Hindawi

Journal of Healthcare Engineering

Volume 2022, Article ID 1601354, 10 pages
https://doi.org/10.1155/2022/1601354

Research Article
Segmentation and Classification of Glaucoma Using U-Net with
Deep Learning Model

M.B. Sudhan,1 M. Sinthuja,2 S. Pravinth Raja,3 J. Amutharaj,4 G. Charlyn Pushpa Latha,5

S. Sheeba Rachel,6 T. Anitha,5 T. Rajendran ,7 and Yosef Asrat Waji 8
1
Department of Artificial Intelligence and Machine Learning, MVJ College of Engineering, Bangalore, Karnataka, India
2
Department of Information Science and Engineering, M. S. Ramaiah Institute of Technology, Bangalore, Karnataka, India
3
Department of Computer Science & Engineering, Presidency University, Bangalore, Karnataka, India
4
Department of Information Science and Engineering, RajaRajeswari College of Engineering, Mysore Road, Bangalore,
Karnataka, India
5
Saveetha School of Engineering, Saveetha Institute of Medical and Technical Sciences (Deemed to be University), Chennai,
Tamilnadu, India
6
Department of Information Technology, Sri Sairam Engineering College (Autonomous), Chennai, Tamilnadu, India
7
Makeit Technologies, Coimbatore, Tamilnadu, India
8
Department of Chemical Engineering, College of Biological and Chemical Engineering Addis Ababa Science and Technology
University, Addis Ababa, Ethiopia

Correspondence should be addressed to T. Rajendran; rajendranthavasimuthuphd@gmail.com and Yosef Asrat Waji; yosef.asrat@
aastu.edu.et

Received 2 December 2021; Revised 7 January 2022; Accepted 12 January 2022; Published 16 February 2022

Academic Editor: Enas Abdulhay

Copyright © 2022 M.B. Sudhan et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Glaucoma is the second most common cause for blindness around the world and the third most common in Europe and the USA.
Around 78 million people are presently living with glaucoma (2020). It is expected that 111.8 million people will have glaucoma by
the year 2040. 90% of glaucoma is undetected in developing nations. It is essential to develop a glaucoma detection system for early
diagnosis. In this research, early prediction of glaucoma using deep learning technique is proposed. In this proposed deep learning
model, the ORIGA dataset is used for the evaluation of glaucoma images. The U-Net architecture based on deep learning
algorithm is implemented for optic cup segmentation and a pretrained transfer learning model; DenseNet-201 is used for feature
extraction along with deep convolution neural network (DCNN). The DCNN approach is used for the classification, where the
final results will be representing whether the glaucoma infected or not. The primary objective of this research is to detect the
glaucoma using the retinal fundus images, which can be useful to determine if the patient was affected by glaucoma or not. The
result of this model can be positive or negative based on the outcome detected as infected by glaucoma or not. The model is
evaluated using parameters such as accuracy, precision, recall, specificity, and F-measure. Also, a comparative analysis is
conducted for the validation of the model proposed. The output is compared to other current deep learning models used for CNN
classification, such as VGG-19, Inception ResNet, ResNet 152v2, and DenseNet-169. The proposed model achieved 98.82%
accuracy in training and 96.90% in testing. Overall, the performance of the proposed model is better in all the analysis.

1. Introduction individuals aged 40 have glaucoma, which ascends to one in

eight by age 80 [1]. Various glaucoma-related risk factors
It is important to diagnose glaucoma early on, which can have been established, where the elevated intraocular
reduce damage and loss of vision and ensure prompt and pressure (IOP) that damages the optic nerves and blood
appropriate care. The worldwide prevalence of glaucoma for vessels is the significant one. It can lead to total damage to
people ages 40 to 80 years is 3.54%. Each one out of 200 the optic nerves and cause vision loss, if glaucoma is left
2 Journal of Healthcare Engineering

