Standard Article

Concurrent Engineering: Research

and Applications
Automated glaucoma detection from 1–12
Ó The Author(s) 2021
fundus images using wavelet-based Article reuse guidelines:
sagepub.com/journals-permissions

denoising and machine learning DOI: 10.1177/1063293X211026620

journals.sagepub.com/home/cer

Sibghatullah I. Khan1 , Shruti Bhargava Choubey1, Abhishek Choubey1,

Abhishek Bhatt2, Pandya Vyomal Naishadhkumar1 and Mohammed
Mahaboob Basha1

Abstract
Glaucoma is a domineering and irretrievable neurodegenerative eye disease produced by the optical nerve head owed
to extended intra-ocular stress inside the eye. Recognition of glaucoma is an essential job for ophthalmologists. In this
paper, we propose a methodology to classify fundus images into normal and glaucoma categories. The proposed
approach makes use of image denoising of digital fundus images by utilizing a non-Gaussian bivariate probability distribu-
tion function to model the statistics of wavelet coefficients of glaucoma images. The traditional image features were
extracted followed by the popular feature selection algorithm. The selected features are then fed to the least square
support vector machine classifier employing various kernel functions. The comparison result shows that the proposed
approach offers maximum classification accuracy of nearly 91.22% over the existing best approaches.

Keywords
Glaucoma fundus images, wavelet transform, feature extraction, feature selection, LS-SVM

Introduction 1991), and Heidelberg retina tomography (Mistlberger

et al., 1999) are generally used imaging approaches for
Glaucoma is a neurodegenerative eye disorder; early the diagnosis of glaucoma.
detection might also stop imaginative and prescient loss Computer-assisted diagnosis (CAD) of glaucoma is
and later stage detection may lead to everlasting loss of a recent field for the study of visual images. Current
vision. It damages the optic nerve head (ONH) due to new techniques of image analysis and high-speed
excessive intra-ocular strain (IOP) in the eye resulting machines have inspired CAD to be a more reliable tool
in permanent blindness (Bock et al., 2010). for the identification of glaucoma (Noronha et al.,
There are two methods of glaucoma reliant on the 2014). Centre-surround data and histograms have been
angle: The elementary form of glaucoma is known used for super-pixel sorting for glaucoma screening
mainly as open-angle glaucoma (Weinreb et al., 2014). (Cheng et al., 2013). Kolar and Jan (2008) have used
In this sympathetic glaucoma, the angle is unfastened fractal dimensions (FDs) and power spectral features
by the drainage canals. It’s growing increasingly and with a support vector machine (SVM) for detecting
it’s not perceptible. Deferment in credentials of glau-
coma may promote total vision loss. Known methods 1
Department of Electronics and Communication Engineering, Sreenidhi
for the diagnosis of glaucoma are anatomical transfor-
Institute of Science and Technology, Hyderabad, Telangana, India
mations in the retina and genetics (Sommer et al., 2
Department of Electronics and Telecommunication, College of
1979), the ONH investigative test (Kavitha et al., Engineering, Pune, India
2014), and retina investigation (Davis et al., 2016).
Computer-based diagnosis for diagnosing glaucoma Corresponding author:
Sibghatullah I Khan, Department of Electronics and Communication
(Anderson et al., 2018) has rendered it simpler like Engineering, Sreenidhi Institute of Science and Technology, Yamnampet,
scanning laser ophthalmoscope (Webb and Hughes, Hyderabad, Telangana 501301, India.
1981), optical coherence tomography (Huang et al., Email: [email protected]
2 Concurrent Engineering: Research and Applications 00(0)

