Automated Glaucoma Detection based on LBP

Representation and GLRLM Feature Extraction

Method

Zakia Ahmed Ahmed Lawgali Mohamed Abdalla

2022 International Conference on Engineering & MIS (ICEMIS) | 978-1-6654-5436-0/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICEMIS56295.2022.9914267

Department of Computer Science Department of Computer Science Department of Software Engineering

Faculty of Sciences, Faculty of Information Technology, Faculty of Information Technology,
University of Omar Al-Mukhtar University of Benghazi University of Benghazi
Albayda, Libya Benghazi, Libya Benghazi, Libya
[email protected] [email protected] [email protected]

Abstract— Glaucoma is a chronic and degenerative disease according to World Health Organization (WHO). By 2040,
that causes irreversible damage to the nerve system of an eye, the number will reach 111.8 million, and is considered the
due to an increase in the Intra-ocular pressure in the retina. It
second cause of blindness after cataracts [3] [4]. Increase the
reduces the vision area and leads to blindness. Early detection
is critical to prevent permanent vision loss. The automated rate of glaucoma patients in Green Mountain Area in Libya
analysis of medical images contributes to the increase in the motivate us to study an automated method to detect this
performance of classifying retinal images to pathological and disease. The early detection of glaucoma is crucial to
nonpathological. This paper presents an automated glaucoma decrease or stop disease progression. Due to the high cost of
diagnosis approach by analyzing the texture of fundus images.
imaging techniques and difficult availability in health
The analysis was performed using Local Binary Pattern
descriptor to represent grayscale and channels of retinal facilities, fundus image is a suitable technique in the context
images and applied Gray Level Run-Length Matrix to describe of affordability and reflects an impairment of the retina [1].
patterns of texture. The classification was performed using two
classifiers. The proposed approach provided the best The biomedical image has been progressed and advanced
classification results with 98%, 91%, and 97% of the three in disease diagnosis, where detecting glaucoma by using
versions of RIM-ONE databases using Support Vector retinal images is one of the active medical areas in research
Machine classifier. Findings showed that local binary pattern [5]. Using the techniques of digital processing of biomedical
representations and concatenated gray level run length matrix image give support to diagnosis, which improves the
features achieved high results. Consequently, proving this
approach is reliable and promising. appearance of texture features and provides more
information for decision-making support.
Keywords— Glaucoma detection, Texture features, LBP
descriptor, GLRLM, Support Vector Machine.

I. INTRODUCTION
Glaucoma is an asymptomatic affected disease causing
of the increase of Intra-Ocular Pressure (IOP) in the retina.
Hence, makes irreversible damage to the optic nerve and
Retinal Nerve Fiber Layer (RNFL). The Optic Disc (OD) is
the region where blood vessels and nerve fibers enter the Fig. 1. The Retinal Image Terminologies.
retina, Optic Cup (OC), on the other hand, is the brighter Structural and Nonstructural features emphasize on
center of the retina, which lies within the OD as shown in external Region Of Interest (ROI) boundary detected and the
Fig. 1. The IOP occurs due to the obstruction of the drainage internal features of a retinal image respectively. Structural
channel of the aqueous humor liquid [1]. The increase of features e.g. Cup to Disc Ratio (ODR), Inferior Superior
IOP causes changes in the structure of eye anatomy to be Nasal Temporal (ISNT) rule, and blood vessels ratio.
enlarged in OC, which is termed as Cupping phenomenon. Nonstructural features as the changes of intensity, texture,
Without treatment applied at early stages, the vision can and shape of an object.
lose [2].
We aim to achieve higher results by defining features
Globally, the statistics indicate an increase in infection vector that can distinguish between pathological and healthy
rates, especially among those people, whose ages are more images. To improve the performance of an automated
than sixty. In 2020, the number of people around the world, glaucoma diagnosis using nonmorphological features
who were suffering from glaucoma reached 80 million, without the need for OD segmentation, we evaluated the
978-1-6654-5436-0/22/$31.00 ©2022 IEEE

Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on July 10,2023 at 10:27:17 UTC from IEEE Xplore. Restrictions apply.
effectiveness of using Local Binary Pattern (LBP) method automated glaucoma identification is required to develop the
and Gray Level Run-Length Matrix (GLRLM) to extract performance of disease detection as stated in [11].
texture features, and emphasize classification using machine
learning approaches. III. METHODOLOGY
In this section, we introduce the proposed methodology
The present work is segmented into Section II, which for classifying pathological and nonpathological images
reviews related studies. Section III describes our using nonstructural features by representing images and
methodology, which involves image acquisition, image recognizing a certain pattern. The proposed methodology is
preprocessing, methods of feature extraction, feature organized into five steps as represented in Fig. 2.
reduction, and classification. The achieved results, their
discussion, and performance metrics are listed in Section
IV. Finally, the conclusion of the study is in Section V.

