Arid Zone Journal of Engineering, Technology and Environment, April, 2017; Vol.

13(2):229-237
Copyright Faculty of Engineering, University of Maiduguri, Maiduguri, Nigeria.
Print ISSN: 1596-2490, Electronic ISSN: 2545-5818, www.azojete.com.ng

RETINAL VESSEL SEGMENTATION BASED ON PRIMAL-DUAL

ASYNCHRONOUS PARTICLE SWARM OPTIMISATION (pdAPSO) ALGORITHM
E. G. Dada
(Department of Computer Engineering, Faculty of Engineering, University of Maiduguri,
Maiduguri, Borno State. Nigeria)
Email address: [email protected]
Abstract
Acute damage to the retina vessel has been identified to be main reason for blindness and impaired vision all
over the world. A timely detection and control of these illnesses can greatly decrease the number of loss of sight
cases. Developing a high performance unsupervised retinal vessel segmentation technique poses an uphill task.
This paper presents study on the Primal-Dual Asynchronous Particle Swarm Optimisation (pdAPSO) method for
the segmentation of retinal vessels. A maximum average accuracy rate 0.9243 with an average specificity of
sensitivity rate of 0.9834 and average sensitivity rate of 0.5721 were achieved on DRIVE database. The
proposed method produces higher mean sensitivity and accuracy rates in the same range of very good
specicity.

Keywords: Retinal Vessel, Segmentation, Asynchronous Particle Swarm Optimisation, Primal Dual

1. Introduction
Image segmentation is one of the primary methods used in image processing. In the last few
years, the image processing methods are broadly used in different medical fields for initial
detection, separation and recognition of diseases. According to Mahalakshmi and Velmuruga
(2015), image processing (IP) can be defined as the evaluation of an image by means of
methods that can recognise shapes, colours and associations among them which the human
eyes cannot notice. The algorithm used for the image processing operates by using an
equation or series of equations to manipulate the pixels (picture element) of the image give is
used to produce a new image. The data from an image are turned to digits and different
computations are performed on the image data to generate an improved image. Image
segmentation algorithms have been broadly applied in different fields of human endeavour
such as medical, remote sensing, education, military, video analysis, robotics, and so on
(Zhang et al., 2015). Retinal vessel segmentation is the process of partitioning a retina image
into sections with related properties such as grey level, colour, texture, brightness, and
contrast (Zhang et al., 2015). The thin layer of the retina which forms the inner part of the
retinal vessel is removed from the original image during image analysis, and image
segmentation algorithms are used to divide the original image into different segments. The
key objective of the retina vessel segmentation is to divide the retina image into equally
exclusive sections so that each section in relation to the pixel concentration is harmonised to
a predetermined benchmark (Zhang et al., 2015).
In the branch of medicine that deals with the diagnosis and treatment of diseases and
disorders of the eye, digital photography and image analysis of retinal vessel are already
gaining ground. Kanski (1989) and Klonoff and Schwartz (2000) observed that these
techniques have been recently recognised as beneficial techniques in the identification and
treatment of some diseases like diabetic retinopathy (DR), retinopathy of prematurity (ROP)
which is pathological disorder of the retina and cardiovascular diseases (Mapayi et al., 2015).
According to World Health Organisation (WHO, 2016), DR and ROP are the chief reasons of

229
Dada: Retinal vessel segmentation based on primal-dual asynchronous particle swarm optimisation
(pdAPSO) algorithm.
AZOJETE, 13(2):229-237. ISSN 1596-2490; e-ISSN 2545-5818, www.azojete.com.ng

eye defect and impaired vision universally, early detection and management of these diseases
will help achieve signicant reduction of blindness cases (Gergely and Gerinec, 2009).
Ophthalmologists, with the aid of identified vessel network, focus on retinal vessel quality
assessment at some stage of the diseases diagnosis. Detection by physical examination and
testing of the retinal vessels in the fundus images has been a very cumbersome and laborious
job that needs competent and skilful people who are not readily available (Varughese et al.,
2008). Nevertheless, according to Marrugo et al. (2012), it is possible for ophthalmologist to
diagnose and effectively control the diseases with the help of automated segmentation and
systematic examination of the arrangement of blood vessels in the retina.