glaucoma from fundus images. Nyul (2009) used pixel Analysis Group, n.d.). The main contribution of the
intensities and fast Fourier transform (FFT) as a fea- present paper is the hybridization of image denoising
ture for classifying glaucoma and normal fundus with machine learning approaches, which is an unex-
image. Raja and Gangatharan (2013) have used com- plored area for fundus image analysis for glaucoma
plex wavelet transform and higher-order spectra for the detection. Thus, the proposed approach is suitable in
analysis of normal and glaucoma fundus images. Raja circumstances where computationally extensive
and Gangatharan (2015) have used wavelet packet approaches like deep learning are not feasible.
decomposition (WPD) combined with entropy and The remaining parts of this paper are given as fol-
energy geographies as a feature to detect glaucoma lows: In section II we have discussed the methodology
from fundus images. Townsend et al. (2008) have uti- which includes, the image database, pre-processing, fea-
lized Heidelberg retinal tomograph (HRT), recursive ture extraction, feature selection, and classification. The
screen, and corrosion tree with SVM classifier for ana- outcomes achieved and their discussions are listed in
lysing and classifying normal and glaucoma fundus section III. Section IV concludes the paper.
images. Kirar and Agrawal (2018) have suggested the
identification of glaucoma from fundus images using
DWT and histogram functions. In Maheshwari et al. Methodology
(2017), the fixed number of modes for variational mode The various steps of the proposed approach are
decomposition (VMD) has been proposed to detect explained in this section. These steps are pre-process-
glaucoma. In Kirar and Agrawal (2019), 2D-DWT has ing, feature selection using the NCA method, and
been proposed to decompose glaucoma and healthy
LS-SVM classifier to process fundus images. The meth-
images and classification has been performed over
odology of the proposed approach using the CAD tool
histogram-based features. Yadav et al. (2014) used
is shown in Figure 1. The input image is a fundus
texture-based feature extraction and classification of
image and data are taken from the MIAG image data-
glaucoma fundus images. To overcome the shortcom-
base. The extraction of glaucoma and healthy images
ings of Empirical mode decomposition (EMD) (Yu
from fundus image is performed in five different
et al., 2020), Agrawal et al. (2019) have proposed glau-
stages as
coma detection using quasi-bivariate variational mode
decomposition (QB-VMD) from digital fundus images.
Maheshwari et al. (2017) decomposed fundus images Stage-1: The R-component, B-component, and G-com-
utilizing empirical wavelet transform (EWT). ponent of the fundus input image are extracted.
Almost all the state-of-the-art techniques require a Stage-2: The pre-processing of R-plane, B-plane, and
tremendous amount of computational resources due to G-plane was performed.
image conversion in the frequency domain or decom- Stage-3: The different feature extraction techniques are
position using computationally extensive methods like used. The details of extracted features of fundus images
EMD and VMD. are shown in Table 1.
In this paper, we have employed spatial features with Stage-4: Various features selection approaches have
a denoising approach for RGB fundus images. The been used, out of which, the neighbourhood compo-
inclusion of all three planes has been carried out to nent analysis (NCA) approach has resulted in better
avoid any possible loss of information. Various feature results.
selection and ranking techniques have been tested to Stage-5: The LV-SVM classifier is used to classify
reduce the dimensionality of the feature space. In the healthy and glaucoma images.
recently reported literature related to glaucoma diagno-
sis from fundus images, the least square support vector
machine (LS-SVM) has been proved to be effective for
classifying fundus images in normal and glaucoma cate-
Image database
gories (Kirar and Agrawal, 2019; Martins et al., 2020; The approach described is tested on 455 images of
Parashar and Agrawal, 2020). Hence in the present RIM-One of MIAG (Medical Image Analysis Group,
study, we have used LS-SVM as a classifier that cate- n.d.). There are 200 fundus images that correspond to
gorizes glaucoma and normal fundus images using the the glaucoma category and 255 fundus images corre-
pull-out features.LS-SVM classifier with linear, polyno- spond to the normal category. It’s worthwhile to note
mial, and RBF kernel has been tested for all the possi- here that the glaucoma category also contains the pre-
ble feature spaces formed by various ranking methods. glaucoma fundus images. The proposed approach is
Accuracy, sensitivity, and specificity were measured implemented using MATLAB 2017 (b) running on a
and tested using tenfold cross-validation. The suggested 2.10 GHz, 12 GB RAM laptop with an Intel i7
approach applies to 455 MIAG images (Medical Image processor.
Khan et al. 3

Figure 1. Methodology of the proposed approach using CAD tool.

Table 1. Feature extraction parameter and corresponding mapping with R-G-B plane (Hira and Gillies, 2015; Khaire and
Dhanalakshmi, 2019).

S.no Parameter Formula Description Feature number

R G B

1 Contrast P
X P
Y It measures local 1 9 17
jx  yj2 P(x, y) intensity variation.
x¼1 y¼1
2 Mean (m) 1 XX XY It is defined as the measure 2 10 18
P(x, y) of average intensity and denoted by m.
XY x¼1 y¼1
3 Homogeneity P
X P
Y P(x, y) It is used to quantify the 3 11 19
1 þ jx  yj distribution of pixel intensities
x¼1 y¼1
in the region of image.
4 Energy (En) X P
P Y It is defined as the measure 4 12 20
P(x, y)2 of uniformity in the range [0, 1].
x¼1 y¼1
5 Entropy (H) P X P Y It is defined as the measure of 5 13 21
 P(x, y)log2 ½P(x, y) uncertainty and randomness.
vx¼1 y¼1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
6 Standard Deviation (s) u X It is used to quantify the closeness 6 14 22
u1 X X Y
t ðPðx, y Þ  mÞ2 of pixel intensities to the mean.
XY x¼1 y¼1
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
7 Root mean square (RMS) u X To quantify the distribution of 7 15 23
u1 X X Y
t ðP(x, y Þ2 pixel intensities of the image.
XY x¼1 y¼1
8 Variance (s2 ) 1 XX X Y It is defined as the measure of 8 16 24
ðPðx, yÞ  mÞ2 the spread of pixels value about
XY x¼1 y¼1
mean and denoted by s2.

Pre-processing present paper, we have proposed DWT based bivariate

Pre-processing is a very critical phase in the processing shrinkage method with the consideration of all three
of images. It eliminates unnecessary inconsistencies that is, RGB plans. To the best of our knowledge, this
such as light, non-uniform lighting and poor contrast. hybridization is not implemented for glaucoma images
In pre-processing, various approaches have been as a pre-processing step.
reported in literature such as discrete wavelets trans- The block diagram of pre-processing steps of the
form (DWT) (Kirar and Agrawal, 2018), EWT (Kirar, proposed system is shown in Figure 2. In this paper we
2018), QB-VMD (Agrawal et al., 2019), filters (Yu have used the original RGB images and resized them to
et al., 2020). In order to achieve better accuracy, in the form (512 3 512), to be of the same size and quality
4 Concurrent Engineering: Research and Applications 00(0)

Figure 2. Block diagram of pre-processing steps of the proposed system and image filtering.