II. LITERATURE REVIEW

Many approaches were stated for automated glaucoma
diagnosis using fundus images. Researchers paid attention
to morphological features such as CDR by performing ROI
segmentation on the retinal image. Author [3] proposed an
automated glaucoma detection by segment ROI to calculate Fig. 2. The Methodology of the Proposed Approach.
CDR, ISNT, blood vessels ratio as structural features in
A. Image Acquisition
retinal images. He used Neural Network (NN), Support
Vector Machine (SVM), and Adaboost. Authors [6] Images are read from public RIM-ONE databases [12].
presented an approach of combining structural features and Open Retinal Image database for Optic Nerve Evaluation
nonstructural features by performing OD segmentation in (RIM-ONE) database is a group of databases (versions) that
order to obtain CDR, Disc Damage Likelihood Scale contain ophthalmic images as presented in Table I. The
(DDLS), and applies various methods to obtain texture database images were designed for glaucoma diagnosis with
features. Needless to say, the use of OD segmentation is normal and glaucoma labels. In addition, it was designed as
difficult, according to Hallem et al. [7] stated that the line an evaluation tool for segmentation algorithms, where it has
between OD and OC visually is not visible. Therefore, other a manual segmentation reference for each image performed
studies used Nonstructural features, and avoided the by ophthalmic experts.
segmentation process. Chaudhary et al. [5] introduced a The images of RIM-ONE version1 (V1) and version2 (V2)
technique for fundus images decomposition and boundary were acquired using Nidek AFC-210 retinal fundus camera
detection, where Machine Learning and Deep Learning under different contrast and brightness conditions. However,
approaches were used for classification based on extracted RIM-ONE version3 (V3) images were taken by Kowa WX
features from sub-images. However, [8] presented an 3D stereo camera that takes two images of retinal from
automated approach to classify glaucoma by determining different angles with high resolution that making one of the
the best collection of features, which classify retinal images automated detection challenge of capturing images with
in a term of homogeneity and heterogeneity of OD and OC different cameras and professionals. It is worth mentioning
color, cup region, and thickness of vessels. Khan et al. [2] that glaucoma images in RIM-ONE V1 contain pre-
used spatial features of denoising R-G-B channels of fundus glaucoma cases in the first stage, which are difficult to
images to detect glaucoma. De Sausa et al. [4] performed distinguish.
geostatistical functions (semivariogram, semimadogram,
covaripgram, and correlogram) to characterize texture TABLE I. PRESENTS THE NUMBER OF IMAGES IN RIM-ONE DATABASE
VERSIONS.
features from represented OD region, genetic algorithm, and
SVM were applied as classifiers. Authors [9] proposed a RIM-ONE No. No. Type Size
computational approach to detect glaucoma by obtaining Version Glaucoma Normal
V1 51 118 bmp Different
nonmorphological features to describe structural changes V2 200 250 jpg Different
that appear from glaucoma. De Sales Carvalho et al. [10] V3 74 85 jpg 2144 x 1424
converted 2D images to volumetric images, followed by
providing them into a three-Dimensional Convolutional
B. Preprocessing
Neural Network (3DCNN) to classify glaucoma images in
spite of the high cost of an entire technical and Eliminating unnecessary inconsistencies is very
computational process. important in medical image preprocessing. Image de-noised
attempts to remove noise or artifacts from an image, which
Review related automated glaucoma detection studies is resulted de-noised image. Firstly, the input images were
important to identify used techniques, features and to resized to (512 x 512) for normalization purposes. We
identify the gab in the research field, where more effort in obtained four images for each image by separating them into
(R, G, B) channels, and grayscale image. We used 2D-

Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on July 10,2023 at 10:27:17 UTC from IEEE Xplore. Restrictions apply.
Discrete Wavelet Transform (2D-DWT) technique to using auxiliary matrices by counting the repetition of gray
enhance the images by performing the de-noising. 2D-DWT level pixel and in which direction. The occurrence of the
was applied on R-G-B channels, and grayscale images as same gray level in the same direction is called run length.
shown in Fig. 3, begin with additive speckle noise to all GLRLM evaluates the spatial distribution of gray level
pixels value in each image. Then applied DWT to transform values, by calculating a set of statistics from local
noisy image into an orthogonal domain. Soft thresholding characteristics distributions [19] [19].
using bivariate shrinkage function was performed on LL Gray level run describes a set of co-linear and successive
sub-band coefficients. Reconstruct de-noised image by pixels with the same value of gray level. Each element in a
computing Inverse DWT (IDWT) and the use of modified matrix represents by calculating the line of pixels that has
coefficients [13]. The denoising method was proposed in [2] equal intensity values and in a certain direction (0ᵒ
as a preprocessing phase to detect glaucoma, where the corresponding to horizontal, 45ᵒ diagonal direction, 90ᵒ
image was separated into R-G-B channels for preprocessing. represent the vertical direction and 135ᵒ). Produced matrix is
Many researchers such as [3] [6] [8] [9] used grayscale the mean of four direction matrices.
images in glaucoma detection. We proposed to use In Computer Aided Diagnosis (CAD) systems of
grayscale image with separated R-G-B planes to obtain glaucoma detection, GLRLM method considers an
more information that could help in glaucoma diagnosis. uncommon feature extracted method. In the literature
review, two studies were extracted statistical features using
GLRLM [6] [8]. In [6], only five features were extracted.
Due to feature ranking, GLRLM features were not involved
in the feature vector that adapted for classification. Where in
Input image Red image Green image Blue image Grayscale image [8] eleven features were extracted from GLRLM method
Fig. 3. The Original input image and resulted de-noising channels/grayscale
without preprocessing the images, which cause decreasing
images of 2D-DWT.
in the method performance.
C. Feature Extraction
The extracted features of GLRLM consist of five
Feature extraction methods aim to describe an image by
measures of [20], two further features were denoted in [19],
using extracted attributes that are used to recognize a pattern
and Dosarathy and Holder determined the other four
and employed to make an accurate decision during
statistical measures in [19]. We added the normalized of two
classification. We obtain nonstructural features from
feature values: (gray level nonuniformity normalized and
preprocessed (R, G, B, and grayscale) images after
run length nonuniformity normalized) with calculated gray
representing images by performing LBP method, followed
level variance, run variance, and run entropy.
by extract 16 features from each image after applying
GLRLM method on them.
a) Local Binary Pattern (LBP) is a simple method to
describe a local image pattern and is adapted in many
applications such as face recognition [14] and shape
localization [15]. LBP represents local features of an image
by providing information of a local difference and provides Red channel Green channel Blue channel Grayscale image
a binary presentation of an image texture, which also
Fig. 4. LBP-representation of R-G-B channels and Grayscale image.
represents the image shape [16].
LBP applied on each pixel in an image. The image splits The less effected by grayscale changes, preserves the
to 3x3 block, pixels in each block are thresholded using the intensity order of pixel, and illuminate the variant changes
center pixel. If the neighborhood pixel value is lower than are strengths of LBP. However, GLRLM method evaluates
center pixel value, then pixel value gets zero value. the spatial distribution of gray level values by calculating a
Otherwise, it is assigned one. The vector of eight set of statistics from local characteristics distributions.
neighborhood pixels is multiplied by two power pixel value
D. Feature Reduction
in a clockwise direction. Results are binary numbers as
shown in Fig. 4. The common way to construct a fast and robust model in
a large space of dimensionality would be the use of one of
LBP descriptor is an efficient invariant texture, which the dimensionality reduction techniques by removing less
integrates structural and statistical texture features. significant and redundant attributes from a large number of
Typically, histogram features (mean, standard deviation, feature set to obtain important features and simplify the
variance, skewness, and kurtosis) as in [9] [17] [18] are prediction model mentioned. We have performed LDA
extracted from LBP representation of an image. We technique for dimensionality reduction and discriminant
employed GLRLM method to extract from LBP among normal and abnormal features to improve classifier
representation. performance.
b) Gray Level Run-length Matrix method (GLRLM) is Linear Discriminant Analysis (LDA) is a supervised
a matrix that aims to describe texture features of an image technique for reducing dimensionality, classification, and

Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on July 10,2023 at 10:27:17 UTC from IEEE Xplore. Restrictions apply.
interpretation of the significant features. We provide Results of RIM-ONE version1 (V1) using sensitivity,
extracted feature vectors of images and their labels to LDA specificity, accuracy, and 10-fold cross-validation accuracy
method. Firstly, LDA performs min-max normalization are presented in Table II.
followed by computing features that have the best
TABLE II. THE RESULTS OF RIM-ONE (VERSION1) DATABASE USING SVM
discriminant among classes, where the method works on AND RF CLASSIFIERS.
minimizing the interclass variance and at the same time Image Clas Param Speci Sensit Accu 10-fold
maximizing the distance between classes [21]. sifier eters ficity ivity racy Accuracy
Grays SVM Linear 43% 97% 80% 79%
E. Classification cale poly 31% 100% 78% 77%
It decision making of assign image to its label using RF 68% 88% 82% 74%
R SVM Linear 68% 100% 90% 82%
specific image features pattern. Classification is categorized poly 37% 100% 80% 79%
into supervised and unsupervised classification. We RF 62% 80% 74% 70%
employed supervised methods to perform glaucoma G SVM Linear 25% 100% 76% 77%
poly 25% 100% 76% 74%
identification, where pre-defined inputs were adapted to
RF 37% 80% 66% 67%
learn classifiers. The discriminant features of LDA are B SVM Linear 50% 88% 76% 80%
provided into a classifier to detect normal and abnormal poly 25% 100% 76% 76%
cases. The effectiveness of SVM and RF classifying have RF 56% 74% 68% 67%
RGB SVM Linear 87% 100% 96% 92%
been proved in related literature of glaucoma detection [5] poly 87% 100% 96% 94%
[4], which motivates us to use them and compare the results RF 81% 100% 94% 87%
in order to notice which classifier will have a high RGB, SVM Linear 93% 100% 98% 97%
Grays poly 93% 100% 98% 95%
performance. 10-fold cross-validation method was applied
cale RF 87% 100% 96% 94%
for classifiers validation and overcome the overfitting.
a) Support Vector Machine (SVM) is supervised The results of 10-fold cross validation observed that
technique that uses a hyperplane to separate a dataset in green channel images obtained the lowest results using the
feature space based on its class, where the maximum margin two classifiers compared to red and blue channels that had
among two classes is the best hyperplane [6]. SVM has the better results than G component and grayscale images.
ability to learn with small and large size databases [5] [22], Where red component images obtained the highest values of
with performing linear and nonlinear classification with accuracy compared to other separated channels and
kernel functions. For classifier parameters selection, various grayscale images.
parameters were adapted to achieve optimal performance,
When we combined the features of RGB component in
we adapted linear and polynomial (with gamma=1.25) one vector, the performance of classifiers improved
kernel functions were adapted in SVM classifier. compared to use extracted features of each component
b) Random Forest (RF) is an ensemble learning alone. Where the accuracy developed to 94% using
technique for classification, with bagging and bootstrap polynomial SVM classifier.
techniques to train RF and make tree learns. Each tree
The sufficient improvement was occurred after we
performs classification with the help of features that are
combined GLRLM features of RGB channels with grayscale
arranged randomly to get the result. RF works effectively
features, where the performance of the classifiers was
with numeric, categorical, and missing data [23]. The RF developed and the accuracy value reached 97% using linear
parameter is the number of trees that needed to learn, we SVM classifier.
proposed 100 trees.
The sensitivity values reached 100% with RGB features
IV. RESULTS AND DISCUSSION and the features combination of RGB and grayscale images,
The performance of extracted features and the used where the high value of sensitivity is important in the
classifiers are discussed in this section. The three versions domain of health care that implies the disease cases are
of RIM-ONE database were evaluated separately. Each rarely misclassifies.
version was divided into 70% training and 30% for testing The results of classifiers performance that applied on
using the split method without overlapping among training version 2 (V2) of RIM-ONE database are listed in Table III.
and testing sets.
In opposite to V1 of the database results, red component
After perform preprocessing method on the database obtained the lowest results compared to green, blue and
images, LBP descriptor was applied on images to obtain grayscale images, where the features of grayscale images
binary images (LBP representation images). GLRLM had the highest results with 78% with linear SVM compared
method conjunction with LBP descriptor by extracting to the results of each component alone.
GLRLM features from LBP represented images. SVM and
RF classifiers were applied to analyze the GLRLM features Such as V1 of the database, the accuracy results were
performance. To evaluate the performance of classifiers, developed when we combine the extracted features of RGB
accuracy, sensitivity, and specificity are metrics that have components in one vector to train the classifiers, where the
been applied to attain classification accuracy, these metrics accuracy result reached to 86% with linear SVM. The
are calculated by using a confusion matrix. highest accuracy value of V2 was achieved resulting of
combine the features of RGB components with grayscale

Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on July 10,2023 at 10:27:17 UTC from IEEE Xplore. Restrictions apply.
 features, where the performance of linear SVM increased to each plane was processed alone than when we concatenated
89%. the features of RGB components.

TABLE III. THE RESULTS OF RIM-ONE (VERSION2) DATABASE USING SVM As we proposed when we used grayscale image with
AND RF CLASSIFIERS. RGB components of images, we achieved the highest
accuracies compared to previous results. In V1, V2, and V3
Image Class Param Speci Sensit Accu 10-fold
ifier eters ficity ivity racy Accuracy
of RIM-ONE database, the highest accuracies achieved
Grays SVM Linear 69% 81% 75% 78% when RGB features vector was concatenated with grayscale
cale poly 41% 94% 70% 70% features that extracted from LBP representations and
RF 63% 71% 67% 65% GLRLM method.
R SVM Linear 63% 78% 71% 69%
poly 30% 91% 63% 65% The results of linear SVM had higher performance than
RF 58% 59% 56% 62% polynomial SVM classifier in V1 and V2 database versions.
G SVM Linear 58% 81% 70% 74% Wherein V3 images, the polynomial classifier was the
poly 36% 64% 67% 70% highest performance, where RF classifier had the lowest
RF 63% 71% 67% 64% accuracy performance compared to linear and polynomial
B SVM Linear 65% 79% 72% 77% SVM classifiers when they learned using extracted features.
poly 33% 94% 66% 70%
RF 57% 68% 63% 67% Although many researchers used grayscale images in
RGB SVM Linear 88% 85% 86% 86% their methodologies [3] [6] [8] [9] or used R-G-B channels
poly 69% 74% 83% 81% to detect glaucoma [2], we obtained the best results when
RF 74% 81% 78% 80%
SVM Linear
we combined the grayscale image with image components.
RGB, 90% 93% 91% 89%
Grays poly 73% 95% 85% 83%
It can be conclude that a combination set of texture features
cale RF 90% 82% 86% 85% has improved the accuracy compared to dealing with each
component as an independent unit. Although many
researchers commonly used grayscale images in their
Table IV observed the results of accuracy metrics of glaucoma detection studies, the image channels can provide
classifiers performance that trained using version 3 (V3) of more information for glaucoma identification.
RIM-ONE database.
Comparing the results with other related literature:
TABLE IV. THE RESULTS OF RIM-ONE (VERSION3) DATABASE USING SVM
AND RF CLASSIFIERS.
Many approaches were proposed for glaucoma
diagnosis, often they used common methods to analyze
Image Class Param Speci Sensit Accu 10-fold image texture, and some of them have used Deep Learning
ifier eters ficity ivity racy Accuracy techniques. The same classifier and database with different
Grays SVM Linear 52% 79% 68% 69%
cale poly 52% 100% 62% 59%
versions are used in our proposed and in related works. The
RF 73% 62% 66% 63% results of previous studies, feature extraction methods, and
R SVM Linear 63% 72% 68% 69% classifiers showed in Table V.
poly 56% 56% 62% 69%
RF 63% 55% 58% 60%
The results of the proposed method are promising
G SVM Linear 78% 79% 79% 67% compared to current literature. With same classifier and
poly 58% 93% 77% 63% database version, khan et al. [2], shinde [3] and Thakur and
RF 78% 55% 64% 67% Juneja [6] were inline with us also in the same homogenous
B SVM Linear 73% 86% 81% 71% size of database images. In [3], morphological features
poly 36% 100% 75% 64% were used to train the model, but the proposed study proved
RF 73% 55% 62% 59% that nonmorphological features can obtain high results.
RGB SVM Linear 89% 89% 89% 86%
poly 84% 96% 91% 83% Although, Carlo et al. [8] used a common method to
RF 84% 93% 89% 80% extract statistical features (Grey level co-occurrence matrix)
RGB, SVM Linear 100% 96% 97% 89% (GLCM) and used seven Convolutional Neural Network
Grays poly 89% 96% 93% 95%
cale
(CNN) architectures to obtain features that are not visible
RF 94% 93% 93% 86%
for the human eye, without preprocessing the grayscale
image, which could limit the performance of the approach.
Such as V1, green channel had the lowest accuracy Krishnamoorthi and Chinnababu [9] and Chaudhary et al.
results, where blue channel obtained the highest results [5] used various feature extraction methods on grayscale
compared to other separated channels and grayscale images. images with the same classifier. Although [6] used structural
As in V1 and V2 versions of database, the classifiers and nonstructural features to train SVM, however, using
performance improved after combine RGB components channel separation and spatial features contributed to
features in one vector, where the accuracy values improved obtaining better results in our work and without the need for
to 86% with linear SVM. In addition, the results become OD segmentation as in De Sausa et al. [4]. We performed
more accurately after combine grayscale features to the the de-noising method as in [2]. R-G-B channels were
feature vector of RGB components, where the accuracy preprocessed and used to extract statistical features in [2],
developed to 95% with polynomial SVM classifier. we suggested using a grayscale image with R-G-B channels
of an image and that contributes to capturing more detailed
We noted that extracted features of R-G-B channels or information about glaucoma images and contribute to
grayscale images separately could not describe glaucoma increasing the performance of the classifiers.
pattern, where all previous accuracies have low values, due
to the classifiers do not obtain meaningful features when

Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on July 10,2023 at 10:27:17 UTC from IEEE Xplore. Restrictions apply.
 TABLE V. COMPARISON WITH RELATED WORKS. ● Testing our method with other public glaucoma
Paper Feature Classifie RIM- Accu databases.
Extraction r ONE racy ● Adapted our methodology to other eye diseases,
method Version such as cataracts.
Shinde [3] CDR + ISNT + Voting V1 96%
blood vessel (SVM,
ratio NN and
REFERENCES
Adaboost) [1] T. R. V. Bisneto, A. O. de Carvalho Filho and D. M. V.
Magalhães, "Generative adversarial network and texture features applied
Carlo et al. GLCM + CNNs SVM, V1 93% to automatic glaucoma detection," Applied Soft Computing, vol. 90, p.
[8] MLP, RF 106165, 2020.
Krishnamoo LBP + HOG + SVM V1 91% [2] S. I. Khan, et al., "Automated glaucoma detection from fundus
rthi and Fractal images using wavelet-based denoising and machine learning," Concurrent
Engineering, vol. 30, pp. 103-115, 2022.
Chinnababu features
[3] R. Shinde, "Glaucoma detection in retinal fundus images using
[9] U-Net and supervised machine learning algorithms," Intelligence-Based
Chaudhary GLCM + Chip SVM V1 90% Medicine, vol. 5, p. 100038, 2021.
et al. [5] features + V2 90% [4] J. A. de Sousa, et al., "Texture based on geostatistic for
invariant glaucoma diagnosis from fundus eye image," Multimedia Tools and
Applications, vol. 76, pp. 19173-19190, 2017.
moment
[5] P. K. Chaudhary and R. B. Pachori, "Automatic diagnosis of
features glaucoma using two-dimensional Fourier-Bessel series expansion based
De Sausa LBP + SVM V2 90% empirical wavelet transform," Biomedical Signal Processing and Control,
et al. [4] Geostatistical vol. 64, p. 102237, 2021.
features [6] N. Thakur and M. Juneja, "Classification of glaucoma using
hybrid features with machine learning approaches," Biomedical Signal
Thakur and Structural and SVM V3 93% Processing and Control, vol. 62, p. 102137, 2020.
Juneja [6] Nonstructural [7] M. S. Haleem, L. Han, J. Van Hemert and B. Lie, "Automatic
features extraction of retinal features from colour retinal images for glaucoma
khan et al. Statistical LS-SVM V3 91% diagnosis: a review," Computerized medical imaging and graphics, vol. 37,
pp. 581-596, 2013.
[2] features [8] M. Claro, et al., "An hybrid feature space from texture
Proposed LBP SVM V1 98% information and transfer learning for glaucoma classification," Journal of
Methodology representations V2 91% Visual Communication and Image Representation, vol. 64, p. 102597,
+ GLRLM V3 97% 2019.
[9] N. Krishnamoorthi and V. K. Chinnababu, "Hybrid feature
vector based detection of Glaucoma," Multimedia Tools and Applications,
vol. 78, pp. 34247-34276, 2019.
The proposed method improves the state-of-the-art [10] N. R. de Sales Carvalho, M. C. L. C. Rodrigues, A. O. de
glaucoma diagnosis results. The results demonstrate the Carvalho Filho and M. J. Mathew "Automatic method for glaucoma
diagnosis using a three-dimensional convoluted neural network,"
extracted GLRLM features from LBP-represented images Neurocomputing, vol. 438, pp. 72-83, 2021.
had better performance compared to extracted statistical [11] G. An, et al., "Comparison of machine-learning classification
models for glaucoma management," Journal of healthcare engineering, vol.
features from it [9] [8] [18]. 2018.
[12] F. Fumero, S. Alayón, J. Sanchez , J. Sigut and M. Gonzalez
V. CONCLUSION Hernandez, "RIM-ONE: An open retinal image database for optic nerve
