International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248

Volume: 4 Issue: 3 421 – 428

_______________________________________________________________________________________________
Robust Retinal Vessel Segmentation using ELM and SVM Classifier

T.Sumathi, Dr.P.Vivekanandan, Dr.Ravikanth Balaji,

Professor Of Eminence, Senior Consultant, MRI and PET
Research scholar, Department Of
Department Of Chemical Engineering, Divisions,
Computer Science,
Anna University, Chennai. Apollo Specialty Hospital, Chennai.
Anna University, Chennai,
[email protected] [email protected]
[email protected]

Abstract—The diagnosis of retinal blood vessels is of much clinical importance, as they are generally examined to evaluate and monitor both the
ophthalmological diseases and the non-retinal diseases. The vascular nature of retinal is very complex and the manual segmentation process is
tedious. It requires more time and skill. In this paper, a novel supervised approach using Extreme Learning Machine (ELM) classifier and Support
Vector Machine (SVM) classifier is proposed to segment the retinal blood vessel. This approach calculates 7-D feature vector comprises of green
channel intensity, Median-Local Binary Pattern (M-LBP), Stroke Width Transform (SWT) response, Weber’s Local Descriptor (WLD) measure,
Frangi’s vesselness measure, Laplacian Of Gaussian (LOG) filter response and morphological bottom-hat transform. This 7-D vector is given as
input to the ELM classifier to classify each pixel as vessel or non-vessel. The primary vessel map from the ELM classifier is combined with the
ridges detected from the enhanced bottom-hat transformed image. Then the high-level features computed from the combined image are used for
final classification using SVM. The performance of this technique was evaluated on the publically available databases like DRIVE, STARE and
CHASE-DB1. The result demonstrates that the proposed approach is very fast and achieves high accuracy about 96.1% , 94.4% and 94.5% for
DRIVE, STARE and CHASE-DB1 respectively.
Keywords: Retinal vessel segmentation, Stroke width transform, Frangi’s vesselness, Bottom-Hat Transform, Local Binary Pattern, Weber’s
Local Descriptor, ELM classifier.
__________________________________________________*****_________________________________________________