for improved interpretation of data on the same scale. X, Y denotes total number of pixels in an image.
All planes in the images are conserved to retain maxi-
mum information. Denoising is important in the area (d) Feature Selection: It is important to select sig-
of medical imaging. In general, optical coherence tomo- nificant features from a large feature set (Li
graphy (OCT) images are corrupted by speckle noise et al., 2020). This procedure overwhelms the
(Sahu et al., 2019). Speckle noise is a multiplicative problem of overfitting, lessens the data size,
noise. The value of the noisy pixel is the multiplication which diminishes the complexity of the scheme
of the noise value and the original pixel value. As a and hence growths speed and classification cor-
pre-processing step denoising procedure is as follows. rectness of the system. The details of feature
We have proposed the use of simple non-Gaussian selection and ranking methods used in the pres-
bivariate probability distribution function to model the ent work are shown in Table 2. For getting
statistics of wavelet coefficients of the fundus images. more insight into these methods, the readers
The model captures the dependence between a wavelet are encouraged to refer (Hira and Gillies, 2015;
coefficient and its parent. Using Bayesian estimation Khaire and Dhanalakshmi, 2019).
theory, we derive from this model a simple non-linear
shrinkage function for wavelet denoising, which gener-
alizes the soft thresholding approach. Least squares support vector machine
Briefly, the algorithm is summarized as,
The support vector machine (SVM) classifier is a popu-
lar classification algorithm used in pattern recognition
1. Calculate the noise variance.
tasks (Suykens and Vandewalle, 1999). SVM constructs
2. For each wavelet coefficient,
optimal hyperplane in higher dimensional feature space
(a) Calculate signal variance,
to separate various classes
(b) Estimate each coefficient using the bivariate
For a feature space containing K data points
shrinkage function.
fxk , yk gKk= 1 , where xk 2 Rn and yk 2 R is kth input data
The algorithmic details of the mentioned method can and the k th output class label respectively, addition-
xendur and Selesnick, 2002).
be found at (S ally, yk can take the value of either + 1 or 21, repre-
(c) Feature Extraction: Feature extraction is an senting two different classes. The SVM classifier
important step towards the accurate classifica- function to discriminate two classes is given as (Chen
tion of normal and glaucoma images. In the et al., 2020),
present study, to keep the computational com-
 
plexity at a minimum, we use basic spatial vð xÞ = sign OT uð xÞ + b ð1Þ
domain methods for image feature extraction.
These features for each R-G-B plane include Where O and b are the weight vector and bias term in
contrast, mean, variance, energy, entropy, d-dimensional feature space, respectively. The function
homogeneity, standard deviation and root uð xÞ maps x into d dimensional feature space. The core
mean square (Chowdhary and Acharjya, 2020), principle of SVM is to determine the optimal hyper-
which are shown in Table 1. In Table 1, the plane that maximizes the distance of data points to that
exact terminologies are as follows: of hyperplane for the respective class. This maximiza-
tion problem stated in SVM may be termed as an opti-
P(x, y) denotes instantaneous pixel values at coordinates mization problem with inequality constraints. For
(x, y). classifying biomedical signals, the least square SVM
Khan et al. 5

Table 2. Feature selection parameter (Yan et al., 2008).

S.no Parameter Description

1 T-Test (T) This method analyses the features and checks whether the means
of two independent classes (samples) coming from the population
(normal distributions) are statistically different from each other.
2 Bhattacharyya Distance method (B) This method can be used as a class reparability measure to
evaluate the statistical dependence between two classes
of a feature.
3 Entropy Ranking Method (E) Here the divergence of two classes is tested by calculating a
measure called Relative Entropy.
4 Neighbourhood Component Analysis It is a non-parametric technique of feature selection wherein,
the stochastic nearest neighbour is used to find optimum set of
features for achieving maximum classification accuracy.
5 Receiver Operating Characteristics (ROC) Method (R) It is a normally used method to assess the presentation of a
classification procedure. ROC curve is a feature ranking
technique that events the individual implication of structures in
the feature vector. The benefit of the ROC curve system is that
it provides evidence around the overlapping of classes.
6 Wilcoxon rank-sum test Method (W) It compares two unpaired classes of the feature set. It is a
non-parametric test in which no assumptions about the
distribution of classes are needed.

(LS-SVM) has been frequently used [41] (Kumar et al., In equation (7), s is the hyperparameter.
2019). In LS-SVM, the optimization problem can be
stated as (Suykens and Vandewalle, 1999): 2. The polynomial kernel function (Medical Image
Analysis Group, n.d.):
1 T gX K
Minimize
|fflfflfflfflffl{zfflfflfflfflffl} = ðO, b, z Þ = O O + z 2 ð2Þ
2 2 k=1 k 
O, b, z K xi , xj = (xi  xj + 1)p ð7Þ
subjected to equality constraints as where p is the degree of the polynomial.
  For evaluating the effectiveness of the classifier, we
yk OT uðxk Þ + b = 1  zk k = 1, 2, 3, :::, K: ð3Þ use performance metrics, namely, sensitivity, specificity,
accuracy, precision, recall and F-score. To compute
where z = (z1 , z2 , ::::zk )T :For equation (3), the
these metrics, four parameters, namely true-positive
Lagrangian multiplier ak can be defined as
(TP), true-negative (TN), false-positive (FP) and false-
negative (FN), are required. These parameters are
X
K
ðO, b, z; aÞ = =ðO, b, zÞ  ak obtained from the resultant confusion matrix. TP and
k=1 ð4Þ TN denote correctly classified positive and negative
  T   instances, respectively, whereas incorrectly classified
yk O uðxk Þ + b  1 + zk
positive and negative instances are denoted by FP and
Solving equation (5) results in decision hyperplane FN, respectively. The performance measures are shown
function (Sharma and Pachori, 2015) as, in Table 3.
" #
X
K
nð xÞ = sign ak yk K(x, xk ) + b (5) ð5Þ Results and discussion
k=1
In the proposed work, automatic glaucoma detection
In equation (6) the Kðx, xk Þ is a kernel function. In this from fundus images has been proposed. In total, 455
work, the following kernel functions are used: images have been used from RIM-MIAG image data-
base (Agrawal et al., 2019; Kirar and Agrawal, 2019;
1. The radial basis function (RBF) kernel (Sharma Parashar and Agrawal, 2020). After extensive pre-
and Pachori, 2015): processing of the images, the eight features were
extracted for three planes of the images. We extracted a
jjxxk jj total of 24 features from three (RGB) planes. External
Kðx, xk Þ = e 2s2 ð6Þ synthetic speckle noise has been added to verify the
6 Concurrent Engineering: Research and Applications 00(0)