1.1. Primal-Dual Asynchronous Particle Swarm Optimisation (pdAPSO) Algorithm

PSO was originally proposed by James Kennedy and Russell Eberhart in 1995 (Kennedy et
al., 2000). The algorithm is made up of particles which have position and velocity. Each of
the particles of a swarm epitomizes a possible solution in PSO. The particles explore the
problem search space seeking for the best or at least a solution that is suitable. Each of the
particles changes their movement according to their own accumulated knowledge of moving
in the environment and that of their neighbours. In PSO (Xi) represent the position of a
particle, and (Vi) the velocity of the particle. The particles number is i, where i = 1,,N, and
N is the number of particles in the swarm. The velocity is the degree at which the subsequent
position is varying as regards the present position represent the velocity for the particle i. As
the algorithm begins, the position and velocity of the particles are given numerical values
haphazardly. This is followed by using Equations (1) and (2) to update the position and
velocity of the particles after successive iterations are conducted throughout the search.

As opined by Shi and Eberhart (1995), to prevent commotion, the value vi(,tm1) is fixed at vmax.
The reason is that the value of vmax is going to be extremely large if the scope of search is too
broad. Also, if vmax is very narrow, the extent of the search will be unreasonably reduced
thereby forcing the particles to do local exploration. The inertia weight is represented as w
(constriction factor) is the inertia parameter; this regulates algorithms searching properties.
Shi and Eberhart (1995) opined that it is better to commence the search using a larger inertia
value (a more global search) that is automatically decreased to the end of the optimization (a
more local search). Using inertia weight with smaller values mostly ensures fast convergence
as little time is wasted on the exploration of the global space (Aziz and Ibrahim, 2012). The
inclusion of w in the equation is to provide equilibrium between the global and local search
capability of the particles. Where the value of varies from 1 to 0. Judging from the results
of experiments that have been performed, the performance of the dynamic method in term of
convergence is superior to the one of linear method. Shi and Eberhart (2004) showed that
PSO having different swarm population has practically alike but not identical performance.
The values c1and c2 are two positive constants representing the cognitive scaling and social

230
Arid Zone Journal of Engineering, Technology and Environment, April, 2017; Vol. 13(2):229-237
ISSN 1596-2490; e-ISSN 2545-5818; www.azojete.com.ng

scaling factors which according to Kennedy et al. (2001) are usually set to 2. The stochastic
variable rand1() and rand2() has the distribution U (0, 1). These random variables are stand-
alone functions that infuse momentum to the particles. The most ideal position located so far
by the particle is denoted as pbesti,m. The best position attained by the neighbouring particles
is denoted as gbestm. There are two types of particles neighbourhood in PSO, and the type of
neighbourhood is what determines the value of gbestm.
Equation 2 is used in updating the position of the particles whereby the velocity is added
together with the earlier position and a new search is started from its former position.
Eberhart and Shi (2000) in their work bounded xi(,tm1) to avoid a situation whereby particles are
spending too much time in infeasible region. A problem dependent fitness function is used in
evaluating the superiority of xi(,tm1) . Assuming the present solution is superior to the fitness of
pbesti,m or gbestm then the new position will replace pbesti,m or gbestm accordingly. Unless the
condition for ending the search (either the iteration has reached its peak or we have gotten the
desired solution) this updating process will continue. The optimal solution is the best particle
found when the stopping criterion is satisfied (Aziz and Ibrahim, 2012). In the Asynchronous
PSO (APSO), pbesti,m or gbestm of a particle, its velocity and position are updated
immediately after computing their fitness and, as a consequence, they update it having
incomplete or imperfect information about the neighbourhood. This result into varieties in the
swarm since some of the information is from the previous iteration while some is from the
current iteration.
The primal-dual interior-point (PDIP) method is an excellent example of an algorithm that
uses the constraint-reduction methods. Mehrotra (1992) in his research work developed the
Mehrotras Predictor-Corrector PDIP algorithm, which has been executed in the majority of
the interior-point software suite for solving both linear and convex-conic problems (Frisch,
1955). The primal-dual methods are a new category of interior-point methods that have of
recent been practically employed for solving large-scale nonlinear optimization problems
(Wright, 1997). Primal-dual algorithm uses considerably less iteration compared to the
simplex method and the algorithm is able to generate the ideal solutions for a linear
programming problem in less than 100 iterations irrespective of the huge number of variables
involved in nearly all its implementations (Glavic and Wehenkel, 2004). The flowchart of the
pdAPSO algorithm is shown in Figure 1 below.

