VISVESVARAYA TECHNOLOGICAL UNIVERSITY

“Jnana Sangama”,Belagavi-590018

PROJECT REPORT
On
“A HIGH PERFORMANCE GLAUCOMA SCREENING TECHNIQUE USING
CNN ARCHITECTURE”
Submitted in partial fulfilment of the requirements for the award of the degree of
BACHELOR OF ENGINEERING
In
ELECTRONICS AND COMMUNICATION

Submitted by
DEEKSHITHA S 1CK16EC009
DHIYA N S 1CK16EC010
SRINIVASA J 1CK15EC047
SUSHIL KAVIN U 1CK15EC049

Under the Guidance of

Mr. CHETHAN KUMAR N.S, M.Tech,(Phd)

Asst.Proffessor
Dept of ECE,CBIT,Kolar.

DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

C.BYREGOWDA INSTITUTE OF TECHNOLOGY
Thoradevandahalli post,Kolar-Srinivasapura
road,Kolar-563101,Karanataka
2019-2020
 C.BYREGOWDA INSTITUTE OF TECHNOLOGY
Department of Electronics & Communication Engineering

Certificate
This is to certify that the Project entitled “A HIGH
PERFORMANCE GLAUCOMA SCREENING TECHNIQUE USING
CNN ARCHITEC-
TURE” is a bonafide work carried out by
DEEKSHITHA S 1CK16EC009
DHIYA N S 1CK16EC010
SRINIVASA J 1CK15EC047
SUSHIL KAVIN U 1CK15EC049
In partial fulfillment for the award of Bachelor of Engineering in Electron-
ics and Communication Engineering of the Visvesvaraya Technological
University, Belagavi during the year 2019-2020. The report has been approved
as it satisfies the academic requirements with respect to project work prescribed
by the VTU of the Bachelor of Engineering degree.

Signature of the Guide Signature of the HOD Signature of the Principal

Mr.Chethan Kumar N.S Mrs.Sujani G Dr.Sreerama Reddy G.M
Asst.Professor,Dept.of ECE HOD,Dept.of ECE Principal
CBIT,Kolar. CBIT,Kolar. CBIT,Kolar.

Name of the Examiners: Signature of the Examiners:

(1)
(2)

NAME: USN:
 ACKNOWLEDGMENT

First and foremost, we would like to thank the ALMIGHTY for giving us
the strength, knowledge, ability and opportunity to undertake this Project and to
persevere and complete it satisfactorily.

We would like to thank Dr. SREERAMA REDDY G.M, Principal, CBIT,

Kolar for providing support for all Project related activities.

We sincerely thank by heart and soul to Mrs. SUJANI G, Head of the

Department, Dept. of ECE, CBIT for her advice, support and constructive sug-
gestions that were useful in building the Project work.

We extend our special in-depth, sincere gratitude to our Internal Guide, Mr.
CHETHAN KUMAR N.S, Asst. Professor, Dept. of ECE, CBIT Whose in-
spiration, patience, valuable suggestions and guidance helped us in finishing the
Project work with a great Endeavour.

We would like to thank UG Project Coordinator, Mr. SUKUMAR B.S,

Asst. Professor, Dept. of ECE, CBIT whose timely support and suggestions went
a long way in the completion of Project work.

Finally, we would like to thank all the teaching and the non-teaching staffs
of the Department of Electronics and Communication Engineering,
CBIT, Kolar, for their valuable suggestions and support during the tenure of the
project.

DEEKSHITHA S 1CK16EC009
DHIYA N S 1CK16EC010
SRINIVASA J 1CK15EC047
SUSHIL KAVIN U 1CK15EC049

i
 DECLARATION

We, DEEKSHITHA S (1CK16EC009), DHIYA N S (1CK16EC010),

SRINIVASA J (1CK15EC047) and SUSHIL KAVIN U (1CK15EC049),
students of 8th Semester B.E, in ELECTRONICS & COMMUNICATION
Engineering, C.Byregowda Institute of Technology hereby declare that the Project
entitled “A HIGH PERFORMANCE GLAUCOMA SCREENING
TECH-
NIQUE USING CNN ARCHITECTURE”, has been carried out by us under
the guidance of Mr. CHETHAN KUMAR N.S, Asst. Professor Depart-
ment of ECE, CBIT as a partial fulfillment of the requirements for the award
of the degree of Bachelor of Engineering in Electronics and Communi-
cation Engineering of Visvesvaraya Technological University, Belagavi
during academic year 2019-2020.

We also declare that, to the best of our knowledge and belief the matter em-
bodied in this dissertation has not been submitted previously by us for the award
of any Degree or Diploma to any other university.

Place : Kolar DEEKSHITHA S 1CK16EC009

Date : DHIYA N S 1CK16EC010
SRINIVASA J 1CK15EC047
SUSHIL KAVIN U 1CK15EC049

ii
 ABSTRACT

Glaucoma is a second irreversible eye disease across the world that

can lead to blindness, if not treated early. At an early stage, it is very
tedious to detect the disease as it does not show any symptoms. So, we
build a supervised method for convolutional neural network to detect
the disease as early as possible. In our model ACRIMA data set of
705 images are used where 396 are glaucomatous images and 309 are
non glaucomatous images. The complete data set is divided into 80%
for training and 20% for testing. The data set is preprocessed before
giving to the CNN model. Our CNN model consists of 8 layers where, 4
convolutional layers for feature extraction and 4 is fully connected
layer to classify between glaucomatous and non-glaucomatous images.
The best performance is obtained at the learning rate and epochs of
0.0001 and 100 . In this work we evaluated the performance of
Glaucoma- Deep system by plotting confusion matrix, the sensitivity,
specificity, accuracy, precision (PRC) statistical measures and AUC
of 0.99

Keywords —CNN, ACRIMA, AUC, precision, sensitivity, specificity

.

iii
Contents

Acknowledgment..............................................................................................i
Declaration......................................................................................................ii
Abstract...................................................................................................iii
Table of Contents.....................................................................................v

List of Figures vi
List of Figures.........................................................................................vi

List of Abbrevations vii

1 INTRODUCTION 1
1.1 Types of Glaucoma . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Clinical diagnosis of Glaucoma . . . . . . . . . . . . . . . . . 3
1.1.2 CAD for Glaucoma . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.3 Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Organization Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 LITERATURE SURVEY 12

3 PROPOSED METHODOLOGY 16
3.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.3 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

iv
4 WORK FLOW 20

5 SOFTWARE AND HARDWARE REQUIRE-

MENTS 26
5.1 Google colab.................................................................................................26
5.2 GPU........................................................................................................27
5.3 Deep learning frame work......................................................................28
5.3.1 Tensor Flow......................................................................................28
5.3.2 Keras.................................................................................................28
5.4 Python.....................................................................................................29
5.5 Operating system....................................................................................30
5.6 Hardware requirements...........................................................................30

6 EXPERIMENTAL RESULTS 31

7 APPLICATIONS, ADVANTAGES AND DIS-

ADVANTAGES
33
7.1 Applications...................................................................................................33
7.2 Advantages.............................................................................................33
7.3 Disadvantages................................................................................................34