Table 3. Performance measures parameters (Chen et al., 2020; Zhang and He, 2020).

S.no Parameter Formula Remarks

1 Sensitivity (SEN) TP It quantifies the ability of the classifier model in

TPþFN 3 100%
predicting positive class labels correctly.
2 Specificity (SPE) TN Contrary to sensitivity, it is the ability of the
TNþFP 3 100%
classifier model to predict negative instances
correctly.
3 Accuracy (ACC) TPþTN Out of total samples, the ability of the classifier
TPþTNþFPþFN 3 100%
model to predict correct positive and negative
classes is quantified by accuracy.
4 Positive predictive TP It quantifies the ability of the classifier model in
TPþFP 3 100%
value (PPV) identifying positive instances from total positive
space.
5 Negative predictive TP On the contrary to PPV, it quantifies the ability of
TPþFN 3 100%
value (NPV) the classifier model in identifying negative instances
from total negative space.
6 Matthews correlation pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
TP:TNFN:FP ffi 3 100% Large values of MCC indicate better classification
(TPþFN)(TPþFP)(TNþFN)(TNþFP)
coefficient (MCC) performance.

Table 4. Feature selection on noisy images with respect to RGB planes using the proposed method.

Test Red plane Green plane Blue plane

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

NCA

robustness of the proposed method. The process is selected features using NCA have been tested using
repeated for noisy and denoised images. Figures 3 and Kruskal Wallis statistical test and the p-value is com-
4 show the noisy and denoised fundus images of nor- puted. Figures 5 and 6 show the comparison of selected
mal and glaucoma category respectively. The signifi- feature sets for noisy and denoised images respectively.
cant features were selected using various feature The selected best features for noisy and denoiy
selection and ranking methods. The detailed outcomes images have been described and tabulatd in Tables 6
for the feature selection techniques are shown in Table and 7 respectively. With respect to the classification
4 for noisy images and in Table 5 for denoised images. problem in hand, from Tables 6 and 7, we can conclude
It is observed that from Table 5, that neighbourhood that image entropy, standard deviation, variance and
component analysis (NCA) has resulted in a smaller RMS value are the prominent features for classifying
number of features than its other counterparts. The fundus images into normal and glaucoma category.
Khan et al. 7

Table 5. Feature selection on denoised images with respect to RGB planes using the proposed method.

Test Red plane Green plane Blue plane

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

NCA

Figure 3. Normal fundus images: (a) orignal, (b) noisy and (c) denoised.

Figure 4. Glucoma fundus images: (a) orignal, (b) noisy and (c) denoisy.

Figure 5. Feature selected using NCA for the noisy image using the proposed method.
8 Concurrent Engineering: Research and Applications 00(0)

Figure 6. Feature selected using NCA for the denoised image using the proposed method.

Table 6. Description of the selected feature using NCA method for noisy fundus images.

Feature number Corresponding plane Description

2 Red Mean
5 Entropy
10 Green Mean
13 Entropy
14 Standard deviation
15 RMS
16 Variance
20 Blue Energy
21 Entropy
23 RMS
24 Variance

Table 7. Description of the selected feature using NCA method for denoised fundus images.

Feature number Corresponding plane Description

4 Red Energy
12 Green Energy
13 Entropy
14 Standard deviation
15 RMS
16 Variance
19 Blue Homogeneity
20 Energy
21 Entropy
24 Variance

The p-value for both the cases is less than 0.05, over the computed feature set. Tables 8 and 9 show the
which indicates the good discrimination ability of the performance of the LS-SVM classifier for various ker-
said feature set in classifying normal and glaucoma nels as well as the feature set obtained through various
fundus images. LS-SVM has been widely used in the feature selection techniques for noisy and denoised fun-
classification of glaucoma images (Azar and El-Said, dus images. From Table 9, it is evident that the maxi-
2014). Therefore, taking motivation from previously mum performance of LS-SVM is achieved with radial
reported work, in the present work we have used LS- basis function over the feature set provided by the
SVM with various kernels namely linear, polynomial NCA method. The improved performance by the LS-
(with the degree of 3), and radial basis function (with SVM classifier may be attributed to the non-linear
s = 1) to classify normal and glaucoma fundus images separability of the constructed feature space. This may
Khan et al. 9

Table 8. Classification performance of LS-SVM for various FS techniques (noisy images).

