Journal of Ambient Intelligence and Humanized Computing (2021) 12:3633–3645

https://doi.org/10.1007/s12652-019-01617-3

ORIGINAL RESEARCH

Exudate characterization to diagnose diabetic retinopathy using

generalized method
R. Valarmathi1 · S. Saravanan2

Received: 6 September 2019 / Accepted: 3 December 2019 / Published online: 19 December 2019
© Springer-Verlag GmbH Germany, part of Springer Nature 2019

Abstract
Globally in recent days, the potential risk for patients with diabetes mellitus is the prevalence of diabetic retinopathy, which
is a silent disease with no early symptoms and is the imperative cause of vision loss. An early diagnosis can be used to pre-
vent for visual loss and blindness. In the regular screening process, assistance of computerized diagnosis can considerably
minimize an ophthalmologists work and improve inter and intra viewer variability. A generalized method of semi automatic
exudates characterization to diagnose diabetic retinopathy with exudates screening system of retinal image is presented in
this paper. This system uses morphological processing based retinal blood vessel suppression, Semi automatic masking of
optic disc structure and morphological component analysis based texture enhancement followed by segmentation and Adap-
tive Neuro-Fuzzy Inference System (ANFIS) based classification method to discriminate between normal and pathological
retinal structures. The novelty of this system relies on the appropriate sequential application of exclusive image processing
techniques in synergy with ANFIS classifier to improve the accuracy of exudate lesions characterization. The performance
of the system has been evaluated by comparing it with various state of the art existing methods in terms of several perfor-
mance metrics such as Accuracy, Average error rate, F-Score and Kappa value. The obtained numerical results prove that
the proposed system with ANFIS classifier demonstrated superior performance in identification of exudate lesions.

Keywords Exudate · ANFIS · Morphological component · Retinal

1 Introduction second and third place, respectively. According to the

prevalence report of International Diabetes Federation
According to the consensus reported in 2000, India and World Health Organization, diabetes is predicted to
(31.7 million) topped the world with the highest num- double globally in the year 2030 with a maximum increase
ber of people with diabetes mellitus followed by China in India. Diabetes disease is common for human beings in
(20.8 million) with the United States (17.7 million) in the world and it affects people irrespective of age. Due to
the death rate, disability and blindness, the diabetic dis-
ease has been marked as a leading among all other dis-
The proposed works deals with the identification and grading
ease in the world (Li et al. 2011). Commonly, the diabetic
of exudates lesions from retinal images and the images for
this proposed work were selectively taken from STARE* and people are affected by retinopathy problem which is the
MESSIDOR* databases. Data are used for only the research most common eye disease. Diabetic Retinopathy is the
purpose. Therefore there is no need to get declaration from the most widely recognized reason for visual deficiency for
patient.
the diabetic patients before 50 years of age (Klein et al.
* R. Valarmathi 1984; Sjolie et al. 1997). The main problem associated
[email protected] with the Diabetic Retinopathy is that the patient might not
S. Saravanan be alerted by any of the warning symptoms or indications
[email protected] at the early phase of disease. The level and severity of this
disease can be pathological identified by types and number
1
Department of Electrical and Electronics Engineering, of lesions present in the surface of the retina. Diabetic
Marutam Nelli Polytechnic, Dharmapuri, India
retinopathy is extensively segregated into two major types
2
Department of Electrical and Electronics Engineering, namely non proliferative diabetic retinopathy (NPDR) and
Muthayammal Engineering College, Rasipuram, India

