic loss, logistic loss, and logarithmic gain are other
names for function. The desired outcome for each class, which can be either 0 or 1, is
compared to each class's anticipated probability, and a score/loss is computed which
regulates the likelihood depending on the extent to which it diverges from its true
expected value. Significant variations close to 1 receive a large score because of the
penalty's logarithmic nature, whereas slight variations close to 0 receive a small score.

In order to change the model weights during training, cross-entropy loss is used. The goal
is to reduce loss; hence, the better the model, the smaller the loss. A cross-entropy loss of
zero indicates a flawless model.

The following is the equation of cross entropy loss function

33
 Fig: 3.19 Mathematical equation for cross entropy loss function

This is also the equation for the sparse categorical entropy loss function which is used in
the proposed model. In case of sparse categorical entropy truth labels are integer encoded.

34
 Chapter- 4: PERFORMANCE ANALYSIS

In machine learning, it is crucial to be able to accurately assess the model being created in
order to ensure that the predictions (such as those for disease prediction or cost estimation
in the future) adequately describe the intended phenomenon. But with so many options
available (Accuracy, Precision, Recall, etc.), it can be difficult for data scientists to decide
which performance statistic to utilise. To be able to measure the performance of the
model objectively and in the appropriate context, the suitable metric must be chosen for a
given model. This chapter discusses the performance parameters that have been chosen
for the purpose of evaluation of the proposed CNN model.

4.1 Selection of best performance parameters :

One can assess the effectiveness of ML, DL algorithms, classification algorithms, and
regression algorithms using a variety of metrics. One must exercise caution while
selecting the measures to measure the effectiveness of ML because

● The statistic that one selects will determine exactly how the effectiveness of ML
algorithms is assessed and compared.

● The metric that one selects will have a significant influence on how one weighs
the relative value of different variables in the outcome.

Classification is a process in machine learning and data science in which a model is built
for the purpose of differentiating among different classes of input data, and the input data
can be binary as well muli class, depending on its nature, during classification process the
model comprehends from various different feature and identify features that make certain
value map to its respective class. In machine learning, deep learning and AI there exists a
lot of algorithm which are used to classify a dataset into its apt category, and this is done

35
with the help of a pre labelled or pre classified set of data which a programmer makes
before training, and is called a training set. All ML algorithms then train upon this
training data to learn hidden features and patterns in order to classify input test sets into
their respective classes.

This study makes use of multilabel classification using deep learning based convolutional
neural networks.

Class-balance and anticipated results are two factors that should be taken into account
when choosing the optimal metrics to assess a given classifier's performance on a certain
dataset. One particular performance metric may only assess a classifier from one angle
while frequently underestimating others. As a result, it is impossible to choose a single
statistic to assess a classifier's effectiveness in general. For multiclass classification there
are various metrics which can be used: Classification report, confusion matrix, Class
Prediction Error, Cohen's kappa, Cross entropy, Area Under Curve, Log Loss, etc.

For this Study confusion matrix and classification report are picked for the purpose of
evaluation.

4.1.1 Confusion Matrix

It is the simplest approach for determining how well a proposed model is performing on a
classification, and its output can include two or more different types of classes. A
confusion matrix is just a square matrix or a table having two main dimensions, Actual
values and Predicted values, as well as True Positives (TP), True Negatives (TN), False
Positives (FP), and False Negatives (FN) for each dimension. The actual values are the
values with real labels and the predicted values are the values that our model predicts. For
Binary Classification the following diagram depicts a simple confusion matrix:

36
 Fig 4.1: A simple binary class classification Confusion matrix

Recall, specificity, precision, accuracy, and—most importantly—AUC-ROC curves may

all be measured with great success with the help of this tool.

The following are the descriptions for TP, FP, FN, TN:
● True Positive: It is the case, when the actual and predicted values are both same,
or the predicted values match their true labels.
● False Positive: It is the case, when the predicted values come out to be true, but in
actual they were labelled as false.
● False Negatives: It is the case, when the predicted values come out to be negative
but in actual they were labelled as true.
● True Negative: It is the case, when the predicted values are negative and their
corresponding actual values are also negative.

