An Efficient Ocular Disease Recognition System

Implementation using GLCM and LBP based

Multilayer Perception Algorithm
2022 IEEE 21st Mediterranean Electrotechnical Conference (MELECON) | 978-1-6654-4280-0/22/$31.00 ©2022 IEEE | DOI: 10.1109/MELECON53508.2022.9843023

1st Lakindu Induwara Mampitiya 2nd Namal Rathnayake

Department of Electrical and Electronic School of Systems Engineering
Sri Lanka Institute of Information Kochi University of Technology
Technology Kochi, Japan
Malabe, Sri Lanka namalhappy@gmail.com
lakinduinduwara21@gmail.com

Abstract—This research study is focused on the classification of There are numerous eye diseases associated symptoms.
ocular diseases by referring to a well-known dataset. The data is Those can be identified as,
divided into seven classes: diabetes, glaucoma, cataract, normal,
• Severe pain in the eyes
hypertension, age-related macular degeneration, pathological my-
opia, and other diseases/abnormalities. A Neural Network is used • Sudden loss of the vision of one eye or both
for the classification of diseases. In addition, the GLCM and • Vision get blur
LBP feature extracting methods have been used to carry out the • Red eyes
feature extraction for the fundus images. This study compares • Watery eyes
five different ocular disease recognizing techniques. Moreover, the
proposed model was evaluated regarding precision, recall, and Considering the severe effect of eye diseases on the human
accuracy. The proposed solution outperformed existing state-of- lifestyle, this research was carried out to identify eye diseases
the-art algorithms, achieving 99.58% accuracy. at a higher accuracy using the fundus image.
Index Terms—MLP, GLCM, LBP, Ocular, Fundus, Classifica- This research mainly focuses on the accurate recognition
tion
of ocular disease considering the GLCM (Gray Level Co-
occurrence Matrix) and LBP (Local Binary Pattern) features. It
I. I NTRODUCTION
aims to develop a combination of feature extraction methods
Globally nearly 2.2 billion individuals suffer from vision and Neural Networks to recognize common types of visual
impairment, according to the World Health Organization disorders using the fundus images.
(WHO), and at least 1 billion of these instances could have
been avoided [1]. Over the years, there has been a rise in II. R ELATED W ORKS
ocular diseases, with one of the reasons being the changes There are many outstanding pieces of research carried out
in the human behavioral pattern due to technology and the in the field of recognition of eye diseases. Most of them
development of technical devices. With that impact, ocular are highly accurate and use machine learning models and
diseases have severely affected the current human life. Some neural networks in classification. To increase the accuracy of
common eye diseases are diabetes, glaucoma, hypertension, the models, the researchers have carried out many exciting
cataracts, pathological Myopia, etc., due to which blindness developments in the algorithms and implemented them.
can occur. Although the consequences of eye diseases can The research [3], carried out on the diabetic retinopathy
be very severe and lead to blindness, early detection of the recognition using an ensemble model, undergoes two steps.
diseases can help eradicate the severity of the disease [2]. First step,outputs the corresponding diagnostic keywords using
Eye diseases can also be caused due to aging, exposure to a convolution neural network. Second step is designed to check
UV light, and genetic problems. With digitizing the prediction diabetic retinopathy. This method has an accuracy of 99.0%
of the ocular disease, the model that could be used for the for detecting diabetic retinopathy and normal eye.
recognition of the disease should be at a very high accuracy The research [4], was carried out to predict Age-related
and efficiency. Macular Degeneration (AMD) with the analysis of the fundus
The eye is one of the leading human organs. Vision mainly images. According to the proposed method, a novel vessel-
helps to identify and detect objects in 3D form. Losing one aware ensemble network for the fundus disease classification
eyesight or both eyesight may lead the human life to a model has been developed. This research has achieved an
disturbing lifestyle as the decisions made by the human in day- accuracy of 79.11%.
to-day lives depend on the things they see. The impact of eye To directly diagnose fundus disorders, Jing Wang et al. cre-
diseases can affect human life personally and economically. ated a CNN-based multi-label classification ensemble model.