13
Vol.:(0123456789)
3634 R. Valarmathi, S. Saravanan

proliferative diabetic retinopathy. NPDR arises when the rid of vision loss. Many of the existing research work ponder
veins get harmed inside the retinal and release liquid over on automating the process of exudate lesions detection from
the retina (Ronald and Peng 2003) making the retina wind retinal images.
up inflamed and damped. In NPDR diverse indications Due to the high cost of treatment and examination and
of retinopathy can exists such as haemorrhages, exu- the lack of ophthalmologists, the diabetic patients are not
dates, micro-aneurysms and inter retinal micro vascular doing regular checkup particularly in rural areas (Mookiah
abnormalities. et al. 2013). Generally, examinations for eye can be taken
Among the various problems of NPDR, exudates in the out once 12 months to observe the early changes of spots by
retina can be identified and it has a possible treatment at its doing capturing and analyzing of retinal images (Saleh and
early stage, but the treatment is impossible when the level Eswaran 2012). Figure 1 illustrated normal retinal image
is at essential stage. The presence of exudates lesion indi- and abnormal retinal image with different abnormal signs.
cates harm to the retinal micro blood vessels and in spite of Few of the existing pioneering automated retinal image
the fact that the pathogenic instruments are not completely analysis system to diagnose diabetic retinopathy with exu-
comprehended, the clinical bio features remains distinctive dates, haemorrhages and micro-aneurysms. Those system
and effortlessly perceived. Basically exudates are of three were based on the concept that they were trained in such
type’s namely hard exudates, encircled plaques of exudates a way as haemorrhages and micro-aneurysms belongs to
and soft exudates. Hard exudates show up as intense yel- single class and moreover they have utilized only less num-
low spots lying externally or somewhere down in the retina. ber of retinal fundus images (Sinthanayothin et al. 2002)
Plaque exudates are characterized in the form of diffused implemented appropriate Hue Saturation Value (HSV)
gathering of lipoprotein and having variety in size. Cushy model in coordination with Mahalanobis distance meas-
or soft exudates have the white woolen appearance and paler ure to carry out the detection of Haemorrhages (Hatanaka
yellow tend to lie all the more externally in the tangible et al. 2008). The implemented an automated diagnostic
retina. Exudates are a class of lipid form of lesion noticeable system for diabetic retinopathy with extensive use of 1273
through retinal imaging. Exudates are extending in shading number of retinal images for intense training and testing
from white to yellow with shifting examples, size, differen- phase (Usher et al. 2004). In an another research work, hard
tiation and shapes. To put simply, the lesions of exudates exudate lesions were identified by making use of dynamic
are the intense spots having maximum pixel intensity with thresholding and median filtering segmentation techniques
moderately unmistakable edges (Joshi and Karule 2018). (Kayal and Banerjee 2014). Otsu’s thresholding techniques
Prominent and major symptoms of diabetic retinopathy and Sobel edge detector methods were used to detect the
are exemplified in the non proliferative diabetic retinopathy. bright exudate lesions (Wisaeng et al. 2015). Quellec et al.
Screening of this disease at an early stage is a difficult and (2008) implemented the Wavelet transform based matched
time consuming process for the pathologists. As this disease filtering approach associated with Gaussian template. The
may not show any early symptoms, regular checkup visits detection of hard exudates was identified by morphological
for screening out diabetic retinopathy is the only way to get techniques (Nayak et al. 2008) implemented Fuzzy C-Means

Fig. 1  Retinal images: a normal, b abnormal fundus

13
Exudate characterization to diagnose diabetic retinopathy using generalized method 3635

segmentation technique to identify the exudates lesions requires manually annotated individual lesions. Providing
(Sopharak et al. 2009). It used gray level co-occurrence these is a tedious and time-consuming activity, which has
matrix based texture feature set to characterize the exudate hampered the application of the algorithms to digital reti-
lesion followed by support vector machine classifier to auto- nal images with varying image formats, e.g., compressed
matically the retinal images (Chand and Dheeba 2015) the images, images of different sizes, etc. Generalized texture
bright exudate lesions were characterized by set of Texture modeling techniques that avoid manual segmentation would
features and used SVM classifier to discriminate the classes greatly enhance progress toward an automated screening of
(Senapati 2016). retinal images. Another common problem when applying
image processing methods to retinal images is the need
for correction of uneven illumination. The first step in the
2 Literature review analysis of retinal images has commonly been to process the
images to remove lighting artifacts, which enhances detec-
Feature extraction based on edge strength, centroid, mean tion of lesions. Heneghan et al. (2002) and Winder et al.
and standard deviation were used along with machine learn- (2009) used color normalization and local contrast enhance-
ing classifiers to discriminate the exudates and non-exudate ment as an initial step for detecting exudates. Zana and Klein
regions (Asha and Karpagavalli 2015). As the current (1997) removed the slow gradients in the background of the
research focus ponders on the development of algorithms green channel of each image resulting in a shade-corrected
for computer-aided diagnosis (CAD) systems for medical image. Other approaches are focused on the development of
diagnosis, many of the current research supports the use of preprocessing steps for the detection and subsequent removal
CAD system in automatic detection of diabetic retinopathy of normal anatomical “background” structures in the image.
(Aqeel 2014; Rokade and Manza 2015; Youssef and Sol- Zana and Klein (1999) applied a 3 × 3 pixel median filter to
ouma 2012). The presence of anatomical structure of reti- reduce this variation.
nal optic disc generally hinders the accuracy of automated They convolved the retinal image with a Gaussian filter,
retinal lesion identification algorithms such as detection of and then normalized the image for the detection of exudates.
exudates, haemmorhages, micro-aneurysms etc. Other methods segment and remove retinal vessels that sim-
Plethora of literature supports many algorithms devel- ulate red lesions. Our method is like the approach described
oped and applied to remove the optic disc from retinal image by Starck et al. (2005), which does not require any preproc-
which eases the process of exudate lesion detection. An optic essing. In our method, the green channel of the image is the
disc based removal algorithm was proposed by for matching input for applying the feature extraction technique. Much
histogram. In this algorithm, initially, from the database, of the published literature on retinal lesion detection has
histogram of color component can be extracted by using focused either on the detection of red lesions such as MAs
optic disc of few images and the colored average histograms and hemorrhages or detection of bright lesions such as exu-
in the image were determined, followed by localizing the dates and cotton wool spots. Starck et al. (2004), described
optic disc’s center. The optic disc removal was performed by methods for detecting red lesions. Imani et al. (2015) also
implementing Hough transform (Ruggeri et al. 2003). The proposed an automated method for the differentiation and
proposed a novel method of geometrical parametric model detection of exudates, cotton wool spots, which are char-
to localize optic disc (Jain et al. 2015). Thus, a geometrical acteristics of DR, and age-related macular degeneration,
parametric model was proposed to describe the direction of respectively. Similarly, Xinbo Gao et al. (2012) and Zhang
these vessels and two of the model parameters are just the and Guo (2009) developed methods for the detection of
coordinates of the optic disc center. Using samples of ves- exudates.
sels directions (extracted from fundus images by the track-
ing procedure) as experimental data, model parameters were
identified by means of a simulated annealing optimization 3 Methods
technique. These estimated values provide the coordinates
of the center of optic disc. A ­Matlab® prototype implement- The prominent objective of this research work deals with the
ing this method was applied to a set of 40 images of both classification and grading the pathological retinal images
normal and pathological subjects. In all these images, the in terms of Class 1: healthy image and Class 2: diabetic
optic disc position was correctly identified, even in rather retinopathy image. The major contributions of this research
difficult pathological situations. An extensive validation on work includes
a set of 81 images (STARE project data set) is currently in
progress to assess the robustness of the proposed technique. 1. An generalized semi-automated system to discriminate
To apply feature segmentation, various researchers find exudate lesions which encompasses firstly the pre-pro-
it necessary to train the algorithm on reference images. This cessing module to suppress blood vessel removal and