8 CONCLUSION AND FUTURE SCOPE 35

References 37
List of Figures

1.1 Anatomy of Healthy Eye & Glaucoma Suspecious Eye.............................2

1.2 Primary Open Angle Glaucoma...............................................................2
1.3 Primary Angle Closure Glaucoma...........................................................3
1.4 Tonometry................................................................................................4
1.5 Ophthalmoscopy..............................................................................................4
1.6 Perimetry test...........................................................................................4
1.7 Gonioscopy......................................................................................................5
1.8 Pachymetry...............................................................................................5
1.9 Artificial Intelligence & its Subsets..........................................................7
1.10 Glaucoma Regional Treatment Market Across the World.......................8

3.1 CNN Architecture...................................................................................17

4.1 Work Flow....................................................................................................20

6.1 AUC- ROC Curve..................................................................................32

vi
List of Abbrevations

ACC Accuracy

ADAM Adaptive Moment Estimation

AI Artificial Intelligence

AMD Age Related Macular Degeneration

ANN Artificial Neural Network

AUC Area Under the Curve

CAD Computer Aided Diagnosis

CDR Cup to Disc Ratio

CNN Convolutional Neural Network

CPU Central Processing Unit

CUDA Compute Unified Device Architecture

DCNN Deep Convolutional Neural Network

DL Deep Learning

DR Diabetic Retinopathy

FN False Negative

FP False Positive Rate

GB Giga Byte

vii
GPU Graphics Processing Unit

HPGST High Performance Glaucoma Screening Tecnique

IOP Intra Ocular Pressure

ML Machine Learning

OCT Optical Coherence Tomography

ONEIRO Open-ended Neuro-Electronic Intelligent Robot Operating

System

ONH Optic Nerve Head

ORIGA Online Retinal Fundus Image Database for Glaucoma

PC Personal Computer

PRC Precision

RAM Random Acess Memory

ReLU Rectified Linear Unit

RGB Red Green Blue

RNFL Retinal Nerve Fibre Layer

ROC Receiver Operating Characteristic Curve

ROI Region of Interest

ROM Read Only Memory

SN Sensitivity

SP Specificity

TB Tera Byte

TN True Negative

TP True Positive

viii
TPR True Positive Rate

TPU Tensor Processing Unit

TWSVM Twin Support Vector Machine

VGG Visual Geometry Group

ix
Chapter 1

INTRODUCTION

Glaucoma is a chronic eye disease that occurs as a result of optic nerve damage
due to intraocular pressure inside the eye [2]. Glaucoma is a leading cause of
blindness in people aged older than 60 years [3]. The World Health Organization
has declared Glaucoma to be the second largest cause of blindness all over the
world [4]. The number of people with glaucoma worldwide was estimated as 64.3
million in 2013, and 80 million in 2020 [5]. As it may be asymptomatic, early
detection and treatment are important to prevent vision loss [1].

Glaucoma is characterized by damage to the optic nerve through increasing

degeneration of the nerve fibers. This condition progressively elevates the intra-
ocular pressure (IOP), affecting the optic nerve and resulting in thickening of
retinal nerve fibre layer (RNFL) and permanent blindness if left untreated, which
is usually known as ‘cupping’ [10]. The decrease in the healthy neuro retinal
tissues can be easily noticed by measuring cup to disc ratio (CDR), which is an
indication of glaucomatous change. Typical value of CDR for a healthy eye is 0.3.
This measurement is commonly used as a glaucoma indicator, which expresses
the vertical diameter proportion of the optic disc and the cup. However, CDR
measurement implies a great effort to obtain a proper optic disc and optic cup
segmentation [1, 10]. Increase in IOP causes no early symptoms or pain which
makes the early detection of glaucoma complex and thus often called “Silent thief
of sight” [8]. As glaucoma may be asymptomatic in its early stages, at least half
of patients with this eye pathology remain undiagnosed, while more than half of
those who are undergoing treatment do not have the disease. Hence it is essential

1
A HPGST Using CNN Architecture 2019-2020

to have a reliable early detection system for glaucoma onset [1, 10].

Figure 1.1: Anatomy of Healthy Eye & Glaucoma Suspecious Eye.

1.1 Types of Glaucoma

There are four main types of Glaucoma:

1. Primary Open Angle Glaucoma : It is the most common form of glau-

coma, accounting for at least 90% of all glaucoma causes & is caused by the
slow clogging of the drainage canals, resulting in increased eye pressure it
has a wide and open angle between the iris and cornea it develops slowly
and is a long life condition its symptoms and damages are not noticed.
Open- angle means that the angle where the iris meets the cornea is as wide
and open as it should be Open-angle glaucoma is also called primary or
chronic glaucoma [13].

Figure 1.2: Primary Open Angle Glaucoma.

Dept.of ECE,CBIT,Kolar 2
 2. Primary Angle Closure Glaucoma : It is a less common form of glau-
coma & is caused by blocked drainage canals, resulting in a sudden rise in
intraocular pressure it has a closed or narrow angle between the iris and
cornea develops very quickly it has symptoms and damage that are usu-
ally very noticeable demands immediate medical attention. It is also called
acute glaucoma or narrow angle glaucoma. Unlike open-angle glaucoma, angle-
closure glaucoma is a result of the angle between the iris and cornea closing
[13].

Figure 1.3: Primary Angle Closure Glaucoma.

3. Secondary Glaucoma : Secondary glaucoma arises due to certain compli-

cated conditions like serious eye injury, tumour, diabetes, etc. Neo-vascular
glaucoma is a type of secondary glaucoma which is a resultant of Diabetic
Retinopathy [13].

4. Developmental Glaucoma : This type of glaucoma occurs in babies

when there is incorrect or incomplete development of the eye’s drainage
canals during the parental period. This is a rare condition that may be
inherited. It is also referred as childhood glaucoma, pediatric or infantile
glaucoma. It is usually diagnosed within the first year of baby life [13].

1.1.1 Clinical diagnosis of Glaucoma

The clinical diagnosis of glaucoma includes a series of various tests that is car-
ried out by the ophthalmologist. The key to prevent glaucoma is to have regular
eye check-ups after 40 years.
 The following different tests are commonly performed to diagnose glaucoma
are described as follows:

1. Tonometry : It measures the pressure within the eye. During tonometry,

eye drops are used to numb the eye. Then a doctor or technician uses a
device called a tonometer to measure the inner pressure of the eye. If it
exceeds ¿21mm Hg, the person is diagnosed with glaucoma [10].

Figure 1.4: Tonometry.

2. Ophthalmoscopy : The doctor examine the optic nerve for glaucoma dam-
age. Eye drops are used to dilate the pupil so that the doctor can see through
the eye to examine the shape and color of the optic nerve [10].

Figure 1.5: Ophthalmoscopy.

3. Perimetry test : It examines the optic nerve if the intraocular pressure is

not within the normal range or if the optic nerve looks unusual, the patient
has to go through Perimetry which is a test to chart the visual field [10].

Figure 1.6: Perimetry test.

 4. Gonioscopy : It is a simple test which is used to inspect the drainage
angle using a gonio lens with a slit lamp [10].

Figure 1.7: Gonioscopy.

5. Pachymetry : It is a simple test, which measures cornea thickness. It

helps in clinical diagnosis as it has great influence on eye pressure reading

[10].

Figure 1.8: Pachymetry.

