Measurement: Sensors 25 (2023) 100566

Contents lists available at ScienceDirect

Measurement: Sensors
journal homepage: www.sciencedirect.com/journal/measurement-sensors

Effect of multi filters in glucoma detection using random forest classifier

Alice K a, *, Deepa N b, Devi T b, BeenaRani B B b, Bharatha Devi N b, Nagaraju V b
a
Department of Computing Technologies, SRMIST, KTR Campus, Tamilnadu, India
b
Department of Computer Science & Engineering, Saveetha School of Engineering, Saveetha Institute of Medical and Technical Sciences, Chennai, India

A R T I C L E I N F O A B S T R A C T

Keywords: Glaucoma is an eye disease that damages the optic nerve which connects the eye to the brain. When the fluid
Gluacoma pressure inside the eye (intraocular pressure) increases, the optic nerve get impaired and has doubled the chance
Retinal fundus image for diabetic patients resulting in irreversible loss of vision if not detected in early stages. In developing countries,
Edge histogram filter
due to the scarcity of ophthalmic experts and lab facilities, the needs for eye disease detecting automation system
Fuzzy color and texture histogram filter
Pyramid histogram of gradients filter
are increased without saying. The field of artificial intelligence is providing many solution’s especially in health
Random forest classifier care domain. The proposed work generate models for recognizing the presence of glaucoma based on open access
Multi filters public dataset of retinal fundus images using machine learning algorithms with the help of image feature de­
scriptors. It classifies the given retinal fundus image as normal or abnormal in two stages. Firstly it extracts image
features using appropriate filters and then it is trained through tree based ensemble classifier to classify the given
input image and then the same is tested to get the better accuracy performance. The above two steps are iterated
by varying over the three effective filters like edge histogram, fuzzy color and texture histogram and pyramid
histogram of gradients. The proposed experiment based on this approach reveals that the use of Edge histogram
filter in combination with fuzzy color and texture histogram with Random forest classifier yields maximum
accuracy of 80.43% and AUC 0.884. The results obtained by applying multi filters is better than that obtained by
applying single filter.

1. Introduction the set of input images for maximum possible accuracy. Many different
machine learning algorithms like Support Vector Machine [7], Random
Machine learning, a subset of Artificial Intelligence is a study of Forest [1], Artificial Neural Networks [4], or Decision Tree [14] can be
computer vision representative algorithms that improves its efficiency applied on image datasets efficient. In weka 3.8.1 environment [8]
through experience gained from training with known data and thus the implemented in java with many different packages for image filters that
built model is used in making predictions about unknown data. The are used to extract the content of the image and represented as a
training data can be either image, numerical or textual. Classification spreadsheet of attributes and their filtered values corresponding to the
through the training should be accurate by not only varying the machine pixel distribution for each image.
learning algorithms but also tested them parallel. However the inputs Retinal fundus images are widely used in diagnose of many
data whether image, textual or numeric should be preprocessed before ophthalmic diseases like glaucoma, senile maculopathy, diabetic reti­
entering into loop of mining iteration. Image classification is the process nopathy etc. Optic disc, blood vessels and macula are the important
of labelling the image into one or more predefined categories [9]. Ap­ region of retinal image. The shape and size of optic disc provides vital
plications of image classification include Face recognition [10], Hand­ information in detecting many ophthalmic diseases. Glaucoma an opti­
writing detection [15], Authentication systems [2] Healthcare cal disease is caused due to pushing of optic nerve by intra ocular
information system [25] and many more. In Content Based Image pressure cause irreversible loss of vision if not detected in early stage. In
Retrieval (CBIR) process, the image classification algorithm works by a survey by world health organization it is estimated that by the end of
extracting the features of image using image filters and those features 2030 the number of glaucoma patients will raise up to 95 million. There
associated to certain classes deterministically. This mechanism classifies are two types of glaucoma namely, angle closure glaucoma and open

* Corresponding author.
E-mail addresses: [email protected] (A. K), [email protected] (D. N), [email protected] (D. T), [email protected] (B. B B),
[email protected] (B.D. N), [email protected] (N. V).