untreated. This gradual and complete damage to the optic (POAG), primary congenital glaucoma (PCG), normal-
nerves is often followed by only mild or no symptoms, so it is tension glaucoma (NTG), pseudoexfoliative glaucoma
known as the “sneak thief of sight” [2]. (XFG), traumatic glaucoma (TG), uveitic glaucoma (UG),
Glaucoma is one of the most common causes of irre- pigmentary glaucoma (PG), and neovascular glaucoma. The
versible vision loss after cataracts worldwide, accounting for forms vary between different ethnicities in intensity, com-
12 percent of all blindness cases each year. The number of plexity, and occurrence. Open-angle and angle-closure
people affected by glaucoma between the ages of 40 and 80 is glaucoma are the two major forms of glaucoma [4]. Figure 2
expected to rise to 111.8 million by 2040. Furthermore, 2.4 is shown in optic nerve head structure.
percent of all people and 4.7 percent of those aged 70 and up The most common form of glaucoma is open-angle
are at risk of developing the disorder. Glaucoma is defined as glaucoma also referred to as wide-angle glaucoma. It hap-
the degeneration of retinal ganglion cells (RGCs) caused by a pens as a result of partial drainage canal blockage in which
variety of disorders. RGC degeneration may result in two the pressure slowly rises as the fluid is not properly drained.
major health concerns: Symptoms start from vision loss in the periphery and may
not be detected until central vision is impaired. Angle-
(i) Structural changes to the optic nerve head (ONH)
closure glaucoma caused by impulsive and aqueous drainage
and the nerve fiber layer
full blockage is often called acute glaucoma. The pressure
(ii) Concurrent functional failures of the field of vision increases exponentially, which quickly leads to vision loss. It
These two glaucoma side effects might induce peripheral is formed because of the angle of narrow drainage and the
vision loss and, if left unchecked, blindness. Besides early small and droopy iris. The iris is pulled inside the anterior
detection and treatment, there is no cure for glaucoma. It is angle of the eye against the trabecular mesh network
essential in developing automated techniques for detecting (drainage canals) leading to blockage and bulging of the iris
glaucoma early on [3]. A retinal fundus image is an essential forward [5].
tool for documenting the optic nerve’s health, vitreous, In most situations, this damage is caused by abnormal
macula, retina, and blood vessels. Ophthalmologists used a rise of the pressure inside the eye. The secretion rate is
fundus camera to take the retinal image. The retinal image equalised to the drainage rate in healthy eyes. Glaucoma
was used to diagnose eye disease like glaucoma. Glaucoma is occurs when the drainage canal was partially or entirely
a significant cause of global blindness that cannot be cured. blocked, leading to a surge in pressure known as intraocular
Glaucoma disease can change the cup region’s shape, which pressure that affects the optic nerves used to relay signals to
is the center portion of the ONH. The changes can be used as the brain where it is possible to perceive visual information.
a parameter for the early indicator of glaucoma. The ONH If this damage is left untreated, complete blindness will
transmits visual information from retina to the brain [2]. result. Hence, it is essential to diagnose glaucoma in early
Figure 1 shows the retinal fundus images. stage.
There are no initial glaucoma symptoms but will In this research, early prediction of glaucoma using deep
gradually damage the optic nerves and then results in learning technique is proposed. In this proposed deep
blindness. Thus, it is crucial to detect glaucoma as early as learning model, the ORIGA dataset is used for the evaluation
possible so that it can prevent visual damage. Physiologi- of glaucoma images. For segmentation, the U-Net seg-
cally, glaucoma is indicated by increased optic cup exca- mentation model is implemented in this model and a
vation. The increasing size of the optic cup will impact the pretrained transfer learning model, DenseNet-201, is used
size of the optic disc, and this relation is known as a cup-to- for feature extraction along with deep convolution neural
disc ratio (CDR). It means ophthalmologists can diagnose network (DCNN). The DCNN approach is used for the
glaucoma progression using the value of CDR measurement. classification, and the final results will be representing
The optic cup and optic disc segmentation will support to whether the glaucoma infected or not.
calculate the CDR from the retinal image [3]. The most
noticeable symptom of glaucoma is often a loss of side vi- 2. Related Works
sion, which might go unnoticed as the condition progresses.
This is why glaucoma is sometimes referred to as the sneaky Several study models have been developed by various au-
thief of vision. In the case of extreme intraocular pressure thors for the segmentation and classification of glaucoma
levels, headache, sudden eye pain, impaired vision, or the detection, each employing a different methodology and
formation of halos around lights might occur. algorithm from the others. As will be detailed more, the
(i) Loss of vision majority of them are deep learning-based models with
varying levels of performance analysis. The fact that retinal
(ii) Eye redness disease is such a terrible ailment makes it difficult to detect
(iii) Hazy eyes (specifically in infants) and distinguish between the two conditions.
(iv) Vomiting or nausea The most common approach used in most of the studies
to diagnose glaucoma was the acquisition of retinal scans
(v) Vision narrowing (tunnel vision) [4]
using digital capture equipment for visual content, which
There are many forms of glaucoma, including angle- was the most common procedure used in most of the studies.
closure glaucoma (ACG), primary open-angle glaucoma The scan images were then preprocessed to equalize the
Journal of Healthcare Engineering 3

(a) (b) (c) (d)

Figure 1: Retinal fundus images: (a) healthy eye, (b) early glaucoma, (c) moderate glaucoma, and (d) deep glaucoma [2].