978-1-6654-4280-0/22/$31.00 ©2022 IEEE 978

zed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:40:40 UTC from IEEE Xplore. Restrictions
 The model was developed for the dataset ODIR-2019 dataset.
The proposed EfficientNetB3 has achieved an accuracy of 1,200 1,140 1,128
92.00% [5]. Shijian Lu et al. have carried out a fundus 979
1,000
image classification with an accuracy of 96.34%. This higher

Number of Samples
accuracy was gained with the implementation of the linear 800
discriminant classifier [6].
Kanika Verma et al. has carried out a newly developed 600
method to identify diabetic retinopathy with the analysis of the
retinal images. The dataset used for the implementation of the 400
model was a project database, STARE (Structured Analysis 215 212 174
200 164
if the Retina). [7]. This research has been carried out to 103
classify the different levels of the growing of the diabetic
0
retinopathy. They are normal, moderate, and non-proliferative
diabetic retinopathy. Hemorrhages and blood vessel detection N G D C A H M O
are two features that are taken into account in the classification. Disease Name (class)
Research has achieved an accuracy of 90% for the classifica-
tion of the normal cases while the moderate and severe non- Figure 1. statistical Representation of the Dataset
proliferative diabetic retinopathy was 87.5%.
Israa Odeh et al. has proposed a new recognition method
for diabetic retinopathy. The proposed method is composed class comprises 1140 fundus images, and the Hypertension
of a machine learning ensemble and has achieved a higher class consists of 103 fundus images. Furthermore, it is also
level of accuracy for the Diabetic retinopathy classification. worth noting that the dataset shows multiple diseases for one
Researchers have used the MESSIDOR dataset with the con- patient.The sample details file represent the image and the type
sideration of the features that can be useful for the classi- of the disease. According sample describing file, the normal
fication. The model has achieved an accuracy of 82.7% for class seems to have some images from the other diseases.
recognition. [8] Considering that, the normal class of the dataset was not
In the research area of ocular diseases recognition, the considered in this study.
researchers have carried out new approaches to identifying IV. I MAGE PRE - PROCESSING
the diseases at a higher accuracy. The research conducted
by Harshvardhan G et al. describes a method to identify It is a must to focus the image on the fundus of the eye as the
Glaucoma by referring to the thermal images considering the whole model depends on the image quality. However, the noise
GLCM features of the images. Logistic regression has been in the images badly affect the accuracy of the classification
used in the accurate classification, and the proposed method model. Furthermore, the fundus color becomes very similar
has achieved an accuracy of 88.8%, a sensitivity of 60.6%, to the background color when the image is converted to
and specificity of 70.3%. [9] grayscale. This effect causes misguide to the classifier because
some of the images exist with white backgrounds. Therefore,
III. DATASET as the first step of pre-processing, the background of all
The identification of eye diseases using the fundus images images were manually removed to reduce the complication
has been a complex case in the real world, where all the given of identifying the interest points.
medical treatments to eye disease depend on the results from
the eye disease identification models.
The database Ocular Disease Intelligent Recognition
(ODIR-2019) contains fundus images of 5000 individuals with
various eye diseases, and the database is well explained with
the patient’s age and eye-wise condition. The fundus images
were captured using different types of cameras available on
the market, and the annotations were done by well-trained
human readers with quality control management. The dataset
is developed initially by the Shanggong Medical Technology
Co., Ltd.
Figure 2. Fundus images - (A) Original, (B) Background removed, (C) Gray
Figure 1 represents the class-wise sample availability of Scale
the dataset. The dataset is a combination of eight eye disease
categories: Normal (N), Diabetes (D), Glaucoma (G), Cataract The pre-processing of the image dataset affected the results
(C), Age-related Macular Degeneration (A), Hypertension (H), significantly positively. Figure 2 (A) represents the initially
Pathological Myopia (M), Other diseases/abnormalities (O). available image with a black background. Once the back-
This figure 1 shows the unbalances of the dataset. The Normal grounds are removed, a clear image representation was ob-