I. INTRODUCTION materials used in this work are described in section 3. The

methodology and implementation of the proposed system are
The segmentation of retinal vessels plays a vital role in the explained in section 4. The performance metrics and the
diagnosis of many diseases including the non-ophthalmic robustness of the proposed algorithm along with the
diseases. Due to the similar nature of the vasculature of both the experimental setup are presented in section 5. At last, the
brain and the retina, retinal vasculature analysis can also be used conclusion and future directions are explicated in section 6.
to diagnose cardiovascular disease and stroke [1] Apart from
medical imaging applications, the vasculatures in retinal images
may be related to biometric applications. The retinal vasculature II. RELATED WORKS
can be manually segmented using the existing image tools. But In recent years, many retinal vessel segmentation algorithms
the manual segmentation task is very tedious, time-consuming have been deployed which can be approximately categorized
job and also the accuracy of the result is highly dependent on the into five classes. They are machine-learning techniques, model-
physician’s proficiency. So, many automatic methods have been based, tracking-based, matched-filtering and multi-scale
evolved for retinal vessel segmentation to reduce the techniques. The machine learning techniques are further
dependency of the manual intervention and diminish error classified into supervised and unsupervised classification. Some
factors. In this paper, a novel method using the combination of of the latest articles which use machine learning techniques to
ELM classifier and SVM classifier is proposed to segment the extract the retinal vasculature are discussed below.
retinal vasculature. Considering that the green channel of the
retinal image possesses a high contrast between the vessels and Shahab Aslani et al. [4] proposed a Random Forest (RF)
the background [2], the green channel is used as an input for the classifier based retinal vasculature segmentation method. The
proposed segmentation method. The proposed method needs a method generates 17-D feature vector comprises of Gabor filter
pre-processing step to highlight the vessel information in the responses (13D), contrast-enhanced intensity(1D), Frangi’s
green channel image. The ELM classifier utilizes 7-D feature vesselness measure(1D), B-COSFIRE(Bar selective-
vector extracted from the pre-processed image to classify each Combination Of Shifted Filter REsponses) filter response(1D)
pixel in the fundus image as a vessel or non-vessel. The learning and morphological top-hat transformed intensity(1D) and this
speed of ELM classifier is faster when compared to the hybrid feature vector is used to train the RF classifier. The author
traditional feed-forward Neural Network (NN) algorithm [3]. uses multi-scale Gabor wavelet transform to produce the 13
The ridges are calculated from the enhanced bottom-hat Gabor features. The Gabor filter responses for 18 different
transformed image and are combined with the ELM vessel map. configurations are generated and which results in 54-
The final segmented image is obtained by classifying the dimensional feature vector comprises of real, imaginary and
combined image using SVM classifier. The 4-D feature vector magnitude part. Out of this 54-D feature vector , a subset of 13
comprises of the high-level features such as solidity, extent, most discriminative Gabor features are selected by using feature
eccentricity and form-factor. The high-level features are selection algorithm as described in [5] and are used for further
extracted from the regions in the combined image and are used classification. Since the traditional B-COSFIRE filter gives a
as input for the SVM classifier. low response at the vessel endings, the author uses a new
Organization of the paper is as follows: Section 2 discusses enhanced B-COSFIRE filter to circumvent this drawback and
some of the related work on retinal vessel segmentation. The
421
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________
International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248
Volume: 4 Issue: 3 421 – 428
_______________________________________________________________________________________________
efficiently detect vessel endings. A vessel probability map is gradient vanishing in back propagation and improve the
generated from the trained RF classifier and as a final step, a discriminative capability of features in lower layers of the deep
simple threshold value is applied over the probability map to network, an auxiliary deep supervision mechanism is
classify a pixel as a vessel or non-vessel. To measure the incorporated in some of the intermediate layers. The author uses
robustness of the method, the author also performs the cross- pre-trained filters in VGGNet to initialize the parameters of
training method using DRIVE and STARE datasets. The convolution layers [12]. The author proposed a network
performance of the method on 10 pathological images in STARE comprising a single stream deep network with the multiple
is better than the other existing methods. branch outputs which have different receptive fields. The
A new retinal vessel extraction technique based on the predicted vessel segmentation map can be obtained by fusing the
relation of neighbour pixels has been proposed by Yuji Hatanaka multi-scale segmentation map output from each branch. The
et al [5]. The author uses 105 mask patterns to calculate the shift- final segmentation result is obtained by applying the
invariant High-Order Local Auto-Correlation (HLAC) features thresholding scheme over the probability map.
for each pixel. Along with this 105 feature values, maximum
value, minimum value, mean and standard deviation of the pixel III. MATERIALS
values in the Region Of Interest (ROI) are given as input to the Similar to the existing retinal vessel segmentation
first Artificial Neural Network layer (ANN). The ANN-1 is a algorithms, the proposed technique is evaluated on publically
three-layered feed-forward network with twelve neurons in the available databases, namely, DRIVE database, STARE
middle layer and only 10,000 randomly selected pixel samples database and the CHSE-DB1 database.
