APPLIED PROBLEMS

Predictive Diagnosis of Glaucoma Based

on Analysis of Focal Notching along
the Neuro-Retinal Rim Using Machine Learning
Rishav Mukherjeea,*, Shamik Kundua, Kaushik Duttaa,
Anindya Sena, and Somnath Majumdarb
a Department of Electronics and Communication Engineering, Heritage Institute of Technology, Kolkata, India
b
Department of Ophthalmology, Apollo Gleneagles Hospital, Kolkata, India
* e-mail: rishavmukherjee17@gmail.com

Abstract—Automatic evaluation of the retinal fundus image is regarded as one of the most important future
tools for early detection and treatment of progressive eye diseases like glaucoma. Glaucoma leads to progres-
sive degeneration of vision which is characterized by shape deformation of the optic cup associated with focal
notching, wherein the degeneration of the blood vessels results in the formation of a notch along the neuroret-
inal rim. In this study, we have developed a methodology for automated prediction of glaucoma based on fea-
ture analysis of the focal notching along the neuroretinal rim and cup to disc ratio values. This procedure has
three phases: the first phase segments the optic disc and cup by suppressing the blood vessels with dynamic
thresholding; the second phase computes the neuroretinal rim width to detect the presence and direction of
notching by the conventional ISNT rule apart from calculating the cup-to-disc ratio from the color fundus
image (CFI); the third phase uses linear support vector based machine learning algorithm by integrating
extracted parameters as features for classification of CFIs into glaucomatous or normal. The algorithm out-
puts have been evaluated on a freely available database of 101 images, each marked with decision of five glau-
coma expert ophthalmologists, thereby returning an accuracy rate of 87.128%.

Keywords: glaucoma, optic-disc segmentation, cup segmentation, blood vessel suppression, cup-to-disc
ratio, focal notching, retinal fundus image, computer aided diagnoses, ISNT rule, machine learning, support
vector machine
DOI: 10.1134/S1054661819030155

1. INTRODUCTION undiagnosed until the late advanced stages. The anal-

Glaucoma has emerged as one of the leading causes
of blindness in the recent times, causing about 3 mil-
lion cases worldwide [1, 2]. Currently more than
60 million people are suffering from glaucoma and
this is likely to affect more than 80 million by 2020 as
projected by World Health Organization (WHO) [3].
Recent data indicates in India more than 11.2 million
people are affected by glaucoma causing 5.9% of the
country’s total blindness [4]. Glaucoma results from
the degeneration of optical nerve fibres over a period
of time, leading to the structural change of optic nerve
head (ONH) and subsequently causing permanent
visual field loss [5].
Glaucoma can be broadly classified into two
types—open angle glaucoma and angle closure glau-
coma. Open angle glaucoma and few cases with some
components of angle closure is asymptomatic in it’s
early phases making glaucoma detection and treat-
ment difficult, hence most cases go undetected and
Received November 14, 2018; revised April 11, 2019;
accepted April 17, 2019
ysis of CFI is emerging as a preferred modality of glau-
coma assessment due to it’s non-invasive and eco-
nomical nature, making it a potential device for large
scale mass screening programs. Though research is
done on detection of glaucoma from the multi-modal
images obtained from HRT, the HRT is very costly
hence impractical for primary health care centers. The
aim of our study is to develop an automatic computer
aided diagnosis (CAD) algorithm to analyze the reti-
nal fundus image which computes the cup-to-disc
ratio (CDR) and the notching factor along with the
direction of notching. Then these parameters are inte-
grated such that an intelligent decision is made about
the existence of glaucoma, which can be used effec-
tively for primary mass screening.
The qualitative and quantitative analysis of the
shape deformation of the optic disc is used to predict
the progress of glaucoma. The optic disc is divided
into two parts, i.e., (i) the central bright depression
called the cup and (ii) the peripheral region where the
nerve fibres bend into the cup region called the neu-
roretinal rim, as shown in Fig. 1. Larger discs have
large neuroretinal rim areas. These two are positively