FS test/performance measure ACC SEN SPE PPV NPV MCC

Linear Kernel T Test 84.34 86.96 81.00 86.32 81.77 63.03

Entropy 82.58 85.39 79.00 84.77 79.75 59.45
Bhattacharya 83.46 86.57 79.50 85.31 81.02 61.20
ROC 83.90 85.39 82.00 86.67 80.49 62.28
Wilcoxon 84.12 85.39 82.50 87.00 80.61 62.75
NCA 86.98 89.31 84.00 88.66 84.80 68.38
Polynomial Kernel T Test 81.92 90.49 71.00 81.22 83.11 57.89
Entropy 83.24 92.06 72.00 82.08 85.24 60.67
Bhattacharya 82.80 91.27 72.00 81.92 84.29 59.72
ROC 81.92 90.88 70.50 81.04 83.44 57.91
Wilcoxon 82.36 92.06 70.00 81.03 84.75 58.89
NCA 87.42 92.45 81.00 87.29 87.61 69.17
RBF Kernel T Test 88.02 82.94 94.50 95.71 81.01 67.08
Entropy 87.14 82.55 93.00 94.47 80.34 65.18
Bhattacharya 88.46 84.12 94.00 95.52 81.79 67.72
ROC 88.02 82.55 95.00 96.05 80.83 67.21
Wilcoxon 88.46 82.94 95.50 96.51 81.25 68.10
NCA 90.88 84.90 98.50 99.25 83.44 73.05

Table 9. Classification performance of LS-SVM for various FS techniques (denoised images).

FS test/performance measure ACC SEN SPE PPV NPV MCC

Linear Kernel T Test 84.34 84.22 84.50 88.13 80.00 63.42

Entropy 84.56 85.39 83.50 87.66 80.85 63.70
Bhattacharya 82.80 83.43 82.00 86.30 78.68 60.21
ROC 84.56 83.82 85.50 88.75 79.88 63.98
Wilcoxon 84.56 83.82 85.50 88.75 79.88 63.98
NCA 85.66 87.75 83.00 87.75 83.00 65.75
Polynomial Kernel T Test 81.92 92.45 68.50 80.34 84.87 58.04
Entropy 83.46 91.27 73.50 82.74 84.65 61.07
Bhattacharya 83.24 90.88 73.50 82.66 84.19 60.60
ROC 80.82 90.88 68.00 79.74 82.78 55.67
Wilcoxon 83.68 93.24 71.50 82.05 86.60 61.66
NCA 87.20 94.02 78.50 86.07 89.00 68.78
RBF Kernel T Test 88.04 83.37 94.00 95.49 81.20 65.03
Entropy 87.38 82.20 94.00 95.26 80.33 63.89
Bhattacharya 88.26 84.16 93.50 95.26 81.66 65.29
ROC 87.82 83.76 93.00 94.81 81.23 64.40
Wilcoxon 88.26 84.16 93.50 95.26 81.66 65.29
NCA 91.22 85.51 98.50 99.37 83.92 72.63

be the reason for the underperformance of the linear radial basis function (with s = 1) to classify normal
kernel in LS-SVM. From Table 9, it is evident that the and glaucoma fundus images over the computed fea-
LS-SVM classifier performs well for the RBF kernel ture set. From Table 9, it is evident that the maximum
compared to other kernel. The p-value for both the performance of LS-SVM is achieved with radial basis
cases is less than 0.05, which indicates the good dis- function over the feature set provided by the NCA
crimination ability of the said feature set in classifying method. The improved performance by the LS-SVM
normal and glaucoma fundus images. LS-SVM has classifier may be attributed to the non-linear separabil-
been widely used in the classification of glaucoma ity of the constructed feature space. This may be the
images (Azar and El-Said, 2014). Therefore, taking reason for the underperformance of the linear kernel in
motivation from previously reported work, in the pres- LS-SVM. From Table 9, it is evident that the LS-SVM
ent work we have used LS-SVM with various kernels classifier performs well for the RBF kernel compared
namely linear, polynomial (with the degree of 3), and to other kernel. For the LS-SVM with RBF kernel, the
10 Concurrent Engineering: Research and Applications 00(0)

Table 10. Performance comparison of the proposed method and existing methods over MIAG dataset for two-class classification.

Authors and year Methodology Classification accuracy

Nyul (2009) Empirical wavelet transforms and correlation entropy-based 84%

features and LS-SVM classifier.
Raja and Gangatharan (2015) Wavelet packet transform, entropy and energy-based 85%
features with ANN classifier.
Kirar and Agrawal (2018) Continues wavelet transform and higher-order 88.33%
statistics-based features with ANN classifier.
Agrawal et al. (2019) Variational mode decomposition, Entropy and Fractal 86.13%
dimensions-based features and LS-SVM classifier.
Kumar et al. (2019) Rim to disc ratio (Drishti dataset). 84%
Parashar and Agrawal (2020) Flexible analytic wavelet transforms, entropies and 90.76%
fractal-based features, and LS-SVM classifier.
Martins et al. (2020) Convolutional Neural Networks. 87%
Proposed method Denoising images with RGB features and LS-SVM classifier. 91.22%