evaluation," in 2011 24th international symposium on computer-based
We presented an automated glaucoma diagnosis medical systems (CBMS), 2011, pp. 1-6.
[13] L. Şendur and I. W. Selesnick, "A bivariate shrinkage function
system using texture features of LBP and GLRLM for wavelet-based denoising," in 2002 IEEE International Conference on
Acoustics, Speech, and Signal Processing, 2002, pp. II-1261-II-1264.
descriptors from grayscale and R-G-B fundus images. We [14] T. Ahonen, A. Hadid and M. Pietikäinen, "Face recognition with
adapted LBP method to generate representations of image local binary patterns," in European conference on computer vision, 2004,
pp. 469-481.
channels and grayscale of an image, GLRLM method [15] X. Huang, S. Li and Y. Wand, "Shape localization based on
extracted statistical texture features from LBP represented statistical method using extended local binary pattern," in Third
International Conference on Image and Graphics (ICIG'04), 2004, pp. 184-
images, SVM and RF were applied to get classification 187.
[16] Z. Guo, L. Zhang and D. Zhang, "A completed modeling of
models of glaucoma detection. The results showed that the local binary pattern operator for texture classification," IEEE transactions
SVM classifier achieves a promising success rate of 98%, on image processing, vol. 19, pp. 1657-1663, 2010.
[17] R. F. Mansour and A. Al-Marghilnai, "Glaucoma detection
91%, and 97% for RIM-ONE V1, V2, and V3 database using novel perceptron based convolutional multi-layer neural network
classification," Multidimensional Systems and Signal Processing, vol. 32,
respectively, compared to the results of RF classifier and the pp. 1217-1235, 2021.
studies in related literature that used the same database and [18] C. Srinivasan, S. Dubey and T. Ganeshbabu, "Complex texture
features for glaucomatous image classification system using fundus
classifier. Our proposed method achieved the highest results images," International Journal of Engineering Research & Science, vol. 2,
pp. 106-113, 2016.
by using grayscale image with R-G-B channel images [19] A. Chu, C. Sehgal and J. Sehgal, "Use of gray value distribution
followed by concatenating LBP-represented images with of run lengths for texture analysis," Pattern Recognition Letters, vol. 11, pp.
415-419, 1990.
GLRLM method to obtain the features. We conclude that [20] M. M. Galloway, "Texture analysis using gray level run
using image channels with grayscale provides more lengths," Computer graphics and image processing, vol. 4, pp. 172-179,
1975.
information, while combining different feature extraction [21] P. Xanthopoulos, P. M. Pardalos, T. P. Trafalis, Robust data
mining: Springer Science & Business Media, 2012.
methods leads to better results, extracted GLRLM features [22] R. Gholami and N. Fakhari, "Support vector machine:
from LBP representations are better to obtain characteristics principles, parameters, and applications," in Handbook of neural
computation, ed: Elsevier, 2017, pp. 515-535.
of image patterns than traditional histogram features that [23] L. Breiman, "Random forests," Machine learning, vol. 45, pp. 5-
used in the literature review [9] [8]. Furthermore, 32, 2001.
preprocessing and dimensionality reduction methods have a
critical role of improve the performance by achieving
discriminant features vector. For future work, we believe:

Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on July 10,2023 at 10:27:17 UTC from IEEE Xplore. Restrictions apply.