Communications on Applied Electronics (CAE) – ISSN : 2394-4714

Foundation of Computer Science FCS, New York, USA

Volume 1 – No.5, April 2015 – www.caeaccess.org

Automated Segmentation of Optical Nerves by Neural

Network based Region Growing
Z. Faizal Khan, Ph.D. Syed Usama Quadri, Ph.D.
Department of Computer and Network Department of Computer and Network
Engineering, Engineering,
College of Engineering, Shaqra University, College of Engineering, Shaqra University,
Al Dawadmi, Kingdom of Saudi Arabia Al Dawadmi, Kingdom of Saudi Arabia

ABSTRACT segmentation are statistical methods, geometrical, structural,

Computer Aided Diagnosis (CAD) of retinal image has been a model based, signal processing methods, spatial domain
revolutionary step in the early diagnosis of diseases present in filters, Fourier domain filtering, Gabor and wavelet models
the eye. Developing an efficient and robust algorithm for have also been used in most works present in the literature [6,
optical nerve segmentation has been a demanding area of 7].
growing research of interest during the last two decades. The In this paper, the combination of ESNN along with the region
initial step in computer aided diagnosis of retinal image is growing algorithm have been used for effective segmentation
generally to segment the nerves present in it and then to of the CT retinal image.
analyze each area separately in order to find the presence of
pathologies present in it. This research reports on The remainder of this paper is organized as follows: Section 2
segmentation of the nerves by segmenting the retinal images discusses the features and methods proposed in related works.
using Echo State Neural Networks along with the combination Section 3 explains the problem definition obtained in this
of region growing algorithm. Region growing has been research work and highlights the advantages of segmenting
combined with ESNN in this work since it reduces the number retinal CT images. Section 4 provides the results. Section 5
of steps in segmentation for the process of identifying a tissue gives the conclusion on this work and also provides some
in the CT retinal image. The performance of this proposed possible future works.
segmentation is proved to be better when it is compared with
other existing conventional segmentation algorithms. From 2. RELATED WORKS
the experimental results, it has been observed that the Accurately segmentation should be done since, the
proposed segmentation approach provides better segmentation pathologies present on it may be on the boundary of the retina.
accuracy. Such pathologies will be lost and it reduces the detection
accuracy, if the entire retinal area is not segmented accurately.
Keywords Main goal of retinal region of interest segmentation is to
Contextual clustering, Segmentation Algorithm, Retinal separating the nerves corresponding to pathology region from
image. the voxels corresponding to the surrounding anatomy
Many techniques have been employed to the exudate
1. INTRODUCTION detection. Gardner et al. [8] proposed an automatic detection
Segmentation methods are useful for partitioning an image
of diabetic retinopathy using an artificial neural network. The
into multiple segments inorder to provide an effective
exudates are identified from grey level images. The fundus
representation of the objects present in it. Moreover, Image
image was analyzed using a back propagation neural network.
segmentation is also useful for locating objects and
The technique did not work well on low contrast images.
boundaries (lines, curves, etc.) and to assign a label to every
pixel in an image such that pixels with the same label share The thresholding and RRGS technique were widely used.
certain visual characteristics. Moreover, each of the pixels Sinthanayothin et al. [5] reported the result of an automated
from particular regions are similar with respect to some detection of diabetic retinopathy on digital fundus images by
characteristic or computed property, such as color, intensity, RRGS algorithm where the performance was measured on
or texture. At the same time, the adjacent regions are 10×10 patches rather on the whole image. Usher et al. [7]
significantly different with respect to the same characteristic detected the candidate exudates region by using a combination
[5]. of RRGS and adaptive intensity thresholding. The candidate
regions were extracted and used as input to a neural network.
The various regions obtained by segmenting an image can be
Poor quality images affected the separation result of bright
further used for different types of analysis and interpretations.
and dark lesions using thresholding and exudate feature
Therefore, segmentation of image involves extracting
extraction using RRGS algorithm.
important features and deriving the relevant metrics to
segregate regions of homogeneous intensities. In order to Zheng et al. [9] detected exudates using thresholding and a
achieve this, it is necessary to choose a selective region of region growing algorithm. The fundus photographs were
interest by considering the application requirements. In the taken with a non-mydriatic fundus camera and were then
past, many image segmentation methods have been proposed scanned by a flat-bed scanner. Colour normalization and local
by various researchers for performing successive image contrast enhancement followed by fuzzy C-means clustering
analysis. In addition, many researchers have used the existing and neural networks were used by Osareh et al. [10]. The
thresholding techniques for segmenting the various regions of system works well only on Luv colour space but in the case of
interest. In short, the most frequently used techniques for non-uniform illumination the detection accuracy is low. Mitra