979

zed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:40:40 UTC from IEEE Xplore. Restrictions
 tained as in Figure 2 (B). Finally, the images were separated Homogeneity, This metric measures how near the GLCM
as the classification label into eight different folders for easy element distribution is to the diagonal.
use.
N
X −1
(i − µ)(j − µ))
V. F EATURE E XTRACTION Correlation = Pij (5)
σ2
As the features of the fundus images, GLCM and LBP were i,j=0
extracted, the features were fed to the classifier to learn. The Simply the correlation represent the linear dependency of
feature extraction was carried out after the image preprocess- grey levels of neighboring pixels.
ing was completed. This section presents an introduction of The above features are extracted through the GLCM. The
feature descriptors used in this research. related equation represents the mathematical background of
A. GLCM the above features and how they behave with the parameter
changes.
The Gray Level Co-occurrence Matrix is a well-known
Pij describes the normalized symmetrical GLCM element
approach for extracting image texture information. The GLCM
i, j. The number of gray levels in the image is specified by
determines the textural interaction between pixels by working
N , number of levels under quantization.The µ represent the
on the images’ second-order statistics. To extract an image’s
mean of GLCM pixel intensities. The variance of referenced
GLCM characteristics, initially it should be transformed to a
pixel’s intensities is represented by the σ 2 . In addition, the C,
grayscale image from an RGB image.
represents the correlation feature. The GLCM features were
GLCM extracts the features as Energy, Entropy, Contrast,
extracted from the gray converted image. [11]. The extract
Homogeneity, Correlation, Shade, and Prominence [10]. Then
features for the images are entered into an N × N matrix.
the extracted features will be written in to a N × N square
matrix. Figure 2 (C) represents the image used for the GLCM B. LBP
feature extraction.
Local Binary Pattern is an image feature extracting method
The background behind the scene is the research is mainly
used to extract the texture features of the image. The LBP
based on eye disease recognition. Therefore, the features were
gets the values for each pixel of the image, and image
extracted from 7 classes, excluding the dataset’s normal eye
texture regularity is derived by considering the LBP histogram
conditions (ODIR-2019). The focused classes in the dataset
distribution shape. The LBP has considerable discriminative
are Hypertension, cataract, pathological Myopia, Age-related
power and computational simplicity that makes the LBP more
diseases, Diabetics, Glaucoma, and other diseases.
effective in computer vision processes [12]. LBP is widely
N
X −1 used in the medical image processing field.
Energy = (Pij )2 (1)
i,j=0 VI. I MPLEMENTATION
The statistical data is known as uniformity or the angular A. MLP
second moment. The Energy represents the disorders in the MLP is categorized as a feed-forward artificial neural network
textures, and it tends to reach a maximum value of one. In the integrated with more than a single layer. The combination of
GLCM, the Energy gives out the sum of squared components. many single perceptrons creates the MLP. Moreover, the input
N −1
layer, hidden layer and output layer are the main layers that
X
Entropy = ln(Pij¨ )Pij (2) are considered when designing a neural network [13]. Same
i,j=0
as in the feed-forward neural networks, in the MLP, the output
layer is fed by the data that flow from the input layer [14].
The entropy of the image measures the complexity and the
• Input Layer - The beginning of the model where the
disorder. When the image is not texturally uniform, the entropy
model takes the input variables as the input to the model.
of the image tends to be a higher value. Entropy is significantly
• Hidden Layer - The hidden layers make the interaction
associated with energy and inverse.
between the input layer and the out put layer. MLP is
N
X −1 created with many hidden layers.
Contrast = Pij (i − j)2 (3) • Output Layer - The layer connected to the out of the
i,j=0 model. Where the output variables are created for the
Contrast tends to read the local variations of the GLCM. prediction or the classification.
Simply this statistic tends to measure the spatial frequency The model learning procedure is carried out under three
of an image. The contrast can be defined as, the difference steps. The MLP uses the Back-propagation for better results.
between the highest and lowest values in an adjoining group • Forward propagation, The input layer is the beginning
of pixels. layer of the model. The data is forward propagated for
N −1 the output layer.
X Pij • Error is calculated by considering the predicted class and
Homogeneity = (4)
i,j=0
1 + (i − j)2 the true class.

980

zed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:40:40 UTC from IEEE Xplore. Restrictions
 • Back-propagate the error to minimize the model’s error
by updating the model.
The model follows the above three steps in learning from
the dataset at multiple epochs.
B. Proposed Model
The proposed method is consisted with LBP, GLCM feature
extractors and MLP artificial neural network as the classifier.
The steps carried out were taken to increase the classification’s
accuracy and efficiency.

Figure 4. Confusion matrix for the proposed model

The figure 4 illustrate that the class P (Pathological Myopia)

has gained an accuracy level of 98.68 % while 1.32 % was
misclassified as O (other diseases). The A (Age-related) gained
an accuracy of 96.15% while 3.85% was misclassified as
C (Cataract). All the remaining five diseases attained 100%
accuracy. In addition, the model had a higher rate of accuracy,
recall, and F1 scores for each class.