13
3636 R. Valarmathi, S. Saravanan

optic disc masking, secondly the morphological com- The proposed methodology collectively involves various
ponent analysis module which enhances and segments levels of image processing algorithms applied sequentially
the exudate lesions, thirdly the feature extraction module as shown in the Fig. 2: flowchart of proposed topology.
and finally the ANFIS classifier. The database image collection numerals are clearly men-
2. The first module namely pre-processing module first tioned in Table 1.
extracts the green channel of the retinal fundus images
and then precisely suppresses the structures of blood 3.2 Green channel extraction
vessels which hinder the application of any kind of auto-
mated algorithms by utilizing infinum and supremum The retinal fundus images available in databases are basi-
morphological operations. This pre-processing helps cally color images captured with 8-bits per each R, G, B
in improving the exudate lesion characterization algo- color plane. Among these three color of R, G, B planes, the
rithms. green light plane shows high contrast and dominant visibility
3. The second module demonstrates the morphological of fundus portion with good peak spectral sensitivity in the
component analysis based operation to enhance and spectrum of green to yellow portion. Thus the green plane
segment the lesion structures. The texture enhancement alone is extracted first in order to start with various image
module computes the non-sub sampled Contourlet trans- processing algorithms. Figure 3 shows the image of original
form which precisely approximates the exudate lesion RGB retinal fundus image and green channel extracted from
structures. Once the lesion is enhanced, MCA algorithm original image.
segments that component out superficially.
4. The third module performs the feature extraction using 3.3 Morphological processing level I: separation
four main descriptor parameters to extract the shape, of blood vasculature and retina
color, intensity and texture characteristics of exudates
lesions. The notion of applying retinal exudates extraction directly on
5. The fourth module describes the ANFIS classifier to the retinal color fundus images is a challenging task and thus
diagnose the clinical exudates biomarkers. exudates extraction requires many of the pre-requisite low
level image processing techniques. The presence of blood
vessels and optic nerve disc in the retinal image may hinder
3.1 Image collection from benchmark databases the application of exudates characterization. The blood ves-
sels in retinal image appear as linear curved shapes, whereas
Multitude of standard and benchmark databases are avail- exudates lesion appear as bright creamy spot-like structures.
able online in order to support the research community with The morphological filter based on da Cunha et al. (2006)
extensive collection of retinal fundus images to progress the has been used to extract blood vessels from retinal images.
research works in improving the pathological diagnosis of Recent literature clearly exemplifies that multitude of
retinal image structures. Moreover the database supports research work pondered with dominant use of morphologi-
the image collections inclusive of doctor’s diagnostic infor- cal processing techniques to remove the blood vessels from
mation and highlights of pathological annotations in order retinal fundus images (Burt and Adelson 1983). This mor-
to spend less time in image collection options but rather phological filter algorithm (Amin et al. 2017; Amadasun
spend effective time in implementing fundus image anal- and King 1989) is readily applicable because of the char-
ysis research algorithms. The proposed works deals with acterization of retinal blood vessels with following proper-
the identification and grading of exudates lesions from ties like blood vessels are linear connected curve segments,
retinal images and the images for this proposed work were intensity profile in cross-section shows a Gaussian shape and
selectively taken from structured analysis of the retina and the intensity value along with its direction might not change
(STARE and MESSIDOR) database. The motivation behind rapidly instead appears with continuous gray value. This
the use of retinal fundus images from various sources is algorithm computes the difference between the supremum
to evaluate the robustness of the proposed exudates detec- and infinum of image opening with two different structuring
tion and segmentation method to be independent of images elements.
acquisition variability. The STARE database comprises of
Fundus images captured by camera named Topcorn TRV-50 3.3.1 Assumptions made in morphological processing
fundus camera with image resolution of 605 × 700 pixels. level: I
Similarly the MESSIDOR Database comprises of Fundus
images captured by camera named Topcorn TRC NW6 fun- 1. Input image ­I0.
dus camera with maximum image and multi resolution up 2. IC image with preservation of blood vessel regions.
to 2304 × 1536 pixels. 3. IB image with absence of blood vessel.