https://doi.org/10.1016/j.measen.2022.100566
Received 26 July 2022; Received in revised form 17 October 2022; Accepted 15 November 2022
Available online 17 December 2022
2665-9174/© 2022 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
A. K et al. Measurement: Sensors 25 (2023) 100566

angle glaucoma. The symptoms of former include severe eye pain, VGG 19 architecture to detect affected images and used three different
sudden slight loss, blurred vision and if not addressed at early stage leads datasets of which two are public datasets and one is a private dataset
to irreversible loss of vision. And the latter is even more dangerous with from the region of Parc de Salut Mar in Barcelona.
no symptoms until peripheral sight is completely lost and is referred as Classification based on initially the feature extraction and later
‘sneak thief of sight’ [34]. feature selection is addressed by many researchers. Acharya et al. [31]
Fig. 1 depicts the taxonomy of possible glaucoma detection ap­ extracted statistical features and the most influencing features are
proaches using optic disc fundus images found in literature. Glaucoma selected using principal component analysis. Then the images are clas­
detection using retinal fundus images is done by segmenting the optic sified using SVM classifier based on reduced feature set. Vijayan et al.
disc and analyzing the size and shape using various masks is used by [32] used gabor filters to represent image features and used a set of five
ophthalmic experts in lab. The cup-to disc metric is used to categories classifiers for classification and proved that Random Forest has highest
the glaucoma into mild, moderate and severe depending on the range of accuracy in classifying retinal fundus image. Neha Gour et al. [33] used
0.4–0.5, 0.5–0.7and more than 0.7 respectively. Similar to cup to disc GIST and PHOG filters to represent image features and SVM Classifier for
metric, ISNT, DDLS and GDI are other effective parameters also used in classification.
detecting glaucoma. As it is very tedious and time consuming which is In this proposed work glaucoma fundus images are classified using
demanding a lot of expertise, many researchers provided various auto­ machine learning algorithms by representing the image as image feature
mated solutions by using machine learning algorithms. Factors like poor descriptors and then classifiers were applied. Image filter like FCTH
quality of fundus images and poor illumination of blood vessels inside filter [5,12], Edge Histogram filter [1,13] and PHOG filter [6,11,16] are
the optic disc, results in improper segmentation which in turn leads to considered to add new numeric data to each instance of retinal fundus
incorrect classification. To avoid this problem in segmentation, features image in the dataset, which helps the classification algorithm to develop
of optic disc is extracted using many well-known feature extraction the data model more precisely.
methods like CNN, Gabor filter, and higher order statistical feature after
which machine learning classifier like SVM, Logistic Regression and 3. Materials and methods
Random forest is applied for classification.
In the following pages of this article are organized as: section 2 Fig. 2 depicts the steps and follow for the methodology used in the
elaborates the previous relevant work of selected authors focusing on proposed work. This section starts with dataset description and explains
the theme. Section 3 deals with materials and Methods described for how the dataset is organized to facilitate feature extraction in pre­
improving the readability. Section 4 focuses on details of primary processing stage. Seven different sets of feature vector for the image
experimental setup and finally deals with main results and discussion. dataset are prepared using three image filters namely Edge Histogram
filter (EH), Fuzzy color and texture histogram feature (FCTH) and Pyr­
2. Literature review amid Histograms of Orientation Gradients (PHOG) filter. The features
thus generated are fed into Random forest classifier for classifying the
Many researchers provided solutions for detecting Glaucoma. A. presence or absence of Glaucoma.
Sevastopolsky [17] proposed an automatic optic disc and cup segmen­
tation using U-Net convolutional neural network. Tan et al. [18] sug­
gested a convolutional neural network (CNN) model without any ground 3.1. Dataset description
truth for segmentation of optic disc with weak supervised learning to
train the dataset. S. Phan et al. [20] indicates CNN and M. Al Ghamdi The proposed algorithm works on three different datasets are used
et al. [21] used CNN with semi supervised learning by using transfer namely DB1, DB2 and DB3 taken from RIM_ONE (an open retinal image
learning. R. Asaoka et al. [22] proposed a system with RESNET archi­ database for optic nerve evaluation) dataset available online at rimone.
tecture and used two datasets, one with data augmentation and other isaatc.ull.es for research purpose and the images are taken from three
without. B. An et al. [23] used Optical Coherence Tomography and different hospitals in Spain [19]. Five experts were involved in labeling
transfer learning to evaluate the model and to detect glaucoma color the classes as glaucoma images and normal images. DB1 dataset consists
fundus images, The Random Forest classifier with CNN model improved of 169 images in bmp image format which is taken from RIM_ONE r1
the system performance. Diaz-Pinto et al. [24] applied five different dataset and all the retinal fundus images are classified into two different
datasets and compared their performance. Cerentinia et al. [25] subsets. 118 instances of normal eye images and glaucoma instances
considered GoogleNet Architecture for training the data to classify the images are categorized into 4 different types. They are Early glaucoma:
glaucoma images from retinal fundus images. It uses RIM-ONE r1,r2, r3
[19] all three databases of retinal fundus images. Orlando et al. [26]
developed an automatic glaucoma detection system using two CNN
models. de Moura Lima et al. [27] developed a glaucoma detection
system using the architectures namely InceptionResNetV2, ResNet50,
InceptionV3, VGG-16 and VGG-19 on RIM-ONE dataset. Li et al. [28]
considered VGG network to diagnose Glaucoma visual fields from
non-glaucoma visual fields. The dataset samples are obtained from three
ophthalmic centers in China. Gómez-Valverde et al. [29] pointed out