performance parameters too performed well with accu- dataset. In a recently reported study by Parashar and
racy, sensitivity and specificity of 91.22 %, 85.51% and Agrawal (2020), flexible analytic wavelet transform has
98.50% respectively. been utilized followed by the entropies and fractal fea-
Moreover, the robustness of the noise removal algo- tures. Parashar and Agrawal (2020) have attended an
rithm is verified from Tables 8 and 9, in which the slight impressive classification accuracy of 90.76% using LS-
increase in the performance for denoised images is evi- SVM classifier. A convolution neural network has been
dent. In Table 9 the bold text signifies the highest per- employed to classify glaucoma from fundus images in
formance. Here it is worthwhile to mention that the Martins et al. (2020). The classification accuracy
LS-SVM classifier is used with 10% outlier removal. reported in Martins et al. (2020) is 87%.
Without outlier removal, we observed a decrease in The common aspect of the state-of-the-art methods
classification performance by 4%–5%. proposed in the existing literature is that they involve
A comparison of the proposed technique with exist- time-frequency analysis and deep learning approaches
ing methods is given in Table 10. Nyul (2009) have used for classification. The advantages associated with the
LS-SVM classifier for classifying normal and glaucoma existing methods can be highlighted as follows:
fundus images over the feature space constructed by
empirical wavelet transform and correlation entropy. 1. The use of time-frequency analysis techniques
The authors in (Nyúl, 2009) have achieved a classifica- enables to capture of the non-linearity in the data-
tion accuracy of 84%. In Raja and Gangatharan set, thereby making classification more accurate.
(2015), wavelet packet transform has been used to 2. The deep learning approach does not require
extract the features from glaucoma images. The features hand-crafted features, which in turn makes further
extracted from wavelet packet transform have been research feasible thereby reducing the feature
used to train the artificial neural network (ANN). Raja exploration burden.
and Gangatharan (2015) have achieved a classification
accuracy of 85%. Higher-order statistics obtained from Being said the advantages of existing methodologies,
continuous wavelet transform has been employed by these methods are generally computationally extensive
Kirar and Agrawal (2018) to classify normal and glau- due to the application of time-frequency analysis meth-
coma fundus images. Kirar and Agrawal (2018) have ods and deep learning approaches.
achieved the classification accuracy of 88.33% with the Thus, the proposed method has an edge over existing
ANN classifier. Recently, Agrawal et al. (2019) have methods in the following aspects:
used a variational mode decomposition technique to
extract the mono frequency modes of glaucoma images. a. The analysis is carried out in the time domain;
Various features based on entropy and fractals have hence the computational complexity is less.
been extracted from the modes and subsequently, LS- b. All three RGB planes of the fundus image have
SVM has been employed for classification to get an been used to preserve the maximum information.
accuracy of 86.13%. Kumar et al. (2019) have used rim c. Classical image feature extraction algorithms
to disc ratio to classify fundus images into normal and have been used, making the proposed approach
glaucoma categories. The classification accuracy simpler.
reported in Kumar et al. (2019) is 84% over the Drishti
Khan et al. 11

The proposed method is limited to the MIAG dataset. ORCID iD

This overcomes this limitation of the proposed work, in Sibghatullah I Khan https://orcid.org/0000-0003-1263-
the future, the out-of-sample dataset can be included to 8100
enable real-time implementation. Moreover, in future
studies, advanced image processing and classification
References
methods (Nyúl, 2009) could be utilized for better classi-
fication accuracy. Agrawal DK, Kirar BS and Pachori RB (2019) Automated
glaucoma detection using quasi-bivariate variational mode
decomposition from fundus images. IET Image Processing
13(13): 2401–2408.
Conclusion Anderson RL, de Los Angeles Ramos Cadena M and Schu-
Early detection of glaucoma is an important task as it man JS (2018) Glaucoma diagnosis: from the artisanal to
can aid in early diagnosis thereby stopping further pre- the defined. Ophthalmology Glaucoma 1(1): 3–14.
valence of the disease. In this paper, we presented a sim- Azar AT and El-Said SA (2014) Performance analysis of sup-
port vector machines classifiers in breast cancer mammo-
ple yet effective strategy to detect glaucoma from
graphy recognition. Neural Computing and Applications
fundus images. The input images are pre-processed 24(5): 1163–1177.
using the discrete wavelet transform-based bivariate Bock R, Meier J, Nyúl LG, et al. (2010) Glaucoma risk index:
shrinkage method. The feature extraction involves spa- Automated glaucoma detection from color fundus images.
tial domain parameters namely contrast, mean homoge- Medical Image Analysis 14(3): 471–481.
neity, energy entropy, RMS, variance. The feature Chen Y, He F, Li H, et al. (2020) A full migration BBO algo-
extraction is carried out for all three RGB planes in rithm with enhanced population quality bounds for multi-
order to retain maximum information. The raw feature modal biomedical image registration. Applied Soft
set which comprises a total of 24 features is then fed to Computing Journal 93: 106335.
various feature selection and ranking techniques to Cheng J, Liu J, Yin F, et al. (2013) Superpixel classification
select appropriate feature space. We included T-Test, based optic disc and optic cup segmentation for glaucoma
screening. IEEE Transactions on Medical Imaging 32(6):
Entropy test, Bhattacharya distance, ROC-based
1019–1032.
method, Wilcoxon rank test and neighbourhood com- Chowdhary CL and Acharjya DP (2020) Segmentation and
ponent analysis (NCA) as feature selection methods. feature extraction in medical imaging: A systematic review.
NCA has resulted in improved performance over other Procedia Computer Science 167: 26–36.
methods of feature selection methods. The feature space Davis BM, Crawley L, Pahlitzsch M, et al. (2016) Glaucoma:
selected by the NCA method is tested with Kruskal The retina and beyond. Acta Neuropathologica 132(6):
Wallis statistical test and significant discrimination abil- 807–826.
ity has been observed (p \ 0.005).LS-SVM classifier is Hira ZM and Gillies DF (2015) A review of feature selection
used to classify normal and glaucoma images over the and feature extraction methods applied on microarray
formed feature spaces by the above-mentioned feature data. Advances in Bioinformatics 2015.
Huang D, Swanson EA, Lin CP, et al. (1991) Optical coher-
selection techniques. To establish the robustness of the
ence tomography. Science 254(5035): 1178–1181.
said methodology, we have added speckle-noise fol-
Kavitha S, Zebardast N, Palaniswamy K, et al. (2014) Family
lowed by denoising using a new simple non-Gaussian history is a strong risk factor for prevalent angle closure in
bivariate probability distribution function to model the a south Indian population. Ophthalmology 121(11):
statistics of wavelet coefficients of images. The pro- 2091–2097.
posed approach has outperformed existing approaches Khaire UM and Dhanalakshmi R (2019) Stability of feature
in terms of classification accuracy. However, the pro- selection algorithm: A review. Journal of King Saud Uni-
posed method needs to be verified on the out-of-sample versity – Computer and Information Sciences. https://
dataset to be used in clinical practice. www.sciencedirect.com/science/article/pii/S1319157819304
379
Kirar BS and Agrawal DK (2018) Glaucoma diagnosis using
Declaration of conflicting interests discrete wavelet transform and histogram features from
The author(s) declared no potential conflicts of interest with fundus images. International Journal of Engineering and
respect to the research, authorship, and/or publication of this Technology 7(4): 2546–2551.
article. Kirar BS and Agrawal DK (2019) Computer aided diagnosis
of glaucoma using discrete and empirical wavelet trans-
form from fundus images. IET Image Processing 13(1):
Funding 73–82.
The author(s) received no financial support for the research, Kirar KG (2018) Image denoising using empirical wavelet
authorship, and/or publication of this article. transform. International Journal of Emerging Technology
and Advanced Engineering. Available at: https://
12 Concurrent Engineering: Research and Applications 00(0)