13
Exudate characterization to diagnose diabetic retinopathy using generalized method 3637

Retinal color fundus

Extract the green channel

PRE PROCESSING
Blood vessel suppression MODULE

Optic disc masking

Morphological Component Analysis

exudate enhancement

Morphological Component Analysis

exudate enhancement

FEATURE EXTRACTION EXUDATE

CHARACTERIZATION
Shape MODULE

Color

Texture

Intensity

ANFIS CLASSIFIER
KNOWLEDGE BASE

Training Data set Testing Data set CLASSIFICATION

MODULE

Data set Cross

Validation

DIAGNOSTIC RESULTS
Class 1 – Normal
(Healthy eyes)
Class 2 – Exudate lesion
(Diabetic Retinopathy)

Fig. 2  Flowchart of the proposed topology

13
3638 R. Valarmathi, S. Saravanan

Table 1  Details of image dataset collections with structuring element, meanwhile non-vessel structures
Database name # of normal # of abnormal images collected
are discarded.
images collected
Exudate Cotton wool L1 Sin(𝛼) ≤ W (1)
lesions spot lesions
where 𝛼 is the angle between blood vessel and structuring
Hard Soft Few Many
element. Image opening operation is performed in 12 dif-
STARE 115 47 50 13 40 ferent angular rotations so that it ensures preservation of
MESSIDOR 85 85 75 blood vessel in all angular positions. The usage of 12 differ-
Total count 200 182 128 ent angles rotation represents 1­ 50gap between each rotation
Normal Abnormal images = 310 which covers 180° in 12 swings.
images = 200 Step 1 Perform image opening operation with structuring
element B.
{ }
4. IV is the image highlighted with blood vessel. IC = Supi=1,2…12 𝛾Bi (I0 ) (2)
5. Assume length and width of blood vessel as L = 17 and
W = 8 pixels, respectively (always W < L). where Bi structuring element in each rotation i.
6. B represents structuring element. Step 2 Repeat image opening operation with structuring
7. 𝛼 is the angle between blood vessel and structuring ele- element B to minutely extract smaller blood vessels.
ment. ( { })
IC = RI0 Supi=1,2…12 𝛾Bi (I0 ) (3)
Procedure for blood vasculature separation is summarized
in Algorithm 1

Algorithm – 1: Morphological processing Level

Step 1: Morphological open operation with supremum using length L1

Step 2: Reconstruct image with open operation
Step 3: Morphological open operation with infinum using length L2
Step 4: Image Subtraction

3.3.2 Algorithm‑1 explanation Step 3 To extract non-blood vessel structures. Here the

blood vessels are to be deleted and its area is filled with
Perform image opening operation with structuring element background pixel intensity. The area of blood vessel to be
B. During this operation blood vessels with parameters deleted and filled with background is obtained by
L = 17 pixels and W are preserved when they are parallel

Fig. 3  a, b Original RGB fundus image, c, d extracted green channel

13
Exudate characterization to diagnose diabetic retinopathy using generalized method 3639

Fig. 4  a, d Green channel, b,

e blood vasculature, c, f non
blood vessels regions

with sparse representation in order to enhance or segment

L2 Sin(𝛼) ≥ W (4)
the texture component of interest. The main objective of this
Here image opening operation is performed with length research work is to segment the exudates lesions from the
L2 and Structuring element B. pre-processed retinal images. Instead of using dictionaries,
{ } the proposed work makes use of non-sub sampled contourlet
IB = inf
i=1,2…12
𝛾Bi (I0 ) (5) transform. The reason behind use of contourlet transform is
that, it provides sparse expansion or measurement for spot
Step 4 The process of highlighting the blood vessels alone like image structures.
is given by To put in another way, to optimally represent the area and
exudates lesion structures it is necessary to select appropri-
IV = IC − IB (6)
ate directional transform. In this aspect, the use of contour-
Figure 4 shows the image with green channel, separated let transform provides an appropriate characterization for
blood vessels and images with non-blood vessel regions. exudates with piecewise smooth and differentiable curves
(Kohavi 1995). This transform decomposes the signal in
3.4 Morphological component analysis based terms of multi-scale resolution and direction by utilizing
texture enhancement and segmentation low-pass filter bank and directional filter bank, respectively.
The contourlet transform is replaced with Non-Sub sam-
Recent research work exemplifies that the concept of mor- pled contourlet transform (NSCT), which eliminates the up-
phological component analysis (MCA), predominantly sampling and down-sampling during signal decomposition
does the signal separation with sparsity (Saeys et al. 2007; and reconstruction. NSCT version of contourlet transform
Abdulshahed et al. 2015). The proposed research work using has been used as it highly avoids aliasing and imparts the
MCA based morphological processing to enhance and seg- prominent property of shift invariance while measuring
ment the texture based on the work published by Sridevi spot like characteristics. NSCT comprises of two attributes
and Nirmala (2016). The main notion and principle of MCA namely non-sub sampled Laplacian Pyramid Filter Bank
algorithm deals with the concept that various image texture and non-sub sampled directional filter bank with McClellan
holds distinct morphological diversity. Moreover it is possi- Transform (Donate et al. 2013; Kohavi 1998). The Lapla-
ble to model a signal in the form of linear mixture consisting cian pyramid filter bank does the process of marking point
of several morphological diverse components. Those mor- discontinuities and directional filter bank does the process
phological components can be very simply characterized by of linking the linear structures (Hossin and Sulaiman 2015).
redundant dictionaries. The accuracy of this MCA algorithm
relies in the course of appropriate selection of dictionaries