ISSN 1054-6618, Pattern Recognition and Image Analysis, 2019, Vol. 29, No. 3, pp. 523–532. © Pleiades Publishing, Ltd., 2019.
524
Nauroretinal rim
Main blood vessel
Optics disc
Pallor region
Cup
Fig. 1. An optic disc centric 2-D retinal image.
correlated. The contour of the cup is influenced by
this correlation. Along with this, intra ocular pressure,
retinal nerve fiber layer (RNFL) characteristics, focal
notching, neuroretinal rim width and peri-papillary
atrophy (PPA) serves as important indicators for pro-
gression of glaucoma. The neural rim area appears to
decline with age and increase in intraocular pressure.
As a note of caution, it has been observed that patients
with diabetes mellitus may have an increase in the
neural rim over time, which could be secondary to
nerve swelling.
Our algorithm uses cup-to-disc ratio and neuroret-
inal rim width based ISNT rule as classification
parameters for detection of glaucoma. The most com-
mon clinical indicator used by ophthalmologists is the
cup-to-disc ratio found out by the visual assessment of
the optic disc through ophthalmoscopy. Higher CDR
values signify greater chances of glaucoma. However
the cup-to-disc ratio often fails to address problems
with genetically large optic cup cases and myopic eye
(where the optic cup is inherently large). As a result of
this fallacy, we additionally introduce notching estab-
lished by the ISNT rule which is a technique used to
measure the width of the neuroretinal rim.
ISNT stands for the four regions in which the opti-
cal nerve head can be divided, as shown in Fig. 2.
• I = Inferior, i.e., the bottommost region.
• S = Superior, i.e., the topmost region.
• N = Nasal, i.e., the near nose region.
• T = Temporal, i.e., the opposite of nasal region.
For a normal eye, the inferior region of the neu-
roretinal rim has the maximum width, followed by the
superior, nasal and temporal region, following the
order
I > S > N > T.
Any retinal image which violates this order is sub-
jected to scrutiny, as the probability of glaucoma
PATTERN RECOGNITION AND IMAGE ANALYSIS
MUKHERJEE et al.
S
T N
I
Fig. 2. ISNT representation for a right eye.
increases due to abnormal elongation of cup towards
inferior (I) or superior (S) direction. The difference
between CDR and ISNT rule is the area of focus.
While the CDR focuses on the optic cup size with
respect to the disc, the ISNT focuses on the neuroret-
inal rim width i.e., the area between the boundary of
optic cup and disc. Our work is focused in developing
an automated pre-screening system to analyze the CFI
to compute the classification parameters, which on
integration with machine learning algorithm will help
in providing an insight to the ophthalmologists to clas-
sify CFIs as glaucomatous and normal.
This paper proposes a pixel based feature classifi-
cation which analyzes multiple pixel features including
numeric properties of a pixel and it’s surrounding. Our
proposed algorithm has two phases: a training phase
and a test phase. During the training phase, the algo-
rithm learns to classify pixels from known classifica-
tions, and in the test phase the algorithm is tested on
unknown images using machine learning which uti-
lizes classification factors as parameters. The optic
disc is segmented using a statistical model, where the
disc pixels are separated based on their intensity val-
ues. Repetitive edge detection and correction tech-
nique is employed to suppress the blood vessels. Cup
segmentation is based on average distance vector cal-
culation on the masked disc region with respect to the
maximum energy pixels, for intensities greater than
the RMS value of the pixels. The segmentation of
optic disc and cup is the primary requirement in
extracting the different parameters from the fundus
images. The segmentation process has been elabo-
rately discussed in our previous work [6]. The ISNT
rule is used to determine the presence and direction of
the focal notch from the width of the neuroretinal rim.
The cup-to-disc ratio and the notching characteristics
are the learning parameters which are integrated with
Vol. 29 No. 3 2019
 PREDICTIVE DIAGNOSIS OF GLAUCOMA BASED
the machine learning algorithm to classify the CFIs for
prediction of glaucoma.
The paper is organized as follows. Section 2 reviews
previous related works on retinal feature extraction
from CFIs and classification of the same into glauco-
matous and normal. Section 3 presents our proposed
methods of segmentation followed by extraction of
different retinal image parameters and the classifica-
tion method based on them. In Section 4, the experi-
ment results and the segmentation and classification
accuracy is presented. Section 5 presents the discus-
sions of our works and the conclusion and future
scopes of the paper are presented in final section.
2. BACKGROUND
The optic disc segmentation is the primary step for
automatic processing of the fundus image, hence a
considerable amount of research work is available. The
earliest attempt of optic disc segmentation was made
by shape based template matching in which the disc is
modeled as either circular [7–9] or elliptical [10]. This
method was unable to characterize the irregularities in
shape arising from pathological changes. For this
problem, the approach has shifted to thresholding
based techniques. Nagakawa et al. [11] developed a P-
tile thresholding technique value in gray scale images.
Gradient based active contour model are developed in
which Mendels [12] approaches contour initialization
manually while Lee [13] approaches it automatically.
Recent works on active contours are more focused on
region based approaches [14, 15] inspired by Mum-
ford-Shah model [16]. The disc segmentation has
been succeeded by blood vessel extraction and removal
[11], because the presence of blood vessels makes
detection of optic cup difficult. A fuzzy algorithm was
developed by Tolias [17] to identify the blood vessel
regions. Chaudhari et al. [18] presented a method of
detection of blood vessels by designing 12 different
templates intended to be used in searching for blood
vessel segments in all possible directions. Vermeer et
al. [19] employed a thresholding segmentation based
on Laplace transform to segment retinal vessels.
The optic cup segmentation poses more challenge
than disc segmentation because of the decreased visi-
bility of the boundary and high density of blood vessels
in the boundary of optic cup and disc region. The oph-
thalmologists mainly use three visual aids to detect the
cup, i.e., (i) change in color, that is, the cup appears to
be much paler as compared to neural rim, (ii) bend in
small blood vessels, and (iii) notching of inner margin
of neuroretinal rim. Compared to the disc segmenta-
tion very few methods have been proposed for the cup
segmentation. Liu et al. [20] proposed a method in
which a potential set of pixels from the cup region is
derived based on the reference color obtained from a
manually selected point. Then an ellipse fit to this set
of pixels to obtain cup boundary, but the cup boundary
thus obtained is coarse. A variant of this method
PATTERN RECOGNITION AND IMAGE ANALYSIS
525
obtained the cup pixels via the thresholding of green
color plane.
The small bends in the blood vessels, i.e., the kinks,
automatically mark the cup boundary. Wong [21] et al.
provided a way for detecting the cup boundary [22]
from the kinks in blood vessels by extracting image
patches from the estimated cup boundary obtained in
[20] and the vessel pixels are identified using the edge
and wavelet transform information. The vessel bends
identified by points of direction change in vessel pixels
are detected and used to obtain the cup boundary.
Edge detection and hough transform combined with
statistical deformable model can be used to extract
optic cup from vessel removed disc [23]. Gopal et al.
[24] proposed a method of cup segmentation by using
r-bends information and by producing a local interpo-
lating spline to naturally approximate cup boundaries
in region where r-bends are absent. The ISNT rule is
widely used by the ophthalmologists to measure the
neuroretinal rim width and deduce the diagnosis and
progression of glaucoma. Tan et al. [25] proposed an
algorithm to classify notching based on ISNT rule and
also detect the feature points from the vessel bends and
local image gradients.
3. MATERIALS AND METHODS
3.1. Data
Proposed algorithm has been trained and tested on
a free database of colored fundus images provided by
Arvind Eye Hospital in Madurai, India in collabora-
tion with IIIT Hyderabad [26]. The dataset comprises
of both normal and glaucomatous images that are of a
single demographic population, as all the subjects
whose eye images are a part of this database are all
Indians. In this dataset, the images are classified pri-
marily by four along with an additional expert oph-
thalmologist. The glaucomatous images are denoted
by +1 and the normal images as–1. The final decision
was taken on the basis of majority count.
3.2. Preprocessing
The preprocessing involves three steps: disc seg-
mentation, blood vessel detection, and removal and
cup segmentation.
3.2.1. Optic disc segmentation. The optic disc
region appears brighter than the other parts of the ret-
ina in the retinal fundus image and can be segmented
with a threshold obtained from a selectively chosen
histogram of single row profile across the fundus disc
image. Our algorithm is based on statistical RGB
model using pixel based feature classification by
examining relevant pixels and their surrounding,
which is used in developing an automatic threshold
based on image characteristics. The row with maxi-
mum mean pixel intensity value Rmax will pass through
the optic disc.
Vol. 29 No. 3 2019
526 MUKHERJEE et al.