231
Dada: Retinal vessel segmentation based on primal-dual asynchronous particle swarm optimisation
(pdAPSO) algorithm.
AZOJETE, 13(2):229-237. ISSN 1596-2490; e-ISSN 2545-5818, www.azojete.com.ng

Figure 1: Primal- Dual-APSO (pdAPSO) algorithm.

The proposed algorithm for retina vessel segmentation is outlined below:

232
Arid Zone Journal of Engineering, Technology and Environment, April, 2017; Vol. 13(2):229-237
ISSN 1596-2490; e-ISSN 2545-5818; www.azojete.com.ng

1.1.1 pdAPSO Segmentation algorithm

Step 1: Initialize the number of clusters to K and number of particles to m.
Step 2: Initialize m sets of K random cluster centers to be used by m particles.

Step 3: For each particle, let each pixel x belong to a cluster in which it has the
smallest Euclidean distance to the centroid.

Step 4: Calculate new cluster centers; If the new cluster centers converge to the old
ones, go to the next step. Otherwise, go to Step 3.
Step 5: Save the best solution found so far performed by each particle as the pbest or
personal best solution.

Step 6: Save the best solution among the m personal best solutions found. Call it gbest
or global best solution.

Step 7: Update cluster centers of each particle according to the cluster center values of
the pbest and gbest solution.
Step 8: If the termination criterion is satisfied go to next step. Otherwise, go to Step 3.
Step 9: Output the optimal solution.

2. Experimental Setup
Experiments were carried out using MATLAB 2015a on an AMD A 10-7300 Radeon R6, 10
Compute Cores 4C+6G, 1.90 GHz, 8.00GB of RAM.

2.1 Performance Measures

In this section the experimental setup used to evaluate three algorithms proposed for retinal
vessel segmentation. The sensitivity, specicity and accuracy are the main performance
measures are described in this work. The algorithms were tested on the DRIVE database
described in Section 3.3.
Sensitivity = TP/ (TP + FN) (3)
Specificity = TN/ (TN + FP) (4)
Accuracy = (TP + TN)/ (TP + TN + FP + FN) (5)

where: TP = True Positive, TN = True Negative, FP = False Positive and FN = False

Negative.
The TP is the number of pixels correctly classied as vessel pixels, TN the number of pixels
correctly classied as non-vessel pixels, FN the number of pixels incorrectly classied as
non-vessel pixels and FP is the number of pixels incorrectly classied as vessel pixels. The
sensitivity (SE) is computed by dividing TP by the total number of vessel pixels in the ground
truth segmentation. The specicity (SP) is computed as TN divided by the total number of
non-vessel pixels in the ground truth.