CNN model took this into consideration and made a de-

cision on it. In this model, subnets of attention prediction,
pathological region localization, and classification were
combined to form an overall model. When it comes to
detecting glaucoma, the model has a 96.2 percent accuracy
rate and an AUC of 0.983. In several cases, the ROI was only
partially highlighted, and the minor problematic regions
were not correctly identified [7].
For the purpose of automatically segmenting the glau-
coma images, MacCormick et al. developed a new glaucoma
(a) (b)
detection algorithm based on spatial detection. The method
Figure 2: Structure of optic nerve head: (a) normal and was developed on the basis of four assumptions: segmen-
(b) glaucoma [3]. tation, deformation, shape, and size of the images were all
taken into consideration. After a segmentation of the cup
and disc of the retinal fundus images was completed, an
anomalies. During the preprocessing stage, blood vessels estimation of the cup/disc ratio (CDR) in 24 cross sections
were segmented and depicted in order to create a vessel free was performed to generate the pCDR (CDR profile). The
image. Furthermore, feature extraction was utilized to ef- results were compared between healthy discs and glau-
ficiently reduce the dimensions of an image in order to comatous discs on both external and internal validation,
represent the interesting areas of an image as a compact with the AUROC for internal validation being 99.6 percent
feature vector that could be used for precisely classifying the and for external validation being 91 percent [8].
large amount of data collected. Techniques such as textures, Juneja et al. proposed an artificial intelligence glaucoma
pixel intensity values, FFT coefficients, and histogram expert system that was based on the segmentation of the
models were employed in the process of feature extraction optic cup and disc. In order to automate the identification of
and classification. Data analysis and classification were ac- glaucoma, a deep learning architecture was designed, with
complished through the use of image classification, which CNN serving as the core element. In this model, two neural
involved examining the numerical aspects of an image. The networks were integrated and used for segmenting images of
data set was divided into several classifications based on the the optic disc and cup of fundus, which were taken from
results, such as normal or glaucoma, to facilitate analysis. different cameras. By examining 50 images, the model was
Prastyo et al. applied the U-Net segmentation technique able to segment the cup with 93 percent accuracy and the
to retinal fundus images in order to segment the optic cup. disc with 95.8 percent accuracy [9]. To diagnose glaucoma in
The segmentation of the optic cup and the optic disc aids in retinal fundus images, Diaz-Pinto et al. used five ImageNet
the achievement of improved performance in the detection trained models, including the VGG-16, VGG-19, ResNet50,
of glaucoma disease. The ROI based on the optic disc image Inception-v3, and Xception, all of which were trained using
was cropped and segmented with the help of the ImageNet data. Performance study revealed that the
U-Net algorithm. In order to obtain optimal training, an Xception model outperformed the other models by
adaptive learning rate optimization technique was applied, obtaining better results, and the Xception model was then
and the model attained a dice coefficient rate of 98.42 tested with five publicly accessible datasets for glaucoma
percent and a loss rate of 0.15 percent during testing [6]. A diagnosis to confirm its superiority. The Xception model was
model of attention-based CNN (AG-CNN) for identifying more efficient than other commonly used models [10] due to
glaucoma was proposed by Li et al. and it was tested on a its higher level of computing efficiency.
database known as the large-scale attention-based glaucoma With the help of deep learning, SynaSreng et al. de-
database (LAG). The removal of large levels of redundancy veloped an automated two-stage model for glaucoma di-
from fundus images may result in a reduction in the ac- agnosis and classification. Initially, the optic disc area was
curacy and reliability of glaucoma identification. The AG- segmented using DeepLabv3+ architecture, but the encoder
4 Journal of Healthcare Engineering