(a) (b) (c) (a) (b) (c)

Fig. 3. (a) Binary disc mask, (b) red channel for disc, (c) Fig. 4. (a) Blood vessel profile in RGB, (b) blood vessel
disc mask in RGB suppressed in red layer, (c) blood vessel suppressed disc
mask.

⎡ n X (i, j )⎤ For kth iteration µlocal is calculated for γ × γ neigh-

Rmax = max im=1 ⎢∑
⎣ j =1 n ⎦
⎥, (1) borhood as follows,
M + γ 2−1 N + γ 2−1
where Rmax returns the maximum average pixel value
of the image with dimensions m × n, and X(i, j)
∑ ∑ X (i, j, k − 1)
i = M − γ 2−1 i = N − γ 2−1
denotes the pixel intensity value for the coordinates μlocal = (4)
(i, j). The pixel intensities that are greater than Rmax are γ2
then stored in another array Y . The dynamic row ∀1 ≤ M ≤ m and 1 ≤ N ≤ n,
threshold (Rth) is computed from the set of pixels
stored in array Y as, where M, N are the coordinates of the edge pixels.
The pixel values expected to lie in the cup region
n(Y ) have pixel intensity greater than µlocal and are extracted
∑Y k in the set X.
Rth = k =1
. (2)
n(Y ) Y = {y|y = X (i, j, k − 1), X (i, j, k − 1) ≥ μlocal }.
The Rth acts as RGB threshold for row specific The estimated cup pixel value µlocal for this neigh-
operation, similar process is used to calculate the borhood is calculated from set Y as,
threshold for the column specific operation (Cth). The n(Y )
mean of row and column threshold is computed to
obtain the DiscThreshold. The disc is most prominent
μlocal = ∑ Yn(Y(i)) ∀1 ≤ k ≤ L.
i =1
(5)
when determined in the red layer of the image, hence
it is used over green and blue layer for disc segmenta- The µlocal value is assigned to the X (M , N ) pixel
tion. and this process is iterated L times in all of the RGB
layers. This enables the blood vessels to be suppressed,
⎧ 1 if X (i, j ) ≥ DiscThreshold as represented in Fig. 4c and the corresponding red
PBWD = ⎨ , (3) channel is shown in Fig. 4b.
⎩0 if X (i, j ) < DiscThreshold 3.2.3. Optic cup segmentation. The blood vessel
where PBWD represents the binary image after imple- suppressed disc mask serves as the precursor image for
menting auto-thresholding method which is used for optic cup segmentation. The maximum valued pixel of
further analysis. Figure 3a constitutes the optic disc the mask which lies in the cup region is obtained and
mask, which when integrated on the original image the root mean square values of all the pixel intensities
results in Fig. 3c and corresponding red layer is shown in the disc are calculated.
in Fig. 3b. n( x )
3.2.2. Blood vessel detection and suppression. The ∑ X (i) 2

masked disc shows that irregularities appear on the Prms = i =1

optic cup periphery due to the presence of blood ves-
sels which makes cup segmentation difficult, hence
they are suppressed by morphological approach. The
primary step involves the detection of the edges of the
blood vessels by implementing sobel edge detection
Algorithm [27] on the green channel of the masked
image, as shown in Fig. 4a. A γ × γ neighborhood is
extracted around that pixel and the mean of this
neighborhood is found out.
. (6)
n( X )
The pixel intensity values greater than the RMS
value are subtracted from the maximum valued pixel
and stored in vector Y.
Y = {y|y = Pmax − X (i )∀X (i ) ≥ Prms }.
The mean µdiff is calculated for vector Y to obtain
the threshold for cup segmentation i.e., CupThreshold.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 29 No. 3 2019