9
 Communications on Applied Electronics (CAE) – ISSN : 2394-4714
Foundation of Computer Science FCS, New York, USA
Volume 1 – No.5, April 2015 – www.caeaccess.org

et al. [11] applied naïve Bayes classifier for diagnosis of dynamics of the ESNN is uniquely controlled by the input,
diseases from retinal image. A system can provide a good and the effect of the initial states vanishes. The current design
decision support to ophthalmologist. of ESNN parameters relies on the selection of spectral radius.
There are many possible weight matrices with the same
3. ESNN WITH REGION GROWING spectral radius, and unfortunately they do not perform at the
Region growing (RG) [1] is an iterative technique employed same level of mean square error (MSE) for functional
to identify connected regions of interest (contiguous sets of approximation.
voxels) in images, obeying some inclusion rule (generally
based on threshold values), and according to the notion of The recurrent network is a reservoir of highly interconnected
discrete connectivity [2]. Initially, the region growing starts dynamical components, states of which are called echo states.
by choosing an initial pixel as a seed point which is present in The memory less linear readout is trained to produce the
the region to be grown and, after checking its inclusion of output. Consider the recurrent discrete-time neural network
neighbors in the growing region based on the threshold value. with M input units, N internal PEs, and L output units. The
Each included voxel becomes in turn a seed point for the value of the input unit at time n is u(n) = [u1(n), u2(n), . . . ,
following iteration. The above process continues until all the uM(n)]T ,
pixels are added in the grown region based on the set of rules The internal units are x(n) = [x1(n), x2(n), . . . , xN(n)]T
and threshold. , and
In our approach, a region growing approach along with the Output units are y(n) = [y1(n), y2(n), . . . , yL (n)]T.
clustering is used to fix the threshold inorder to segment the
region of interest present in the CT retinal images. The initial
seed point is a voxel (3x3 or 5x5 pixels) belonging to the The connection weights are given
retinal nerve region, and the fuzzy rule fixes a value by
selecting the voxels with intensity values lower than the given
threshold. In this way the entire nerve is connected which
 in an (N x M) weight matrix W back  Wijback for
present in the image is starting from the bronchi, carena, and connections between the input and the internal PEs,
the trachea. The initial seed point is automatically chosen by
selecting the 3x3 pixel which is present in the central slice of  in an N × N matrix W in  Wijin for connections
the CT image and grows towards the entire region present in
the image. between the internal PEs

Recently, there has been considerable interest among  in an L × N matrix W out  Wijout for connections
researchers in statistical clustering techniques [3] in image
segmentation. In a clustering technique along with the region from PEs to the output units and
growing, each pixel is associated with one of the finite
number of threshold is grown to form disjoint regions. The  In an N × L matrix W back  Wijback for the
contextual clustering method proposed by [4, 9] is a connections that project back from the output to the
supervised algorithm. It uses a 3 X 3 overlapping windows of internal PEs.
pixels to form a segmented image. The quality of segmented
image depends upon 1) A defined threshold value (T=140) by The activation of the internal PEs (echo state) is updated
the user which is used to choose the nearest regions for according to
segmentation, 2) a controlling parameter β which is in the x(n + 1) = f(Win u(n + 1) + Wx(n) +Wbacky(n)),
range of 0< β <1, 3) the median value of the 9 pixels in the
window, 4) the total number of intensity values (u) >1 inside Where f = (f1, f2, . . . , fN) are the internal PEs‟ activation
the window, excluding the already identified median value. functions.