The clinical diagnosis of eye using above techniques is time consuming and
involves inter/intra observer variability. Digital fundus images captured using a
fundus camera can be effectively utilized for observing the progression of the dia-
betic retinopathy (DR), glaucoma and age related macular degeneration (AMD).
The interesting clinical features of eye such as retina, optic disc, blood vessels
etc., can be clearly visualized in fundus images. In addition, fundus camera is
reliable, less expensive and easy to operate and it can be used to measure various
struc- tures such as change in cup to disc ratio, optic nerve head (ONH), cup
diameter etc. Hence, fundus images can be effectively utilized as a cost effective
tool for the diagnosis of retinal health and eye abnormalities (DR, AMD and
glaucoma) using a single fundus image. Computer aided diagnosis (CAD) of
fundus images helps to diagnose the retinal health using various computational
algorithms. It is a cost effective tool which can avoid inter/intra observer
variability which may be encountered in clinical diagnosis[10].
1.1.2 CAD for Glaucoma

An accurate Glaucoma detection is among one of the crucial requirements to

control its progression and reduce the risk of blindness. In biomedical imaging
glaucoma detection is one of the active researches being done in the field of au-
tonomous glaucoma detection systems to provide state of art computer aided de-
sign tool that can aid ophthalmologists in early glaucoma detection. Many image
processing, computer vision and ML techniques & tools are being used to excel in
this research field and come up with more accurate results that might help in more
accurate and early glaucoma diagnosis. An automated system can be used for of-
fering standardized large-scale screening at a lower cost, which may reduce
human errors, provide services to remote areas, as well as free from observer bias
and fatigue. Computers are able to obtain much quicker classifications once
trained, giving the ability to aid clinicians in real-time classification. Machine
learning for glaucoma diagnosis has achieved great development in recent
years[8].

1.1.3 Artificial Intelligence

Artificial intelligence has the potential to revolutionize the screening, diagno-

sis and classification of glaucoma, both through the automated processing of large
data sets and by earlier detection of new disease patterns. In addition, the artificial
intelligence holds promise for fundamentally changing research aimed at under-
standing the development, progression and treatment of glaucoma, by identifying
novel risk factors and by evaluating the importance of existing ones. The applica-
tion of artificial intelligence (AI) algorithms in ophthalmology has improved our
understanding of much retinal, macular, choroidal and corneal pathologies. With
the advent of deep learning, a number of tools for the classification, segmentation
and enhancement of ocular images have been developed. Over the years, several
AI techniques have been proposed to help detect glaucoma by analysis of func-
tional and/or structural evaluations of the eye. Moreover, the use of AI has also
been explored to improve the reliability of ascribing disease prognosis.

There are two main subsets of artificial intelligence namely:

1. Machine Learning : Machine learning is a type of artificial intelligence
(AI) that provides computers with the ability to learn without being explic-
itly programmed. Machine learning focuses on the development of
Computer Programs that can change when exposed to new data. We’ll see
basics of Machine Learning, and implementation of a simple machine
learning algo- rithm using python. Deep learning (DL) has been applied to
produce highly accurate algorithms that can detect eye conditions such as
diabetic retinopa- thy (DR) with accuracy comparable with that of human
experts. The use of this technology may aid screening efforts, and recent
work demonstrates value in assisting ophthalmologists. DL algorithms also
have been developed for other diseases, including glaucoma.

2. Deep learning : It is an artificial intelligence function that imitates the

workings of the human brain in processing data and creating patterns for
use in decision making. Also known as deep neural learning or deep neural
network. Nowadays, image classification is performed using deep learning
technique. It integrates both feature extraction and classification. These
methods can achieve promising results using complex networks built with
large scale data. In this work, we have proposed a novel convolutional neu-
ral network (CNN) architecture to detect the glaucoma automatically with
highest performance. During training our network, the features are extracted
from the input images yielding a robust deep CNN model and it efficiently
classifies the unknown image during the testing phase. The detailed struc-
ture of the proposed method is described in following sections.

Figure 1.9: Artificial Intelligence & its Subsets.

1.2 Motivation
As glaucoma is asymptomatic, the patients are unaware until noticeable vision
loss occurs. It is currently estimated that 64.3 million individuals are affected
by this disease in the world, and this number is anticipated to reach 112 million
by 2040. Blindness in glaucoma can be avoidable if the disease is detected early
enough and treatment strategies commenced. However, ≥ 50 percent of cases
with glaucoma are undiagnosed in most population surveys around the world.
This is attributable to glaucoma being largely asymptomatic until late in the
disease pro- cess. Due to its insidious nature, it is critical to detect glaucoma at the
earliest. Moreover, early detection of glaucoma is particularly significant since it
allows timely treatment to prevent major visual field loss and prolongs the
effective years of usable vision. There are many sub groups of glaucoma,
separated by causes, genetics, or morphology and within each group, there may be
tens of different glaucoma types. Among which two main types are open-angle
and angle-closure glaucoma. These are marked by an increase of Intra Ocular
Pressure (IOP) or pressure inside the eye.

Figure 1.10: Glaucoma Regional Treatment Market Across the World.

1.3 Problem Statement
Glaucoma is a group of diseases that can lead to damage to the eye’s optic
nerve and ends in blindness. Since glaucoma has no symptoms, the only way to
detect its progression is by conducting studies based on the internal structure of
the eye. The risk of blindness depends on the intraocular pressure, as well as
family history of this disease. About half the glaucoma patients in a community
remain undiagnosed worldwide. To detect glaucoma, besides the standard eye
pressure test, doctor will test for optic nerve damage, looking for slight changes
that may reveal the beginnings of the eye disease. The glaucoma eye exam cost
can run about 7,000 Rupees to 12,000 Rupees, unless you have health insurance.
This technology has made great progress in the creation and modification of
instruments used for different diagnoses within the medical area. In the field of
ophthalmology, there is a wide variety of tools for taking pictures that stand out
for their image quality. The use of artificial neural networks in recent years has
given solutions to a wide variety of pattern recognition problems, such as
computer vision, image recognition and speech recognition. The most used tool in
the construction of convolutional neural networks is TensorFlow2. It is a flexible
architecture allows to implement algorithms in one or more CPUs or GPUs or
TPUs. We believe that by extending model with pictures of the fundus of the eye,
the glaucoma detection process could be automated.

1.4 Objectives
The objectives are important to achieve the goal. The main objectives of this
project are:

1. To develop an efficient Pre-processing /Data Augmentation technique.

2. To develop a Novel algorithm to extract features using CNN.

3. To develop high computational classifier to detect the Glaucomatous images.

4. To develop a hybrid system to increase the diagnosis speed of the ophthal-

mologists.
1.5 Organization Thesis
The association of the thesis is the design of an automated glaucoma diagnosis
system using retinal fundus images. This thesis is organized into seven main
chapters.

Chapter 1 : It describes the introduction to glaucoma, types of glaucoma,

clinical diagnosis of glaucoma, CAD for glaucoma and small explanation for deep
learning. In addition to that, the motivation behind the automated glaucoma
diagnosis and the problem statement for automated glaucoma analysis system
and objectives for building up our model are discussed along with the
organization thesis of the report.

Chapter 2 : It describes the literature survey of various glaucoma diagnosis

approaches along with statistics of glaucoma and risk factors for glaucoma.

Chapter 3 : It deals with the model that is been used, the mathematical
background of CNN and the datasets used for training and testing process. The
proposed glaucoma detection system based on CNN and supervised classification
approaches are presented.