from the 20 images are used for the training. To improve the
performance, the author uses one more ANN (ANN2) to classify A. DRIVE Database
the pixel as a vessel or non-vessel pixel. The 5-D feature vector The retinal images of the DRIVE (Digital Retinal Images for
comprises of output from ANN1, Gabor filter response, Double- Vessel Extraction) database were obtained from a diabetic
Ring filter response Black-Top-hat transform value and Green retinopathy screening program in the Netherlands [12]. Out of
channel intensity. This 5-D feature vector is given as input to the the 400 diabetic subjects between the age of 25 and 90, 40
five middle layered ANN2. The HLAC-based algorithm results photographs have been randomly selected and JPEG compressed
in better performance while extracting blood vessels with low to create the DRIVE database. The 7 images in the database
contrast. show the sign of mild diabetic retinopathy whereas the
Zhaohui Tang et al [7] proposed a retinal vessel remaining images are normal. The photos were captured using
segmentation method using a new feature named Influence Canon CR5 camera with 45-degrees Field-Of-View (FOV) and
Degree of Average Intensity (IDAI). The method includes three digitized using 8 bits per color plane at 768 by 584 pixels
stages. The first stage is the candidate regions detection to rule resolution. The dataset has been divided into two sets, namely,
out the false detection of optical interferences and lesions. In the the training set and test set with 20 images in each set. Two
next stage, a new feature IDAI is extracted based on the idea that manual segmentations of the retinal vasculature are available for
the vessel pixel will decrease the average intensity of the local the images in the test set, whereas, a single manual segmentation
region whereas the non-vessel pixel will increase the average result is available for the images in the training set. The DRIVE
intensity of the local region. In the last stage, supervised database can be downloaded at
classification using Support Vector Machine (SVM) classifier is http://www.isi.uu.nl/Research/Databases/DRIVE/.
performed to extract the vessel pixels. The author selects SVM
as the appropriate classifier by evaluating the performance of B. STARE Database
different classifiers like SVM, AdaBoost, K-Nearest Neighbours The STARE (Structured Analysis of the Retina) database
(KNN) and Neural Network over STARE database. Here, the contains two different set of retinal images and out of which the
candidate vessel like region detection using Hessian values second set has 20 retinal images with a reference indicating the
insures a pretty high true positive rate. The newly included IADI retinal vasculature. The first set has 81 images with a reference
feature can easily differentiate the vessel pixel from their indication the optic disk location. The images were captured by
neighbourhood background pixels. So, the segmentation results a TopCon TRV-50 fundus camera at 35-degrees FOV. The slides
are not only with a high sensitivity but also with a high were digitized using 8 bits per color channel at 605×700 pixels
specificity. resolution. The database is available online at
Eva Tuba et al [8] proposed an overlapping-block-based http://cecas.clemson.edu/~ahoover/stare/.
algorithm to extract the retinal vasculature by using SVM
classifier. In this algorithm, the input image is divided into C. CHASE-DB1 Database
blocks and features which are extracted from each block are used The CHASE_DB1 is a retinal vessel reference database comprises
to construct the input feature vector for the classifier. The 4-D 28 images acquired from 14 multi-ethnic school children in the program
input vector contains green channel intensity value of the centre Child Heart And Health Study in England. The database is available
pixel, the average green intensity value of the block, the standard online at
https://staffnet.kingston.ac.uk/~ku15565/CHASE_DB1/assets/CHASE
deviation of the green intensity value of the block and the feature DB1.zip. The images were captured using a hand-held Nidek NM-200-
based on Discrete Cosine Transform (DCT) coefficients. For the D fundus camera at 30-degrees FOV with the resolution of 1280×960
SVM, Radial Basis Function (RBF) is used as kernel function pixels. The database also contains manual segmentations done by two
and the parameters are tuned and set by using hierarchical grid independent observers.
search method. The method provides better accuracy when
compared to the existing local entropy thresholding based IV. METHODOLOGY
approach [9] and region growing based approach [10]. A new supervised approach for blood vessel detection based
A retinal vessel extraction method using deep convolution on ELM and SVM classifier has been proposed in this paper.
network has been proposed by Juan Mo et al [11] by leveraging This work comprises of the following stages.1)pre-processing 2)
multilevel hierarchical features. To overcome the problem of Feature Extraction 3) Primary vessel map generation using ELM
422
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________
International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248
Volume: 4 Issue: 3 421 – 428
_______________________________________________________________________________________________
classifier 4) ridge detection 5) SVM classification to obtain a ELM classifier is considered as the primary vessel map. The
final segmented image. Since the green channel image provides image generated from the bottom-hat filter is enhanced and is
a higher contrast between the vessel and the image background, used for the ridge detection. The detected ridges are combined
it is used as the input image for the proposed approach. The with the primary vessel map generated from the ELM classifier
block diagram of the proposed work is shown in Figure 1. and are further classified using SVM classifier to obtain the final
segmented image. The high-level features extracted from the
The 7-D feature vector is constructed from the features combined image are used for SVM classification. The following
extracted from the green channel image. The feature vector is sections explain about the each stage in detail.
given as input to the trained ELM classifier and the output of the