233
Dada: Retinal vessel segmentation based on primal-dual asynchronous particle swarm optimisation
(pdAPSO) algorithm.
AZOJETE, 13(2):229-237. ISSN 1596-2490; e-ISSN 2545-5818, www.azojete.com.ng

3.2 Dataset used

The proposed technique was tested on the DRIVE (Digital Retinal Images for Vessel
Extraction) publicly available DRIVE (2016) database. The photographs for the DRIVE
database were captured from a diabetic retinopathy testing program in The Netherlands. The
tested population comprised of 400 diabetic patients between 2590 years of age. Forty (40)
photographs were haphazardly chosen. Out of these, 33 do not demonstration any symptom
of diabetic retinopathy while 7 indicated that they have been mildly affected by early diabetic
retinopathy. The images were obtained using a Canon CR5 non-mydriatic 3CCD camera that
has a 45 degree eld of view (FOV). Each image was captured using 8 bits per colour plane
at 768 by 584 pixels. The FOV of each image is circular with a diameter of approximately
540 pixels. For this database, the images have been cropped around the FOV. For each
image, a mask image is provided that delineates the FOV. The set of 40 images were divided
into training and test sets, both them containing 20 images each. For the training images, a
single manual segmentation of the vasculature was available. For the test cases, two manual
segmentations were available; one was used as gold standard, the other one was used to
compare computer generated segmentations with those of an independent human observer.

3. Results and Discussion

The results below (Figure 2 a - e) show the images produced by the pdAPSO algorithm.

Figure 2a: Example of DRIVE data: (a) retinal image; (b) mask image; (c) ground truth
segmentation

Figure 2b: Example of DRIVE data: (a) retinal image; (b) mask image; (c) pdAPSO
segmentation

234
Arid Zone Journal of Engineering, Technology and Environment, April, 2017; Vol. 13(2):229-237
ISSN 1596-2490; e-ISSN 2545-5818; www.azojete.com.ng

Figure 2c: Example of DRIVE data: (a) retinal image; (b) mask image; (c) pdAPSO
segmentation

Figure 2d: Example of DRIVE data: (a) retinal image; (b) mask image; (c) pdAPSO
segmentation

Figure 2e: Example of DRIVE data: (a) retinal image; (b) mask image; (c) pdAPSO
segmentation

The performance of the proposed algorithm which is based on average sensitivity, average
specificity, and average accuracy is depicted in the table 1 below:

Table 1: Performance of pdAPSO Segmentation methods on DRIVE Database

Method Average Sensitivity Average Specificity Average Accuracy
pdAPSO 57.21% 98.34% 92.43%

235
Dada: Retinal vessel segmentation based on primal-dual asynchronous particle swarm optimisation
(pdAPSO) algorithm.
AZOJETE, 13(2):229-237. ISSN 1596-2490; e-ISSN 2545-5818, www.azojete.com.ng

From the Table above, the pdAPSO algorithm has the best average accuracy rate 0.9243 with
an average specificity of sensitivity rate of 0.9834 and average sensitivity rate of 0.5721. In
figure 2a is presented the retina image, mask image, and the ground truth segmentation image
obtained from the DRIVE database. Figure 2(b e) also gives the pictorial depiction of the
output images of the results obtained from vessels segmented using pdAPSO method. The
proposed pdAPSO technique detected a greater part of the large and thin vessels, while a very
few thinner vessels remain undiscovered. The output images also shows that there are some
false detection and other relic near the boundary of the optic disc. The false detection near the
border of the optic disc are however fewer on the segmented vessels when compared with
some output produced by some existing techniques in literature. This is due to the robustness
of the PSO algorithm that enables it to generate segmented image that is an improvement of
the original retinal image.

4. Conclusion
A study on the use of pdAPSO for the segmentation of retinal vessels was presented this
paper. It was demonstrated that the proposed vessel segmentation technique is time efcient
and yields a high average accuracy, specicity rates and very good average sensitivity while
compared to some segmentation methods on DRIVE database. In our future work, the
performance of pdAPSO method will be compare with other state of the art segmentation
algorithm. The proposed algorithm will also be applied to solve the brain tumour
segmentation problem.

