Bonfring International Journal of Man Machine Interface, Vol. 1, Special Issue, December 2011 15
ISSN 2250 – 1061 | © 2011 Bonfring
Abstract--- Diabetic Retinopathy (DR) is one of the most
important ophthalmic pathological reasons of blindness
among people of working age. Previous techniques for blood
vessel detection in retinal images can be categorized into rule-
based and supervised methods. This research presents a new
supervised technique for blood vessel detection in digital
retinal images. This novel approach uses an Extreme Learning
Machine (ELM) approach for pixel classification and
calculates a 7-D vector comprises of gray-level and moment
invariants-based features for pixel representation. The
approach is based on pixel classification using a 7-D feature
vector obtained from preprocessed retinal images and given as
input to a ELM. Classification results (real values between 0
and 1) are thresholded to categorize each pixel into two
classes namely vessel and nonvessel. Ultimately, a post
processing fills pixel gaps in detected blood vessels and
eliminates falsely-detected isolated vessel pixels. The
technique was assessed on the publicly available DRIVE and
STARE databases, widely used for this purpose, as they
comprises of retinal images where the vascular structure has
been precisely marked by experts. Method performance on
both sets of test images is better than other existing solutions
in literature. The approach proves particularly accurate for
vessel detection in STARE images. Its efficiency and strength
with different image conditions, along with its simplicity and
fast implementation, make this blood vessel segmentation
proposal appropriate for retinal image computer analyses
such as automated screening for early diabetic retinopathy
detection.
Keywords— Diabetic Retinopathy, Moment Invariants,
Retinal Imaging, Vessels Segmentation
I. INTRODUCTION
IABETIC retinopathy (DR) is one of the most serious and
most frequent eye diseases in the world; it is the most
common cause of blindness [2] [3] in adults between 20 and
60 years of age. DR is a silent disease in its first stages, i.e.
many patients are not aware of its presence, it often remains
undiagnosed until serious vision impairment occurs. To build
an expert system that is able to perform the diagnoses task
needed to use digital retinal images. By imaging the retina of a
person with a special camera, then using image processing and
pattern recognition technique to analyze that retina make a
specific diagnoses decision. One of the important operations
X. Merlin Sheeba, KS Rangasamy College of Tehnology, Tiruchengode, E-
mail: renachristina@gmail.com
S. Vasanthi, KS Rangasamy College of Tehnology, Tiruchengode
performed on the digital retina is the edge detection of blood
vessels of the retina. A specialized edge detection technique
may be needed since traditional edge detection technique
might not be able to accurately segment the vessels from the
background of the retinal image. Accurate vasculature
segmentation is a difficult task for several reasons: the
presence of noise, the low contrast between vessels and
background and the variability of vessels width, brightness,
and shape.
Several methods have been proposed for this purpose,
matched filer, morphology edge detection, and ridge –based
vessel segmentation which is called pixel classification,
tracking methods, supervised classification and wavelet
transform. Also, due to the reflection on the tiny uneven
surface of the soft tissue in the image, the low contrast
between the vessels and background and the pathological
variations, detecting blood vessels automatically from a retinal
image is a challenging problem. Many methods have been
used for retinal vessel segmentation. These can be divided
into two groups: rule-based methods and supervised method.
Supervised methods are based on pixel classification, which
consists on classifying each pixel into two classes, vessel and
non-vessel. Regarding rule-based methods, vessel tracking
method attempt to obtain the vasculature structure by
following vessel center lines.
The segmentation of vessels can be represented in two
ways: One way is to mark out all vessel pixels. The other way
is to find the vessel centerlines and the radius at each pixel of
the center lines. The main aim of vessel segmentation
algorithm is automated detection of eye diseases.
II. LITERATURE SURVEY
Currently there are number of automatic systems has been
developed for the detection of various eye diseases like
diabetic retinopathy. To segment vessels in retinal images,
seven classes of methods have been commonly used matched
filters, vessel tracking, morphological processing, region
growing, multiscale, supervised and adaptive thresholding
approaches. Some of the reference papers have been explained
below;
Ricci and perfetti [4] uses basic line detector whose
response is thresholded to obtain unsupervised pixel
classification. They employ two orthogonal line detectors
along with the grey level of the target pixel to construct feature
vector for supervised classification using a support vector
machine. The basic line detector; is computationally simple
and gives very good result with respect to existing
unsupervised methods. SVM guarantee good generalization
performance even with a small training set. Performance
An Efficient ELM Approach for Blood Vessel
Segmentation in Retinal Images
X. Merlin Sheeba and S. Vasanthi
D
Bonfring International Journal of Man Machine Interface, Vol. 1, Special Issue, December 2011 16
ISSN 2250 – 1061 | © 2011 Bonfring
estimated both in terms of AUC and accuracy show this
supervised system is slightly superior to other existing system.