 PREDICTIVE DIAGNOSIS OF GLAUCOMA BASED
n(Y )
∑ Y (i)
μdiff = i =1
, (7)
n(Y )
CupThreshold = Pmax − μdiff . (8)
The thresholding techniques gives best results in
red and green layer of the image for cup extraction,
hence the logical union of these components are used
to obtain the cup pixels PBWC from the disc mask. Fig-
ure 5a represents the binary image of the optic cup and
the corresponding cup in RGB is shown in Fig. 5b.
⎧ 1 if X (i, j ) ≥ CupThreshold
PBWC = ⎨ (9)
⎩0 if X (i, j ) < CupThreshold.
3.3. Feature Extraction
A collection of features are used for the classifica-
tion of the retinal images into glaucomatous and non-
glaucomatous.
3.3.1. Cup-to-disc ratio calculation. The segmenta-
tion of the optic cup and the optic disc is followed by
regional property analysis of the segmented binary
images. The two primary features that are used for
classification of retinal images for Glaucoma detec-
tion are: area cup-to-disc ratio (ACDR) and diameter
cup-to-disc ratio (DCDR). The area of the disc and
the cup are calculated by considering maximum num-
ber of connected components. The ACDR is com-
puted by the ratio of the area of the optic cup with
respect to the optic disc. Similarly, the DCDR is com-
puted by the ratio of the length of the major axis of the
optic cup with respect to optic disc. The integrated
result of ACDR and DCDR is used as a decision mak-
ing parameter for the primary classification of the ret-
inal images.
3.3.2. Notching detection from neuroretinal rim
width. After calculating the ACDR and DCDR values
for all the images in the database, they are further ana-
lyzed for better prediction results, by detecting and
predicting the presence of notch in the optic cup
boundary and the direction of the notching, caused by
the irregular distribution of the optic nerve fibers fall-
ing into the cup.
Selective loss of neuroretinal rim tissue in glau-
coma occurs primarily in the infero-temporal region
(more so inferiorly), and to a lesser extent in supero-
temporal region (more so superiorly). As glaucoma
progresses, the temporal neural rim is typically
involved after the vertical poles and the nasal quadrant
is the last to be destroyed.
The ISNT rule (I > S > N > T) [28] has been the
basis of the detection of this notching feature, where
any abnormalities in this pre-established hypothesis
leads to the identification of the notching criterion.
The notching feature extraction is done in a three-
step process, expunging out the different cases in each
PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 29
527
(a) (b)
Fig. 5. (a) Binary mask of segmented cup, (b) Cup marked
in RGB.
step. The major axis and the minor axis lengths of the
cup are computed. Their difference length is calcu-
lated which gives us an estimation of the extent of
deformation of the optic cup boundary. This is termed
as notch factor which yields low values for normal
optic cups, and considerably high values for cups
which have been deformed due to notching.
In the next step, the orientation of the cup with
respect to the optic disc is considered for the detection
of notching. The detailed manifestation of the ISNT
rule depicts that a normal optic cup with no notch is
somewhat elliptical in the horizontal direction and a
cup with a notch and violating the aforesaid rule
becomes elongated in the vertical or the inferior-supe-
rior direction. Thus, it has been empirically concluded
by the ophthalmologists that the effect of notching is
initially and vividly visible in the I and S directions,
followed by the effects in the temporal quadrant and
much later in the nasal neural rim. If a notch is pres-
ent, it has to be present initially in I or S direction.
In order to find out the orientation of the cup, the
image is divided into four quadrants from the centroid
of the optic cup as shown in the Fig. 6. The quadrant
with angle (90° ± 45°) forms the superior quadrant
and the one with angle (90° ± 45°) forms the inferior
quadrant.
If the major axis of the optic cup, which must
include the notch, forms an angle (with respect to the
horizontal) lying within the inferior or superior quad-
rants, the presence of the notch is ascertained. The
images where the notch is present are extracted and
then processed to find the direction of the notch. In
the final step, to estimate the direction of the notch
either in I or S direction, the major axis of the optic
cup carrying the notch is further extended in either
sides to meet the optic disc boundary as shown in
Fig. 7a. The distances from the optic cup boundary to
the optic disc boundary along both the directions of
the major axis is calculated for all the dataset images.
The distance from the cup boundary to the disc
boundary along the superior direction is termed as S-
distance and the same along the Inferior direction is
termed as I-distance, represented in Fig. 7b.
No. 3 2019
528
S
T N
I
Fig. 6. The quadrant division of ISNT for a right eye.
In CFIs, where the I-distance value is considerably
less than the S-distance one, it can be concluded that
the notching is in the I-direction and vice versa, thus
giving rise to a parameter called notch prediction
which represents the predicted directions of the notch
present in the neuroretinal rim. In this way, the pres-
ence of the notch in the optic cup and its direction is
found out by processing the retinal images, which cor-
roborates the presence of glaucoma.
3.4. Decision Making with Machine Learning
Machine learning is a well known prediction sys-
tem which learns the characteristics from a training
dataset and implements its learning for prediction the
results for the test dataset. The resultant model from
the machine learning algorithm can work inde-
pendently on any unclassified image. The features
extracted from the segmented images are used to train
the classification model. In this paper we have used
support vector machine (SVM) [29] for binary classi-
fication between glaucomatous and nonglaucoma-
tous. Though SVM algorithm takes greater time to
train from a training data set and has a fairly high
memory usage, it comes with a very fast prediction
speed especially for a binary classified data. Moreover
the training data set being small it takes less time to
train with high prediction accuracy when SVM is used.
SVM has several kinds of kernel functions with dif-
ferent characteristics. The input features and the time
required for the training process serve as important
factors for selecting kernel functions. Linear kernel
function is used for our screening system, as linear
SVM has the lowest training time and memory usage
when compared to other kernel functions without
much variations in results. This algorithm classifies
the images in the database on the basis of the features
extracted above. Firstly, it classifies the images based
on the two input features area CDR and diameter
PATTERN RECOGNITION AND IMAGE ANALYSIS
MUKHERJEE et al.
(a) (b)
S-distance
I-distance
Fig. 7. The S-distance and I-distance represented in (a)
binary image and (b) RGB image.
CDR. In the second step, the six input features used
for classification are area CDR, diameter CDR, notch
factor, S-distance, I-distance and notch prediction
(using ISNT rule). The generated classifier is verified
by a 5-fold cross validation. Randomly selected data
from the data set is used as test data to verify the clas-
sifier and the results are verified with the 5 expert
markings in the data set. A prediction which is match-
ing with the at least 2 of the 5 experts are considered as
true because the expert decisions show large variations
in certain CFI’s across the dataset.
4. RESULTS
4.1. Segmentation
Five colored fundus images are selected from the
dataset arbitrarily and the segmentation results for the
different preprocessing steps are shown in Fig. 8. In
the 4 × 5 matrix provided below, the first column rep-
resents the original images as provided in the dataset.
The second column consists of the segmented optic
disc mask in RGB. The third column represents the
blood vessel suppressed optic disc mask that has been
used for cup segmentation in the final step. The fourth
column is formed by images, where the disc is demar-
cated by the algorithm in blue and the optic cup in
green. The rows in the matrix represent the successive
preprocessing steps for the corresponding images.
The segmentation results are validated against the
ground truths of the experts provided in the database.
The experts have marked their estimated areas of cup
and disc and those were used as a standard to validate
our segmentation results, as shown in Figs. 9 and 10.
The area of the retina that has been marked as optic
disc, both by the ophthalmologists and by our method
is denoted as true positive (TP). The area that has been
left out by the experts as well as by our method are
termed as True Negative (TN). The area that has been
identified as optic disc by our method but not by the
experts is denoted as false positive (FP). The area that
falls under the disc region according to the experts but
is undetected by our method is represented by False
Negative (FN). The corresponding TP, TN, FP, FN
Vol. 29 No. 3 2019
 PREDICTIVE DIAGNOSIS OF GLAUCOMA BASED 529

