Automatic Glaucoma Detection using Adaptive

Threshold based Technique in Fundus Image

Ayushi Agarwal, Shradha Gulia, Somal Chaudhary, Malay Kishore Dutta, Radim Burget, and Kamil Riha

is more economical and is frequently used for basic eye

Abstract—Glaucoma is a kind of ocular disorder that results examination.
in a damaged optic nerve which is responsible for transmitting The most common way [1] via which glaucoma can be
images to the brain. The conventional methods to detect detected is by calculating Cup-to-disc ratio (CDR). An
glaucoma like Optical Coherence Tomography (OCT) and important feature in detection of glaucoma is the optic disk
Heidelberg Retinal Tomography (HRT) are expensive and need [2] and to segment it accurately is the most crucial and vital
specialized manpower. A digital fundus image can be used to
step. One method [3] in which optic disk is segmented
identify glaucoma. This paper describes an efficient method to
analyze a computer-aided fundus image which can act as a automatically was using morphological processing and
diagnostic tool for detection of glaucoma. A technique based on circular Hough transform. Gaussian Vessel Detector and
histogram of the image is used to study some statistical features Tangent Information Transform [4] are used in detecting optic
of the image such as mean and standard deviation. A relationship disk centre. Detection of glaucoma is also done using texture
between them is established to find a threshold value for feature [5] within the images such as wavelet based energy
segmenting optic disc and optic cup. An adaptive threshold based features. Also, there has been some work done in automatic
method which is independent of image quality and invariant to segmentation [6] of optic disk and optic cup from fundus
noise is used to segment the optic disc, optic cup and the cup-to- images. Arturo Aquino et.al. [7] used morphological
disc ratio CDR which is used to screen glaucoma. The
experimental results obtained are compared with those of the
operations, edge detection and feature extraction technique for
ophthalmologist and are found to have a high accuracy. Also in detecting the optic disk boundary. Glaucoma produces
addition the proposed method is efficient having a low variation in shape, color, depth of optic disk which helps in its
computational cost. identification [8]. Methods using superpixel classification for
glaucoma screening have also been proposed by Cheng et al.
Keywords—Glaucoma, histogram, Medical imaging, [9]. Digital fundus images are analyzed in two phases [10]
optical cup to disc ratio, Threshold based segmentation
one of which is vessel detection and inpainting. The second
phase comprises of CDR calculation. Image inpainting and
I. INTRODUCTION active contour model have been used for correct identification
of optic disk and optic cup. Vessel Kinking technique [11] has
G LAUCOMA, being one of the severe eye diseases is a
reason for blindness in aged people universally. Almost
all the tests for glaucoma are time consuming and
been used for the detection of optic cup boundary with the
help of shifting multi-scale window interval to probe along the
vessels. Albeit some image processing techniques have been
require special skills. The devices used by ophthalmologists proposed, there is still a requirement for a more proficient and
like Heidelberg Retinal Tomography (HRT) and Optical less complex method that is independent of image quality for
Coherence Tomography (OCT) which consist of similar the screening of glaucoma.
information as the images captured by digital fundus camera The main contribution of this paper is Glaucoma detection
are expensive and are not easily available. Therefore digital where optic disk and optic cup are automatically localized
images are used for the screening of glaucoma that consume from digital fundus images using Image adaptive thresholding
less time, have accuracy and require no skilled force. As technique making the proposed method independent of image
compared to other complex devices the digital fundus camera quality and invariant to noise. In this method statistical
features such as mean and standard deviation are considered
Manuscript received February 5, 2015. This work was supported in part
and analyzed with respect to image histogram to track the
by the Grants from Department of Science and Technology, Gov. of India threshold value that depends upon a relationship between the
No. DST/TSG/ICT/2013/37. Also, the authors express their thankfulness to statistical features. Based on these threshold values, optic disk
Dr. S. C. Gupta, of Venu Eye Research Centre for his contribution in and optic cup are detected that makes the method less
Classification of the images. Research described in this paper was financed
by the National Sustainability Program under grant LO1401. International
complex and independent of the quality of the fundus image.
cooperation in frame of COST LD14091 was supported by Czech Ministry The extraneous information is removed after which CDR is
of Education under grant no. LD12012. For the research, infrastructure of the calculated for glaucoma detection.
SIX Center was used. Another significant contribution of this proposed work is
A. Agarwal, S. Gulia, S. Chaudhary and M. K. Dutta are with Amity
School of Engineering and Technology, Amity University, Uttar Pradesh,
that this threshold based method of glaucoma identification is
India.(e-mail: ayushi2011993@gmail.com, shradhagulia more robust and requires less computation cost without
@gmail.com,somalchaudhary30@gmail.com, mkdutta@amity.edu) compromising the accuracy of the results.
R. Burget and K. Riha are with the Brno University of Technology, Further the paper is organized as follows. Section II
faculty of Electrical Engineering, Czech Republic (e-mail:
describes the important features related to fundus image.
burgetrm@feec.vutbr.cz, rihak@feec.vutbr.cz).