Fig. 1. Taxonomy of Glaucoma detection approaches. Fig. 2. Proposed methodology for glaucoma detection.

2
A. K et al. Measurement: Sensors 25 (2023) 100566

12 images, Moderate glaucoma: 14 images, Deep glaucoma: 14 images,

Ocular hypertension (OHT): 11 images. All these four types of glaucoma
images are grouped together to have 51 instances of glaucoma images.
DB2 dataset consists of 455 images in jpg format which is taken from
RIM_ONE r2 dataset with 255 instances of normal retinal images and
200 instances of glaucoma images. In our proposed algorithm, DB3
dataset is also used which contains 624 images and is a combination of
both the DB1 and DB2 dataset to increase the representation capacity of
final classifier. DB3 dataset consists of 373 instances of normal retinal
images and 251 instances of glaucoma images. Details of dataset used for
experiments are as shown in Table 1. Fig. 3 contains image samples on
glaucoma and normal instances respectively.

Fig. 3. Fundus Image from RIM ONE Database (a) Glaucomatous image (b)
3.2. Preprocessing Normal image.

The input fundus image is preprocessed before feature extraction and

The steps in PHOG feature extraction process is as follows [16]:
classification. All the images in DB1 and DB2 are arranged in two
Step-1: for a given sample image edge contours, description about shape
different folders and combined into one dataset and stored in third
information are extracted. Step-2: the image is divided into cells at
folder. All the files in each of the folder are named consequently which is
different pyramid levels. Step-3: for each grid and at each pyramid
done to get the class names of the instances to an arff file. This arff files
resolution level the HOG was calculated. Step-4: the final PHOG
contains only filename and class nominal values for each instances of
descriptor is obtained by concatenating all the HOG vectors generated at
image. The image must be converted to some numeric values before
each pyramid resolution and this gives the spatial information of the
classification. So the next step is to apply image filters which convert the
image. Next each HOG is normalized to unity for all pyramid levels to
pixel intensity values into some numeric information which describes
compensate illuminations. PHOG descriptor is very sensitive to changes
the image characteristics.
in object shape due to clutter. For any applications the number of levels
of pyramid can be limited to 3 to avoid over fitting [11].
3.3. Feature descriptors
3.4. Classification
Fuzzy color and Texture histogram (FCTH) filter is a descriptor
that encodes color and texture information in a single histogram. It is a In proposed work Random forest classifier is used for classification
combination of three fuzzy groups [12]. In first group 20 fuzzy rules are purpose as that amounts to be the classification effort by committee of
applied so that 10 histogram bins (0-bin to 9-bin) are produced. Each bin base classifiers, each classifier being a decision tree. This committee of
represents a preset color as follows: bin 0 - Black, bin 1- Gray, bin 2- classifiers provides good results in classifying retinal fundus image
White, bin 3- Red, (4) Orange, (5) Yellow, (6) Green, (7) Cyan, (8) Blue dataset when compared with other classifiers like J48, KNN and OneR
and (9) Magenta [5]. In the second group the hue value of each color is [32] one at a time. Random forest Classifier is more efficient and widely
imported and histogram expands from 10-bin to 24-bin. In the third used classifier. It is highly effective in handling nonlinear classification
group, set of texture features are exported and the histogram further especially for a huge dataset.
expands to 192 bins. Thus the color, texture and other spatial informa­ It tries to find out the best node in the decision trees and performs
tion are extracted in a single histogram [1]. inbuilt optimization operation. Since Random forest is a collection of
Edge histogram filter is a very powerful filter for extracting MPEG7 more than one decision tree and each tree depends on the random vector
edge features from images. It represents edge distribution in a histogram sampled independently, same data distribution is used for all the
based on local edge distribution and shape information [3]. The edge member trees. This comes under ensemble type of classifier in which
histogram depends on frequency and brightness changes and shows the best solution is obtained by majority voting. It minimizes the problem of
direction of edges in an image. The edge histogram considered five types over-fitting by averaging the result [30] and it always sounds better than
of edges as shown in Fig. 4: horizontal, Vertical, 45◦ diagonal 135◦ di­ using a single decision tree. The default values of hyper parameters used
agonal and non-directional edges [13]. in this classifier are NUMDECIMALPLACES = 2, NUMFEATURE = 0 and
Irrespective of image size the image is divided 4 × 4 non overlapping SEED = 1.
sub blocks totaling 16 blocks and the histogram was generated for each
block which represents the distribution of above mentioned edges in the 3.5. Evaluation metric
sub blocks taken into consideration.
Pyramid Histograms of Orientation Gradients (PHOG) filter is a Evaluation metric like accuracy and area under curve are calculated
descriptor that encodes the orientation gradients of each image region at based on confusion matrix generated by the selected classifier. Accuracy
every pyramid resolution level [6]. PHOG features are spatial shape is the ratio of the total number of correctly classified images to the
descriptors which describes the spatial layout and local shape of an number of images in the dataset taken into consideration.
image. These local shapes are captured using the edge orientation dis­
tribution within a region and spatial layout is captured by tilting on Accuracy = (TP + TN) / (TP + FN + TN + FP)
image into regions at multiple resolutions.
Where TP- Correctly classified glaucoma images.

Table 1 TN-Correctly classified normal images

Experiment dataset details.
FP- Incorrectly classified normal images
Name of Number of Glaucoma Fundus Normal Fundus File FN-Incorrectly classified glaucoma images
Dataset Instances Image Instances Image Instances format

DB1 169 51 118 BMP Sensitivity is the ability of a system to correctly identify images with
DB2 455 200 255 JPG a glaucoma disease. It is also called as true positive rate. Specificity is the
DB3 624 251 373 BMP&JPG
ability of a system to correctly identify images without the glaucoma

3
A. K et al. Measurement: Sensors 25 (2023) 100566

Fig. 4. Types of edge direction in edge histogram filter.