values for the optic cup have been found in a similar

way as that of the disc.
Thus, a quantitative measure of the segmentation
accuracy of both the disc and the cup has been found
out by using those above mentioned terms, and repre-
sented by the parameters Precision Factor, Rectal
Factor, and Fscore, where
Expert Expert

Precision Factor = TP , Proposed Proposed

TP + FP
TP Fig. 8. Qualitative measure of segmentation accuracy in
Recall Factor = , optic disc.
TP + FN

Fscore = 2 Precision × Recall .

Precision + Recall
The parameters for a five images are represented in
Table 1. The mean, range and standard deviation for
each parameter is given below each column of the
table.

4.2. Classification Expert Expert

The SVM based machine learning algorithm is Proposed Proposed
used for the classification of the CFIs into glaucoma-
tous and normal. The prediction accuracy of our algo-
Fig. 9. Qualitative measure of segmentation accuracy in
rithm has been represented by a confusion matrix and optic cup.
validated by the calculation of specificity, sensitivity,
precision factor and negative predictive value. In the
confusion matrix, the horizontal axis represents true
condition and the vertical axis represents predicted
condition. Image no. Original image Segmented disc Vessel suppressed Final markings

In Tables 2 and 3, TP stands for true positive, TN

stands for true negative, FP stands for false positive 5
and FN for false negative. The mathematical expres-
sion for the parameters, depicted in the tables, is
shown below.