Chapter 4 : It deals with the overall workflow with the sequence of tasks
that processes a set of data.

Chapter 5 : The experimental results of the proposed system are demon-

strated to validate the performance of the proposed glaucoma detection
framework. The measured feature values are compared with ground truth features
measures. Also, the performances of the supervised clustering approaches are
analyzed.

Chapter 6 : The software and hardware tools used for building the auto-
matic glaucoma detection models are discussed.

Chapter 7 : It describes the advantages and disadvantages along with the

applications of the automatic glaucoma detection model, in which the disadvan-
tages of the model will be used for betterment of the model future works.
 Chapter 8 : It concludes this thesis, summarizing the main achievements
based on the obtained experimental results, and suggesting areas for further in-
vestigation.
Chapter 2

LITERATURE SURVEY

Andres Diaz Pinto et al.(2019) “CNNs for automatic glaucoma as-

sessment using fundus images: an extensive validation”. In this paper,
five different ImageNet-trained CNN architectures (VGG16, VGG19,
InceptionV3, ResNet50 and Xception) were analyzed and used as glaucoma
classifiers. Using only publicly available databases, the Xception architecture
shows the best per- formance for glaucoma classification. Based on the 1707
images and data aug- mentation technique, an average AUC of 0.9605 with a 95%
confidence interval of 95.92–97.07%, an average specificity of 0.8580 and an
[1]
average sensitivity of 0.9346 were obtained .

Ali Serener et al.(2019) “Transfer Learning for Early and Advanced

Glaucoma Detection with Convolutional Neural Networks”. Early and
advanced glaucoma classifications are performed on fundus images using two deep
learning methods, ResNet- 50 and GoogLeNet. The performances of the two mod-
els are evaluated in terms of accuracy, sensitivity, specificity and the area under
the ROC curve. The results show that for early, advanced as well as overall glau-
[2]
coma detection GoogLeNet outperforms ResNet-50 .

Sang Phan et al.(2019) “Evaluation of Deep convolutional neural

network for glaucoma detection”. Three DCNNs showed areas under the curve
(AUCs) of 0.9 or more. The AUC of the DCNN using glaucoma confirmed eyes
against non-glaucomatous eyes was higher than that using glaucoma-suspected
eyes against non-glaucomatous eyes by approximately 0.1. The image size did

12
not

12
A HPGST Using CNN Architecture 2019-2020

affect the discriminative ability. Heatmap analysis showed that the optic disc area
[3]
was the most important area for the discrimination of glaucoma .

Muhammad Naseer Bajwa et al.(2019) “Two-stage framework for

optic disc localization and glaucoma classification in retinal fundus im-
ages using deep learning”. Proposed a method which is evaluated on seven
publicly available datasets for disc localization and on ORIGA dataset, which
is the largest publicly available dataset with healthy and glaucoma labels, for
glaucoma classification. The results of automatic localization mark new state-of- the-
art on six datasets with accuracy reaching 100% on four of them[4].

Juanj.Gomez-Valverde et al.(2019)“Automatic glaucoma classifica-

tion using color fundus images based on CNN and transfer learning”.
It exploited the application of different CNN schemes to show the influence in the
performance of relevant factors like the data set size, the architecture and the use
of transfer learning vs newly defined architectures. It also compared the perfor-
mance of the CNN based system with respect to human evaluators and explored
the influence of the integration of images and data collected from the clinical his-
[5]
tory of the patients .

Manal Al Ghamdi et al.(2019) “Semi-supervised Transfer Learning

for Convolutional Neural Networks for Glaucoma Detection”. This paper
presented a CNN semi-supervised learning framework for detecting glaucoma in
fundus images. By employing the self-learning strategy to increase the training
samples using the unlabeled data and fine-tuning a pre-trained CNN to define a
glaucoma-specific classifier, the presented approach demonstrated promising per-
formance compared to existing methods[6].

Kaveri A.Thakoor et al. (2019) “Enhancing the Accuracy of Glau-

coma Detection from OCT Probability Maps using Convolutional Neu-
ral Networks”. In this paper they developed a purely OCT-trained (Type A)
as well as transfer-learning based (Type B) CNN architectures, all achieved high-

Dept.of ECE,CBIT,Kolar 13
accuracy and high AUC-score detection of glaucoma from OCT probability map
images. In fact, one model of each Type - one CNN-A model and one CNN-B
[7]
model achieved the best accuracy rates of the entire set .

Nacer Eddine Benzebouchi et al.(2018) “Glaucoma Diagnosis Using

Cooperative Convolutional Neural Networks”. Proposed a convolutional
neural networks (CNN) approach to diagnosing glaucoma using multimodal data
from retinal fundus images and achieved high classification accuracy. Developed
a network with CNN architecture that avoid the classical handcrafted features
extraction step, by processing features extraction and classification at one time
within the same network of neurons and consequently provide a diagnosis au-
tomatically and without user input. They trained this network on the publicly
available RIM-ONE dataset and demonstrated impressive results, particularly for
[8]
a high-level classification task .

Jongwoo Kim et al.(2018) “Region of Interest Detection in Fundus

Images Using DL and Blood vessel Information”. This paper proposes an
automatic method to classify ROI, from fundus images using deep learning. Two
types of image datasets are used to train the CNNs for each dataset using different
sizes of kernel and stride in the first convolutional layer. The test results show
over 98% accuracy, in the case of the MESSIDOR dataset. In the case of the
Open Sources dataset the CNN using the blood vessels shows the best
performance over- all: over 99% accuracy for the MESSIDOR dataset and over
[9]
97% accuracy for the Open Sources dataset .

Raghavendra U et al.(2018) “Deep Convolution Neural Network

for Accurate Diagnosis of Glaucoma Using Digital Fundus Images”. Pro-
posed a novel CAD tool for the accurate detection of glaucoma using deep
learning technique. An eighteen layers of convolutional neural network (CNN) is
effectively trained in order to extract robust features from the digital fundus
images. Finally these features are classified into normal and glaucoma classes
during testing. They have achieved the highest accuracy of 98.13% using 1426
(589: normal and 837:
glaucoma) fundus images[10].

Baidaa Al-Bander et al.(2017) “Automated Glaucoma Diagnosis us-

ing Deep Learning Approach”. The proposed method that is based on
CNN demonstrated promising performance in diagnosing the glaucoma with
consider- ably lower computational cost as compared to existing equivalent
methods. The proposed deep learning method in automatic feature learning and
diagnosis of the glaucoma. In addition, a more efficient data augmentation and
data sampling approaches can also be used to obtain higher performance
[11]
measured in terms of accuracy, sensitivity, specificity and other metrics .

Radia Touahri et al(2017) “A Comparative Study of Convolutional

Neural Network and Twin SVM for Automatic Glaucoma Diagnosis
Embedded & Distributed Systems”. This paper compared two different
method of classification, the first one based is TWSVM method and the sec-
ond one represents two different architectures of Convolutional Neural Network
classifier. They have used them as a computer aided diagnostic system for au-
tomatic glaucoma funds images classification applied on color image of Rim-One
dataset. The obtained results after several empirical tests demonstrated the ro-
bustness and effectiveness of our proposed method comparing with Twin SVM
[12]
and 2-convolutional CNN architecture .