The following figure depicts the confusion matrix for the predicted and actual values.

For the purpose of plotting the matrix, confusion_matrix method of sklearn library is
used, and seaborn and matplot library further helps in better visualisation of the results of
the model.

37
 Fig 4.2: Code for plotting confusion matrix

Fig 4.3: Confusion matrix

Here Classes of retinal images - normal, glaucoma, diabetic retinopathy and cataract are
encoded as - 0, 1, 2, 3 respectively.

The Dark blue coloured diagonal from top left to bottom right corner of the matrix,
depicts the true positive values. Therefore 102 images of ‘class 0’ which is ‘normal’ class
of retinal images are predicted correctly. 112 images of ‘class 1’ which is ‘glaucoma’
class are predicted correctly. ‘150’ images of ‘class 2’ which is ‘diabetic retinopathy’ are
predicted correctly. And 178 images of ‘class 3’ which is ‘cataract’ are predicted
correctly.

38
For class ‘0’ or Normal retinal images

True positives (TP): 102

True negatives (TN): 112 + 10 + 28+ 17+ 150 + 22 + 18 + 0 +178 = 535

False Positives (FP): 51 + 51 + 178 = 196

False Negatives (FN): 18 + 0 + 178 = 196

For class ‘1’ or retinal images with Glaucoma disease

True positives (TP): 112

True negatives (TN): 150 + 22 + 178 + 0 + 28 + 3 + 8 + 51 +102 = 542

False Positives (FP): 18 + 17 + 51 = 86

False Negatives (FN): 28 + 10 + 28 = 66

For class ‘2’ or retinal images with Diabetic Retinopathy disease

True positives (TP): 150

True negatives (TN): 102 + 51 + 51 + 28 + 112 + 28 + 3 +18 + 178 =571

False Positives (FP): 0 + 10 + 8 = 18

False Negatives (FN): 26 + 17 + 22 = 65

For class ‘3’ or retinal images with Cataract disease

True positives (TP): 178

True negatives (TN): 102 + 51 + 8 + 28 + 112 + 10 + 26 + 17 + 150 = 504

39
False Positives (FP): 51 + 28 + 22 = 101

False Negatives (FN): 3 + 18 + 0 = 21

4.1.2 Classification Accuracy

Its use as a performance metric for classification algorithms is very widespread. It can be
described as the proportion of all true or accurate forecasts made to the number of
forecast predictions made. With the help of the following mentioned formula, one can
quickly calculate it using the confusion matrix, once the confusion matrix is generated.

Accuracy = TP+TN
TP+FP+FN+TN

The evaluation is the major process during any model-development process because it
determines whether the model is the best match for the given problem and pertinent data.
The Keras model offers a method called evaluate() that can be used with the object that
stores the model, and as the name suggests is used to perform evaluation of the model. It
gives testing loss and test accuracy of the model.

For the purpose of getting the accuracy of the proposed model, following code was
executed and the final accuracy was around 67.4%

Fig 4.4: Model final accuracy

40
4.1.3 Classification Report

In machine learning and data science, a classification report is a performance metric

which is used for the purpose of evaluation of a model. It is used to denote a trained
model’s classification accuracy and other parameters like - recall, F1 Score, and support.
It gives a clearer picture of a trained model's overall performance.

The following is the detail description of each of the metric of classification report :

Precision : The number of positive forecasts were actually correct can be determined by
precision. To do this, it divides the total number of positive predictions—correct or
incorrect—by the number of samples that were correctly predicted as positive (TP).

The following represents the formula for calculating precision :

Precision = TP
TP+FP

Recall : Parallel to Precision, Recall is the measure of true positives that were accurately
predicted by the model on a test set. This is accomplished by dividing the total number of
samples that were properly forecasted as positive or wrongly predicted as negative (TP,
FN) by the total number of samples that were correctly predicted as positive (TP). Recall
is also referred to as sensitivity, true positive rate, or hit rate.