13
3640 R. Valarmathi, S. Saravanan

3.4.1 MCA texture enhancement and segmentation 3. Using MCA algorithm, each component of XI can be
algorithm recovered by solving the optimization problem with the
constraint of min‖ ‖P
‖𝛼I ‖P
1. Consider the size of the image as N × N and is converted ∑K
‖ ‖
into 1D vector of size N2. ‖Y −
‖ 𝜙I 𝛼I ‖
‖ ≤𝜎 (9)
I=1
2. The 1D vector is modeled as a mixture of K number of ‖ ‖2
different image components. 4. The solution for 𝛼I is obtained by hard thresholding the
∑K marginal residual given by
Y=
I=1
XI + ∈ (7) ∑
rI = Y − 𝜙 𝛼
m≠I m m (10)
 
𝜎∈2 = Var(∈) (8)
  where all components except Ith component is fixed.
where XI = 𝜙I 𝛼I This marginal residual rI contains the salient information
of component XI

13
Exudate characterization to diagnose diabetic retinopathy using generalized method 3641

Fig. 5  Morphological compo-

nent analysis based a, d texture
enhancement, b, e optic disc
localization and removal, c, f
MCA based segmentation

Table 2  Feature descriptors for exudate lesion characterization

Symbol Feature set Explanation

Shape parameters F1—area Sum of pixels in the lesion

F2—eccentricity Distance between foci of lesion divided by major axis
F3—perimeter Sum of pixels in the contour
F4—aspect ratio Ratio of major axis to minor axis in the lesion
F5—compactness Circularity details of the lesion
F6—diameter Diameter of the lesion
Intensity color parameters F7—mean enhanced intensity Describes the color and intensity properties of the lesion
F8—mean gradient magnitude
F9—mean box gradient
F10—energy
Gray tone texture parameters F7—coarseness Describes the spatially visible image texture properties of the lesion
F8—busyness
F9—complexity
F10—strength
Fourier power parameters F11—radial sum Describes the image texture property of the lesion in Fourier domain
F12—angular sum

Figure 5 shows the image of morphological component The following feature sets are selected with utmost impor-
analysis based enhanced image, optic disk masked image tance in order to characterize the features of exudate lesions
and segmented image, respectively. covering the aspects of shape parameter, intensity and color
parameter (Agurto et al. 2010, 2014), texture parameter and
3.5 Feature extraction Fourier power parameter in order to improve the accuracy
of classification. Table 2 illustrates the feature descriptors
The course of feature extraction is carried out to pull out the selected in detailed manner.
diagnostically significant bio-features of the exudate lesions
to characterize and train the automatic classification system.

13
3642 R. Valarmathi, S. Saravanan

3.6 ANFIS classification vector and imparts the IF part of the Fuzzy logic rules. The
second layer generates the fuzzy rules during the course of
Recently, adaptive neuro fuzzy inference system (ANFIS) training. The output of this layer decides the IF–THEN rule
classifier has gained impetus in many of the research appli- strength and thus it executes the premise portion of Fuzzy
cations as it synergizes the merits of both Fuzzy inference rules. The third layer computes the Fuzzy rule strength is
system and Neural Network approach (Acharya et al. 2012; computed as normalized value based on modification of pre-
Olson and Cohen 2019). The proposed method identi- vious layer. The fourth layer defines the consequent param-
fies the exudate lesions by making use of ANFIS classi- eter that means the THEN portion of Fuzzy logic rules. The
fier. ANFIS classifier is built with Sugeno Fuzzy Inference fifth layer generates the cumulative results of all previous
System, which attains acquaintance data base and executes nodes as a single value, which determines the class of input
inference process by framing IF–THEN rules. Rule formula- feature vector. The ANFIS classifier provides the diagnos-
tion is based on two significant parameters namely premise tic results as class belonging to either C1 as normal image
parameters and consequent parameters. The former defines and C2 as abnormal image with exudate lesion with diabetic
the Fuzzy-IF value which is fed as input feature vector and retinopathy. The image feature vector ­[Fi] fed as input to
the later defines the Fuzzy-THEN which will be resulted the first layer represents the premise (IF) parameter and the
as diagnostic class information. The prominent advantage output diagnostic result Cout to be extracted from the out-
of this classifier is that it widens the knowledge capacity put layer of classifier is represented as consequent (THEN)
by observing the training data-set and thereby modifies the parameters. The classifier has the capability to infer the input
premise parameters (membership functions) and consequent feature vector and generate the diagnostic classification by
parameters (fuzzy rules). The structural design of ANFIS using the Takagi–Sugeno FIS fuzzy rules defined by
classifier comprises of five different layers to perform a
RULEN = IF (F1 is 𝜇(1, 1)) AND (Fi is 𝜇(i, m)) THEN (C is Cout )
range of tasks on input signals. (19)
The first layer converts the input feature vector into a
range of 0–1 values belonging to linguistic markers graded where ­F1 to ­Fi signifies the input feature vector, 𝜇(i, 1) signi-
by low (L), medium (M) and high (H). This layer generates fies the fuzzy membership functions, m signifies the count
the fuzzy membership grade values as output. Membership of linguistic values, C signifies the classifier output as diag-
functions are delineated theoretically by means of general- nostic results. The accuracy of the ANFIS classifier and its
ized bell function defined by pattern learning relies on appropriate selection of highly dis-
tinctive feature vector and carrying out intensive training
𝜇(i, m) =
1 process. The training process is carried out by rapid hybrid
{( )2 }bs learning in terms of forward and backward pass learning
x−cs (18)
1+ as scheme which synergizes the gradient descent method and
least squares method IF–THEN parameters.
where ­as is the curve width, ­bs is the curve slope and ­cs is
the curve centre. This curve parameters are automatically
adjusted by the membership functions of the input feature