Table I
Figure 3. Step-wise Overall Model process C LASS WISE RESULTS FOR P RECISION , R ECALL , AND F1 S CORE

Class Precision Recall F1-Score

The selected features, GLCM and LBP, were extracted Age related 1.00 1.00 1.00
from the fundus images after the image’s backgrounds were Cataract 1.00 0.96 0.98
removed. The fundus images were turned to the grayscale Diabetics 1.00 1.00 1.00
Glaucoma 1.00 1.00 1.00
to carry out GLCM and LBP feature extraction. The feature Hypertension 1.00 1.00 1.00
extraction process, up to the MLP classifier development, was Other 0.97 1.00 0.99
carried out using MATLAB [15]. The extracted values were Pathological M 1.00 0.99 0.99
fed to the MLP classifier model, consisting of 5 hidden layers
then the model training was carried out under 70 epochs. Table I represents the precision, recall, and F1 score for
all the seven evaluated classes. The precision represents the
VII. R ESULTS AND D ISCUSSION quality of the positive predictions made by the model. The
This research was carried out to recognize seven ocular recall shows how many true positives were recalled by the
diseases with the analysis of fundus images using the GLCM, model, while the F1 score represents the weighted average
LBP feature extraction methods, and Multilayer Perception of the recall and the precision attained by the model in the
(MLP) classifier algorithm. prediction process. K-fold cross-validation is a resampling
In the process of the proposed model recognizing defined method for evaluating models on small data sets. For a better
ocular diseases, the model should be highly efficient and evaluation of the proposed model, K-fold cross-validation was
accurate. As the dataset (ODIR-2019) was a highly unbalanced carried out for K =10. The accuracies seem to be slightly
dataset designed with eight ocular disease fundus images, the different with the increase of the K value from 1 to 10.
model was evaluated by considering the unbalance of the The statistical graph, Figure 5 represents how the accuracy
dataset. of the model tends to change with the K value. At K = 4 model
Precision, Recall, and F1 accuracy was separately calculated gets the best scores for the accuracy value 99.58%. The model
for each class: Age-related, Cataract, Diabetics, Glaucoma, achieves a minimum accuracy of 98.12% at the K= 10.
Hypertension, Pathological Myopia, and other diseases. The Figure 6 illustrates the variation of the loss with reference to
confusion matrix analyzed the imbalance class to illustrate the K -Fold value. The proposed model achieved the highest
the accuracies for each category according to the weight of loss of 0.078% at the K = 10, with minimum accuracy. 0.018
the class. % loss was achieved at K = 4, where the model performs at the

981

zed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:40:40 UTC from IEEE Xplore. Restrictions
 1
99.5
0.9

Accuracy (100%)
Accuracy

99 0.8

0.7
98.5
0.6

0.5
98
2 4 6 8 10 0 20 40 60
K Fold Epochs

Figure 5. Change of the Accuracy with the K-Fold value Figure 8. Validation accuracy for epochs

·10−2
8 1.5
Accuracy (100%)

6 1
Loss

4 0.5

2 0
0 20 40 60
2 4 6 8 10
Epochs
K Fold
Figure 9. Validation loss for epochs
Figure 6. Change of Loss with the K-Fold value

1 highest accuracy, 99.58%. Accuracy was evaluated concerning

the loss value for each K value.
The training accuracy indicates how well the model classi-
fies the two images during the training process on the training
Accuracy (100%)

0.8 dataset. Figure 7 the training accuracy of the model that

performed at every epoch value, where the epoch value was
set to 70. The training accuracy starts to grow from a value of
47.55% and reaches the highest value 100% at 50 epochs, and
after the 50 epochs, the model maintains the training accuracy
0.6
higher than 99%.
The validation accuracy represents the accuracy attained
with the validation data. According to the figure 8 the accuracy
starts growing from a value of 50.8 % and extends to the
0 20 40 60 maximum value of 99.58 %. The epoch value in the range of
Epochs 38 to 70 model keeps the accuracy at a higher level with a
slight fluctuation.
Figure 7. Training Accuracy for epochs The Validation loss is the value of the loss achieved by the