segment was replaced with several deep CNNs after the network (RNN) classification model that extracts not just the
initial segmentation. For classification, a trained DCNN was spatial features in the fundus images but additionally the
employed with three approaches: transfer learning, feature temporal features inherent in the consecutive images was
descriptors learning using SVM, and constructing an en- developed. Because CNN was designed to diagnose glau-
semble of techniques in transfer learning and feature de- coma, it was built on spatial information encoded in images.
scriptors learning, respectively. It was possible to segment CNN was used with RNN for increased performance in
the optic discs using DeepLabv3+ and MobileNet archi- detecting glaucoma based on both temporal and spatial
tectures because of the integration of the two systems. Five features [16].
separate glaucoma datasets were used in the classification
process, which was done using an ensemble of algorithms. 3. Proposed Methodology
Finally, utilizing the ACRIMA dataset, DeepLabv3+ and
MobileNet were able to achieve an accuracy of 99.7 percent In this research, the deep learning-based models are pro-
for OD segmentation and 99.53 percent for classification posed for segmentation and classification of glaucoma de-
using DeepLabv3+ and MobileNet [11]. tection using retinal fundus images collected from ORIGA
To diagnose diabetic retinopathy, Mateen et al. developed a database. For segmentation, the U-Net architecture is used
fundus image classification model that combined the VGG-19 and a pretrained DenseNet-201 architecture was used to
with principal component analysis (PCA) and singular value extract the features from the segmented image. For classi-
decomposition (SVD) and used the VGG-19. The model’s fication, the DCNN architecture is used to classify the images
performance in region segmentation, feature extraction and for detecting glaucoma.
selection, and classification has been improved by combining
the Gaussian mixture model with the VGG, PCA, and SVD
3.1. Dataset Description. The ORIGA dataset is used in this
[12, 13]. Fu et al. employed two deep learning-based glaucoma
research for evaluation [17]. The data set contains 650 images
detection techniques, multilabel segmentation network (M-
of the color retinal fundus with the extension (.jpg) and ground
Net) and disc-aware ensemble network, to detect the presence
truth with the extension (.mat). The retinal images were col-
of glaucoma (DENet). Initially, M-Net was utilized to solve the
lected by the Singapore Malay Eye Study (SiMES). ORIGA
segmentations of both the optic cup and the disc, and DENet
database shares clinical ground truth retinal images with the
was used to combine the deep hierarchical context of the global
public and provides open access for researchers to benchmark
fundus image with the local optic disc region in the initial
their computer-aided segmentation algorithms. ORIGA
stages. The CDR was calculated based on the segmentation of
dataset is open for online access upon request. After pre-
the optic cup and disc in order to determine the glaucoma risk.
processing, 650 image data were divided into 488 image data as
It is possible to get accurate results from an image without
training data, 162 image data as testing data [1, 6, 8, 11, and 13].
segmenting it using the DENet [13].
Jiang et al. developed a new multipath recurrent U-Net
model for segmenting retinal fundus image. The efficiency of 3.2. Segmentation Using U-Net. The deep learning algo-
the model was validated by the performance of two seg- rithm-based U-Net architecture is implemented for optic
mentation processes like optic cup and disc segmentation cup segmentation. The U-Net architecture is the most widely
and retinal vessel segmentation. The model achieved 99.67% used segmentation architecture for medical images. The
accuracy for optic disc segmentation, 99.50% for optic cup architecture of the U-Net segmentation process is shown in
segmentation, and 96.42% for retinal vessel segmentation by Figure 3. The retinal fundus image is given as input; the ROI
using the Drishti-GS1 dataset [14]. based on optical disc image is cropped and segmented using
Mahum et al. proposed an early-stage glaucoma diag- deep learning algorithm. The output of the segmentation will
nosis model based on deep learning-based feature extrac- be based on the optic cup, where the optic cup outline is
tion. Images were preprocessed in the first phase before the masked as shown in Figure 3.
region of interest was retrieved using segmentation. Then, Before segmenting the image, in preprocessing, the
using the hybrid features descriptors, such as CNN, histo- ground truth (mask) image was changed to (.png) so that an
gram of oriented gradients, local binary patterns, and algorithm could process it. To get the Optic Disk (OD) mask,
speeded up robust features, characteristics of the optic disc we used the equation (disc � double (mask > 0)), while for
were recovered from images including optic cup. Further- Optic Cup (OC), we employed the equation (cup � double
more, HOG was used to extract low-level features, while the (mask > 1)). After that, we took the region of interest (ROI)
LBP and SURF descriptors were used to extract texture from the retinal fundus image by using the ground truth
features. Furthermore, CNN was used to compute high-level from OD to take OC’s closest area.
characteristics. The feature selection and ranking technique A contracting path (left side) and an expansive path
of maximum relevance minimum redundancy was applied. (right side) are included in this architecture, joined by
Finally, multiclass classifiers such as SVM, KNN, and ran- multilevel skip connections (middle). Input to the con-
dom forest were used to determine if fundus images were tracting path is retinal fundus images, and predictions are
healthy or diseased [15]. generated from a final layer following the expansive path.
Gheisari et al. proposed a new method for detecting Each convolution layer has filter banks, each applying 3 × 3
glaucoma that combined temporal (dynamic vascular) and padded convolutions followed by a rectified linear activation
spatial (static structural) data. A CNN and recurrent neural unit whose functional form is f(z) � max (0, z) [18–20].
Journal of Healthcare Engineering 5

Input 1 112 112 224 112 112 1 Predicted Mask

Contraction Path Expansion Path

Encoder Decoder

2562
2562
2562

2562

2562
2562

2562
112 224 224 448 112 112
1282
1282

1282

1282

1282

1282
224 448 448 896 224 224
642

642

642

642

642

642
448 448 448

322 322 322

Conv 3×3, ReLU Dropout, then

conv 3×3, ReLU
MaxPool 2×2
Copy
Up-conv 2×2
Conv 1×1, sigmoid

Figure 3: U-Net architecture of segmentation [6].

There are three convolutional blocks each in the con- as it is implemented to identify the glaucoma images from the
tracting and expansive paths. Two convolutional layers input dataset by classifying the retinal fundus images. To
consist of a block in the contracting path followed by a max- extract features from the dataset, a pretrained DenseNet-201
pooling layer with a pool size of 2 × 2. A block contains a model is used, and the DCNN model is used for classification.
2 × 2 upsampling layer in the expansive path, a concate- 256 × 256 is the input image size. The architecture of the
nation from the contracting path with the corresponding DenseNet-201 with DCNN is shown in Figure 4.
block (i.e., a merged layer), a dropout layer, and two con- DCNN usually performs well with a larger data set over a
volutional layers. The connecting path includes two con- smaller one. TL could be useful in those CNN applications
volutional layers. Finally, a 1 × 1 convolutional layer with a where the data set is not huge. For applications with
sigmoid activation and a single filter to output pixel-wise comparatively smaller datasets, TL’s concept utilizes the
class scores is the final output layer. Every convolution layer learned model from large datasets such as ImageNet. This
in blocks 1, 2, and 3 includes 112, 224, and 448 filters in the removes the need for a large dataset and decreases the
contracting path, while blocks 5, 6, and 7 include 224, 122, lengthy training time as needed when generated from
and 122 filters in the expansive path individually. There are scratch by the deep learning algorithm. TL is a deep learning
448 filters in every convolutional layer in the connecting method that uses a model trained for a single assignment as a
path. The proposed DCNN differs from the original U-Net starting point to train a model for a similar assignment. It is
in the number of filters chosen for the model to fit into the typically much quicker and simpler to fine-tune a TL net-
GPU memory in each convolution layer and the use of work than training a network from scratch. By leveraging
dropouts in the expansive path. common models that have been already trained on large data
sets, it allows the training of models using similar small
labeled data. Training time and computing resources can be
3.3. DenseNet-201 with CNN. A DCNN model with pre- significantly decreased. With TL, the model does not need to
trained DenseNet-201 is proposed in this research [21]. This be trained for as many epochs (a complete training period on
DenseNet-201 model is based on deep transfer learning (DTL) the entire dataset) like a new model.
6 Journal of Healthcare Engineering