Table 3  Confusion matrix

Classifier Predictions
Actual clinical samples

Abnormal Images Normal Images

Abnormal
Images

TN FP
(Correct Rejections) (False alarms)
Normal
Images

FN
TP
(Negative misses)
(Hits)

13
Exudate characterization to diagnose diabetic retinopathy using generalized method 3643

Table 4  Performance Method Classifier types Average F-Score Accuracy Kappa value
comparison of various error rate
classifiers and methods
Proposed ANFIS 0.015 0.968 88 0.791
Proposed SVM 0.023 0.902 82 0.635
Proposed RBFNN 0.026 0.893 79 0.551
Imani et al. (2015) – 0.024 0.905 85 0.535
Agurto et al. (2010) – 0.043 0.898 78 0.451
Acharya et al. (2012) – 0.038 0.832 81 0.382
Long et al. (2019) Dynamic threshold – 0.767 76.5 –
and SVM

4 Results and discussion images and is mentioned as false alarms made in the diag-
nostic decision. FN represents the count of normal images
Clinical diabetic Retinopathy screening requires exclusive classified wrongly as abnormal images and is mentioned as
exposure to identify anatomical structures responsible to negative misses made in the diagnostic decision. TP rep-
worse the disease such as micro aneurysms, exudates, cotton resents the count of normal images classified correctly as
wool spots etc. Moreover, the retinal image sometimes may normal images and is mentioned as hits made in the diag-
be insufficiently illuminated which protrudes the diabetic nostic decision. Accuracy is defined as the ratio of number
retinopathy screening as a time consuming and most difficult of correctly classified samples to total number of samples
task. This research work deals about the implementation of the ratio of number of correctly classified samples to total
generalized method to characterize the exudate lesions from number of samples. As the classification is implemented for
retinal fundus images responsible for diabetic retinopathy. In K number of times to encompass the merit of K-fold cross
order to validate this generalized method of exudates clas- validation, the classifier performance can be measured by
sification with ANFIS classifier, the performance metrics the very familiar parameter namely average error rate. The
taken into consideration are Classifier Average Error rate, average error rate E is given by
F-score, Kappa value and Accuracy. K
The accuracy and consistent performance of the classifier 1∑ FPn + FNn
E= (20)
exclusively depends on the cross-validation of training and K n=0 TPn + TNn + FPn + FNn
testing data-set by using K-fold cross validation (Sridevi and
Nirmala 2016). The prominent merit on using this cross vali- There are certain other statistical classifier performance
dation scheme is that classifier faces the dataset effectively measures namely Kappa statistical coefficient and F-Score
in equal probability of training and testing data set. The trial (Kohavi 1998). Kappa value ensures the performance of the
of cross validation is performed for over K times with data- classifier in addition to compare one classifier with another
set divided as 90% under training set and 10% under testing classifier also. Similarly, F-Score is a measure made by two
set. The selected image data-set were arbitrarily separated parameters namely precision P and recall R denoted by
into K (K = 5) disjoint folds of equivalent size where each TP
part has generally similar class dissemination. In each trail P= (21)
TP + FP
of cross validation, the dataset will be repeated K-1 times so
as every dataset has the possibility of being as testing data TP
(Kohavi 1995). R= (22)
TP + FN
One of the prominent performance metric of the diagnos-
tic classifier is accuracy of the classification. To ensure the Higher the F-score value, higher is the classification accu-
measurement of accuracy four other parameters are to taken racy of classifier under evaluation. F-Score is formulated as
into consideration. They are true negative (TN), false posi- 2PR
tive (FP), false negative (FN) and true positive (TP), which F= (23)
P+R
can be clearly depicted by a confusion matrix in Table 3
(Donate et al. 2013). Table 4 illustrates the performance comparison with
TN represents the count of abnormal images classified various statistical measures and classifier metrics. The per-
correctly as abnormal images and is mentioned as correct formance measures obtained for the proposed method with
rejections made in the diagnostic decision. FP represents ANFIS classifier is comparatively higher than other clas-
the count of abnormal images classified wrongly as normal sifiers like support vector machine (SVM) and radial basis