978-1-4799-8498-5/15/$31.00 ©2015 IEEE 416 TSP 2015

Section III gives the description of proposed method that is A. Image Feature extraction
used for segmenting optic disc and optic cup from the retinal To make the method adaptive of image quality, local
image. Section IV gives the experimental results of cup to features such as mean and standard deviation are extracted
disc ratio. Finally section V concludes the paper. from it. Mean X and variance σ2 are obtained as:

X= 1/N Ʃ xi (1)
II. IMPORTANT FEATURES OF RETINAL IMAGE
The fundus image represents the interior surface of an eye σ2 = 1/N Ʃ (xi – X)2 (2)
and includes optic disk, optic cup, blood vessels, as its main
features. Fig.1 depicts colored fundus image consisting of Standard deviation is defined as square root of variance and
some of its main features like macula, blood vessels, optic given by
disk etc. Since Glaucoma affects the optic disk via changing
the cup to disk ratio therefore proper segmentation is crucial σ= √ σ2 (3)
for detection of glaucoma efficiently.
The flowchart of the proposed method is described below
in Fig 3. Each block is separately explained in this section.

Fig. 1: The colored RGB fundus image

III. PROPOSED METHOD

In this paper, RGB fundus image is used as an input. This
image consists of three channels namely red, green and blue.
Fig.2 shows these images in different components. It is seen
that in fig. 2a (Red image), it is difficult to differentiate
between optic disc and optic cup and in fig 2c (Blue image),
the over all contrast of the image is not high enough. Fig. 2c
(green image) provides best contrast and the boundary of
optic disk and optic cup are also differentiated easily.
Therefore, fig. 2b is chosen for further processing.
Fig. 3: Image processing system to calculate optic disc area and cup area.

B. Histogram analysis
Histogram of the green channel image is examined and it is
shown in Fig. 4. It is a graphical representation of number of
pixels with respect to gray levels in a digital image. The 8-bit
(a) (b) grayscale digital image consists of 255 gray levels. It
represents the probability distribution function of the image
as:

(4)

Where rk is the kth gray level, nk is the number of pixels

(c) corresponding to kth gray level and n is the total number of
Fig. 2: RGB components of Fundus image (a) Red channel (b) Green pixels in the grayscale image.
channel (c) Blue channel

417
 C. Adaptive Image Thresholding
The thresholding method is one of the robust and effective
method of segmenting objects from the image. After selecting
particular intensity value the image is thresholded in the
following way. If f(x,y) is the image and I is the threshold
(a) (b)
Fig. 4 (a) : The green channel fundus image (a) Sample 1(b) Sample 2.
value then the output image g( x,y ) is obtained as:

12
x 10
4

g (x,y) = { 1 if f(x,y) ≥ I
10
0 otherwise (7)
8
By using the above mentioned thresholding a value, the 8-
6
bit green channel image is converted to binary image to obtain
optic disk and optic cup. Some samples are presented in Fig.
4