Original Image Output Image Groundtruth Image Here, the input layer channels are given by H0. A 1 × 1
convolution layer preceding each 3 × 3 convolution layer is
added to increase computational performance, which re-
duces the input feature maps that were usually higher than
the feature maps of output H. The 1 × 1 conv layer was
known as the bottleneck layer and generates feature maps.
FC layers act as a classifier in the classification stage. It uses
extracted features and assesses the probability of a segment
in the image. The architecture of DenseNet-201 is shown in
Figure 6.
To create nonlinearity and to reduce overfitting, the
activation function and dropout layer are typically used. Two
dense layers of 128 and 64 neurons were implemented for
classification. The DenseNet-201 feature extraction model
was used for binary classification preceded by the sigmoid
activation function to replace the softmax activation func-
tion utilized in the DenseNet-201 design. In the FC dense
layer, every neuron was FC in the prior layer. The FC layer
“i” whose input 2D feature map was extended to a 1D feature
vector can be mathematically described.
vi−1 � Bernoulli(p),
Figure 4: U-Net output image compared with ground truth image. zi−1 � vi− 1 ∗ di−1 , (3)
zi � f􏼐xk zi−1 + ui 􏼑.
Because of the feature reuse possibility by various layers,
the DenseNet-201 uses the condensed network that provides Here, the Bernoulli function produces a vector vi � 1
simple to train and highly parametrical effective models and randomly with a certain probability that obeys the 0-1
expands variety in the following layer input and enhances distribution. The dimension of the vector is di−1. The
the execution. On various data sets, such as CIFAR-100 and dropout strategy is used by the initial two layers of the FC
ImageNet, DenseNet-201 has shown remarkable results. layer to randomly block some neurons based on a defined
Direct connections from each previous layer to every sub- probability, which efficiently avoids overfitting situations in
sequent layer are added to boost connectivity in the Den- deep networks. The terms “x” and “u” describe the FC layer’s
seNet-201 model as shown in Figure 5. respective weighting and offset parameters. The function of
The concatenation of feature can be mathematically sigmoid activation was to convert nonnormalized outputs to
expressed as 0 or 1 as binary outputs. Therefore, it helps to classify the
images as nonglaucoma or glaucoma. The sigmoid function
fci � NLi 􏼐fc0 , fc1 , . . . , fci−1 􏼑. (1)
can be expressed as
Here, NLi(∙) was a nonlinear transformation that could 1
be described as batch normalization (BN) composite S� , (4)
function, accompanied by a rectified linear unit function 1 + e− 􏽐 xi ·zi 􏼁
(ReLU) and a (3 × 3) convolution layer.
For ease of implementation, [fc0, fc1, . . ., fci − 1] indicates where the neuron output is S. The weights and inputs, re-
the feature map concatenation according to layers 0 to i − 1 spectively, represent xi and zi.
are combined into a single tensor. Dense blocks are generated
in the network architecture for downsampling purposes, 4. Performance Analysis
divided by layers known as transition layer consisting of BN
followed by a 1 × 1 convolution layer and an average 2 × 2 The performance analysis of the proposed DCNN with the
pooling layer. DenseNet-201’s growth rate defines how dense U-Net and DenseNet-201 model is assessed using the dataset
architecture produces better results, and the “H” hyper- in this section. The model is evaluated using parameters such
parameter denotes it. Because of its structure, where feature as accuracy, precision, recall, specificity, and F-measure.
maps were regarded as a network’s global state, DenseNet-201 Also, a comparative analysis is conducted for the validation
performs adequately well even with a minimal growth rate. of the model proposed. The output is compared to other
Therefore, all function maps of the preceding layers have current deep learning models used for CNN classification,
access to each successive layer. Each layer includes “H” feature such as VGG-19, Inception ResNet, ResNet 152v2, and
maps to the global state where each count of input feature DenseNet-169. On the MATLAB 2019a Simulink toolbox, all
maps at ith layers (fm)i was expressed as the experiments are implemented and carried out. The
dataset is split into 75% for training and 30% for validating
(fm)i � H0 + H(i − 1). (2) the performance analysis.
Journal of Healthcare Engineering 7

Average Pooling
(2 × 2)
Segmented
Image
n1 channels
H1 H2 H3 H4 (3 × 3 × n1)
X1 X2 X3 X4

Fully Connected
(Dense)
ReLU Activation Fully Connected
(Dense)
Droput 0.2 ReLU Activation Fully Connected
Droput 0.3 (Dense)
Sigmold
Glaucoma
Flatten

Non-Glaucoma

OUTPUT

Figure 5: Feature extraction using pretrained DenseNet-201 model and classification using DCNN [21].