Some false positive are found around the border of the optic
disk and in proximity of the pathological regions. Some
problems occur to segment the thinnest vessels or when the
local contrast is quite low, in this case, vessels can be captured
only at the expense of high FPR. Because of this, segmentation
image tends to be quite noisy. Maria mendonca [5] combines
the retinal vascular network with differential filters, for vessel
centerline extraction with morphological operators, used for
filling vessel segments. Several intensity and morphological
properties of vascular structures, such as linearity,
connectivity, and width are considered. The main type of
errors are 1)detection of other retinal structure ,like the optic
disc, several pathological areas and background structures,2)
under segmentation of some vessel segments,3) partial or
complete missing if thin vessel branches. Staal [6] used a
system which is based on extraction of image ridges, which
coincide approximately with vessel centerlines. The ridges are
used to compose primitives in the form of line elements. With
the line elements an image is partitioned into patches by
assigning each image pixel to the closest line element. Every
line element constitutes a local coordinate frame for its
corresponding patch. For every pixel, feature vectors are
computed that make use of properties of the patches and the
line elements. The feature vectors are classified using a NN-
classifier and sequential forward feature selection. Two types
of errors can be distinguished. The first type is over- and under
segmentation of the vessels. This is important in applications
where determination of the vessel width is needed. The second
type of error is the missing or erroneous detection of vessel
branches. Marin [1] proposed a new methodology for blood
vessel detection is presented. It is based on pixel classification
using a 7-D feature vector extracted from preprocessed retinal
images and given as input to a neural network. The NN was
then fed with the values of these pixel windows for classifying
each pixel into vessel or not. Classification results are
thresholded to classify each pixel into two classes: vessel and
nonvessel. Classifiers are trained by supervised learning with
data from manually-labeled images. Finally, a post processing
fills pixel gaps in detected blood vessels and removes falsely-
detected isolated vessel pixels. Despite of its simplicity, the
high accuracy achieved by this method in blood vessel
detection.
III. MATERIAL
To evaluate the vessel segmentation methodology
described in the next section, two publicly available databases
containing retinal images, the DRIVE [8] and STARE [9]
databases, were used. These databases have been widely used
by other researchers to test their vessel segmentation
methodologies since,apart from being public, they provide
manual segmentations for performance evaluation. The
DRIVE database comprises 40 eye-fundus color images (seven
of which present pathology) taken with a Canon CR5
nonmydriatic 3CCD camera with a 45 field-of-view (FOV).
Each image was captured at 768 X 584 pixels; 8 bits per color
plane and, in spite of being offered in LZW compressed TIFF
format, they were originally saved in JPEG format. The
database is divided into two sets: a test set and a training set,
each of them containing 20 images. The test set provides the
corresponding FOV masks for the images, which are circular
(approximated diameter of 540 pixels) and two manual
segmentations generated by two different specialists for each
image. The selection of the first observer is accepted as ground
truth and used for algorithm performance evaluation in
literature. The training set also includes the FOV masks for the
images and a set of manual segmentations made by the first
observer. The STARE database, originally collected by
Hoover et al. [7], comprises 20 eye-fundus color images (ten
of them contain pathology) captured with a TopCon TRV-50
fundus camera at 35 FOV. The images were digitalized to 700
X 605 pixels, 8 bits per color channel and are available in PPM
format. The database contains two sets of manual
segmentations made by two different observers. Performance
is computed with the segmentations of the first observer as
ground truth.
IV. VESSEL SEGMENTATION METHOD
This paper proposes a new supervised approach for blood
vessel detection based on a ELM for pixel classification. The
necessary feature vector is computed from preprocessed retinal
images in the neighborhood of the pixel under consideration.
The following process stages may be identified: 1) original
fundus image preprocessing for gray-level homogenization
and blood vessel enhancement, 2) feature extraction for pixel
numerical representation, 3) application of a classifier to label
the pixel as vessel or nonvessel, and 4) post processing for
filling pixel gaps in detected blood vessels and removing
falsely-detected isolated vessel pixels. Input images are
monochrome and obtained by extracting the green band from
original RGB retinal images. The green channel provides the
best vessel-background contrast of the RGB-representation,
while the red channel is the brightest color channel and has
low contrast, and the blue one offers poor dynamic range.