Precision Factor (PREC) = TP , 58

TP + FP

Negative Predictive Value (NPV) = TN ,

TN + FN 72

Specificity (SPEC)= TN ,
TN + FP

Sensitivity (SENS)= TP , 82

TP + FN
TP + TN
Accuracy (ACC) = .
TP + FP + TN + FN 97

Table 2 represents the confusion matrix where the

dataset is classified by using two input parameters:
Area CDR and diameter CDR, achieving sensitivity of
76.54% and specificity of 65%, with an accuracy of
74.25%. Table 3 represents the confusion matrix where
the dataset is classified by using six input parameters:
Fig. 10. First column: original image; second column: seg-
mented optic disc in RGB; third column: blood vessel sup-
pressed optic disc; fourth column: marked optic cup and
disc (cup marked in green and disc marked in blue).

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 29 No. 3 2019

530 MUKHERJEE et al.

Table 1. Precision factor, recall factor and Fscore of the images shown in Fig. 8
Image Disc Cup
number precision factor recall factor Fscore precision factor recall factor Fscore
05 0.9556 0.9904 0.9727 0.9306 0.9359 0.9332
58 0.8799 0.9960 0.9343 0.8492 0.9349 0.8900
72 0.9728 0.9726 0.9727 0.9613 0.6988 0.8099
82 0.9162 0.9983 0.9555 0.9069 0.7366 0.8129
97 0.9495 0.9548 0.9521 0.9662 0.6726 0.7931
Mean 0.9348 0.9824 0.9575 0.9228 0.7957 0.8478
Range 0.0929 0.0435 0.0384 0.1170 0.2637 0.1401
Std. Dev. 0.0369 0.0184 0.0161 0.0477 0.1295 0.0607

area CDR, diameter CDR, notch factor, S-distance, eters along with the traditional CDR parameters helps
I-distance and notch prediction, achieving sensitivity to improve the accuracy of the algorithm for detection
of 86.42% and specificity of 90%, with an accuracy of of glaucoma.
87.128%. The integration of the focal notching param-
5. DISCUSSIONS
Table 2. Confusion matrix representation for images with
only CDR The dataset contains certain images where the
True
glaucoma phenomenon is not so vivid, and the deci-
sion of whether that CFI can be called as glaucoma-
condition
tous or normal varies among the experts. In such
total cases, the algorithms decision is validated against the
glaucoma normal opinions of at least two of the five expert ophthalmol-
population
ogists, so that the algorithms classification can fetch a
Predicted Predicted TP FP PREC six opinion going in line with those two varying expert
condition condition 62 7 89.85% decisions. Thus, our algorithm successfully evaluates
Glaucoma 76.54% 35% 88 of 101 images present in the database, hereby etch-
ing out a cumulative success rate of 87.128%. Also,
Predicted FN TN NPV there are observer variations in evaluation of CDR.
condition 19 13 40.62% The difference between the estimated cup to disc ratio
Normal 23.46% 65% for the same eye at different times seldom exceeds 0.2.
SENS. SPEC. ACC Hence, any difference greater than 0.2 over time
should be viewed with suspicion.
76.54% 65% 74.25%
The technique developed in our work has the fol-
lowing advantages over the existing techniques. Pro-
Table 3. Confusion matrix representation for images with posed method considered all the RGB layers for the
CDR and notching processing of CFIs, instead of the vastly used grayscale
True approximation which involves loss of information.
condition Furthermore, we have applied an image property
based thresholding approach to correctly extract the
total retinal features. Other works have used shape based
glaucoma normal
population template matching technique which is mostly database
specific, and also image specific. Also, we have used
Predicted Predicted TP FP PREC linear SVM based machine learning algorithm which
condition condition 70 2 97.22% makes our method adaptive, instead of being database
Glaucoma 86.4% 10% specific.
Predicted FN TN NPV Finally, we have used a unique method of detecting
condition 11 18 62.06% a focal notch to corroborate the presence of glaucoma,
Normal 13.6% 90% along with the traditional CDR evaluation techniques.
SENS. SPEC. ACC
This addition of notching parameters increases our
algorithm classification accuracy, which is clearly
86.42% 90% 87.128% depicted by the Tables 2 and 3.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 29 No. 3 2019