4.1. Performance Metrics. The primary objective of this re- compared with other techniques. The testing accuracy is
search is to detect the glaucoma using the retinal fundus 96.90%, which is 1.36% to 5.26% increased performance
images, which can be useful to determine if the patient was than the other existing compared models. The graphical
affected by glaucoma or not. The result of this model can be chart of the comparison is plotted in Figure 7.
positive or negative based on the outcome detected as in- Precision is a positive predictive value. It is the measure
fected by glaucoma or not. The true positive, true negative, of the cumulative predictive positive value of the correctly
false positive, and false negative are properly analyzed to predicted positive observation. The lower precision value
estimate the outcome of this model. reflects that a large number of false positives have affected
the classification model. The measure of precision can be
TP: it indicates the total predictions correctly obtained
computed using the following equation.
in positive cases
FP: it indicates the total predictions incorrectly ob- TP
Precision � . (6)
tained in positive cases TP + FP
TN: it indicates the total predictions correctly obtained The estimation of precision is tabulated in Table 2,
in negative cases which shows that the proposed model has achieved better
FN: it indicates the total incorrect predictions in precision value than the compared models. The model
negative cases obtained 98.63% precision rate in training, which was 1.1%
to 4.8% improved compared with other techniques. The
Accuracy is the model’s estimation of the performance precision rate in testing was 96.45%, which was 1.08% to
subset. It is the primary output metric used to calculate the 4.9% increased performance than the other existing
efficiency of the classification process. It is usually used to compared models. Figure 8 shows the comparison of
estimate when both the positive and negative classes are precision analysis.
equally important. It is calculated using the following The sensitivity is also referred to as recall. It is the ratio of
equation. properly predicted positive evaluation of the overall positive
TP + TN predictive value. The lower recall value reflects that a large
Accuracy � . (5) number of false negative values have affected the classifi-
TP + TN + FP + FN
cation model. The recall estimation can be calculated using
As shown in Table 1, the proposed model achieved better the following equation.
classification accuracy in both training and testing for
classifying the glaucoma fundus images. The model obtained TP
Recall � . (7)
98.82% training accuracy, which is 1.09% to 3.96% improved TP + FN
8 Journal of Healthcare Engineering

INPUT 99

Conv. (7×7), stride 2 98

97

Accuracy
Max Pool. (3×3), stride 2 96

Conv. (1×1) 95
Dense Block 1 ×6
Conv. (3×3) 94

93
Conv. (1×1)
92
Transition Layer 1
Avg Pool. (2×2), stride 2 VGG-19 Inception ResNet ResNet 152v2 DenseNet169 Proposed
Models
Conv. (1×1) Training
Dense Block 2 × 12 Testing
Conv. (3×3)
Figure 7: Graphical plot of accuracy.
Conv. (1×1)
Transition Layer 2 Table 2: Performance analysis of precision.
Avg Pool. (2×2), stride 2
Models Training Testing
Conv. (1×1) VGG-19 97.30 94.70
Dense Block 3 × 48 Inception ResNet 93.81 91.52
Conv. (3×3)
ResNet 152v2 97.28 93.02
Conv. (1×1)
DenseNet169 97.49 95.37
Proposed 98.63 96.45
Transition Layer 3
Avg Pool. (2×2), stride 2
99
Conv. (1×1) 98
Dense Block 4 × 32
Conv. (3×3)
97
Precision

Global Avg Pool. (7×7) 96

Classification Layer 95
Softmax 94

OUTPUT 93

92
Figure 6: DenseNet-201 architecture [21].
VGG-19 Inception ResNet ResNet 152v2 DenseNet169 Proposed
Table 1: Performance analysis of accuracy. Models
Training
Models Training Testing Testing
VGG-19 97.73 95.54 Figure 8: Graphical plot of precision.
Inception ResNet 94.86 91.64
ResNet 152v2 97.56 93.21
DenseNet169 97.14 95.45 Table 3: Performance analysis of recall.
Proposed 98.82 96.90
Models Training Testing
VGG-19 97.84 95.62
The proposed model has gained better recall or sensi- Inception ResNet 94.90 91.97
tivity rate as tabulated in Table 3. The model obtained ResNet 152v2 97.62 94.05
98.95% recall rate in training, which was 1.1% to 4.05% DenseNet169 97.35 95.69
improved compared with other techniques. The recall rate in Proposed 98.95 97.03
testing was 97.03%, which was 1.3% to 5.06% better per-
formance than the other existing compared models. The
comparison graph is plotted, as shown in Figure 9. TN
As per this model, specificity is the prediction that Specificity � . (8)
TN + FP
healthy subjects do not have the disease. It is the percentage
of subjects with no illness that is tested as negative. The As shown in Table 4, the proposed model has obtained a
specificity estimation can be calculated using the following better specificity rate than the other comparative models of
equation. deep learning.
Journal of Healthcare Engineering 9