disease or normal images. It is also called as false positive rate. Receiver out using Random forest (RF) learning model. This model is represented
operating characteristics curve is a graph between true positive and false by a family of decision trees and acts like an ensemble to predict the
positive cases. glaucoma with expected accuracy. A confusion matrix is generated after
Area under curve gives a measure of area under the ROC curve. The RF classification based in the feature vector generated for each dataset.
closer the curve towards its top left corner the better is the classifier The results obtained from the proposed system is summarized in
performance. Table 2 for DB1, DB2 and DB3 dataset respectively for seven different
combinations of filters selected to represent the retinal fundus images in
4. Results and discussion the dataset.
From Table 2 it is clear that the performance of Random forest
The proposed system was experimented using WEKA 3.8.1 in win­ classifier is higher than the rest in terms of accuracy namely edge his­
dows 10 equipped with core i7. The mining tool used for getting the togram oriented features yield up to 73.96% for DB1 when applied
results is implemented in Java and widely used in the research com­ separately than with combination of other features like FCTH and/or
munity for around 3 decades. The performance evaluation of the glau­ PHOG. However the performance of the model increases for DB2 with
coma diagnosis system is measured using three different datasets having the combined filters than the single filters. The combination of FCTH
624 retinal fundus images in total. The images in first RIM ONE dataset and PHOG features is found to be increase in DB3 than in others. The
available at rimone. isaatc.ull.es and all these retinal fundus images are performance of model with combined features of both EH and FCTH
captured using KOWA WX 3D camera with focus at optic nerve head works better for DB2 with 80.43% slightly higher than the model that
with a resolution of 2144 × 1424. Similarly RIM ONE r2 and r3 are uses EH separately and DB3 it is 75.48%. Fig. 5 depicts the variation in
captured. accuracy of model for varying combination of filters.
For each of these dataset seven different sets of feature vector are In any binary classifier, the accuracy of the classification model de­
extracted using three popular filters: Edge Histogram, Fuzzy Color and pends upon the total number of instances in each class label. The bias­
Texture Histogram and Pyramid Histogram for Oriented Gradients. ness is inevitable in an imbalanced distributed dataset as the
These filter facilities are available in software package version 3.8.1 performance executed by the majority class always dominates the mi­
Weka tool as an addendum. The image filter package in Weka may be nority class. Unlike accuracy metric the receiver operating characteristic
available under tab ‘un-supervised instance ‘filters from 10 different curve is a visualization tool used to specify in a cost effective manner
types of image filters listed in the dropdown box. From this list three whether the used model is appropriate or not. The area under ROC Curve
most suitable image filters which performs well for image dataset is used is given as AUC.