3.1 Echo state neural network e x  e x

Segmentation of the CT image has been done with Echo state Here, all fi‟s are hyperbolic tangent functions x . The
neural network [3]. ESNN is a recurrent neural network which e  e x
contains at least one cyclic path, where the same input output from the readout network is computed according to y(n
information repeatedly influences the activity of neurons on + 1) = fout(Woutx(n + 1)), .
the cyclic path. Such networks are more closely related to
biological neural networks, which are also mostly recurrent. Where f
out
 ( f1 , f 2 ,...., f L ) are the output unit‟s
out out out

The outputs of this internal PEs (echo states) are fed to a nonlinear functions. Generally, the readout is linear so fout is
memory less but adaptive readout network that produces the identity [14].
network output. The interesting property of ESNN is that only
the memory less readout is trained, whereas the recurrent ESNN algorithm consists of two phases namely Training
topology has fixed connection weights. This reduces the and Testing (segmentation)
complexity of RNN training to simple linear regression while Phase 1: Training ESNN
preserving a recurrent topology, but obviously places
important constraints in the overall architecture that have not Step 1: Read 31st slice in DICOM format and separate the
yet been fully studied. image into 3 x 3 overlapping blocks of pixels.
The echo state condition is defined in terms of the spectral Step 2: Decide the number of reservoirs.
radius (the largest among the absolute values of the
Step 3: Decide the number of nodes in the input layer = 3.
eigenvalues of a matrix, denoted by (|| ||) of the reservoir‟s
weight matrix (|| W || < 1). This condition states that the

10
 Communications on Applied Electronics (CAE) – ISSN : 2394-4714
Foundation of Computer Science FCS, New York, USA
Volume 1 – No.5, April 2015 – www.caeaccess.org

Step 4: Decide the number of nodes in the output layer = Step 1: Adopt step 1 and step 2 mentioned in Training.
number of target values=1.
Step 2: Calculate state_vector= tanh ((ILhL)weights*Inputpattern
Step 5: Initialize state vector (number of reservoirs)=0. +(hL)weights* state vectorpresent+(hLoL)weights * Targetpattern).
Step 6:Initialize random weights between input layer (I L) and Step 3: Estimated output= state_vector * Wout
hidden layer(hL). Initialize weights between output
layer(oL) and hidden layer (hL). Initialize weights in Step 4: Assign 0 (black) or 255(white) in the new matrix
the reservoirs. which will be the segmented image.

Step 7:Calculate state_vectornext = tanh 4. EXPERIMENTAL RESULTS AND

((ILhL)weights*Inputpattern +(hL)weights* state DISCUSSION
vectorpresent+(hLoL)weights * Targetpattern). Retina images have been considered in this paper. Computed
Step 8:Calculate, a = Pseudo inverse (State vector all patterns). tomography images of different patients have been taken from
the Drive database. The ESNN segmented results for different
Step 9:Calculate, Wout = a * T and store Wout for region growing threshold along with its original images has
segmentation. been presented in Table 1
Phase II: Testing (Segmentation)
Table 1 Segmented Results of Block Size 3x3

No Original Image Ground Truth Segmented Result

11
 Communications on Applied Electronics (CAE) – ISSN : 2394-4714
Foundation of Computer Science FCS, New York, USA
Volume 1 – No.5, April 2015 – www.caeaccess.org

Table 1 shows the segmented images obtained from the original algorithm are block size of the moving window is 3x3.
retina image. The parameters given for the variables of ESNN

500
Block size=3x3

NUMBER OF OBJECTS
400

300

200

100

0
80 100 120 140
THRESHOLD IN RG

Figure 1 Segmented Objects for Different Threshold at Block Size 3x 3

Figure 1 shows the number of objects calculated using region object is obtained when the threshold is less than 120. Correct
properties of Matlab for a block size of 3x3 moving window.. number of objects that is 67 is obtained when the threshold is
More number of segmented objects is present when the image 140.
grown to a threshold of 120 for region growing. Only one