Pavitra G et al.(2017) “Detection of Secondary Glaucoma in Human

Eyes using Sophisticated Bio-Medical Image Processing Algorithms”. In
this paper the work was also explored & arrived at the definition of the problem
that had to be tackled with. Methodology is proposed in the form of a block
diagram to solve the above defined problem using Matlab and to arrive at the
[13]
expected results .
Chapter 3

PROPOSED METHODOLOGY

We develop an algorithm with CNN architecture that avoid the classical hand-
crafted features extraction step, by processing features extraction and
classification at one time within the same network of neurons and consequently
provide a diag- nosis automatically and without user input. The Convolutional
Neural Network is currently the most powerful models for classifying images.
They have two distinct parts. At the input, an image is provided in the form of a
matrix of pixels. It has 2 dimensions of a grayscale image. The color is
represented by a third dimension, depth 3 to represent the fundamental colors
(RGB). The first part of a CNN is the conventional part itself. It functions as an
extractor of image characteristics. An image is passed through a succession of
filters, or convoluted nuclei, creating new images called convolution maps. Some
intermediate filters reduce the resolution of the image by a local maximum
operation. In the end, the convolution maps are flattened and concatenated into a
feature vector, called a CNN code. This code CNN got out of it from the
convolutive party is then connected in the entry of a second part, constituted by
completely connected layers (multilayer perceptron). The role of this part is to
combine the characteristics of the code CNN to classify the image. The output is
a last layer with one neuron per category.

In input layer the images with random pixels are resized into the pixel size of
224 x 224 x 3 by pre-processing. Then, in convolutional layer the image pixels
224 x 224 x 3 are convolved with the weights and bias terms. An activation
function (ReLU) is used to convert all the negative values obtained after

16
convolution to 0 and to retain the positive values as it is. This activation function
decides whether

16
A HPGST Using CNN Architecture 2019-2020

Figure 3.1: CNN Architecture.

to activate a neuron or not. Then these images are given to max-pooling layer
in order to reduce the dimension of the image and to increase the computation
speed. Again, this pooled image is given for batch normalization which
normalizes the image within the range from 0 to 1 in each channel (RGB). And
this process is carried out for 3 more times with dropout layer in the alternate
time, the dropout randomly drops some units in our model so as to reduce the
model complexity and increase the speed of performance. We are also using a
regularizer where it sums all the squared weight values of the weight matrix to
calculate the loss function so that the weights get updated to reduce the model
loss. Now after all this steps the image pixel values are given to the flattening
layer which helps to convert to 1D from 2D format to feed the data to the fully
connected network for classification.

3.1 Preprocessing
In the preprocessing step, we processed the images from the different data sets
to a common and standard format in order to train the networks in a homogeneous
way. In input layer the images with random pixels are resized into the pixel size
of 224 x 224 x 3 by pre-processing. In our model, we are using a data
augmentation technique that involves creating a transformed versions of images
in the training dataset that belong to the same class as the original image. In our

Dept.of ECE,CBIT,Kolar 17
 A HPGST Using CNN Architecture 2019-2020
paper we use Transforms which include a range of operations from the field of
image manipu-

Dept.of ECE,CBIT,Kolar 18
lation, such as shifts, flip, and rotation. A shift to an image means moving all
pixels of the image in one direction, such as horizontally or vertically while
keeping the image dimensions same. The “width shift range and height shift
range” argu- ments to the ImageDataGenerator constructor control the amount
of horizontal and vertical shifts respectively. A rotation augmentation
unsystematically rotates the image clockwise by a given number of degrees from
0 to 360. The rotation will likely rotate pixels out of the image frame and leave
areas of the frame with no pixel data that must be filled in Image data
augmentation which is typically only applied to the training dataset, and not to the
test dataset. The “Keras” deep learning library which provides the ability to use
data augmentation automatically when training a model. It is achieved using
ImageDataGenerator class.

3.2 Feature Extraction

Feature extraction block consists of the input layer followed by convolutional
layers, max-pooling pooling layers, batch normalization layers, and dropout layer,
in this block we use Regularization to avoid the overfitting. The training data has
been fed into the Convolutional layer, where it is the key component of Convo-
lutional neural networks to detect the presence of a set of features in the images
received as input. We are using 8 layers which consist of 4 convolution layers and
4 fully connected layers in Convolutional neural network. The several
convolution kernels are applied to the image and returns the feature maps. We
apply the filter to features maps obtained with new kernels, which gives us new
features maps to normalize and resize. Finally, the values of the last feature maps
are arranging into a series. This series defines the output of the feature extraction
and the input to Classification.

3.3 Classification
Classification block is also knows as fully connected layer, this layers is
always last layer of all neural networks. The classifier which we are using is
Artificial neu- ral network (ANN), the specialty of this network is that it connects
each and every
neuron of its layer to the neuron of the next layer which yields better accuracy
when compared to other classifiers. In our model we are using 4 fully connected
layers, in first 3 fully connected layers it works along with dropout layer and the
last fully connected layer works along with an activation (sigmoid) function
which activates within the range of 0 to 1. We are also using an optimizer called
ADAM, to optimize the weight values during the back propagation. The back-
propagation process is nothing but it updates the weights of the model in reverse
order (from last step to the first) based on the error rate calculated in the model
using the optimizer. This back-propagation helps in increasing the accuracy of our
model. Learning rate is scheduled in our model for tuning the parameter in an
optimiza- tion algorithm which is used to determine the step size at each iteration
while moving toward a minimum of a loss function. Epoch is defined before
training our model; one epoch is when an entire dataset is passed in both forward
and backward through the neural network only once. For our model epoch is
given for 100 times (100 epochs). Using all the extracted features the binary
classifier classifies the images into infected and non-infected where, 0 indicates
glaucoma infected and 1 indicates non-infected glaucoma.
Chapter 4

WORK FLOW

Overall sequence of tasks that processes the set of data in a high performance
glaucoma screening technique using CNN architecture are as follows:

Figure 4.1: Work Flow.

20
A HPGST Using CNN Architecture 2019-2020

1. Database

The publicly available database called ACRIMA is used for the evalua-
tion of glaucoma classification methods. It consists of 705 images in which
396 images are glaucomatous images and 309 images are non-glaucomatous
images.

2. Labeling the database

The available database is then given for Labeling, in which glaucomatous

images (396) are annotated with the label ‘0’ and the non-glaucomatous
images (309) are annotated with the label ‘1’. A properly labeled dataset
provides a ground truth that the DL model uses to check its predictions for
accuracy and to continue refining its algorithm.

3. Splitting the database

The labeled database is then undergone Splitting, in which the labeled

database is randomly divided into 80% (564 labeled images) for training
purpose and 20% (141 labeled images) for testing purpose.

4. Preprocessing

The defined training dataset is then undergone Data augmentation as the

pre-processing technique, which enables to increase the diversity of data for
training the model by rotating, shifting and flipping operations. It is used
to overcome the problem of limited quantity and limited diversity of data.

5. Defining CNN model

The representation of the CNN model by defining the network layers

and network hyper-parameters. The CNN model mainly consists of 4
convolu- tional layers, 4 maxpooling layers and 4 fully connected layers.

(a) Convolutional layer

A convolutional layer is a fundamental component of the CNN

archi- tecture that performs feature extraction. It consists of a
combination of linear and nonlinear operations, that is convolution
operation and activation function.