Thus, blood containing elements in the retinal layer (such as
vessels) are best represented and reach higher contrast in the
green channel [10]. All parameters described below were set
by experiments carried out on DRIVE images with the aim of
contributing the best segmentation performance on this
database Therefore, they refer to retinas of approximately 540
pixels in diameter. The application of the methodology to
retinas of different size (i.e., the diameter in pixels of STARE
database retinas is approximately 650 pixels) demands either
resizing input images to fulfil this condition or adapting
proportionately the whole set of used parameters to this new
retina size.
A. Preprocessing
Color fundus images often show important lighting
variations, poor contrast and noise. In order to reduce these
imperfections and generate images more suitable for extracting
the pixel features demanded in the classification step, a
preprocessing comprising the following steps is applied: 1)
vessel central light reflex removal, 2) background
homogenization, and 3) vessel enhancement. Next, a
description of the procedure, illustrated through its application
Bonfring International Journal of Man Machine Interface, Vol. 1, Special Issue, December 2011 17
ISSN 2250 – 1061 | © 2011 Bonfring
to a STARE database fundus image (Fig. 1), is detailed.
1) Vessel Central Light Reflex Removal: Since retinal
blood vessels have lower reflectance when compared to other
retinal surfaces, they appear darker than the background.
Although the typical vessel cross-sectional gray-level profile
can be approximated by a Gaussian shaped curve (inner vessel
pixels are darker than the outermost ones), some blood vessels
include a light streak (known as a light reflex) which runs
down the central length of the blood vessel. To remove this
brighter strip, the green plane of the image is filtered by
applying a morphological opening using a three-pixel diameter
disc, defined in a square grid by using eight-connexity, as
structuring element. Disc diameter was fixed to the possible
minimum value to reduce the risk of merging close vessels.
denotes the resultant image for future references. An example
of vessel central light reflex and its removal from a fundus
image by means of opening filtering operation is shown in Fig.
1(a) .
2) Background Homogenization: Fundus images often
contain background intensity variation due to nonuniform
illumination. Consequently, background pixels may have
different intensity for the same image and, although their gray-
levels are usually higher than those of vessel pixels (in relation
to green channel images), the intensity values of some
background pixels is comparable to that of brighter vessel
pixels. Since the feature vector used to represent a pixel in the
classification stage is formed by gray-scale values, this effect
may worsen the performance of the vessel segmentation
methodology. With the purpose of removing these background
lightening variations, a shade-corrected image is accomplished
from a background estimate. This image is the result of a
filtering operation with a large arithmetic mean kernel, as
described below.
Firstly, a 3x3 mean filter is applied to smooth occasional
salt-and-pepper noise. Further noise smoothing is performed
by convolving the resultant image with a Gaussian kernel of
dimensions mxm= 9x9, mean = 0 and variance
2
= 1.8
2
,
Secondly, a background image IB, is produced by applying a
69x69 mean filter [Fig. 1(b)]. When this filter is applied to the
pixels in the FOV near the border, the results are strongly
biased by the external dark region. To overcome this problem,
out-of-the FOV gray-levels are replaced by average gray-
levels in the remaining pixels in the square. Then, the
difference D between
and
is calculated for every pixel
,
=
,
(, ) (1)
Figure 1: Illustration of the Preprocessing Process: (a) Green
Channel of the Original Image. (b) Background Image
Finally, a shade-corrected image
is obtained by
transforming linearly values into integers covering the whole
range of possible gray-levels ([0-255] , referred to 8-bit
images). Fig. 1(d) shows the corresponding to a nonuniformly
illuminated image. The proposed shade-correction algorithm is
observed to reduce background intensity variations and
enhance contrast in relation to the original green channel
image.Besides the background intensity variations in images,
intensities can reveal significant variations between images
due to different illumination conditions in the acquisition
process.
Figure 2: Two Examples of Application of the Preprocessing
on Two Images with Different Illumination Conditions. (a),
Green channel of the Original Images. (b), Homogenized
Images. (c), Vessel-Enhanced Images
In order to reduce this influence, a homogenized image
[Fig. 1(a)] is produced as follows: the histogram of is
displaced toward the middle of the gray-scale by modifying
pixel intensities according to the following gray-level global
transformation function:
=
0, < 0
255, > 255
,
(2)
Where
=
+ 128
_
(3)
and
and
are the gray-level variables of input
and output images (and , respectively). The variable denoted
by defines the gray-level presenting the highest number of
pixels in (x,y). By means of this operation, pixels with gray-
level, which are observed to correspond to the background of
the retina, are set to 128 for 8-bit images. Thus, background
pixels in images with different illumination conditions will
standardize their intensity around this value. Fig. 2 (a), (b) and
(d), (e), shows this effect for two fundus images in the STARE
database.