982

zed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:40:40 UTC from IEEE Xplore. Restrictions
 model on the validation data. With the consideration of the image,” in 2020 IEEE International Conference on Image
above figure 9 it can be clearly understood that the model is Processing (ICIP), 2020, pp. 320–324.
rightly fitted to the dataset without underfitting or overfitting. [5] J. Wang, L. Yang, Z. Huo, W. He, and J. Luo, “Multi-
The fitting of the model could be identified by the training loss label classification of fundus images with efficientnet,”
and validation loss graphs, where both graphs reach a minimal IEEE Access, vol. 8, pp. 212 499–212 508, 2020.
point with higher stability and the minimal gap between the [6] S. Lu, J. Liu, J. H. Lim, Z. Zhang, T. N. Meng, W. K.
last loss values. The graph is the plot for decreasing the loss Wong, H. Li, and T. Y. Wong, “Automatic fundus image
value with the epoch values. classification for computer-aided diagonsis,” in 2009 An-
nual International Conference of the IEEE Engineering
VIII. C ONCLUSION AND F UTURE WORK in Medicine and Biology Society, 2009, pp. 1453–1456.
The proposed method is for ocular diseases recognition [7] K. Verma, P. Deep, and A. G. Ramakrishnan, “Detection
using the fundus images. The discussed process comprises two and classification of diabetic retinopathy using retinal
feature extractors, GLCM and LBP, while the data classifica- images,” in 2011 Annual IEEE India Conference, 2011,
tion was carried out using the MLP. Since the available dataset pp. 1–6.
was entirely imbalanced, the model was evaluated according [8] I. Odeh, M. Alkasassbeh, and M. Alauthman, “Diabetic
to the weighted class values. The model was able to achieve retinopathy detection using ensemble machine learning,”
a higher accuracy value of 99.58% with a loss of 0.018%. in 2021 International Conference on Information Tech-
With the consideration of the confusion matrix and other nology (ICIT), 2021, pp. 173–178.
plots mentioned above, it can be concluded that the proposed [9] H. G, V. N, and P. N, “Assessment of glaucoma with
model is functioning at higher accuracy and efficiency. Several ocular thermal images using glcm techniques and logistic
research have been carried out in past years to determine the regression classifier,” in 2016 International Conference
ocular disease. on Wireless Communications, Signal Processing and
Networking (WiSPNET), 2016, pp. 1534–1537.
Table II [10] P. K. Mall, P. K. Singh, and D. Yadav, “Glcm based
ACCURACY COMPARISON OF THIS STUDY AGAINST SIMILAR WORKS . feature extraction and medical x-ray image classification
Reference Study Algorithm Accuracy (%) using machine learning techniques,” in 2019 IEEE Con-
This study Multilayer Perceptron 99.58 ference on Information and Communication Technology,
Kanika Verma, et al. [7] Random Forest 87.5 2019, pp. 1–6.
Israa Odeh , et al. [8] Ensemble Classifier 82.7
Yadeeswaran K S, et al. [3] CNN 99.0 [11] F. Mirzapour and H. Ghassemian, “Using glcm and gabor
Dihao Luo, et al. [4] Ensemble Classifier 79.11 filters for classification of pan images,” in 2013 21st
Jing Wang, et al. [5] EfficientNetB3 Ensemble 92.00 Iranian Conference on Electrical Engineering (ICEE),
2013, pp. 1–6.
As shown in the table II, it can be identified that the [12] A. Varghese, R. R. Varghese, K. Balakrishnan, and J. S.
proposed model performs well compared with the models Paul, “Level identification of brain mr images using
mentioned above with a higher accuracy level of 99.58% by histogram of a lbp variant,” in 2012 IEEE International
classifying seven well-known ocular diseases. Conference on Computational Intelligence and Comput-
Future studies might include developing a real-time visual ing Research, 2012, pp. 1–4.
disease detection system using a Multilayer Perceptron model [13] J. Qin, H. Wang, K. Li, Y. Qi, X. Jia, S. Xu, and C. Dong,
with GLCM and LBP feature extractors. “The quantitative prediction of auxiliary sliding distance
of freestyle skiing based on mlp neural network,” in
R EFERENCES 2019 IEEE 2nd International Conference on Information
[1] “Vision impairment and blindness.” [Online]. Communication and Signal Processing (ICICSP), 2019,
Available: https://www.who.int/news-room/fact-sheets/ pp. 469–472.
detail/blindness-and-visual-impairment [14] L. Abhishek, “Optical character recognition using ensem-
[2] U. R. Acharya, N. Kannathal, E. Y. K. Ng, L. C. Min, ble of svm, mlp and extra trees classifier,” in 2020 Inter-
and J. S. Suri, “Computer-based classification of eye national Conference for Emerging Technology (INCET),
diseases,” in 2006 International Conference of the IEEE 2020, pp. 1–4.
Engineering in Medicine and Biology Society, 2006, pp. [15] M. Moradi, M. A. Poormina, and F. Razzazi, “Fpga
6121–6124. implementation of feature extraction and mlp neural
[3] Y. K S, N. Mithra, V. KS, and K. R, “Classification of network classifier for farsi handwritten digit recognition,”
diabetic retinopathy through identification of diagnostic in 2009 Third UKSim European Symposium on Computer
keywords,” in 2021 Third International Conference on Modeling and Simulation, 2009, pp. 231–234.
Inventive Research in Computing Applications (ICIRCA),
2021, pp. 716–721.
[4] D. Luo and L. Shen, “Vessel-net: A vessel-aware en-
semble network for retinopathy screening from fundus

983

zed licensed use limited to: Vignan's Foundation for Science Technology & Research (Deemed to be University). Downloaded on June 15,2024 at 04:40:40 UTC from IEEE Xplore. Restrictions