99 Table 5: Performance analysis of F-measure.

98 Models Training Testing
97 VGG-19 97.52 95.39
Inception ResNet 94.79 91.55
96
Recall

ResNet 152v2 97.35 93.14

95 DenseNet169 97.07 95.09
94 Proposed 98.50 96.28

93

92
98
VGG-19 Inception ResNet ResNet 152v2 DenseNet169 Proposed
97
Models
Training 96

F-Measure
Testing
95
Figure 9: Graphical plot of recall. 94

93
Table 4: Performance analysis of specificity. 92

Models Training Testing VGG-19 Inception ResNet ResNet 152v2 DenseNet169 Proposed
VGG-19 97.24 95.67 Models
Inception ResNet 94.05 89.92
Training
ResNet 152v2 97.28 92.73 Testing
DenseNet169 97.00 94.89
Proposed 98.15 96.33 Figure 11: Graphical plot of F-measure.

2 × Precision × Recall
F − measure � . (9)
98 Precision + Recall
97
The F-measure estimation is tabulated in Table 5, which
96
represents that the proposed model has achieved better
95 F-measure value than the compared models. The model
Specificity

94 obtained 98.50% F-measure rate in training, which was 0.9%

93 to 3.7% improved compared with other techniques. The
92 F-measure rate in testing was 96.28%, which was 0.8% to
91
4.7% better performance than the other existing compared
models. Figure 11 shows the comparison of F-measure
90
analysis.
VGG-19 Inception ResNet ResNet 152v2 DenseNet169 Proposed In this research, by comparing all the models like VGG-
Models 19, Inception ResNet, ResNet 152v2, and DenseNet-169, the
Training proposed model has achieved better performance in both the
Testing training and testing stages. The least performance achieved
Figure 10: Graphical plot of specificity. model is Inception ResNet and DenseNet-169 has some
close performance to the proposed model.

The model obtained 98.15% specificity rate in training, 5. Conclusion

which was 0.8% to 4.1% improved compared with other
techniques. The specificity rate in testing was 96.33%, which In this research, early prediction of glaucoma detection
was 0.6% to 6.4% better performance than the other existing model using deep learning technique was proposed. In this
compared models. Figure 10 represents the comparison of proposed deep learning model, the ORIGA dataset was used
specificity estimated. for the evaluation of glaucoma images. 75% of the data was
The F-measure estimates the accuracy of the test and is used for training and 25% of data was used for testing. For
defined as the weighted harmonic mean of the precision of segmentation, the U-Net segmentation model was imple-
the test and the recall. The accuracy does not take into mented in this model and a pretrained transfer learning
account how the data was distributed. The F-measure is then model, DenseNet-201, was used for feature extraction along
utilized to manage the distribution problem with accuracy. with DCNN. The DCNN approach was used to classify the
When the data set has imbalance classes, it was useful. The images for glaucoma detection. The primary objective of this
F-measure estimation can be calculated using the following model was to detect the glaucoma using the retinal fundus
equation. images, which can be useful to determine whether the
10 Journal of Healthcare Engineering