in proposed system. These filters are applied either as a single filter (first Table 3 gives a summary of area under ROC curve of the proposed
three rows in Table 2) or as multiple filter in order to study the variation systems for DB1, DB2 and DB3.
in the performance. After applying these filters to images, a set of fea­ From Table 3 it is clear that the model with edge histogram feature
tures is generated and is stored as numeric values in ARFF file which can and random forest classifier yields better results with AUC 0.721 for DB1
then be used in classification. The generated ARFF file contains records dataset compared with all other set of feature vector taken into
corresponding to the number of input images converted as a set of at­ consideration. Similarly the model with Edge histogram feature and
tributes depending up on the filter applied along with filename and class fuzzy color and texture histogram feature and random forest classifier
label as follows where n depends on the type of filter used. works better with AUC 0.888 and 0.850 for DB2 and DB3 respectively
Attribute Attribute Attribute ………… Attribute Class when compared with other feature set combinations.
1 2 3 …… n Label Fig. 6 shows that the results of AUC for DB2 are high for all type of
2 1 5 0 NO
3 1 0 5 YES

After this filtering or preprocessing the classification step is carried

Table 2
ACCURACY of proposed method on DB1, DB2 and DB3.
FILTER NO OF Accuracy Accuracy Accuracy
ATTRIBUTES DB1 DB2 DB3

EH 80 73.96 80.21 75.44

FCTH 192 69.23 71.64 70.67
PHOG 630 68.05 55.82 58.65
EH + FCTH 272 63.9 80.43 75.48
FCTH + PHOG 822 57.39 67.42 72.75
PHOG + EH 710 64.49 76.48 67.62
EH + FCTH + 902 66.86 76.7 74.36
PHOG
Fig. 5. Accuracy of Random Forest model for various filter combination.

4
A. K et al. Measurement: Sensors 25 (2023) 100566

Table 3
AUC of proposed method on DB1, DB2 and DB3.
FILTER NO OF ATTRIBUTES AUC DB1 AUC DB2 AUC DB3

EH 80 0.721 0.878 0.832

FCTH 192 0.658 0.782 0.742
PHOG 630 0.486 0.562 0.551
EH + FCTH 272 0.575 0.888 0.850
FCTH + PHOG 822 0.501 0.707 0.698
PHOG + EH 710 0.584 0.854 0.816
EH + FCTH + PHOG 902 0.596 0.861 0.837

Fig. 8. ROC plot for DB2 classified under Random Forest for EH and
FCTH Filter.

Fig. 6. Area under curve of Random Forest model for various filter
combination.

filter combination with maximum for EH and FCTH combined filters.

The AUC for DB1 and DB3 is higher when EH filter is applied to image
dataset.
Fig. 7 The ROC plot for maximum values of area under curve for DB1
is shown. Fig. 8 The ROC plot for maximum values of area under curve
for DB2 and Fig. 9 the ROC plot for maximum values of area under curve
for DB3 is shown. The more towards top left corner the graph is the
better is the classifier performance which is reported the best in Fig. 6
with AUC 0.888.
The results obtained for RIM one database is summarized in Table 4
and compared with recent technologies in the literature. Syna sreng
et al. [34] used optic disc segmentation based glaucoma detection and
used eleven pre trained networks as feature descriptors in which
inceptionV3 with SVM classifier has an accuracy of 63.16% and for Fig. 9. ROC plot for DB3 classified under Random Forest for EH and
Inception Resnet with SVM classifier it is 78.95% for RIM one database. FCTH Filter.