Dept.of ECE,CBIT,Kolar 21
(b) Convolution
Convolution is a specialized type of linear operation used for feature
extraction, where a small array of numbers, called a kernel, is applied
across the input, which is an array of numbers. An element-wise prod-
uct between each element of the kernel and the input is calculated at
each location and summed to obtain the output value in the corre-
sponding position of the output, called a feature map. This procedure
is repeated applying multiple kernels to form an arbitrary number of
feature maps, which represent different characteristics of the inputs,
different kernels can thus, be considered as different feature extractors.
There are 4 important key hyper-parameters of the convolutional
layers

i. Number of filters - The number of different filters is used to

extract the different features from the images.
ii. Kernel size - It is also called as filter size refers to the
dimensions of the sliding window over the input. Small kernel
sizes are able to extract a much larger amount of information
containing highly local features from the input. Conversely, a
large kernel size extracts less information, which leads to a faster
reduction in layer dimensions.
iii. Padding – Zero-padding is the technique involves adding zeros
symmetrically around the edges of an input. This padding is used
when the kernel extends beyond the feature map. Padding con-
serves data at the borders of feature maps, which leads to better
performance, and it can help preserve the input’s spatial size,
which allows to build deeper, higher performing networks.
iv. Stride - It indicates how many pixels the kernel should be shifted
over at a time. As stride is decreased, more features are learned
because more data is extracted, which also leads to larger output
layers. Conversely, as stride is increased, this leads to more limited
feature extraction and smaller output layer dimensions.
(c) Nonlinear activation function
 The outputs of a linear operation such as convolution are then
passed through a nonlinear activation function ReLU, where negative
values are set to zero. The ReLU is used to overcome the vanishing
gradi- ent problem, and allows the models to learn faster and increase
the efficiency.

(d) Max pooling layer

The each convolutional layer is followed by Maxpooling layer,

which extracts patches from the input feature maps, outputs the
maximum value in each patch, and discards all the other. A pooling
layer provides a typical down sampling operation which reduces
spatial dimensionality of the feature maps in order to decrease the
number of subsequent learnable parameters.

(e) Fully connected layer

The output feature maps of the final convolution or pooling layer is

typically flattened, that is transformed into a one-dimensional array of
numbers and it is connected to one or more fully connected layers, also
known as dense layers, in which every input is connected to every
output by a learnable weights. Once the features extracted by the
convolution layers and down sampled by the pooling layers are
created, they are mapped by a subset of fully connected layers to the
final outputs of the network, such as the probabilities for each class in
classification tasks. Each fully connected layer is followed by a
nonlinear function, such as ReLU. The final fully connected layer
contains a single neuron in order to make predictions. It uses sigmoid
activation function in order to produce a probability output in the range
of 0 to 1 that can easily and automatically be converted to class
values.

Further, to boost up the performance, techniques like batch nor-

malization which normalizes the output of previous activation layer,
Dropout which randomly drop the nodes to reduce the learnable pa-
rameters and L2 Regularization which decreases the weight matrices,
all these techniques are used to avoid overfitting.
6. Compiling

Finally, we are using the logarithmic loss function (binary crossentropy)

during training, the preferred loss function for binary classification
problems. The model also uses the efficient Adam optimization algorithm
for gradient descent and accuracy metrics will be collected when the model
is trained. The output of this will be the summary of the model.

7. Scheduling learning rate

The learning rate will be specified for the optimizer which is the amount
of change to the model during each step of this search process, or the step
size where we systematically drop the learning rate after specific epochs
during training and it is considered to be the most important
hyperparameter to tune the neural network in order to achieve good
performance of the model.

8. Training CNN model

After creating the CNN model, then the model is undergone Training
with the training data sets, validating data sets and the hyperparameters
like batch size and epochs are tuned to get the better performance of the
model.

9. Model accuracy

Once the model is trained we can visualize training/validation accuracy

and loss, the model accuracy plots the graph of training accuracy versus
validation accuracy over the number of epochs and the model loss plots the
graph of training loss versus validation loss over the number of epochs from
which it is able to evaluate underfitting or overfitting of the model.

10. Determining the predictions

After training the CNN model, the testing datasets are given to make
predictions which helps to determine how well the CNN model is trained.

11. Calculating confusion matrix

A confusion matrix is a summary of prediction results on a classification

problem. It is a table with 4 different combinations of predicted and actual
 values that is True Positive, False Positive, True Negative and False
Negative classes. It is extremely useful for measuring Recall, Precision,
Specificity, Accuracy, f1-score and most importantly AUC-ROC Curve.

12. Plotting AUC-ROC curve

AUC - ROC curve is a performance measurement for classification prob-

lem at various thresholds settings. ROC is a probability curve and AUC rep-
resents degree or measure of separability. The ROC curve plots the graph of
True Positive Rate (TPR) over the False Positive Rate (FPR). It tells how
much the model is capable of distinguishing between classes. Higher the
AUC, better the model is at distinguishing between patients with glaucoma
and non-glaucoma.

13. Obtaining the ground truth

Obtaining the ground truth allows image data to be related to real

features and materials on the ground.
Chapter 5

SOFTWARE AND HARDWARE

REQUIREMENTS

5.1 Google colab

Colab is a Google research project created to help disseminate machine
learning education and research. Colab is a free Jupyter notebook environment
that runs entirely in the cloud. Most importantly, it does not require a setup and
the note- books that we create can be simultaneously edited by us. Colab supports
many popular machine learning, deep learning libraries which can be easily
loaded in colab notebook.

It supports Python, but not R or Scala yet. Google colab supports free graph-
ics processing unit. The types of graphics processing units that are available in
Colab vary over time. This is necessary for Colab to be able to provide access to
these resources for free. The graphics processing units available in Google Colab
often include Nvidia K80s, T4s, P4s and P100s. There is no way to choose what
type of graphics processing unit you can connect to in Google Colab at any given
time.

Colab is able to provide free resources in part by having dynamic usage limits
that sometimes fluctuate, and by not providing guaranteed or unlimited resources.
This means that overall usage limits as well as idle timeout periods, maximum
VM lifetime, graphics processing unit types available, and other factors vary over
time.
26
A HPGST Using CNN Architecture 2019-2020

Colab does not publish these limits, because they can vary quickly.

Notebooks run by connecting to virtual machines that have maximum

lifetimes that can be as much as 12 hours. Notebooks will also disconnect from
virtual ma- chines when left idle for too long. Maximum virtual machine lifetime
and idle timeout behaviour may vary over time, or based on our usage. This is
necessary for Colab to be able to offer computational resources for free.

Colab is ideal for everything from improving our Python coding skills to
work- ing with deep learning libraries, like PyTorch, Keras, TensorFlow, and
OpenCV. We can create notebooks in Colab, upload notebooks, store notebooks,
share note- books, mount our Google Drive and use whatever is stored in our
drive, we can upload our personal Jupyter Notebooks, upload notebooks directly
from GitHub, upload Kaggle files, download our notebooks, and use it.