5.
2

0
D. Cup to Disk ratio
0 50 100 150 200 250 300

(a)
12
x 10
4
The segmented optic disc and cup obtained in B contains
10
some amount of noise in the form of extraneous information
which needs to be removed for accurate results. This noise is
8

removed by selecting the maximum area from the segmented

6
images obtained previously for both optic disk and cup.
4
Total number of white pixels is calculated individually for
segmented optic disk and optic cup image and further cup to
2

disc ratio is obtained. If CDR is greater than or equal to 0.3

0
0 50 100 150 200 250 300
then image is classified as glaucomatous else image is
(b) glaucoma free.
Fig.4 :Histogram of green channel fundus images (a) Sample 1(b) Sample 2

The first peak in the histogram represents the black and IV. EXPERIMENTAL RESULTS
dark portion in the image. The middle region in the histogram The adaptive threshold based method used in this paper for
indicates grey portion and right side represents light and pure automatic segmentation of optic disk and optic cup to detect
white portions in the image. The optic disk and optic cup are glaucoma has been tested over a large database of around 110
segmented by providing a threshold value that has a high images. These samples are collected and labeled by doctors of
intensity in the histogram as it is observed from the fundus Venu Eye Research Centre, New Delhi. Fig. 5 gives some
image that the optic disk and cup are brighter regions of the sample results of threshold images of segmented optic disk
image. The histogram of each image is examined to determine and cup.
threshold level for segmenting optic disk and optic cup. From Table I represents statistical features, mean and standard
the statistical mathematics of histogram, it is known that the deviation, CDR, results computed by the proposed method
mean is the central tendency and the standard deviation is the and their ground truth. It was observed that the proposed
dispersion from that central value. The addition of mean and method is 90% accurate in which glaucoma was detected
standard deviation gives an intensity level that points to the correctly except for two images which were detected as
highest number of pixels in the grey region of the image. After glaucomatic but were normal.
analyzing the images it is determined experimentally that
optic disk lies at an intensity level which is given by:

Tdisc = 1.2 x sum(mean, standard deviation) (5)

This intensity level is selected as the threshold value to

obtain the optic disk. Since the optic cup is brighter than optic
disc, the threshold value for obtaining optic cup is greater
than threshold value of optic disc. The threshold value for
optic cup is also obtained in an experimental manner given
as:

Tcup= 1.25 x (thresholdoptic disk) + diff (mean, standard

deviation) (6)

418
 Table II represents the time taken to compute the proposed
method. The required computational time needed to segment
optic disk and optic cup using adaptive threshold based
technique and thereby calculating cup to disk ratio is very
less. Therefore, the proposed method can be considered to be
to be an effective development in the detection of glaucoma.
Table III gives a comparative study between different
methods and proposed method for detection of glaucoma. As
compared to other methods, the proposed method is better in
terms of both accuracy and speed.

(a) (b) (c)

Fig. 5: Results containing (a) Fundus images, (b) Segmented optic disk, (c)
Segmented optic cup

Table I: Table II:

STATISTICAL FEATURES, CDR, RESULTS AND THEIR GROUND TRUTH FOR COMPUTATION TIME REQUIRED FOR CDR CALCULATION BY PROPOSED
GLAUCOMA DETECTION BY PROPOSED METHOD
METHOD
Hardware: CPU@1.9 GHz, 4Gb RAM, 64 bit operating system, Software: Hardware: CPU@1.9 GHz, 4Gb RAM, 64 bit operating
MATLAB R2010b system, Software: MATLAB R2010b
Sam Mean SD CDR Exp. Ground Compa Sample Computational time (in
ple Result Truth red seconds)
Results
1 33.02 26.08 0.50 Glaucoma Glaucoma Yes Sample1 5.57
2 29.07 23.09 0.59 Glaucoma Glaucoma Yes Sample2 5.45
3 70.49 49.52 0.4 Glaucoma Glaucoma Yes Sample3 5.50
4 70.0 46.5 0.52 Glaucoma Glaucoma Yes Sample4 5.28
5 52.37 45.78 0.74 Glaucoma Glaucoma Yes Sample5 5.26
6 47.88 34.92 0.56 Glaucoma Glaucoma Yes Sample6 5.54
7 61.18 40.50 0.29 Normal Normal Yes Sample7 5.48
8 61.85 42.48 0.41 Glaucoma Glaucoma Yes Sample 8 5.47
9 77.4 49.81 0.33 Glaucoma Glaucoma Yes Sample 9 5.67
10 49.83 37.86 0.39 Glaucoma Normal No Sample 10 5.35
11 33.82 29.79 0.24 Normal Normal Yes Sample11 5.49
12 69.82 45.82 0.18 Normal Normal Yes Sample 12 5.76
13 50.9 40.77 0.42 Glaucoma Normal No Sample 13 5.45
14 76.18 56.04 0.18 Normal Normal Yes Sample 14 5.52
15 57.57 39.63 0.16 Normal Normal Yes Sample 15 5.31
16 67.51 52.38 0.29 Normal Normal Yes Sample 16 5.52
17 54.96 40.97 0.23 Normal Normal Yes Sample 17 5.62
18 69.82 45.82 0.18 Normal Normal Yes Sample 18 5.25
19 71.88 48.28 0.22 Normal Normal Yes Sample 19 5.37
20 54.86 49.05 0.73 Glaucoma Glaucoma Yes Sample 20 5.30

Table III:
A COMPARATIVE STUDY WITH DIFFERENT METHODS FOR GLAUCOMA DETECTION
Results
Reference Methodology Used Database Used Features Used Computation
Accuracy
Time (in secs)
Vertical CDR estimation
Region based active contour error -0.09/0.08
G D Joshi Region Based Active Contour Local database-
model, Hough transform, (mean/standard deviation) Not reported
et al [2] Model and r-bends information 138 images
Vessel Bend detection CDR area ratio
-0.12/0.10
Active shape Model, Edge
F Yin et al Circular Hough Transform and ORIGA-Light
detection, Circular Hough CDR error-0.10% Not reported
[6] optimum channel selection 650 images
Transform
Morphological, Edge detection MESSIDOR
A Aquino 86% accuracy in
and Feature Extraction database-1200 Circular Hough transform 5.69 s
et al [7] segmentation algorithm
Techniques images
A Hoover,
M Fuzzy convergence of blood Local database- Blood Vessel convergence,
89% accuracy Not reported
Goldbaum vessels 81 images Illumination equalization
[12]
Local Local image features such
Proposed
Adaptive Thresholding database-110 as mean, standard 90% accuracy 5.45 s
Method
images deviation, Cup-disk ratio