13
3644 R. Valarmathi, S. Saravanan

function neural network (RBFNN) and state of the art exist- Donate JP, Cortez P, SáNchez GG, De Miguel AS (2013) Time series
ing methods. forecasting using a weighted cross-validation evolutionary artifi-
cial neural network ensemble. Neurocomputing 109:27–32
Hatanaka Y, Nakagawa T, Hayashi Y, Hara T, Fujita H (2008) Improve-
ment of automatic hemorrhages detection methods using bright-
5 Conclusion ness correction on fundus images. In: 30th Annual international
IEEE EMBS conference Vancouver, British Columbia, Canada,
August 20–24, p 5429
Despite diabetic retinopathy screening from retinal fun- Heneghan C, Flynn J, O’Keefe M, Cahill M (2002) Characterization
dus image hangs out to be a most challenging task, the use of changes in blood vessel width and tortuosity in retinopathy of
of computer aided diagnosis will improve the diagnostic prematurity using image analysis. Med Image Anal 6(4):407–429
decision. In view of the fact that the diagnosis of diabetic Hossin M, Sulaiman MN (2015) A review on evaluation metrics for
data classification evaluations. Int J Data Min Knowl Manag Pro-
retinopathy can be executed accurately only by experienced cess 5(2):1
pathologists, there exists a prominent demand for algorith- Imani E, Pourreza HR, Banaee T (2015) Fully automated diabetic
mic computerized diagnosis. The proposed work deals about retinopathy screening using morphological component analysis.
the generalized semi-automated method by incorporating Comput Med Imaging Graph 43:78–88
Jain AB, Prakash VJ, Bhende M (2015) Techniques of fundus imaging.
various image processing tasks to precisely identify the Med Vis Res Found 33(2):100
exudates lesions from retinal fundus images. Moreover the Joshi S, Karule PT (2018) A review on exudates detection methods for
classification evaluation process has been performed with diabetic retinopathy. Biomed Pharmacother 97:1454
the use of data-set handling with fivefold cross validation Kayal D, Banerjee S (2014) A new dynamic thresholding based tech-
nique for detection of hard exudates in digital retinal fundus
methodology. Comparative performance analysis results image. In: 2014 International conference on signal processing
provide the evidence that the proposed method with ANFIS and integrated networks, p 141
classifier has the latent ability with high classification accu- Klein R, Klein BEK, Moss SE (1984) Visual impairment in diabetes.
racy than other classifiers and existing methods. Ophthalmology 91(1):1
Kohavi R (1995) A study of cross-validation and bootstrap for accu-
racy estimation and model selection. Int Jt Conf Artif Intell
14(2):1137–1145
References Kohavi RF (1998) Provost: glossary of terms. Mach Learn
30(2/3):271–274
Abdulshahed AM, Longstaff AP, Fletcher S (2015) The application of Li HK, Ma Horton, Bursell S-E, Cavallerano J, Zimmer-Galler I, Ten-
ANFIS prediction models for thermal error compensation on CNC nant M (2011) Telehealth practice recommendations for diabetic
machine tools. Appl Soft Comput 27:158–168 retinopathy. 2nd edition. Telemed e-Health 17(10):814
Acharya UR, Ng EYK, Tan JH, Sree SV, Ng KH (2012) An integrated Long S, Huang X, Chen Z, Pardhan S, Zheng D (2019) Automatic
index for the identification of diabetic retinopathy stages using detection of hard exudates in color retinal images using dynamic
texture parameters. J Med Syst 36(3):2011–2020 threshold and SVM classification: algorithm development and
Agurto C, Murray V, Barriga E, Murillo S, Pattichis M, Davis H, Soliz evaluation. Hindawi BioMed Res Int 2019:1–13
P (2010) Multiscale AM-FM methods for diabetic retinopathy Mookiah MRK, Acharya UR, Martis RJ, Chua CK, Lim CM, Ng EYK
lesion detection. IEEE Trans Med Imaging 29(2):502–512 (2013) Evolutionary algorithm based classifier parameter tuning
Agurto C, Murray V, Yu H, Wigdahl J, Pattichis M, Nemeth S, Bar- for automatic diabetic retinopathy grading: a hybrid feature extrac-
riga ES, Soliz P (2014) A multi scale optimization approach to tion approach. Knowl Based Syst 39(2):9
detect exudates in the macula. IEEE J Biomed Health Inform Nayak J, Bhat PS, Acharya R, Lim CM, Kagathi M (2008) Automated
18:1328–1336 identification of diabetic retinopathy stages using digital fundus
Amadasun M, King R (1989) Textural features corresponding to tex- images. J Med Syst 32:107–115
tural properties. IEEE Trans Syst Man Cybern 19(5):1264–1274 Olson GB, Cohen M (2019) A general mechanism of martensitic nucle-
Amin J, Sharif M, Yasmin M, Ali H, Fernandes SL (2017) A method ation: part II. FCC → BCC and other martensitic transformations.
for the detection and classification of diabetic retinopathy using Ambient Intell Humaniz Comput 10:1897–1914
structural predictors of bright lesions. J Comput Sci 19:153–164 Quellec G, Lamard M, Josselin PM, Cazuguel G, Cochener B, Roux
Aqeel AF (2014) Automated algorithm for retinal image exudates C (2008) Optimal wavelet transform for the detection of microa-
and drusens detection, segmentation and measurement. In: IEEE neurysms in retinal photographs. IEEE Trans Med Imaging
international conference on electro/information technology, 27(9):1230
pp 206–215 Rokade PM, Manza RR (2015) Automatic detection of hard exudates
Asha PR, Karpagavalli S (2015) Diabetic retinal exudates detection in retinal images using haar wavelet transform. Eye 4:402–410
using machine learning techniques. In: Advanced computing and Ronald PC, Peng TK (2003) A text book of clinical ophthalmology: a
communication systems, international conference on, pp 1–5 practical guide to disorders of the eyes and their management, 3rd
Burt PJ, Adelson EH (1983) The Laplacian pyramid as a compact edn. World Scientific Publishing Company, Singapore
image code. IEEE Trans Commun 31:532–540 Ruggeri A, Forrachia M, Grisan E (2003) Detecting the optic disc in
Chand CR, Dheeba J (2015) Automatic detection of exudates in color retinal images by means of a geometrical model of vessel network.
fundus retinopathy images. Indian J Sci Technol 8:1–6 In: Proceedings of the 25th annual international conference of the
da Cunha L, Arthur JZ, Do Minh N (2006) The non sub sampled con- IEEE engineering in medicine and biology society (IEEE Cat. No.
tourlet transform: theory, design, and applications. IEEE Trans 03CH37439). IEEE, vol 1, pp 902–905
Image Proc 15:3089–3101 Saeys Y, Inza I, Larrañaga P (2007) A review of feature selection tech-
niques in bioinformatics. Bioinformatics 23(19):2507–2517