Block size=3x3
120000
NUMBER OF PIXELS

100000
80000
60000
40000
20000
0
80 100 120 140
THRESHOLD IN RG

Figure 2 Segmented Pixels for Different Threshold at Block Size 3x 3

Figure 2 shows the number of pixels calculated using region types of retina datasets inorder to validate the efficiency of the
properties of Matlab for a block size of 3x3 moving window. proposed algorithm. In this proposed framework,
More number of segmented pixels is present when the image segmentation of the normal tissues is not degraded since the
is segmented with threshold less than 100. All the segmented unwanted section other than the region of interest is removes
pixels correspond to one object only. Keeping the ground exactly.
truth image into consideration, the number of pixels
corresponding 67 objects are found to be the correct number 6. REFERENCES
of segmented pixels (999). Matlab „Region props‟ function [1] B. S. Morse, Lecture 18: Segmentation (Region Based),
has been used and the number pixels and number of objects 1998-2000.
are shown in figure 1 and 2. Earlier researchers had used
[2] Giorgio De Nunzio, Eleonora Tommasi, Antonella
different metrics to evaluate the segmentation accuracy. In
Agrusti, Rosella Cataldo, Ivan De Mitri, Marco Favetta,
this report, we have used the „Region props‟ function to
Silvio Maglio, Andrea Massafra, Maurizio Quarta,
evaluate the accuracy of segmentation and it has been found
Massimo Torsello, Ilaria Zecca, Roberto Bellotti, Sabina
that the proposed segmentation approach is much better when
Tangaro, Piero Calvini, Niccolò Camarlinghi, Fabio
compared to that of remaining segmentation approaches
Falaschi, Piergiorgio Cerello, and Piernicola Oliva,
mentioned in the literature.
“Automatic retinal Segmentation in CT Images with
5. CONCLUSION Accurate Handling of the Hilar Region”, Journal of
In this paper we have proposed a new method for segmenting digital imaging, Vol 24, No 1, pp 11-27, 2011.
retina images using the supervised contextual clustering based [3] Dr.Z. Faizal Khan, Dr.G.Nalini Priya, Dr A. Kannan, „A
region growing method. The main purpose of proposing this novel Approach for Segmenting Computer Tomography
approach is to improve the segmentation accuracy by reducing Lung Images Using Echo State Neural Networks‟,
the false segmentation. Main features of the proposed Article in press, Journal of Theoretical and Applied
algorithm is the use of region growing along with the Echo Information technology.
State Neural network. This method is applied to different

12
 Communications on Applied Electronics (CAE) – ISSN : 2394-4714
Foundation of Computer Science FCS, New York, USA
Volume 1 – No.5, April 2015 – www.caeaccess.org

[4] Sinthanayothin C, Boyce JF, Williamson TH, Cook HL, [8] Zheng Liu, Opas C, Krishnan SM. Automatic image
Mensah E, Lal S. Automated detection of diabetic analysis of fundus photograph. In: Proceedings of the
retinopathy on digital fundus image. J Diabet Med International Conference on Engineering in Medicine
2002;19:105–12. and Biology, vol. 2. 1997. p. 524–5.
[5] Niemeijer M, van Ginneken B, Staal J, Suttorp-Schulten [9] Osareh A, Mirmehdi M, Thomas B, Markham R.
MS, Abramoff MD.Automatic detection of red lesions in Automated identification of diabetic retinal exudates in
digital color fundus photographs. IEEE Trans Med Imag digital colour images. Br J Ophthalmol 2003;87:1220–3.
2005;24:584–92.
[10] Mitra SK, Te-Won Lee, Goldbaum M. Bayesian network
[6] Usher D, Dumskyj M, Himaga M, Williamson TH, based sequential inference for diagnosis of diseases from
Nussey S, Boyce J. Automated detection of diabetic retinal images. Pattern Recogn Lett 2005;26:459–70.
retinopathy in digital retinal images: a tool for diabetic
retinopathy screening. Diabet Med 2004;21:84–90. [11] Purushothaman S and Suganthi D, 2008, fMRI
segmentation using echo state neural network,
[7] Gardner GG, Keating D, Williamson TH, Elliott AT. International Journal of Image Processing, Vol. 2, Issue
Automatic detection of diabetic retinopathy using an 1, pp.1-9.
artificial neural network: a screening tool. Br J
Ophthalmol 1996;80:940–4.

13