patient is affected by glaucoma or not. By segmenting the [6] P. H. Prastyo, A. S. Sumi, and A. Nuraini, “Optic cup seg-
fundus images, the optic cup region was segmented and mentation using U-net architecture on retinal fundus image,”
compared with ground truth images from the dataset. After Journal of Information Technology and Computer Engineering,
segmentation, the features were extracted from the images vol. 4, no. 2, pp. 105–109, 2020.
using DenseNet model and classified using DCNN. The [7] L. Li, M. Xu, H. Liu et al., “A large-scale database and a CNN
model for attention-based glaucoma detection,” IEEE
proposed model obtained 98.82% training accuracy, which
Transactions on Medical Imaging, vol. 39, no. 2, pp. 413–424,
was 1.09% to 3.96% higher compared with other models and 2020.
the testing accuracy was 96.90%, which was 1.36% to 5.26% [8] I. J. C. MacCormick, B. M. Williams, Y. Zheng et al., “Ac-
higher than the compared models. By analyzing the per- curate, fast, data efficient and interpretable glaucoma diag-
formance analysis, the results obtained by the proposed nosis with automated spatial analysis of the whole cup to disc
model are efficient and the reason for not achieving 100% profile,” PLoS One, vol. 14, no. 1, Article ID e0209409, 2019.
results was due to the false positives and false negatives. In [9] M. Juneja, S. Singh, N. Agarwal et al., “Automated detection of
future, this imbalance issue will be sorted out by improving Glaucoma using deep learning convolution network (G-net),”
the classifier and reducing the threshold. This model can be Multimedia Tools and Applications, vol. 79, no. 21-22,
useful for various medical image segmentation and classi- pp. 15531–15553, 2020.
[10] A. Diaz-Pinto, S. Morales, V. Naranjo, T. Köhler, M. Jose, and
fication processes like diabetic retinopathy, brain tumor
A. Navea, “CNNs for automatic glaucoma assessment using
detection, breast cancer detection, etc. fundus images: an extensive validation,” BioMedical Engi-
neering Online, vol. 18, no. 29, pp. 1–19, 2019.
Data Availability [11] N. M. SynaSreng, K. Hamamoto, and K. Y. Win, “Deep learning
for optic disc segmentation and glaucoma diagnosis on retinal
The datasets used and/or analyzed during the current study images,” Applied Sciences, vol. 10, no. 4916, pp. 1–19, 2020.
are available from the corresponding author on reasonable [12] M. Mateen, J. Wen, S. Sun, and Z. Huang, “Fundus image
request. classification using VGG-19 architecture with PCA and SVD,”
Symmetry, vol. 11, no. 1, pp. 1–12, 2019.
[13] H. Fu, J. Cheng, Y. Xu, and J. Liu, “Glaucoma detection based
Conflicts of Interest on deep learning network in fundus image,” in Deep Learning
and Convolutional Neural Networks for Medical Imaging and
The authors declare no conflicts of interest.
Clinical Informatics, Advances in Computer Vision and Pat-
tern Recognition, L. Lu, X. Wang, G. Carneiro, and L. Yang,
Authors’ Contributions Eds., Springer, Cham, pp. 119–137, 2019.
[14] Y. Jiang, F. Wang, J. Gao, and S. Cao, “Multi-path recurrent
Authors Sudhan M B and Sinthuja M are responsible for U-net segmentation of retinal fundus image,” Applied Sci-
project design and concept. Authors Pravinth Raja S and J ences, vol. 10, no. 3777, pp. 1–17, 2020.
Amutharaj are responsible for surveys and content writing [15] R. Mahum, S. U. Rehman, O. D. Okon, A. Alabrah, T. Meraj,
and proofreading. Authors Sheeba Rachel S and Charlyn and H. T. Rauf, “A novel hybrid approach based on deep CNN
pushpaLatha G are responsible for algorithm design, de- to detect glaucoma using fundus imaging,” Electronics, vol. 11,
velopment, and proofreading. no. 1, Article ID 26, 2022.
[16] S. Gheisari, S. Shariflou, J. Phu et al., “A combined con-
volutional and recurrent neural network for enhanced
References glaucoma detection,” Scientific Reports, vol. 11, no. 1945,
pp. 1–11, 2021.
[1] M. C. V. Stella Mary, E. B. Rajsingh, and G. R. Naik, “Retinal [17] Z. Zhang, F. S. Yin, J. Liu et al., “ORIGA(-light): an online
fundus image analysis for diagnosis of glaucoma: a com- retinal fundus image database for glaucoma analysis and
prehensive survey,” IEEE Access, vol. 4, pp. 4327–4354, 2016. research,” Annual International Conference of the IEEE En-
[2] S. Anindita and H. Agus, “Automatic glaucoma detection gineering in Medicine and Biology Society, vol. 2010, no. 1,
based on the type of features used: a review,” Journal of pp. 3065–3068, 2010.
Theoretical and Applied Information Technology, vol. 72, no. 3, [18] E. A. M. Yasir and H. I. Bashir, “Classification of diabetic
pp. 366–375, 2015. retinopathy using stacked autoencoder-based deep neural
[3] A. Sarhan, J. Rokne, and R. Alhajj, “Glaucoma detection using network,” Journal of Computational Science and Intelligent
image processing techniques: a literature review,” Comput- Technologies, vol. 1, no. 1, pp. 9–14, 2020.
erized Medical Imaging and Graphics: The Official Journal of [19] C. Narmatha, A. Hayam, and A. H. Qasem, “An analysis of
the Computerized Medical Imaging Society, vol. 78, Article ID deep learning techniques in neuroimaging,” Journal of
101657, 2019. Computational Science and Intelligent Technologies, vol. 2,
[4] B. N. Kumar, R. P. Chauhan, and N. Dahiya, “Detection of no. 1, pp. 7–13, 2021.
glaucoma using image processing techniques: a review,” in [20] P. Manimegalai and P. K. Jayalakshmi, “A study on diabetic
Proceedings of the 2016 International Conference on Micro- retinopathy detection using image processing,” Journal of
electronics, Computing and Communications (MicroCom), Computational Science and Intelligent Technologies, vol. 2,
pp. 1–9, Durgapur, India, January 2016. no. 1, pp. 21–26, 2021.
[5] D. M. S. Barros, J. C. C. Moura, C. R. Freire, A. C. Taleb, [21] A. Jaiswal, N. Gianchandani, D. Singh, V. Kumar, and M. Kaur,
R. A. M. Valentim, and P. S. G. Morais, “Machine learning “Classification of the COVID-19 infected patients using
applied to retinal image processing for glaucoma detection: DenseNet201 based deep transfer learning,” Journal of Bio-
review and perspective,” BioMedical Engineering Online, molecular Structure and Dynamics, vol. 39, no. 1, pp. 1–8, 2020.
vol. 19, no. 20, pp. 1–21, 2019.