Read the input image

Extract the Green channel

Central light reflex

Contrast Enhancement Border Extension
removal

Feature Extraction

Green LBP SWT WLD

Hessian LoG Bot-Hat filter Response Enhance the filtered image

Ridge detection
Generate ELM vessel map

Combine the ELM map along with the ridges

Remove the isolated non-vessel regions using SVM classifier

Generate the final segmented image

Figure 1. Block diagram of the proposed approach

A. Pre-Processing
• Central light reflex removal
Colour retinal images often contain artifacts like lightning Since the green channel of the input image shows the highest
variations, poor contrast and noise. A pre-processing method contrast between the vessel pixels and the background, the green
comprising the following steps is used to reduce the layer is isolated from the input color image and is used for further
imperfections and generate the image more suitable for processing. Although the retinal vessels appear darker when
extracting the pixel features. compared to other retinal surfaces, a vessel with a light streak
which runs down the central length (central light reflex) may be
423
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________
International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248
Volume: 4 Issue: 3 421 – 428
_______________________________________________________________________________________________
misunderstood as two vessels. To remove the light streak, a its removal is shown in Fig. 2. To avoid the over-amplification
morphological opening using three- pixel diameter, disc defined of noise and uniformly distributes the grey-level values, contrast
in eight connectivity square grid as structuring element is limited adaptive histogram equalization (CLAHE) algorithm is
performed. An example of a vessel with central light reflex and applied over the central light reflex removal image.