References
Aziz, NAA. and Ibrahim, Z. 2012. Asynchronous Particle Swarm Optimization for Swarm
Robotics, International Symposium on Robotics and Intelligent Sensors (IRIS 2012).
Procedia Engineering, vol. 41, pp. 951 957.
Eberhart, RC. and Shi, Y. 2000. Comparing Inertia Weights and Constriction Factors in
Particle Swarm Optimization, In: Proceedings of the Congress on Evolutionary Computation,
Vol. 1 pp. 84 88.

Frisch, KR. 1995. The logarithmic potential method of convex programming, Technical
Report, University Institute of Economics, Oslo, Norway.

Gergely, K. and Gerinec, A. 2009. Retinopathy of prematurity epidemics, incidence,

prevalence, blindness, Bratislavske lekarske listy, vol. 111, no. 9, pp. 514 - 517.

Glavic, M. and Wehenkel, L. 2004. Interior Point Methods: A Survey, Short Survey of
Applications to Power Systems, and Research Opportunities, Technical Report. University of
Lige Electrical Engineering and Computer Science Department Sart Tilman B-28 4000
Liege, Belgium.

Kanski, JJ. 1989. Clinical Ophthalmology: A Systematic Approach. Published by

Butterworth-Heinemann Ltd, ISBN 10: 0750614293, Dunfermline, United Kingdom.

236
Arid Zone Journal of Engineering, Technology and Environment, April, 2017; Vol. 13(2):229-237
ISSN 1596-2490; e-ISSN 2545-5818; www.azojete.com.ng

Kennedy, J. and Eberhart, RC. 1995. Particle swarm optimization, in: Proceedings of the
International Conference on Neural Networks, IEEE Press, Piscataway, NJ, vol. 4, pp. 1942-
1948.

Kennedy, J., Eberhart, RC. and Shi, Y. 2001. Swarm Intelligence, Morgan Kaufmann
Publishers Inc. San Francisco, CA, USA.

Klonoff, DC. and Schwartz, DM. 2000. An economic analysis of interventions for diabetes,
Diabetes Care, Vol. 23, no. 3, pp. 390-404.

Mahalakshmi, S. and Velmuruga, T. 2015. Detection of Brain Tumor by Particle Swarm

Optimization using Image Segmentation. Indian Journal of Science and Technology, vol 8,
no. 22, pp. 13 19.

Mapayi, T., Tapamo, JR. and Viriri, S. 2015. Retinal Vessel Segmentation: A Comparative
Study of Fuzzy C-means and Sum Entropy Information on Phase Congruency International
Journal of Advanced Robotic Systems, vol. 12, no. 133, pp 1-11.
Marrugo, AG., Milln, MS., Cristbal, G., Gabarda, S., Sorel, M., and Sroubek, F. 2012. Image
analysis in modern ophthalmology: from acquisition to computer assisted diagnosis and
telemedicine, In SPIE Photonics Europe, International Society for Optics and Photonics,
June, 2012, pp. 84360C-84360C.

Mehrotra, S. 1992. On the implementation of a primal-dual interior point method, SIAM

Journal on Optimization, vol. 2, pp. 575-601.

Research Section Digital Retinal Image for Vessel Extraction (DRIVE) Database 2016.
Utrecht, The Netherlands, Univ. Med. Center Utrecht, Image Sci. Inst. [Online]. Available
on: http://www.isi.uu.nl/Research/Databases/DRIVE.

Varughese, S., Gilbert, C., Pieper, C. and Cook, C. 2008. Retinopathy of prematurity in South
Africa: an assessment of needs, resources and requirements for screening programmes,
British Journal of Ophthalmology, vol. 92, no. 7, pp. 879- 882.

World Health Organization 2016. Prevention of Blindness and Visual Impairment. Available
at: http://www.who.int/blindness/causes/priority/en/index8.html

Zhang, W., Zhou, C. and Bao, X. 2015. Investigation on digital media image processing
algorithm based on asynchronous and inertia adaptive particle swarm optimization.
International Journal of Signal Processing, Image Processing and Pattern Recognition, vol. 8,
no. 2, pp. 6576.

237