The following is the formula for recall :

Recall = TP
TP+FN

41
Specificity : Similar to recall (also known as sensitivity), specificity measures the
percentage of actual negatives that were accurately identified. It calculates this by
dividing the number of negative samples that were properly projected to be negative or
wrongly forecasted to be positive (TN, FP) by the total number of negative samples.

The following is the formula used to calculate specificity:

Specificity = TN
TN+FP

To evaluate the ML and DL models created, precision and recall are frequently used in
the data science field. Medical professionals frequently consider specificity and
sensitivity while assessing medical testing. These ideas are quite dissimilar yet similar at
the same time. This disparity may lead to numerous misconceptions between the medical
community and the data science community when the two worlds collide, as when a
medical test is a machine learning model.

Precision: This indicates the proportion of instances that positive predictions came true.
Recall, on the other hand, indicates how many times the model accurately anticipated
based on actual positive evidence. False negatives should be avoided at all costs in any
medical diagnosis, for example cancer disease detection, since they can have fatal
repercussions. Here, recall serves as a more useful metric than precision. False negatives
are less significant if a person has to optimise suggestions on YouTube because just a
tiny percentage of recommendations is displayed in any case.

Since the model proposed in the study is based on medical images dataset, and aims to
classify eye diseases among four different categories, emphasis is laid on why precision is
important and as a evaluation parameter for this study, a detailed research has been
carried out to find the importance and use of the same for model evaluation, and why
doctors must consider precision before recall and other parameters before evaluating the
trained model.

42
Precision and recall frequently have the opposite relationships. A model is likely to have
higher precision but lower recall if it predicts fewer positive cases. However, because
false positives can have more serious repercussions than false negatives, precision is
typically given precedence over recall in medical classification tasks. As a result,
accuracy aids in striking a compromise between the model's overall performance and
accurate positive forecasts.

In case of imbalanced Datasets that is the case of medical datasets which often suffer
from class imbalance, one class may have a disproportionately low number of instances
compared to another. Precision is crucial in these circumstances. For medical diagnostic
and treatment decisions, a high accuracy score guarantees that the model is correctly
recognising the positive class even in the presence of skewed data.

Misclassification can have negative effects in medical situations. Precision reduces the
likelihood of making an incorrect diagnosis of a patient as positive when they are actually
negative, which helps to minimise false positives. By minimising unnecessary procedures
or interventions, we can improve patient safety by maximising precision.

Cost effectiveness is aided by precision as Misclassifications may result in time- and

money-consuming and needless medical procedures, examinations, or treatments.
Precision guarantees highly accurate positive forecasts, cutting down on wasteful
expenses related to false positives.

Since precision guarantees patient safety, Misclassification can have negative effects in
medical situations. Precision reduces the likelihood of making an incorrect diagnosis of a
patient as positive when they are actually negative, which helps to minimise false
positives. By minimising unnecessary procedures or interventions, we can improve
patient safety by maximising precision.

High precision in a medical categorization model suggests a high degree of dependability

and credibility. The predictions made by the model must be trusted by patients and
healthcare providers. By putting precision first, we can create models that deliver reliable,
precise outcomes, strengthening the relationship between the model and its end users.

43
Support : Support in classification report defines the presence of each particular instance
of the different classes of dataset in the test set or the set which consists of true responses.

F1 Score : It generates the harmonic mean of precision and recall. Mathematically F1

score is weighted average of precision and recall. It is more useful than accuracy during
unevenly distributed classes of dataset.

F1 = 2 * (precision*recall)
(precision + recall)

F1 score 1 represents its highest value and reveals that models precision and recall are
perfect. Whereas if F1 score is 0, this represents its lowest value and indicates that either
recall is 0 or precision is 0.

F1 scores generally have almost always an equal amount of contribution of both precision
and recall during model evaluation.

The classification report of the model can be calculated using sklearn’s metrics module.
The following figure shows the classification report for the proposed model and the
model trained:

Fig 4.5: Classification report for the model

44
 Chapter-5: CONCLUSIONS

5.1 Conclusion

In recent years, numerous models for detecting eye diseases have been developed. This
study has employed deep neural networks to help in reaching the model’s aim, along with
several essential libraries like keras, TensorFlow, pandas, and NumPy, and through the
model’s final conclusion and result this study is able to meet all of its goals. CNN is used
to create a deep learning model that can categorise eye illnesses into four separate groups.
The model will be able to help patients identify their eye state in terms of cataract,
diabetic retinopathy, glaucoma, or normal. It is also cost-effective and easy to use.