 PREDICTIVE DIAGNOSIS OF GLAUCOMA BASED
6. CONCLUSIONS
In this paper, we aimed at designing a predictive
analysis for estimating glaucoma by an intelligent
decision making algorithm. Although our method is 11.
not full proof, incorporating a few more classifying
parameters like age, sex, race of the patient, etc.,
which are not available in our current database, can
help in better prediction of glaucoma. As glaucoma is
a progressive phenomenon, a database that consists of
the fundus images for a patient over a period of time
will surely help to predict better, whether the case is 12.
tending towards glaucomatous or not. We also intend
to extend our work on practical field by using a porta-
ble fundus camera for easy collection of CFIs, so that
this technology can be established as a key step
towards primary assessment of glaucoma. 13.
REFERENCES
1. H. A. Quigley and A. T. Broman, “The number of peo- 14.
ple with glaucoma worldwide in 2010 and 2020,” Br. J.
Ophthalmol. 90 (3), 262–267 (2006).
2. S. Resnikoff, D. Pascolini, D. Etya’Ale, I. Kocur,
R. Pararajasegaram, G. P. Pokharel, and S. P. Mariotti,
“Global data on visual impairment in the year 2002,”
Bull. World Health Org. 82 (11), 844–851 (2004). 15.
3. Global Initiative for the Elimination of Avoidable Blind-
ness (World Health Organization, Geneva, 2000);
WHO/PBL/97.61 Rev2. 16.
4. R. George, R. S. Ve, and L. Vijaya, “Glaucoma in In-
dia: estimated burden of disease,” J. Glaucoma 19 (6),
391–397 (2010).
5. G. Michelson, J. Hornegger, S. Warntges, and B. Lau- 17.
sen, “The papilla as screening parameter for early diag-
nosis of glaucoma,” Dtsch. Aerztebl. Int. 105 (34–35),
583–589 (2008).
18.
6. K. Dutta, R. Mukherjee, S. Kundu, T. Biswas, and
A. Sen, “Automatic evaluation and predictive analysis
of optic nerve head for the detection of glaucoma,” in
Proc. 2018 2nd Int. Conf. on Electronics, Materials Engi-
neering & Nano-Technology (IEMENTech) (IEEE, Kol- 19.
kata, 2018), pp. 1–7.
7. R. A. Abdel-Ghafar and T. Morris, “Progress towards
automated detection and characterization of the optic
disc in glaucoma and diabetic retinopathy,” Med. Inf. 20.
Internet Med. 32 (1), 19–25 (2007).
8. R. Chrástek, M. Wolf, K. Donath, G. Michelson, and
H. Niemann, “Optic disc segmentation in retinal imag-
es,” in Bildverarbeitung fur die Medizin 2002, Ed. by
M. Meiler, D. Saupe, F. Kruggel, H. Handels, and
T. M. Lehmann, (Springer-Verlag, Berlin, 2002),
pp. 263–266.
9. M. Lalonde, M. Beaulieu, and L. Gagnon, “Fast and 21.
robust optic disc detection using pyramidal decomposi-
tion and Hausdorff-based template matching,” IEEE
Trans. Med. Imaging 20 (11), 1193–1200 (2001).
10. P. M. D. S. Pallawala, W. Hsu, M. L. Lee, and 22.
K.-G. A. Eong, “Automated optic disc localization and
contour detection using ellipse fitting and wavelet
transform,” in Proc. 8th European Conference on Com-
puter Vision “Computer Vision – ECCV 2004,” Ed. by
PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 29
531
T. Pajdla and J. Matas, Lecture Notes in Computer Sci-
ence Series (Springer-Verlag, Berlin, 2004),
Vol. 3022, Part 2, pp. 139–151.
T. Nakagawa, Y. Hayashi, Y. Hatanaka, A. Aoyama,
Y. Mizukusa, A. Fujita, M. Kakogawa, T. Hara, H. Fu-
jita, and T. Yamamoto, “Recognition of optic nerve
head using blood-vessel-erased image and its applica-
tion to production of simulated stereogram in comput-
er-aided diagnosis system for retinal images,” IEICE
Trans. Inf. Syst. J89-D (11), 2491–2501 (2006) [in Jap-
anese].
F. Mendels, C. Heneghan, and J.-F. Thiran, “Identifi-
cation of the optic disk boundary in retinal images using
active contours,” in Proc. Irish Machine Vision and Im-
age Processing Conf. (IMVIP) 1999, (IEEE, Piscataway,
1999), No. EPFLCONF-86621, pp. 103–115.
S. Lee and M. Brady, “Optic disk boundary detection,”
in Proc. British Machine Vision Conf. (BMVC 1991), Ed.
by P. Mowforth (Springer-Verlag, London, 1991),
pp. 359–362.
T. Chan and L. Vese, “An active contour model without
edges,” in Proc. 2nd Int. Conf. Scale-Space’99 “Scale-
Space Theories in Computer Vision,” Ed. by M. Nielsen,
P. Johansen, O. F. Olsen, and J. Weickert, Lecture
Notes in Computer Science (Springer-Verlag, Berlin,
1999), Vol. 1682, pp. 141–151.
T. F. Chan and L. A. Vese, “Active contours without
edges,” IEEE Trans. Image Process. 10 (2), 266–277
(2001).
D. Mumford and J. Shah, “Optimal approximations by
piecewise smooth functions and associated variational
problems,” Commun. Pure Appl. Math. 42 (5), 577–
685 (1989).
Y. A. Tolias and S. M. Panas, “A fuzzy vessel tracking
algorithm for retinal images based on fuzzy clustering,”
IEEE Trans. Med. Imaging 17 (2), 263–273 (1998).
S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and
M. Goldbaum, “Detection of blood vessels in retinal
images using two-dimensional matched filters,” IEEE
Trans. Med. Imaging 8 (3), 263–269 (1989).