Table 4
Performance comparison.
Method Features Classifier Accuracy AUC Database

Syna sreng InceptionV3 SVM 63.16% – RIM one

et al., 2020 Inception SVM 78.95% – RIM one
[34] ResNet
Raghvendran Radon + SVM 78.80% 0.72 HRF
et al., 2018 Gabor
[36]
Neha Gour PHOG SVM 75.25% 0.72 Drishti-
et al., 2019 GS1
[33] PHOG SVM 73.00% 0.80 HRF
VijayanT et al., Gabor Random 70.15% 0.86 Kaggle-
2020 [32] Forest retinal
fundus
images
Proposed EH + FCTH Random 80.43% 0.884 RIM one r2
system Forest
EH + FCTH Random 75.48% 0.850 RIM one
Forest r1+r2

Fig. 7. ROC plot for DB1 classified under Random Forest for EH Filter.

5
A. K et al. Measurement: Sensors 25 (2023) 100566

Raghvendran et al. [36] used feature extraction based glaucoma [5] S.A. Chatzichristofis, Y.S. Boutalis, Fcth: fuzzy color and texture histogram-a low
level feature for accurate image retrieval, in: Proceedings of Ninth International
detection with non CNN based method for HRF database and a combi­
Workshop on Image Analysis for Multimedia Interactive Services (Klagenfurt,
nation of Radon and Gabor Transform is used as image feature Austria, IEEE, May , pp. 191–196. WIAMIS’08.
descriptor with SVM classifier has an accuracy of 78.80% and AUC 0.72. [6] A. Chauhan, D. Chauhan, C. Rout, Role of Gist and PHOG Features in Computer-
Neha Gour et al. [33] used two different dataset for glaucoma detection Aided Diagnosis of Tuberculosis without Segmentation. PLoS ONE 9 (11) (2014),
e112980, https://doi.org/10.1371/journal.pone.0112980.
and achieved an accuracy of 75.25% and AUC 0.72 when used PHOG as [7] C. Cortes, V. Vapnik, Support vector networks, Mach. Learn. 20 (1995) 273–297.
filter to represent image with SVM classifier for Drishti-GS1database but [8] WEKA. http://www.cs.waikato.ac.nz.
for HRF for the same architectural setup they achieved 73% accuracy [9] P. Kamavisdar, S. Saluja, S. Agrawal, A survey on image classification approaches
and techniques, International Journal of Advanced Research in Computer and
and 0.80 AUC. VijayanT et al. [32] used Gabor filter and Random forest Communication Engineering 2 (1) (2013) 1005–1009. Jan. 2013.
classifier to build a model to classify eye disease using kaggle database [10] K. More, P. Kadam, A. Jadhav, D. Dalgade, Face authentication application for
with high resolution retina images to achieve an accuracy of 70.15% and social networking site, Int. J. Comput. Sci. Mobile Comput. 4 (3) (2015) 430–433.
Mar. 2015.
AUC 0.862. The proposed system outperforms the techniques discussed [11] Partha Pratim Sarangi, B.S.P. Mishra, Sachidanada Dehuri, Pyramid histogram of
above in terms of accuracy and AUC when edge histogram and fizzy oriented gradients based human ear identification, Int. J. Control Theor. Appl. 10
color and texture features are combined and used as feature descriptors (15) (2017), 0974-5572,International science press.
[12] Radmila Jankovic,Classifying Cultural Heritage Images by Using Decision Tree
with random forest classifier is used with maximum accuracy of 80.43% Classifiers in WEKA,1st International Workshop on Visual Pattern Extraction and
and AUC 0.884 for DB2 and 75.48% accuracy and 0.850 AUC for DB3. Recognition for Cultural Heritage Understanding (VIPERC) Co-located with 15th
Italian Research Conference on Digital Libraries (IRCDL 2019) Jan2019.
[13] Radmila Jankovic, Machine learning models for cultural heritage image
5. Conclusion classification: comparison based on attribute selection, Information 11 (1) (2020)
12, https://doi.org/10.3390/info11010012. Dec 2019.
In this article seven set of various combinations of filters are applied [14] L. Rokach, O. Maimon, Decision trees, in: The Data Mining and Knowledge