(a) (b)

Figure 2. Central light reflex removal (a) Fragment of an input image containing central light reflex.
(b)After reflex removal.

captures the line-like structures to some extent and hence

• Border Extension considered for the classification.
The border extension technique proposed by George • Weber’s Local Descriptors (WLD)
Azzopardi et al., [14] is used to remove the artifacts produced WLD is a simple and robust local descriptor based on the
near the border of the camera aperture. Initially, every black Weber’s law [16]. In this proposed work, the differential
pixel which lies just on the exterior boundary of the FOV-mask excitation component in the WLD is used as a feature. The
is identified. For each black pixel, the mean value of the eight differential excitation component D_x for a pixel x is
neighbours inside the ROI is calculated and the pixel values are calculated using 3x3 neighborhood pixels as follows:
replaced with this average value. So, this initial iteration 𝐼𝐼(𝑚𝑚)−𝐼𝐼(𝑥𝑥)
increases the radius of ROI by one. This iteration is repeated for 𝐷𝐷𝑥𝑥 = 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 �∑𝑚𝑚∊𝑁𝑁(𝑥𝑥) 𝐼𝐼(𝑥𝑥)
� (1)
50 times and the radius of ROI is incremented by 50.
where I is the input image, N(x) is the set of neighborhood
B. Feature Extraction pixels.
The main aim of the feature extraction phase is to extract • Local Binary Pattern (LBP)
some quantifiable measurements from the input image which The LBP is a rotation and gradient invariant measure used
may be easily used in the classification stage to determine for classification in many diverse applications. Instead of the
whether the pixels belong to the blood vessel or not. The 7-D classical LBP, the LBPmed is used in this work to segment the
feature vector comprises of green channel intensity, Median- blood vessel. LBPmed is less sensitive to noise when
Local Binary Pattern (M-LBP), Stroke Width Transform (SWT) compared to the classical measure[17].The central pixel
response, Weber’s Local Descriptor (WLD) measure, Frangi’s value is replaced with the median of itself and the neighbors.
vesselness measure, Laplacian Of Gaussian (LOG) filter LBPmed for a central pixel (Pc) is calculated as follows :
8

∑ s( g
response and morphological Bottom-hat transform and it is used
for classification. All the feature values other than the green LBPmed (Pc ) = h− g~ ) 2h
h =0 (2)
channel intensity are calculated from the preprocessed image.
~
• Features Where g represents the median of the neighbours and the
• Green Channel Intensity central pixel, gh is the value of the neighbor pixel and the
The green channel image extracted from the colored input comparison function s(z) is given as follows:
image provides better vessel-background contrast. So, the 1 if z ≥ 0
green channel intensity of the pixel is considered as one of s( z ) = 
the features and is included in the feature vector. 0 otherwise
(3)
• Stroke Width Transform (SWT) The LBPmed operator assigns a label to every pixel of an
SWT is a local descriptor that computes per pixel the width
of the most likely stroke-like structure containing the pixel image by thresholding the 3*3 neighborhood of each pixel
[15]. The size of the SWT image is same as the input image with the median value and the result is considered as a binary
and each pixel contains the width of the stroke associated number.
with the corresponding input pixel. First, the edge of the • Hessian vesselness measure
input image is computed using the canny edge detection The Hessian-based vessel enhancement filters use the eigen
method. Then, for each edge pixel p, a ray along its gradient values from the Hessian matrix to determine the geometrical
direction is detected until it reaches another edge pixel q .The structures which are tubular in nature [18]. The eigen vectors
length of the ray k i.e. k = |p-q| is assigned as stroke width and the eigen values of Hessian matrix are closely related to
value to all the pixels which lie between p and q. If several the vascular direction and intensity. The vesselness function
rays intersect at a pixel, then the smaller value among the
stroke width values will be assigned to that pixel. The SWT v0 ( σ ) can be defined as follows [19]:

424
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________
International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248
Volume: 4 Issue: 3 421 – 428
_______________________________________________________________________________________________
0 , if λ2 > 0 conventional methods. In the existing gradient-based

  2   techniques, the issues like over-fitting and local-minima are
v0 ( σ ) =   2 
 RB  ×  1 − exp − S   , otherwise
exp − 2    2γ 2   needed to be resolved by using weight decay or early stopping