3) Vessel Enhancement: The final preprocessing step
consists on generating a new vessel-enhanced image , which
proves more suitable for further extraction of moment
invariants- based features (see Section IV-B).
Vessel enhancement is performed by estimating the
complementary image of the homogenized image,
, and
Bonfring International Journal of Man Machine Interface, Vol. 1, Special Issue, December 2011 18
ISSN 2250 – 1061 | © 2011 Bonfring
subsequently applying the morphological Top-Hat
transformation [Fig. 1(f)]
=
(
) (4)
where is a morphological opening operation using a disc of
eight pixels in radius. Thus, while bright retinal structures are
removed (i.e., optic disc, possible presence of exudates or
reflection artifacts), the darker structures remaining after the
opening operation become enhanced (i.e., blood vessels, fovea,
possible presence of microaneurysms or hemorrhages).
Samples of vessel enhancement operation results are shown
in Fig. 2(c) and (f) for two fundus images with variable
illumination conditions.
B. Feature Extraction
The aim of the feature extraction stage is pixel
characterization by means of a feature vector, a pixel
representation in terms of some quantifiable measurements
which may be easily used in the classification stage to decide
whether pixels belong to a real blood vessel or not. In this
paper, the following sets of features were selected.
• Gray-level-based features: features based on the
differences between the gray-level in the candidate pixel
and a statistical value representative of its surroundings.
• Moment invariants-based features: features based on
moment invariants for describing small image regions
formed by the gray-scale values of a window centered on
the represented pixels.
Gray-Level-Based Features: Since blood vessels are
always darker than their surroundings, features based on
describing gray-level variation in the surroundings of
candidate pixels seem a good choice. A set of gray-level-based
descriptors taking this information into account were derived
from homogenized images considering only a small pixel
region centered on the described pixel (, ).
,
stands for the
set of coordinates in a × sized square window centered
on point (, ). Then, these descriptors can be expressed as
1
,
=
,
min
(,)
,
(, )
(5)
2
,
= min
,
,
,
,
(6)
3
,
=
,
min
(,)
,
,
(7)
4
,
= std
(,)
,
,
(8)
5
,
=
,
(9)
2) Moment Invariants-Based Features: The vasculature in
retinal images is known to be piecewise linear and can be
approximated by many connected line segments. For detecting
these quasi-linear shapes, which are not all equally wide and
may be oriented at any angle, shape descriptors invariant to
translation, rotation and scale change may play an important
role. Within this context, moment invariants proposed by Hu
[60] provide an attractive solution and are included in the
feature vector. In this paper, they are computed as follows.
Given a pixel (x,y) of the vessel-enhanced image
, a
subimage is generated by taking the region defined by
,
17
.
The size of this region was fixed to 17x17 so that, considering
that the region is centered on the middle of a “wide” vessel (8-
9-pixel wide and referred to retinas of approximately 540
pixels in diameter), the subimage includes an approximately
equal number of vessel and nonvessel pixels.
A set of seven moment invariants under size, translation, and
rotation, known as Hu moment invariants, can be derived
from combinations of regular moments. Among them, our tests
have revealed that only those defined by
1
=
20
+
02
(10)
2
=
20
+
02
2
+ 4
11
2
(11)
constitute the combination providing optimal performance
interms of average accuracy (see Section V-B). The inclusion
of the remainder moments result in decreasing classification
performance and increasing computation needed for
classification.
Moreover, the module of the logarithm was used instead of
its values themselves. Using the logarithm reduces the
dynamic range and the module prevents from having to deal
with the complex numbers resulting from computing the
logarithm of negative moment invariants.
In conclusion, the following descriptors were considered to
be part of the feature vector of a pixel located at
6
,
=
log
1
(12)
7
,
=
log
2
(13)
where
1
and
2
are the moment invariants
C. Classification
In the feature extraction stage, each pixel from a fundus
image is characterized by a vector in a 7-D feature space.