5.2 Application of the Project

The following are the applications of the this deep neural network model:

● Early detection and diagnosis is one of the main uses of this deep neural network
model for classifying eye diseases. Doctors can act and treat the illness before it
worsens thanks to the model's ability to analyse photographs of the eye and spot
any signs of sickness or abnormality.
● Drug development for eye illnesses can also make use of this deep neural network
model. The programme can find prospective drug targets and forecast the
effectiveness of new treatments by analysing vast quantities of data on illness
development and treatment outcomes.
● Another use is in screening programmes, especially in places with poor access to
medical personnel. The model can examine eye scans and categorise them
according to the likelihood of disease, assisting in the selection of patients for
additional testing and treatment.
● Last but not least, studies can make use of this deep neural network model to
comprehend the root causes of eye illnesses. The model can uncover patterns and

45
 correlations that might not be obvious to human researchers by analysing
enormous volumes of patient data, leading to fresh ideas and discoveries.

5.3 Limitation of the Major Project

One of the drawbacks of the proposed deep learning model is that it needs a lot of
high-quality data to perform well. The model's accuracy might be enhanced by including
more information on eye illnesses by adding more retinal images to our dataset. Deep
learning models are frequently referred to as "black boxes" due to the fact that it can be
challenging to comprehend how they make their predictions. Creating tools for
deciphering the choices made by this proposed model could aid physicians and
ophthalmologists in comprehending its output and gaining confidence in its outcomes.

5.4 Future Work

Incorporating more data is one of the major future work that needs to be done for the
model to enhance its overall accuracy and make it a more trustworthy model in the field
of medicine. The proposed methodology can be used to apply transfer learning. It entails
modifying a pre-trained deep learning model for a particular job. This method, which has
been proven successful in other medical imaging activities, might also be used to identify
eye diseases. Many eye illness identification models, including this study’s model, focus
on identifying only a single disease or condition, but in practice, individuals may have
numerous disorders at once. As a result, adding a multi-class categorization to this model
can be helpful. Hence It would be beneficial to create a multi-class classification model
that can recognise several circumstances at once. Any model for identifying eye diseases
must ultimately be evaluated in a clinical environment to ascertain its applicability. Future
study should focus on creating a strategy for how the model might be included into
clinical workflows and assessing how well it works in practical situations.

46
 REFERENCES

[1] M. S. Khan et al., “Deep learning for ocular disease recognition: An inner-class
balance,” Comput. Intell. Neurosci., vol. 2022, p. 5007111, 2022.

[2] S. K. Sattigeri, N. Harshith, G. N. Dhanush, K. A. Ullas, and M. S. Aditya, “Eye

disease identification using deep learning,” Irjet.net. [Online]. Available:
https://www.irjet.net/archives/V9/i7/IRJET-V9I7185.pdf.

[3] N. Badah, A. Algefes, A. AlArjani, and R. Mokni, “Automatic eye disease detection
using machine learning and deep learning models,” in Pervasive Computing and Social
Networking, Singapore: Springer Nature Singapore, 2023, pp. 773–787.

[4] G. Meller, “Ocular disease recognition using convolutional neural networks,” Towards
Data Science, 04-Aug-2020. [Online]. Available:
https://towardsdatascience.com/ocular-disease-recognition-using-convolutional-neural-ne
tworks-c04d63a7a2da.

[5] H. Gu et al., “Deep learning for identifying corneal diseases from ocular surface
slit-lamp photographs,” Sci. Rep., vol. 10, no. 1, p. 17851, 2020.

[6] P. K. Upadhyay, S. Rastogi, and K. V. Kumar, “Coherent convolution neural network

based retinal disease detection using optical coherence tomographic images,” J. King
Saud Univ. - Comput. Inf. Sci., vol. 34, no. 10, Part B, pp. 9688–9695, 2022.