  2 β     techniques. But, in the ELM, the input-hidden weights are
(4)
In this work, the threshold factors β and ϒ are chosen as 0.5 randomly assigned and the hidden-out weights are calculated by
and 20 respectively. The parameters RB and S in the least square regression or ridge regression. ELM contains an
vesselness equation are calculated using eigen values λ1, λ2 input layer, hidden layer and an output layer. The structure of
of Hessian matrix as follows the ELM with ‘N’ input nodes and ‘P’ hidden nodes is shown
in Figure 3. The ELM algorithm works as follows:
λ1 Given an input of training set of ‘N’ samples, activation
𝑅𝑅𝐵𝐵 =
λ2 function g(x) and the number of hidden neurons ‘P’.
• Assign a random value to the input weight vectors ai
𝑆𝑆 = λ1 2 + λ 2 2 (5) and hidden node impact factor bj where i = 1...N and
j = 1..P.
The filter is applied at multiple scales varying from 1 to 10, • Calculate the hidden layer output matrix
and the maximum response is selected to be a final estimate 𝐻𝐻 =
of vesselness and included in the feature vector.
𝑔𝑔(𝑎𝑎1 , 𝑏𝑏1 , 𝑥𝑥1 ) … 𝑔𝑔�𝑎𝑎𝑝𝑝 , 𝑏𝑏𝑝𝑝 , 𝑥𝑥1 �
• Bottom-Hat Transform
Bottom hat transformation very well enhances the retinal � ⋮ … ⋮ �
blood vessels and so included in the feature vector and 𝑔𝑔(𝑎𝑎1 , 𝑏𝑏1 , 𝑥𝑥𝑁𝑁 ) … 𝑔𝑔�𝑎𝑎𝑝𝑝 , 𝑏𝑏𝑝𝑝 , 𝑥𝑥𝑁𝑁 � 𝑁𝑁𝑁𝑁𝑁𝑁
considered for classification. The bottom hat transform ( ) is • Calculate the output weight β using β = H+T where
applid over the preprocessed image and is calculated as H+ is the Moore-Penrose generalised inverse of H
follows: and T is the output of the training sample.
Tb( g ) = g ⊗ b − g (6) The ELM gives a result with a minimum training error and the
where g is the preprocessed image, b is the structuring best generalization performance. The minimum norm least-
square solution of Hβ = T is also unique.
element, ⊗ is a morphological closing operation. The
morphological operation is performed using a three-pixel
diameter disc as structuring element. I O
I
• Laplacian Of Gaussian(LoG) H
The Laplacian of an image is often used for edge detection n
since it highlights the rapid intensity change and it gives a p O
better result if it is applied over the smoothed image. Mostly, u I
the Laplacian is applied to an image which has been t H
previously smoothed with the Gaussian filter in order to
reduce its sensitivity to noise. The output image is obtained V O
by convolving the filtered image with an input image and it e I
can be expressed as follows: c
O ( x , y ) = L( x , y ) ∗ I ( x , y ) (7)
t H
o
 x + y −σ   
 exp − x + y
2 2 2 2 2
L( x , y ) =   r
 σ4   2σ 2  I O
   (8)
where O(x,y), L(x,y), σ(scaling factor) and I ( x , y ) are the
output image, Log filter image, standard deviation, and the
input image respectively. The preprocessed image is ELM Feature ELM Learning
convolved with LoG filters that vary in the value of their Figure 3. Structure Of the ELM
scale parameter σ i.e. from 0.2 to 0.8 and retain the
maximum response at every pixel as a feature value.
C. Extreme Learning Machine classifier D. Classification using ELM
Extreme learning machine is a simple machine learning The ELM classifier is trained with different configurations
algorithm for Single Hidden Layer Feed Forward Neural and the model which gives better accuracy is used for
Networks (SFLNs) [20] which randomly select the input classification. Some commonly used activation functions in
weights and analytically determines the output weights. Though ELM model are sigmoidal, fourier, hard-limit, wavelet,
it was initially proposed for SLFNs, it was then enhanced to the Gaussian function etc. In the proposed work the ELM model
generalized SLFNs.In ELM, the hidden node parameters are with 930 hidden neurons and the sigmoidal activation function
completely independent from the training data. There is no need is used for classification. In the training phase, it is found that
to spend much time in tuning and training learning machines the model with 930 hidden neurons using sigmoidal activation
except the predefined network architecture. The learning speed function performs well and produces results with a high
of the ELM is much faster when compared with the other accuracy. The performance comparison between ELM models
with different activation function is shown in Figure 4.
425
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________
International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248
Volume: 4 Issue: 3 421 – 428
_______________________________________________________________________________________________
E. Ridge Detection and SVM Classifier
0.955
0.954 sig
Accuracy