Discovery Handbook, Springer US, 2010, https://doi.org/10.1007/978-0-387-
to three different datasets for detection of glaucoma and the classifica­
09823-4.
tion performance are compared. Filters were applied individually and [15] P. Yadav, N. Yadav, Handwriting recognition system - a review, 0975–8887, Int. J.
then combination of filters is used to check whether the performance Comput. Appl. 114 (19) (2015) 36–40. Mar. 2015.
measure is increased. Random forest is chosen for the classification part [16] Bai Yang, Lihua Guo, Lianwen Jin, Qinghua Huang, A novel feature extraction
method using Pyramid Histogram of Orientation Gradients for smile recognition,
because it has proven to be better for classifying similar type of dataset in: 2009 16th IEEE International Conference on Image Processing (ICIP), 2009,
[32]. Different features represent different properties of the input image https://doi.org/10.1109/icip.2009.5413938.
so different filters will have different impact on the eventual output. [17] A. Sevastopolsky, Optic disc and cup segmentation methods for glaucoma detection
with modification of U-net convolutional neural network, Pattern Recogn. Image
From the Area under curve information the performance of Random Anal. 27 (3) (2017) 618–624.
forest model in classifying the images in DB1 is not that good and this [18] J.H. Tan, et al., Segmentation of optic disc, fovea and retinal vasculature using a
may be due to the less number of instances of images and also the single convolutional neural network, J. Comput. Sci. 20 (2017) 70–79.
[19] F. Fumero, et al., RIM-ONE: an open retinal image database for optic nerve
imbalance in dataset when compared with the performance of image evaluation, Int. Symp. Comput.-Based Med. Syst. 1 (2011) 1–6.
classification in DB2 and DB3 which has 0.80 accuracy. Images when [20] S. Phan, S. Satoh, Y. Yoda, K. Kashiwagi, T. Oshika, Evaluation of deep
subjected to filter combination of Edge Histogram filter and Fuzzy color convolutional neural networks for glaucoma detection, Jpn. J. Ophthalmol. 63
(Feb. 2019) 276283.
and texture histogram filter or even if edge histogram filter is applied [21] M. Al Ghamdi, M. Li, M. Abdel-Mottaleb, M.A. Shousha, Semisupervised transfer
individually then the Random forest classifier yields a better perfor­ learning for convolutional neural networks for glaucoma detection, in: Proc. IEEE
mance. The last two rows in Table 4 clearly demonstrate the improve­ Int. Conf. Acoust., Speech Signal Process., ICASSP), May 2019, 38123816.
[22] R. Asaoka, M. Tanito, N. Shibata, K. Mitsuhashi, K. Nakahara, Y. Fujino,
ment of the filters the author proposed. In future the research may be
M. Matsuura, H. Murata, K. Tokumo, andY. Kiuchi, Validation of a deep learning
extended for other combinations of filters with reduced feature. model to screen for glaucoma using images from different fundus cameras and data
augmentation,’’ Ophthalmol, Glaucoma 2 (4) (Jul. 2019) 224231.
[23] G. An, K. Omodaka, K. Hashimoto, S. Tsuda, Y. Shiga, N. Takada, T. Kikawa,
Credit author statement
H. Yokota, M. Akiba, T. Nakazawa, Glaucoma diagnosis with machine learning
based on optical coherence tomography and color fundus images, J. Healthcare
K.Alice: Conceptualization of Proposed Methodology, N.Deepa Eng. 2019 (Feb. 2019) 19.
[24] A. Diaz-Pinto, S. Morales, V. Naranjo, T. Köhler, J.M. Mossi, A. Navea, CNNs for
Dataset Collection, Devi.T: Implementation and overall Supervision,
automatic glaucoma assessment using fundus images: an extensive validation,
BeenaRani B B: Implementation and Evaluation, N. Bharatha Devi: Biomed. Eng. Online 18 (1) (Dec. 2019) 29.