[7] Researchgate.net. [Online]. Available:

https://www.researchgate.net/figure/Diabetic-retinopathy-affected-eye-as-compared-to-th
at-of-the-normal-eye_fig1_275029242.

[8] “Common symptoms of cataracts,” Eyecheck.com, 18-Mar-2019. [Online]. Available:

https://learn.eyecheck.com/common-symptoms-of-cataracts.

[9] “What is glaucoma? Learn about glaucoma,” InMed Pharmaceuticals, 02-Apr-2020.

[Online]. Available: https://www.inmedpharma.com/learn/what-is-glaucoma/.

[10] G. V. Doddi, “eye_diseases_classification.'' 28-Aug-2022. Available:

https://www.kaggle.com/datasets/gunavenkatdoddi/eye-diseases-classification

[11] “What are Convolutional Neural Networks?,” Ibm.com. [Online]. Available:

https://www.ibm.com/topics/convolutional-neural-networks.

47
[12] P. Mandal, “System Flow Diagram,” EDUCBA, 31-May-2020. [Online]. Available:
https://www.educba.com/system-flow-diagram/.

[13] A. J. Mortimer, “DATA FLOW DIAGRAMS,” in Information Structure Design for

Databases, Elsevier, 1993, pp. 119–133.

[14] “Levels in data flow diagrams (DFD),” GeeksforGeeks, 18-Mar-2019. [Online].

Available: https://www.geeksforgeeks.org/levels-in-data-flow-diagrams-dfd/.

[15] “Machine Learning - Performance Metrics,” Tutorialspoint.com. [Online]. Available:

https://www.tutorialspoint.com/machine_learning_with_python/machine_learning_algorit
hms_performance_metrics.html.

[16] E. Zuccarelli, “Performance metrics in ML - part 1: Classification,” Towards Data

Science, 29-Dec-2020. [Online]. Available:
https://towardsdatascience.com/performance-metrics-in-machine-learning-part-1-classific
ation-6c6b8d8a8c92.

[17] H. Goonewardana, “Evaluating multi-class classifiers - apprentice journal -

medium,” Apprentice Journal, 03-Jan-2019. [Online]. Available:
https://medium.com/apprentice-journal/evaluating-multi-class-classifiers-12b2946e755b.

[18] M. Banoula, “Classification in Machine Learning: What it is & Classification

Models,” Simplilearn.com, 02-Feb-2021. [Online]. Available:
https://www.simplilearn.com/tutorials/machine-learning-tutorial/classification-in-machine
-learning.

[19] S. Narkhede, “Understanding confusion matrix,” Towards Data Science,

09-May-2018. [Online]. Available:
https://towardsdatascience.com/understanding-confusion-matrix-a9ad42dcfd62.

[20] A. Kharwal, “Classification report in machine learning,” the clever programmer,

07-Jul-2021. [Online]. Available:
https://thecleverprogrammer.com/2021/07/07/classification-report-in-machine-learning/..

[21] C. Dilmegani, “What is Data Augmentation? Techniques & Examples in 2023,”

AIMultiple, 30-Apr-2021. [Online]. Available:
https://research.aimultiple.com/data-augmentation/.

[22] C. Gallo, “Artificial Neural Networks Tutorial,” in Encyclopedia of Information

Science and Technology, Third Edition, IGI Global, 2014, pp. 6369–6378.

48
[23] D. Johnson, “Back propagation in neural network: Machine learning algorithm,”
Guru99, 10-Feb-2020. [Online]. Available:
https://www.guru99.com/backpropogation-neural-network.html.

[24] S. A. Hicks et al., “On evaluation metrics for medical applications of artificial
intelligence,” bioRxiv, p. 2021.04.07.21254975, 2021.

[25] L. Gupta, “Precision-recall tradeoff in real-world use cases - analytics Vidhya -

medium,” Analytics Vidhya, 19-Feb-2021. [Online]. Available:
https://medium.com/analytics-vidhya/precision-recall-tradeoff-for-real-world-use-cases-c
6de4fabbcd0.

49