0.953 sin
0.952 hardlim

0.951 rad

0.95 tan
650 850 1050 tri
Number of hidden neurons

(a)
Figure 4. Performance comparison – ELM model with different configuration
sig-sigmoidal, sin-sinusoidal ,hardlim-hard-limit, rad – radial basis,
tan – hyperbolic tangent, tri-triangular
The classification time for ELM with a different configuration
is given in Figure. 5.It is found that the processing time for ELM
model with triangular activation function is low when compared
to the other functions. So, an analysis has been made to check
whether the ELM model with more number of neurons using
triangular function produces a better result with less processing
time. But, there is no significant improvement in performance
even by increasing the number of neurons up to 1500.So, the
model with sigmoidal function is finalized for classification.
The classification result obtained from the ELM classifier is (b)
considered as the primary vessel map.

9
Classification time (seconds)

8
sig
7 sin

6 hardlim
rad
5
tan
4
tri
650 850 1050 (c)
Figure 6. Ridge Detection (a)Bottom-Hat filtered image.
Number of hidden neurons (b)Enhanced Bottom-Hat image (c)Ridge image

Figure.5. Processing time for ELM model with different configuration.

( sig-sigmoidal, sin-sinusoidal ,hardlim-hard-limit, rad – radial basis,
tan – hyperbolic tangent, tri-triangular) Though the ELM classifier results with better performance,
it fails to detect the thin vessels. So, in order to cover the thin
vessels, the ridges detected from the enhanced bottom-hat
filtered image are combined with the primary vessel map. The
background lightening variation in the bottom-hat filtered
image is removed by subtracting the background image from
itself. The background image is generated by applying mean
filter over the bottom-hat filtered image. In this proposed work,
the mean filter of size 45 * 45 is used to generate the background
image. The ridge image for a sample image is shown in Figure
6.
The primary vessel map generated by the ELM classifier and
the ridge detected image is combined and is further classified
with the SVM classifier to produce the final segmented image.

426
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________
International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248
Volume: 4 Issue: 3 421 – 428
_______________________________________________________________________________________________
The high-level features namely, solidity, eccentricity, form how nearly the region is circular. The measure extent tells how
factor and extent are extracted from the regions in the combined the shape is close to the rectangle. The solidity measure tells
image and given as input to the SVM classifier. The SVM how the shape is concave and convex. The output generated by
classifier with radial basis function (RBF) kernel with sigma the SVM classifier is the final segmented image and the
value as 1.0 is used in this proposed work for final performance measure is calculated for that image. The final
classification. The form-factor and eccentricity measure shows segmented image for a sample image is given in Figure 7.

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

Figure 7 Final classification (a)ELM primary vessel map (b)Ridge image (c) Combined image (d)Final segmented image

V. EXPERIMENT AND RESULT sensitivity. The proposed algorithm is very fast and the average
processing time for the databases is given in Table 2.
The proposed approach is implemented in Matlab 2015 on a
PC with 360 GHz Intel Core i7 processor system and 20 GB VI. CONCLUSION AND FUTURE WORK
RAM. The ridgedetection plugin in ImageJ 1.5 software tool is In this proposed work, the retinal vasculature is extracted by
used to detect the ridges in the enhanced bottom-hat filtered using the combination of ELM, ridge detector and the SVM
image. The performance of the proposed method is qualified by classifier. The ELM classifier uses pixel-wise 7-D feature vector
comparing the results with the manually segmented results to generate the primary vessel map.
available for DRIVE, STARE and CHASE-DB1. The
performance is evaluated on the basis of sensitivity, accuracy, TABLE II. PROCESSING TIME FOR DRIVE, STARE AND CHASE-DB1
specificity and Area Under the Curve (AUC).The average value
of performance metrics is given in Table 1. The performance Processing Time
DRIVE STARE CHASE-DB1
(seconds)
result for each image in DRIVE and STARE databases are
shown in Table 3 and Table 4 respectively. Classification
16.22566 20.71069 38.8058059
Time
TABLE I. PERFORMANCE MEASURES FOR DRIVE, STARE AND
CHASE-DB1 Feature
24.84628 32.28129 67.0208153
Extraction Time
Measure DRIVE STARE CHASE-DB1
Whereas, the SVM classifier uses high-level features to remove
Accuracy 0.9615 0.9444 0.9450 the non-vessel regions from the combined image. The pixel-wise
Sensitivity 0.7120 0.6392 0.6060 features are generated from the well pre-processed image to
overcome the problem of homogeneity variations in the input
Specificity 0.9857 0.9827 0.9706 image. The proposed algorithm demonstrates the performance
AUC 0.8488 0.811 0.7883 advantage in terms of accuracy. But, some non-vessel structures
located near the blood vessels and are connected to the vessel
Though the final segmented image includes most of the major network and hence they are detected as vessel segments. So,
and thin vessels, the sensitivity of the resultant image is low due especially the pathological images results in low sensitivity. This
to the presence of false positives. The future work will be problem has not been resolved in the proposed work and will be
carried out including a postprocessing stage to remove the false considered as future work by including post-processing step by
positives and hence improve the performance in terms of employing some morphological operations.