www.academia.edu/36867547/Image_Denoising_Usin- glaucoma diagnoses. International Journal of Automation

g_Empirical_Wavelet_Transform (accessed 13 January and Computing 12(4): 393–401.
2021). Sahu S., Singh H.V., Kumar B., Singh A.K., Kumar P.
Kolar R and Jan J (2008) Detection of glaucomatous eye via (2019) Image Processing Based Automated Glaucoma
color fundus images using fractal dimensions. Radioengi- Detection Techniques and Role of De-Noising: A Techni-
neering 17(3): 109–114. cal Survey. In: Singh A., Mohan A. (eds) Handbook of
Kumar JRH, Seelamantula CS, Kamath YS, et al. (2019) Multimedia Information Security: Techniques and Applica-
Rim-to-disc ratio outperforms cup-to-disc ratio for glau- tions. Springer, Cham. https://doi.org/10.1007/978-3-030-
coma prescreening. Scientific Reports 9(1): 1–9. 15887-3_16
Li H, He F, Liang Y, et al. (2020) A dividing-based many- xendur L and Selesnick IW (2002) Bivariate shrinkage with
S
objective evolutionary algorithm for large-scale feature local variance estimation. IEEE Signal Processing Letters
selection. Soft Computing 24(9): 6851–6870. 9(12): 438–441.
Maheshwari S, Pachori RB and Acharya UR (2017) Auto- Sharma R and Pachori RB (2015) Classification of epileptic
mated diagnosis of glaucoma using empirical wavelet seizures in EEG signals based on phase space representa-
transform and correntropy features extracted from fundus tion of intrinsic mode functions. Expert Systems with
images. IEEE Journal of Biomedical and Health Infor- Applications 42(3): 1106–1117.
matics 21(3): 803–813. Sommer A, Pollack I and Maumenee AE (1979) Optic disc
Maheshwari S, Pachori RB, Kanhangad V, et al. (2017) Itera- parameters and onset of glaucomatous field loss: I. Meth-
tive variational mode decomposition based automated ods and progressive changes in disc morphology. Archives
detection of glaucoma using fundus images. Computers in of Ophthalmology 97(8): 1444–1448.
Biology and Medicine 88: 142–149. Suykens JAK and Vandewalle J (1999) Least squares support
Martins J, Cardoso JS and Soares F (2020) Offline computer- vector machine classifiers. Neural Processing Letters 9(3):
aided diagnosis for glaucoma detection using fundus 293–300.
images targeted at mobile devices. Computer Methods and Townsend KA, Wollstein G, Danks D, et al. (2008) Heidel-
Programs in Biomedicine 192: 105341. berg retina tomograph 3 machine learning classifiers for
Medical Image Analysis Group (n.d.) Available at: https:// glaucoma detection. British Journal of Ophthalmology
medimrg.webs.ull.es/ (accessed 13 January 2021). 92(6): 814–818.
Mistlberger A, Liebmann JM, Greenfield DS, et al. (1999) Webb RH and Hughes GW (1981) Scanning laser ophthalmo-
Heidelberg retina tomography and optical coherence scope. IEEE Transactions on Biomedical Engineering 28(7):
tomography in normal, ocular-hypertensive, and glauco- 488–492.
matous eyes. Ophthalmology 106(10): 2027–2032. Weinreb RN, Aung T and Medeiros FA (2014) The patho-
Noronha KP, Acharya UR, Nayak KP, et al. (2014) Auto- physiology and treatment of glaucoma: A review. Journal
mated classification of glaucoma stages using higher order of the American Medical Association 311(18): 1901–1911.
cumulant features. Biomedical Signal Processing and Con- Yadav D, Sarathi MP and Dutta MK (2014) Classification of
trol 10(1): 174–183. glaucoma based on texture features using neural networks.
Nyúl LG (2009) Retinal image analysis for automated glau- In: 2014 7th international conference on contemporary com-
coma risk evaluation. Medical Imaging, Parallel Processing puting, IC3 2014, Noida, India, 7–9 August, pp.109–112.
of Images, and Optimization Techniques 7497: 74971C. New York: IEEE.
Parashar D and Agrawal DK (2020) Automated classification Yan Z, Wang Z and Xie H (2008) The application of mutual
of glaucoma stages using flexible analytic wavelet trans- information-based feature selection and fuzzy LS-SVM-
form from retinal fundus images. IEEE Sensors Journal based classifier in motion classification. Computer Meth-
20(21): 12885–12894. ods and Programs in Biomedicine 90(3): 275–284.
Raja C and Gangatharan N (2013) Glaucoma detection in Yu H, He F and Pan Y (2020) A scalable region-based level
set method using adaptive bilateral filter for noisy image
fundal retinal images using trispectrum and complex
segmentation. Multimedia Tools and Applications 79(9–10):
wavelet-based features. European Journal of Scientific
5743–5765.
Research 97(1): 159–171.
Zhang S and He F (2020) DRCDN: learning deep residual
Raja C and Gangatharan N (2015) Appropriate sub-band
convolutional dehazing networks. Visual Computer 36(9):
selection in wavelet packet decomposition for automated
1797–1808.