5.2 GPU
A graphics processing unit (GPU) is a specialized electronic circuit designed
to rapidly manipulate and alter memory to accelerate the creation of images in
a frame buffer intended for output to a display device. The Graphics Card is re-
sponsible for rendering an image to your monitor, it does this by converting data
into a signal your monitor can understand. This device is made freely available
through cloud based google colab with limited access session. Google Colab pro-
vides a single 12GB NVIDIA Tesla K80 graphics processing unit that can be used
up to 12 hours continuously. Recently, Colab also started offering free TPU.

To use the google colab in a GPU mode, make sure the hardware accelerator
is configured to graphics processing unit. Sometimes, all graphics processing
units are in use and there is no graphics processing unit available. If this is the
case, we will get the following alert, and have to wait for a while and try again.

Dept.of ECE,CBIT,Kolar 27
5.3 Deep learning frame work

5.3.1 Tensor Flow

TensorFlow is an open source library developed by Google for its internal use.
TensorFlow is Google Brain’s second-generation system. Its main usage is in ma-
chine learning and dataflow programming. The name TensorFlow derives from
the operations that such neural networks perform on multidimensional data arrays,
which are referred to as tensors. TensorFlow, as the name indicates, is a frame-
work to define and run computations involving tensors. A tensor is a
generalization of vectors and matrices to potentially higher dimensions.
Internally, TensorFlow represents tensors as n-dimensional arrays of base
datatypes. TensorFlow com- putations are expressed as stateful dataflow graphs.
Its flexible architecture al- lows for the easy deployment of computation across a
variety of platforms (CPUs, GPUs, TPUs), and from desktops to clusters of
servers to mobile and edge devices.

TensorFlow is the most famous deep learning library these recent years. A
practitioner using TensorFlow can build any deep learning structure, like CNN,
RNN or simple artificial neural network. TensorFlow is mostly used by
academics, start-ups, and large companies. Google uses TensorFlow in almost all
Google daily products including Gmail, Photo and Google Search Engine. They
built a framework called Tensor Flow to let researchers and developers work
together on an AI model. Once developed and scaled, it allows lots of people to
use it.

5.3.2 Keras

Keras is an open-source neural-network library written in Python. It is capa-

ble of running on top of TensorFlow, Microsoft Cognitive Toolkit, R, Theano, or
PlaidML. Designed to enable fast experimentation with deep neural networks, it
focuses on being user-friendly, modular, and extensible. It was developed as part
of the research effort of project ONEIROS (Open-ended Neuro-Electronic
Intelligent Robot Operating System). It is an interface rather than a standalone
machine learning framework. It offers a higher-level, more intuitive set of
abstractions that make it easy to develop deep learning models regardless of the
computational
backend used. Microsoft added a Microsoft cognitive tool kit backend to Keras as
well, available as of CNTK v2.0.

Keras contains numerous implementations of commonly used neural-network

building blocks such as layers, objectives, activation functions, optimizers, and a
host of tools to make working with image and text data easier to simplify the
coding necessary for writing deep neural network code. The code is hosted on
GitHub, and community support forums include the GitHub issues page, and a
Slack channel.

In addition to standard neural networks, Keras has support for convolutional

and recurrent neural networks. It supports other common utility layers like
dropout, batch normalization, and pooling. Keras allows users to productize deep
models on smartphones (iOS and Android), on the web, or on the Java Virtual
Machine. It also allows use of distributed training of deep-learning models on
clusters of Graphics processing units (GPU) and tensor processing units (TPU)
principally in conjunction with CUDA.

5.4 Python
Python is an interpreted, high-level, general-purpose programming language.
Created by Guido van Rossum and first released in 1991, Python’s design philos-
ophy emphasizes code readability with its notable use of significant whitespace.
Its language constructs and object-oriented approach aim to help programmers
write clear, logical code for small and large-scale projects. Python is dynamically
typed and garbage-collected. It supports multiple programming paradigms, in-
cluding structured (particularly, procedural), object-oriented, and functional pro-
gramming. Python is often described as a “batteries included” language due to
its comprehensive standard library.

The Python 2 language was officially discontinued in 2020 (first planned for
2015), and “Python 2.7.18 is the last Python 2.7 release and therefore the last
Python 2 release.” No more security patches or other improvements will be released
for it. Python interpreters are available for many operating systems.

5.5 Operating system

Windows 10 is a series of operating systems produced by the American multi-
national technology company Microsoft and released as part of its Windows NT
family of operating systems. It is the successor to Windows 8.1. One of Windows
10’s most notable features is its support for universal apps, an expansion of the
Metro-style apps first introduced in Windows 8. Universal apps can be designed
to run across multiple Microsoft product families with nearly identical code- in-
cluding PCs, tablets, smart phones, embedded systems, Xbox One, Surface Hub
and Mixed Reality. The Windows user interface was revised to handle transitions
between a mouse-oriented interface and a touch screen-optimized interface based
on available input devices—particularly on 2-in-1 PCs, both interfaces include an
updated Start menu which incorporates elements of Windows 7’s traditional Start
menu with the tiles of Windows 8. Windows 10 also introduced the Microsoft
Edge web browser, a virtual desktop system, a window and desktop management
feature called Task View, support for fingerprint and face recognition login, new
security features for enterprise environments, and Direct X12.

Microsoft aimed to have Windows 10 installed on over a billion devices

within three years of its release; this goal was ultimately reached in March 2020.
By April 2020, an estimated 70.98% of traditional PCs running Windows are
running Windows 10, and thus 55% of all traditional PCs use it. Across all
platforms (PC, mobile, tablet and console), 35% of devices run some kind of
Windows, Windows 10 or older.

5.6 Hardware requirements

A pc/laptop with an intel i5 processor of 64-bit operating system which has
8GB random access memory (RAM) and 1TB of read only memory (ROM).
Chapter 6

EXPERIMENTAL RESULTS

Our automatic glaucoma detection using retinal fundus images model was built
and tested on personal laptop with the processor of intel core i5 CPU @ 1.60
GHZ to 1.80 GHZ and random-access memory (RAM) of 8GB. The model was
imple- mented in deep learning frame work which is an open source library
(Tensor flow and Keras) @ python IDE (Google colab).

Using the 705 glaucoma and non-glaucoma retinal fundus images from ACRIMA
data set our model was randomly trained and tested. Where, 80% of the images
(396 images consisting of both glaucoma and non-glaucoma) were randomly chosen
for training and 20% of images (309 images consisting of both glaucoma and non-
glaucoma) were randomly chosen for testing our model. The model extracted the
features and trained the model to classify images into glaucoma and non-glaucoma.
The performance of our model was evaluated using statistical measures like preci-
sion, recall, f1-score, accuracy, AUC-ROC curve for three different learning rates
0.1, 0.001, 0.0001 with 100 epochs for each learning rate. A comparison of three
different learning rates with the statistical measures are shown in the table.

The mathematical expressions for statistical measures are as follows:

TP + TN
Accuracy =
TP + FN + TN + (6.1)
FP

Precision = TP FP
TP +

31
(6.2)

31
A HPGST Using CNN Architecture 2019-2020

Recall/Sensitivity =
TP
(6.3)
Specificity = TP +
FN (6.4)
TP
TN + FP
F1− precision recall
(6.5)
precision +
recall
Where,
TP = True Positive values
TN = True Negative
values FP = False Positive
values FN = False
Negative values

Our model obtained best performance at 0.0001 learning rate for 100 epochs
of accuracy 0.99, precision of 0.99, recall of 0.99, f1- score of 0.99 and we also
achieved the AUC-ROC curve area of 0.99 s shown in the graph

Figure 6.1: AUC- ROC Curve.