427
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________
International Journal on Future Revolution in Computer Science & Communication Engineering ISSN: 2454-4248
Volume: 4 Issue: 3 421 – 428
_______________________________________________________________________________________________
TABLE III. PERFORMANCE RESULT ON DRIVE DATABASE [5] Hall,M.A.,Smith,L.A., "Practical feature subset selection for machine
learning", Computer Science ’98 Proceedings of the 21st Australasian
Computer Science Conference ACSC’98 (1998) pp.181–191.
Image [6] Yuji Hatanaka, Kazuki Samo, Mikiya Tajima, Kazunori Ogohara,
Accuracy Sensitivity Specificity AUC
No. Chisako Muramatsu, Susumu Okumura and Hiroshi Fujita,
1 0.9654 0.7840 0.9832 0.8836 “Automated blood vessel extraction using local features on retinal
2 0.9600 0.7212 0.9873 0.8543 images”, Proceedings SPIE – Conference ,Medical Imaging 2016:
3 0.9533 0.6828 0.9833 0.8330 Computer-Aided Diagnosis , vol. 9785, 2016.
4 0.9621 0.6540 0.9933 0.8237 [7] Zhaohui Tang, Jin Zhang and Weihua Gui ,“Selective Search and
5 0.9580 0.6518 0.9896 0.8207 Intensity Context Based Retina Vessel Image Segmentation”, journal
6 0.9571 0.6471 0.9905 0.8188 of medical systems,vol. 41, no. 3, 2017
7 0.9562 0.6598 0.9860 0.8229 [8] Eva tuba, Lazar Mrkela, Milan tuba, ”Retinal Blood Vessel
Segmentation by Support Vector Machine Classification”, 27 th
8 0.9536 0.6687 0.9805 0.8246
international conference Radioelektronika, 2017.
9 0.9622 0.6708 0.9879 0.8293
[9] Singh N.P, Kumar.R, and Srivastava.R, “Local entropy thresholding
10 0.9622 0.6848 0.9871 0.8360
based fast retinal vessels segmentation by modifying matched filter,”
11 0.9572 0.6924 0.9833 0.8378 in International Conference on Computing, Communication
12 0.9634 0.7119 0.9871 0.8495 Automation, pp. 1166–1170, 2015.
13 0.9550 0.6645 0.9865 0.8255 [10] Palomera-Perez .M, Martinez-Perez.M.E, Benitez-Perez.H, and
14 0.9607 0.7607 0.9783 0.8695 Ortega-Arjona.J, “Parallel multiscale feature extraction and region
15 0.9643 0.7513 0.9808 0.8660 growing: Application in retinal blood vessel detection,” IEEE
16 0.9638 0.7458 0.9854 0.8656 Transactions on Information Technology in Biomedicine, vol. 14, no.
17 0.9628 0.7025 0.9868 0.8447 2, pp. 500–506, 2010.
18 0.9678 0.7648 0.9853 0.8751 [11] Juan Mo1 and Lei Zhang “Multi-level deep supervised networks for
19 0.9750 0.8371 0.9875 0.9123 retinal vessel segmentation”, International Journal of Computer
20 0.9698 0.7840 0.9846 0.8843 Assisted Radiology and Surgery, pp 1-13, 2017.
Average 0.9615 0.7120 0.9857 0.8489 [12] Simonyan,K, Zisserman,A.,"Very deep convolutional networks for
large-scale image recognition", International Conference on Learning
Representation (ICLR), 2015.
TABLE IV. PERFORMANCE RESULT ON STARE DATABASE [13] Staal,J.J., Abramoff,M.D., Niemeijer,M., Viergever,M.A., van
Ginneken,B., "Ridge based vessel segmentation in color images of the
Image retina", IEEE Transactions on Medical Imaging, 2004, vol. 23, pp. 501-
Accuracy Sensitivity Specificity AUC
No. 509.
1 0.9346 0.6918 0.9593 0.8256 [14] George Azzopardi , Nicola Strisciuglio, Mario Vento and Nicolai
2 0.9544 0.6900 0.9699 0.8300 Petkov, “Trainable COSFIRE filters for vessel delineation with
3 0.9384 0.7134 0.9521 0.8328 application to retinal images”, Medical Image Analysis, vol. 19, pp 46-
57, 2015.
4 0.9440 0.5499 0.9747 0.7623
5 0.9366 0.6586 0.9693 0.8140 [15] Epshtein, B., Ofek, E.,Wexler, Y.: Detecting text in natural scenes with
stroke width transform. In: Proceedings of Computer Vision and
6 0.9406 0.6429 0.9790 0.8110 Pattern Recognition, IEEE Conference, 2010, pp. 2963–2970.
7 0.9467 0.6837 0.9860 0.8349
[16] Chen .J, Shan, S., He, C., Zhao, G., Pietikainen, M., Chen, X.,
8 0.9379 0.6144 0.9895 0.8019 Gao,W.:Wld: a robust local image descriptor. Pattern Anal. Mach.
9 0.9471 0.6771 0.9848 0.8309 Intell. IEEE Trans. 32(9), 1705–1720 (2010)
10 0.9178 0.5624 0.9819 0.7721 [17] Rodrigo Nava,Gabriel Cristóbal and Boris Escalante-Ramírez , "A
11 0.9510 0.6642 0.9880 0.8261 comprehensive study of texture analysis based on local binary
12 0.9609 0.7060 0.9943 0.8501 patterns", proceedings of SPIE 8436, Optics, Photonics, and Digital
13 0.9335 0.5833 0.9910 0.7872 Technologies for Multimedia Applications II, 2012.
14 0.9406 0.6217 0.9894 0.8055 [18] Daniel Jimenez-Carretero, Andres Santos, SjoerdKerkstra, R
15 0.9492 0.6273 0.9953 0.8113 DewiRudyanto, Maria J Ledesma-Carbayo, "3D Frangi-based lung
16 0.9214 0.5324 0.9913 0.7619 vessel enhancement filter penalizing airways , IEEE 10th International
17 0.9446 0.6323 0.9917 0.8120 Symposium on Biomedical Imaging (ISBI),pp. 926-929,Apr 2013
18 0.9749 0.6703 0.9949 0.8326 [19] Frangi,AF., Wiro J. Niessen, Koen L. Vincken and Max A. Viergeve,
19 0.9669 0.6563 0.9877 0.8220 "Multiscale vessel enhancement filtering" ,Medical Image Computing
20 0.9473 0.6061 0.9851 0.7956 and Computer-Assisted Intervention (MICCAI’98), Lecture Notes in
Computer Science, Springer, Berlin, Germany, vol. 1496, pp. 130–
Average 0.9444 0.6392 0.9828 0.8110
137,October 1998
[20] Qin-Yu Zhu,Chee-KheongSiew,”Extreme LearningMachine: A New
REFERENCES Learning Scheme of Feedforward
[1] Kanski ,J.J., “Clinical Ophthalmology: A Systematic Approach”, NeuralNetworks”,Neurocomputing.Vol.70,pp.489-501, 2006.
London, U.K.: Butterworth-Heinemann, 1989.
[2] Peng Fenga, Yingjun Pana, Biao Weia, Wei Jinb and Deling Mia,
"Enhancing retinal image by the contour let transform",Pattern
Recognition Letters, vol. 28,no. 4, pp.516–522,2007.
[3] Diego Marín, Arturo Aquino, Manuel Emilio Gegúndez-Arias, and
José Manuel Bravo, "A New Supervised Method for Blood Vessel
Segmentation in Retinal Images by Using Gray- Level and Moment
Invariants-Based Features", IEEE Transactions on Medical imaging,
Jan 2011. pp. 146-158
[4] Shahab Aslani and Haldun Sarnel, “A new supervised retinal vessel
segmentation method based on robust hybrid features”, Biomedical
Signal Processing and Control, vol. 30, pp. 1-12,2016.

428
IJFRCSCE | March 2018, Available @ http://www.ijfrcsce.org
_______________________________________________________________________________________