,
= (
1
,
, ,
7
(, )) (14)
Now, a classification procedure assigns one of the classes
(vessel) or (nonvessel) to each candidate pixel when its
representation is known. In order to select a suitable classifier,
the distribution of the training set data (described below) in the
feature space was analyzed. The results of this analysis
showed that the class linear separability grade was not high
enough for the accuracy level required for vasculature
segmentation in retinal images. Therefore, the use of a non
linear classifier was necessary. The following nonlinear
classifiers can be found in the existing literature on this topic:
the kNN method [6] and [11], support vector machines [17],
Bayesian classifier [12], or neural networks [13], [14].
Extreme Learning Machine classifier was selected in this
paper.
Extreme Learning Machine
ELMs parameters can be analytically determined rather
Bonfring International Journal of Man Machine Interface, Vol. 1, Special Issue, December 2011 19
ISSN 2250 – 1061 | © 2011 Bonfring
than being tuned. This algorithm provides good generalization
performance at very fast learning speed. From function
approximation point of view ELM[15] is very different
compared to the traditional methods. ELM shows that the
hidden node parameters can be completely independent from
the training data.
In conventional learning theory, the hidden node
parameters cannot be created without seeing the training
data.
In ELM, the hidden node parameters can be generated
before seeing the training data.
D. Salient Features of ELM
Compared to popular Back propagation (BP) Algorithm and
Support Vector Machine (SVM), ELM has several salient
features:
Ease of use: Except predefined network architecture, no
other parameters need to be manually tuned. Users need
not have spent much time in tuning and training learning
machines.
Faster learning speed: The time taken for most of the
training will be in milliseconds, seconds, and minutes.
Other conventional methods cannot provide such a fast
learning speed.
Higher generalization performance: The generalization
performance of ELM is better than SVM and back
propagation in most cases.
Applicable for all nonlinear activation functions:
Discontinuous, differential, non-differential functions can
be used as activation functions in ELM.
Applicable for fully complex activation functions:
Complex functions can also be used as activation
functions in ELM.
Extreme Learning Machine (ELM) meant for Single Hidden
Layer Feed-Forward Neural Networks (SLFNs) will randomly
select the input weights and analytically determines the output
weights of SLFNs. This algorithm tends to afford the best
generalization performance at extremely fast learning speed.
The structure of ELM network is shown in figure 1. ELM
contains an input layer, hidden layer and an output layer. The
ELM has several interesting and significant features different
from traditional popular learning algorithms for feed forward
neural networks. These include the following:
The learning speed of ELM is very quick when compared
to other classifier. The learning process of ELM can be
performed in seconds or less than seconds for several
applications. In all the previous existing learning
techniques, the learning performed by feed-forward
network will take huge time even for simple applications.
The ELM has enhanced generalization result when
compared to the gradient-based learning techniques. The
existing gradient-based learning technique and a few other
learning techniques may encounter several problems like
local minima, not proper learning rate and over fitting, etc.
To overcome these problems, some techniques like weight
decay and early stopping techniques must be utilized in
these existing learning techniques.
Figure 3: Structure of ELM Network
The ELM will attain the results directly without such
difficulties. The ELM learning algorithm is much simple
than the other learning techniques for feed-forward neural
networks. The existing learning techniques can be applied
to only differentiable activation functions, whereas the
ELM learning algorithm can also be used to train SLFNs
with many non-differentiable activation functions.
E. Extreme Learning Machine Training Algorithm
If there are N samples (x
i
, t
i
), where x
i
= [x
i1
, x
i2
… x
in
]
T
R
n
and t
i
= [t
i1
, t
i2
, … , t
im
]
T
R
n
, then the standard SLFN[16]
with N hidden neurons and activation function g(x) is defined
as:
=1
.
+
= 0
, = 1, . , ., (15)
where w
i
= [w
i1
, w
i2
, … , w
in
]
T
represents the weight vector
that links the ith hidden neuron and the input neurons, ß
i
= [ß
i1
,
ß
i2
, … , ß
im
]
T
represents weight vector that links the ith neuron
and the output neurons, and b
i
represents the threshold of the
ith hidden neuron. The “.” in w
i
. x
j
indicates the inner product
of w
i
and x
j
. The SLFN try to reduce the difference between o
j
and t
j
. This can be given as:
=1
.
+
=
, = 1, . , ., (16)
or, more in a matrix format as H ß = T, where
1
, . ,
,
, ,
,
1
, . ,
=
1
,
1
+
1
,
+
1
,
1
+
1
,
+
=
1
and =
1
(17)
The matrix H is the hidden layer output matrix of the
neural network. If the number of neurons in the hidden layer is
equal to the number of samples, then H is square and
invertible. Otherwise, the system of equations requires to be
solved by numerical methods, concretely by solving