Dept.of ECE,CBIT,Kolar 32
Dept.of ECE,CBIT,Kolar 32
Chapter 7

APPLICATIONS, ADVANTAGES AND

DISADVANTAGES

7.1 Applications
1. It is used to screen the glaucoma at earlier stage.

2. CNN can learn appropriate features by themselves automatically.

3. Since it reduces the cost of Glaucoma screening technique, everyone can

afford to pay and take a test.

4. It reduces the work of the ophthalmologists in finding the cause of eye dis-
ease.

7.2 Advantages
1. The algorithm reduce the processing time taken by the manual computer
based algorithm without compromising.

2. It helps the ophthalmologists to start the treatment earlier.

3. The high computational classifier is developed to detect the Glaucoma and

non-glaucoma at early stage which helps to reduce the risk of vision loss.

33
A HPGST Using CNN Architecture 2019-2020

7.3 Disadvantages
1. In our project we are using ACRIMA data which contains Less number of
images.

2. By implementing our model in real time there is a chance of unemployment

for the lab technicians.

3. In convolutional layer if the images contains some degree of tilt/rotation

then CNN usually have the difficulty in classification.

4. The Convnet requires large dataset to process and to train the neural net-
work.

Dept.of ECE,CBIT,Kolar 34
Chapter 8

CONCLUSION AND FUTURE SCOPE

In our model, glaucoma can be diagnose using advance deep learning archi-
tecture using retinal funds images without using the hand-crafted features. To
develop a Glaucoma deep learning algorithms, a convolutional neural network
(CNN) supervised model is applied on 705 funds images to extract the features
through multilayer from raw pixel intensities. As we examine the positive effect
of architecture, the training strategy, the size of the Data set and collected data
from the online clinical. We have used data set of ACRIMA which is publicly
available in online. In this implemented algorithm contains 4 convolution layer
and 4 fully connected layer, we take on batch-normalization and max-pooling
layer. To in- crease our computation power and to avoid over-fitting we adopt
dropout and data augmentation. Precision (PRC), Sensitivity ( SN), Accuracy
(ACC), Specificity (SP) and statistical measure are utilized to evaluate the
performance of glaucoma deep learning algorithm. On average, the Precision
(PRC) of 99%, Accuracy (ACC) of 99%, Specificity (SP) of 99%, Sensitivity
(SN) of 99% were achieved. Conclusively, we evaluated the performance of the
integration of the data from the clinical history with color fundus images. The
result show a amendment in specificity and sensitivity with homogeneous AUCs.
Further tests with more data and incipient architectural approaches should be
developed and assessed to attest this line of work.

In future, Development of data desensitization methods will promote data

standardization and normalization. Tensor processing unit (TPU) can also be
used to accelerate the computational speed of matrix computations, denser vector

35
A HPGST Using CNN Architecture 2019-2020

processing, massive datasets and huge models. Further, simulation will be carried
out for clinical data, for real-time hardware implementation using raspberry pi
will also be investigated.

Dept.of ECE,CBIT,Kolar 36
References

[1] Andres Diaz Pintol, Sandra Morales, Valery Naranjo, Thomas K¨ohler, Jose
M. Moss and Amparo Navea. “CNNs for automatic glaucoma assessment
using fundus images: an extensive validation”. Diaz Pinto et al. Bio Med Eng
OnLine March-2019.

[2] Ali Serener and Sertan Serte “ Transfer Learning for Early and Advanced
Glaucoma Detection with Convolutional Neural Networks” IEEE Interna-
tional Conference January-2019 .

[3] Sang Phan, Shin’ichi Satoh, Yoshioki Yoda, Kenji Kashiwagi and Tetsuro
Oshika. “Evaluation of Deep convolutional neural network for glaucoma de-
tection” Research Center for Medical Bigdata (RCMB), National Institute of
Informatics, jan-2019.

[4] Muhammad Naseer Bajwa, Muhammad Imran Malik, Shoaib Ahmed Sid-
diqui, Andreas Dengel, Faisal Shafait, Wolfgang Neumeier and Sheraz Ahmed,
“Two-stage framework for optic disc localization and glaucoma
classification in retinal fundus images using deep learning”. BMC Medical
Informatics and Decision Making, july-2019.

[5] Juanj.Gomez-Valverde, Alfonso Anton, Gian Luca Fatti, Bart Liefers, Ale-
jandra Herranz, Andres Santos, Clara I.Sanchez and Maria J.Ledesma-
Carbayo,“Automatic glaucoma classification using colour fundus images
based on convolutional neural networks and transfer learning”. researchgate
published 25 Jan 2019.

[6] Al Ghamdi M, Li M, Abdel-Mottaleb M, & Shousha M.A (2019).“Semi-

supervised Transfer Learning for Convolutional Neural Networks for Glau-

37
A HPGST Using CNN Architecture 2019-2020

coma Detection”. ICASSP 2019 - 2019 IEEE International Conference on

Acoustics, Speech and Signal Processing (ICASSP).
[7] Thakoor K.A, Li X, Tsamis E, Sajda P, & Hood D.C (2019). “Enhancing the
Accuracy of Glaucoma Detection from OCT Probability Maps using Convo-
lutional Neural Networks”. 2019 41st Annual International Conference of the
IEEE Engineering in Medicine and Biology Society (EMBC).

[8] Nacer Eddine Benzebouchi, Nabiha Azizi and Seif Eddine Bouzaine“ Glau-
coma Diagnosis Using Cooperative Convolutional Neural Networks”. Inter-
national Journal of Advances in Electronics and Computer Science, ISSN:
2393-2835, Jan-2018.

[9] Jongwoo kim, Sema Candemir and George R.Thoma, Emily Y.Chew.
“Region of Interest Detection in Fundus Images Using DL and Blood vessel
Informa- tion”. IEEE 31st International symposium of computer-based
medical system jan-2018.

[10] Raghavendra U, Hamido Fujita, Sulatha V Bhandary, Anjan Gudigar, Jen

Hong Tan and U Rajendra Acharya“ Deep Convolution Neural Network for
Accurate Diagnosis of Glaucoma Using Digital Fundus Images”. Article in
Information Sciences , researchgate February 2018

[11] Baidaa Al Bander, Waleed Al-Nuaimy, Majid A.Al-Taee and Yalin Zheng.
“Automated Glaucoma Diagnosis using Deep Learning Approach”.14th In-
ternational MultiConference on Systems, Signals & Devices (SSD) 2017.

[12] Radia Touahri, Nabiha Azizi, Nacer Eddine Benzebouchi, Nacer Eddine
Ham- mami and Ouided Moumene. “A Comparative Study of Convolutional
Neural Network and Twin SVM for Automatic Glaucoma Diagnosis
Embedded & Distributed Systems”. (EDiS)IEEE, pp.16,2017.

[13] Pavithra G, Dr T.C Manjunath, Dr Krishnananda & Dr Dharmanna Lamani.“

Detection of Secondary Glaucoma in Human Eyes using Sophisticated Bio-
Medical Image Processing Algorithms”.IEEE International Conference on
Power, Control, Signals and Instrumentation Engineering (ICPCSI-2017)

Dept.of ECE,CBIT,Kolar 38