K. A. Vermeer, F. M. Vos, H. G. Lemij, and
A. M. Vossepoel, “A model based method for retinal
blood vessel detection,” Comput. Biol. Med. 34 (3),
209–219 (2004).
J. Liu, F. S. Yin, D. W. K. Wong, Z. Zhang, N. M. Tan,
C. Cheung, M. Baskaran, T. Aung, and T. Y. Wong,
“A-levelset-based automatic cup-to-disc ratio mea-
surement for glaucoma diagnosis from fundus image,”
in Image Analysis and Modeling in Ophthalmology, Ed.
by E. Y. K. Ng, U. R. Acharya, J. S. Suri, and
A. Campilho (CRC Press, Boca Raton, 2014), pp. 129–
142.
D. W. K. Wong, J. Liu, J. H. Lim, H. Li, and T. Y. Wong,
“Automated detection of kinks from blood vessels for
optic cup segmentation in retinal images,” Proc. SPIE
7260, 72601J (2009).
G. D. Joshi, J. Sivaswamy, K. Karan, R. Prashanth, and
S. R. Krishnadas, “Vessel bend-based cup segmenta-
tion in retinal images,” in Proc. 2010 20th Int. Conf. on
Pattern Recognition (ICPR 2010), Istanbul (IEEE, Pis-
cataway, 2010), pp. 2536–2539.
No. 3 2019
532 MUKHERJEE et al.
23. F. Yin, J. Liu, D. W. K. Wong, N. M. Tan, C. Cheung,
M. Baskaran, T. Aung, and T. Y. Wong, “Automated
segmentation of optic disc and optic cup in fundus im-
ages for glaucoma diagnosis,” in Proc. 2012 25th Inter-
national Symposium on Computer-Based Medical Sys-
tems (CBMS), Rome, Italy (IEEE, Piscataway, 2012),
pp. 1–6.
24. G. D. Joshi, J. Sivaswamy, and S. Krishnadas, “Optic
disk and cup segmentation from monocular color reti-
nal images for glaucoma assessment,” IEEE Trans.
Med. Imaging 30 (6), 1192–1205 (2011).
25. M. H. Tan, Y. Sun, S. H. Ong, J. Liu, M. Baskaran,
T. Aung, and T. Y. Wong, “Automatic notch detection
in retinal images,” in Proc. 2013 IEEE 10th Internation-
al Symposium on Biomedical Imaging (ISBI), San Fran-
cisco, CA, USA (IEEE, Piscataway, 2013), pp. 1440–
1443.
26. J. Sivaswamy, S. R. Krishnadas, A. Chakravarty,
G. D. Joshi, Ujjwal, and T. A. Syed, “A comprehensive
retinal image dataset for the assessment of glaucoma
Rishav Mukherjee received the
B.Tech degree from Heritage Insti-
tute of Technology affiliated to
Maulana Abul Kalam Azad Univer-
sity of Technology in 2018 in Elec-
tronics and Communication Engi-
neering. He has successfully quali-
fied GATE, examination held by
IIT. He has been the joint first
author to a previous research paper
on Glaucoma detection from CFIs.
His research interests include image
processing, biomedical image analy-
sis, networking and cognitive radio networks (CRN).
Shamik Kundu received his
B.Tech degree in Electronics and
Communications Engineering
from Heritage Institute of Tech-
nology affiliated to Maulana
Abul Kalam Azad University of
Technology in 2018. He has
been the joint first author to a
previous research paper on
Glaucoma detection from CFIs.
His research interests include
signal and image processing,
communications, and networks.
PATTERN RECOGNITION AND IMAGE ANALYSIS
from the optic nerve head analysis,” JSM Biomed. Im-
aging Data Pap. 2 (1), 1004 (2015).
27. H. Scharr, PhD Thesis (Ruprecht-Karls-Universität,
Heidelberg, 2000).
28. N. Harizman, C. Oliveira, A. Chiang, C. Tello,
M. Marmor, R. Ritch, and J. M. Liebmann, “The
ISNT rule and differentiation of normal from glauco-
matous eyes,” Arch. Ophthalmol. 124 (11), 1579–1583
(2006).
29. M. A. Hearst, S. T. Dumais, E. Osuna, J. Platt, and
B. Scholkopf, “Support vector machines,” IEEE Intell.
Syst. Their Appl. 13 (4), 18–28 (1998).
Kaushik Dutta received his B.Tech
degree in Electronics and Commu-
nications Engineering from Heritage
Institute of Technology affiliated to
Maulana Abul Kalam Azad Univer-
sity of Technology in 2018. He has
been the joint first author to a previ-
ous research paper on Glaucoma
detection from CFIs. His research
interests include biomedical image
processing, computer vision, and
cognitive radio networks.
Anindya Sen is a Professor at the
department of Electronics and
Communication Engineering, Heri-
tage Institute of Technology, a pri-
vate autonomous engineering col-
lege in Anandapur, Kolkata, India.
He received his B.E. from Jadavpur
University, India in the year 1980,
PhD from University of Minnesota,
Twin Cities in 1996, and got his
Post-Doctoral training from Uni-
versity of Chicago from 1996 to
2000. He currently holds one US
patent and thirty research paper publications. His research
interests include, medical image processing, internet of
things, artificial intelligence and VLSI design.
Somnath Majumdar is a senior con-
sultant ophthalmologist in Kolkata,
India for last 20 years. He is cur-
rently a consultant ophthalmologist
for Apollo Hospitals, Kolkata and
Fortis Hospitals. He had done post
graduation from Dr. R.P. Centre for
Ophthalmic Sciences, AIIMS, New
Delhi and completed his FRCS
(Edin. and Glasgow) in 2000. He is
an expert in fields of Glaucoma,
Cataract, and Retinal surgery.
Vol. 29 No. 3 2019