419
 [5] Dua S., Acharya U.R.,Chowriappa P.,Sree S.V., “Wavelet based energy
features for glaucomatous image classification” IEEE transaction on
V. CONCLUSION information technology in biomedicine, Vol. 16, no.1,pp 80-87,
January 2012.
The adaptive threshold based technique for the automated
[6] Fengshou Yin, Jiang Liu, Damon Wing Kee Wong, Ngam Meng Tan,
segmentation of optic disk and cup can be useful in the early Carol Cheung, Manibhaskaran, Tien Yin Wong “Automated
detection of glaucoma to curb permanent loss of vision. In this segmentation of optic disk and optic cup in fundus images for
paper an adaptive thresholding method is proposed which glaucoma diagnosis”, 25th international Symposium on computer based
medical system pp.1-6, June 2012.
incorporates image features like mean, variance and standard [7] Aquino A., Gegundez-Arias M.E, Marin D. “Detecting the optic disk
deviation to segment the optic disk and optic cup from the boundary in digital fundus images using morphological, edge detection
fundus image which makes the method independent of the and feature extraction techniques” IEEE transactions on medical
image quality and invariant to noises. imaging, Vol.29, pp1860-1869, November 2010.
The thresholding technique used in this paper is robust and [8] Li.H, Chutatape O. “A model based approach for automated feature
extraction in fundus images”, proceedings of 9th IEEE international
effective because it uses the image features like mean and conference on computer vision (ICCV’03),Vol.1,pp 394-399, October
standard deviation and does not depend on factors such as 2003.
image contrast, brightness, intensity of light etc. Extraneous [9] Cheng J., Liu J., Xu Y., Yin F., Damon Wing Kee Wong, Ngan Meng
and unwanted information is removed from the image after Tan, Tao D., Ching Yu Cheng, Tin Aung, Tien Yin Wong “Superpixel
classification based optic disk and optic cup segmentation for
segmentation which enhances the results. When tested on a glaucoma screening”, IEEE transactions on medical imaging, vol.32,
large data set of fundus images, the proposed method gives No. 6, pp 1019-1032, June 2013.
promising results over 90% accuracy. Future work may [10] Chih-Yin Ho, Tun-Wen Pai, Hau-Teng Chang, Hsin-Yi Chen “An
involve the revival of a meager number of pixels that are lost automatic fundus image analysis system for clinical diagnosis of
glaucoma”, IEEE, International conference on complex, intelligent and
from the optic cup and optic disk during the removal of noise software intensive systems (CISIS), pp 559-564,July 2011.
from the image. [11] Wong wing kee Damon, Jimmy Liu, Tan Ngan Meng, Yin fengshou,
Wong Tien Yin “Automatic detection of optic cup using vessel kinking
in digital retinal fundus images”, IEEE, ISBI 2012, pp 1647-1650,May
REFERENCES 2012.
[1] Inoue N., Yanashima K., Magatani K., Kurihara T. “Development of a [12] Hoover A., Goldbaum M., “Locating the optic nerve in a retinal image
simple diagnostic method for the glaucoma using ocular fundus using the fuzzy convergence of the blood vessels”,IEEE transactions on
pictures” proceedings of 2005 IEEE, Engineering in medicine and Medical Imaging, Vol.22, No.8, pp 951-958,August 2003.
biology 27th annual conference, shanghai, China, pp 3355-3358, [13] Nayak J., Acharya R., Bhatt P.S., Nakul Shetty,Lim T.C. “Automated
January 2006. diagnosis of Glaucoma using fundus images”, Springer science +
Business media, LLC 2008.
[2] Joshi G., Sivaswamy, Krishnadas S.R “Optic disk and cup
[14] Dutta M.K., Mourya A.K., Singh A., Parthasarathi M., Radim Burget
segmentation from monocular retinal images for glaucoma
& Kamil Riha, “Glaucoma Detection by Segmenting the Super Pixels
assessment”, IEEE transactions on medical imaging, vol.30, pp.1192-
from Fundus Colour Retinal Images” International Conference on
1205, June 2011.
Medical Imaging, m- Health & Emerging Communication System
[3] Shekhar S., Al-Nuaimy W., Nandi A.K. “Automated localization of
(MEDCOM 2014), pp. 86-90, November 2014.
retinal optic disk using Hough transform”, IEEE, ISBI, pp 1577-1580,
[15] Yadav D., Parthasarathi M., Dutta M.K., “Classification of Glaucoma
May 2008.
based on Texture Features using Neural networks”, 2014 Seventh
[4] Yang X., Hamaguchi S., Sun Y., Xiao S. “Detect of optic disk centre
based on Gaussian vessel detector and tangent information transform” International Conference on Contempory Computing (IC3), pp 109-
IEEE 2011, 4th international conference on biomedical engineering and 112, August 2014.
informatics (BMEI),Vol.1,pp 250-254, October 2011. [16] Singh A., Dutta M.K., Parthasarathi M., Burget R., Riha K., “ An
efficient automatic method of optic disk segmentation using region
growing techniques in retinal images”,pp 480-484, November 2014.

420