Author biographies
Sibghatullah I. Khan. completed his B.E, M.E and Ph.D in Electronics and
Telecommunication Engineering from SGB Amravati University, India in the years 2007,
2012, and 2018 respectively. Presently he is working as an Associate Professor in the
Department of Electronics and Communication Engineering at Sreenidhi Institute of
Science and Technology Hyderabad, INDIA. His research interests focus on Bio-Medical
Signal Processing.
Khan et al. 13

Dr.Shruti Bhargava Choubey received BE with honors from RGPV Bhopal and M.
Tech.degree in Digital Communication Engineering from RGPV Bhopal subsequently she
carried out her research from Dr.K.N. Modi University Banasthali Rajasthan and awarded
Ph.D. in 2015. Presently she is working as an Associate Professor in the Department of
Electronics and Communication at Sreenidhi Institute of Science and Technology,
Hyderabad she has published more than 100 papers of national and international repute.
She has been a Member of many selection committees for the recruitment of staff and
faculty. her research areas include signal processing Image processing and Biomedical
Engineering. He has produced 18 M.Tech degrees & guided more than 70 B.Tech Project.
She is a member of IETE, New Delhi and International Association of Engineers (IAENG)
and IEEE. She worked in different positions like Dean academic & HOD with numerous
capacities. she is awarded by MP Young scientist Award 2015 & Received mp council fellow-
ship in year 2014 & 2015’’ for her contribution to Research.

Abhishek Choubey received a Ph.D. degree in the field of VLSI for digital signal processing
from Jaypee University and technology Guna MP, in 2017. He is currently associated with
Sreenidhi institute of science and technology, Hyderabad, as an Associate Professor. He has
published nearly 70 technical articles. His research interest includes reconfigurable architec-
tures, approximate-computation, algorithm design, and implementation of high-performance
VLSI systems for signal processing applications. He was a recipient of the Sydney R. Parker
and M. N. S. Swamy Best Paper Award for Circuits, Systems, and Signal Processing in 2018.

Dr. Abhishek Bhatt, is working as an associate professor in the department of Electronics

and Telecommunication Engineering at the College of Engineering Pune (COEP). He
received his B.Tech in Electronics and Communication and M.Tech in Digital
Communication during 2003 and 2008, respectively, and his PhD degree in Image Processing
and Remote Sensing from the Indian Institute of Technology Roorkee in 2016. He is the
author of more than 50 journal/conference papers and has written one book in signal pro-
cessing using MATLAB. His current research interests include Image Processing, Computer
Vision, Machine Learning, Remote Sensing, and Pattern Recognition.

Pandya Vyomal completed his B.E. in Electronics Engineering from Veer Narmad South
Gujarat University in 2007, M.E. in Electronics and Communication Systems from
Dharmsinh desai University, Nadiad, Gujarat in 2010 and Ph.D. from Gujarat
Technological University, Chandkheda, Gujarat in 2016. He is working as an Associate
Professor in the Department of Electronics & Communication Engineering. He has a total 8
years of experience in academia. He has published 15 technical papers in international jour-
nal. He guided 12 UG projects and 8 M.Tech. Thesis. He worked in various engineering col-
leges of Gujarat. He is a life time member of IEEE. He is also a member of ISTE. He has
very strong research foundation. His areas of research are Wireless Sensor Network, Delay
Tolerant Networks and Computer Networking.

M. Mahaboob Basha obtained his Master of Technology from the SK University,

Ananthapuramu and PhD in the area of low power VLSI design from the JNTUA
University, Ananthapuramu, India. His research interests are signal processing, low power
VLSI circuits and energy efficient computing.