13
Exudate characterization to diagnose diabetic retinopathy using generalized method 3645

Saleh MD, Eswaran C (2012) An automated decision-support system Winder RJ, Morrow PJ, McRitchie IN, Bailie JR, Hart PM (2009)
for non-proliferative diabetic retinopathy disease based on MAs Algorithms for digital image processing in diabetic retinopathy.
and HAs detection. J Comput Methods Progr Biomed 108(1):186 Comput Med Imaging Graph 33(8):608–622
Senapati RK (2016) Bright lesion detection in color fundus images Wisaeng K, Hiransakolwong N, Pothiruk E (2015) Automatic detection
based on texture features. Bull Electr Eng Inform 5:92–100 of exudates in retinal images based on threshold moving average
Sinthanayothin C, Boyce JF, Williamson TH, Cook HK, Mensah E, Lal models. Biofizika 60(2):360
S, Usher D (2002) Automated detection of diabetic retinopathy on Xinbo Gao WL, Tao Dacheng, Li Xuelong (2012) Image quality assess-
digital fundus images. Diabet Med 19(2):105 ment based on multiscale geometric analysis. IEEE Trans Image
Sjolie AK, Stephenson J, Aldington S, Kohner E, Janka H, Stevens L, Proc 2009(18):1409–1423
Fuller J (1997) EURODIAB Complications Study Group, retin- Youssef D, Solouma NH (2012) Accurate detection of blood ves-
opathy and vision loss in insulin-dependent diabetes in Europe. sels improves the detection of exudates in color fundus images.
Ophthalmology 104(2):252 Comput Methods Progr Biomed 108(3):1052–1061. https​://doi.
Sopharak A, Uyyanonvara B, Barman S (2009) Automatic exudate org/10.1016/j.cmpb.2012.06.006
detection from non-dilated diabetic retinopathy retinal images Zana F, Klein JC (1997) Robust segmentation of vessels from retinal
using fuzzy c-means clustering. Sensors 9:2148–2161 angiography. In: Proceedings of international conference digital
Sridevi S, Nirmala S (2016) ANFIS based decision support system for signal processing, Santorini, Greece, pp 1087–1091
prenatal detection of Truncus Arteriosus congenital heart defect. Zana F, Klein JC (1999) A multimodal registration algorithm of eye
Appl Soft Comput 46:577–587 fundus images using vessels detection and Hough transform. IEEE
Starck J-L, Elad M, Donoho D (2004) Redundant multiscale transforms Trans Med Imaging 18:419–427
and their application for morphological component separation. Zhang Q, Guo B-l (2009) Multifocus image fusion using the nonsub-
Adv Imaging Electron Phys 132:287–348 sampled contourlettransform. Signal Proc 89(7):1334–1346
Starck J-L, Elad M, Donoho DL (2005) Image decomposition via the
combination of sparse representations and a variational approach. Publisher’s Note Springer Nature remains neutral with regard to
Image Proc IEEE Trans 14:1570–1582 jurisdictional claims in published maps and institutional affiliations.
Usher D, Dumsky M, Himaga M, Williamson T, Nussey S, Boyce
J (2004) Automated detection of diabetic retinopathy in digital
retinal images: a tool for diabetic retinopathy screening. Diabet
Med 21(1):84

13