overall Supervision, V.Nagaraju: Additional Dataset Collection [25] A. Cerentinia, D. Welfera, M.C. d’Ornellasa, C.J.P. Haygertb, G.N. Dottob,
Automatic identication of glaucoma sing deep learning methods, in: Proc. 16th
World Congr. Med. Health Informat.Precision Healthcare through Informat.
Declaration of competing interest (MEDINFO), vol. 245, 2018, p. 318.
[26] J.I. Orlando, E. Prokofyeva, M. del Fresno, M.B. Blaschko, Convolutional neural
network transfer for automated glaucoma identication, Proc. 12th Int. Symp. Med.
The authors declare that they have no known competing financial Inf. Process. Anal. 10160 (2017) 101600U.
interests or personal relationships that could have appeared to influence [27] A.C. de Moura Lima, L.B. Maia, R.M.P. Pereira, G.B. Junior, J.D.S. de Almeida, A.
the work reported in this paper. C. de Paiva, Glaucoma diagnosis over eye fundus image through deep features, in:
Proc. 25th Int. Conf. Syst.,Signals Image Process, IWSSIP), Jun, 2018, p. 14.
[28] F. Li, Z. Wang, G. Qu, D. Song, Y. Yuan, Y. Xu, K. Gao, G. Luo, Z. Xiao, D.S.C. Lam,
Data availability H. Zhong, Y. Qiao, X. Zhang, Automatic differentiation of glaucoma visual eld from
non-glaucoma visual led using deep convolutional neural network, BMC Med.
Imag. 18 (1) (Dec. 2018) 35.
No data was used for the research described in the article. [29] J.J. Gómez-Valverde, A. Antón, G. Fatti, B. Liefers, A. Herranz, A. Santos, C.
I. Sánchez, M.J. Ledesma-Carbayo, Automatic glaucoma classification using color
References fundus images based on convolutional neural networks and transfer learning,
Biomed. Opt Express 10 (2) (2019), 892913.
[30] Muhammad Arsalan, Muhammad Owais, Aiding the diagnosis of diabetic and
[1] Abidin, D. Effects of image filters on various image datasets. Proceedings of the
hypertensive retinopathy using artificial intelligence-based semantic segmentation,
2019 5th International Conference on Computer and Technology Applications -
J. Clin. Med. 8 (1446) (2019) 1–27, https://doi.org/10.3390/jcm8091446.
ICCTA 2019. doi:10.1145/3323933.3324056.
[31] U.R. Acharya, H. Fujita, S. Bhat, U. Raghavendra, A. Gudigar, F. Molinari, A. Vi-
[2] K. Alice, N. Ramaraj, S.P. Rajagopalan, Rotation Invariant Image Authentication
jayananthan, K.H. Ng, Decision support system for fatty liver disease using gist
Using Haralick Features” , Multimedia Tools and Applications, Springer, March
descriptors extracted from ultrasound images, Inf. Fusion 29 (2016) 32–39.
2020, pp. 17211–17225.
[32] T. Vijayan, M. Sangeetha, A. Kumaravel, B. Karthik, Gabor filter and machine
[3] R. Balasubramani, V. Kannan, Efficient use of MPEG-7 color layout and edge
learning based diabetic retinopathy analysis and detection”, Microprocess.
histogram descriptors in CBIR systems, Global J. Comput. Sci. Technol. 9 (5)
Microsyst. (2020) https://doi.org/10.1016/j.micpro.2020.103353.
(2009) 157–163.
[4] M. Caudill, Neural Network Primer: Part I vol. 1989, AI Expert, 1989.

6
A. K et al. Measurement: Sensors 25 (2023) 100566

[33] Neha Gour, Pritee Khanna, Automated glaucoma detection using GIST and pyramid [36] U. Raghavendra, S.V. Bhandary, A. Gudigar, U.R. Acharya, Novel expert system for
histogram of oriented gradients, PHOG) descriptors” Pattern Recognition Letters 3 glaucoma identification using non-parametric spatial envelope energy spectrum
(May 22, 2019) 57. with fundus images, Biocybern. Biomed. Eng. 38 (1) (2018) 170–180.
[34] Syna Sreng, Noppadol Maneerat, Kazuhiko Hamamoto, Khin YadanarWin, Deep
learning for optic disc segmentation and glaucoma diagnosis on retinal images,
Appl. Sci. 10 (4916) (2020) 1–19.