Retinal Blood Vessel Segmentation by a MAS Approach

Carla Pereira

, Jason Mahdjoub

, Zahia Guessoum

, Luis Gonçalves

and Manuel Ferreira

Algoritmi Center, University of Minho, Campus de Azurém, 4800-058, Guimarães, Portugal

CReSTIC – MODECO, University of Reims, Rue des Crayères, 51100, Reims, France

Ophthalmology Service, Centro Hospitalar do Alto Ave, Guimarães, SA., 4835-044, Guimarães, Portugal

Keywords: Diabetic Retinopathy, Fundus Image Analyzing, Blood Vessels Segmentation, Multi-Agents System.

Abstract: Retinal blood vessels segmentation by color fundus images analysis has got huge importance for the

diabetic retinopathy early diagnosis. Several interesting computational approaches have been done in this

field, but none of them has shown the required performance due to the use of global approaches. Therefore,

a new approach is proposed based on an organization of agents enabling vessels detection. This multi-agent

approach is preceded by a preprocessing phase in which the fundamental filter is a Kirsch derivative

improved version. This first phase allows an environment construction where the agents are situated and

interact. Then, blood vessels segmentation emerges from agents’ interaction. According to this study,

competitive results were achieved comparing to those found in the present literature. It seems to be that a

very efficient system for the diabetic retinopathy diagnosis can be built using MAS mechanisms.

1 INTRODUCTION

Diabetic retinopathy (DR) has been presented as the

most common cause of blindness among working

age people. Retinal blood vessels segmentation has

got huge importance for the DR early diagnosis.

Numerous research efforts have been done in

segmenting blood vessels using image processing

techniques applied to fundus images. Some of the

approaches include matched filtering (Zhang et al.,

2010); machine-learning methods (Staal et al.,

2004), morphological operators (Mendonça and

Campilho, 2006). The main difficulties in vessels

accurate segmentation are pathologies presence,

noise, the low contrast between vasculature and

background, vessels width, brightness and shape

variability. To solve the variability problem, it is

important to adapt image interpretations, in loco,

instead of only applying one algorithm on the entire

image. A multi-agent system (MAS) is thus

proposed as a solution since agents allow several

algorithms cohabitation.

There are some studies reported in literature that

associate MAS to image processing in medical

images (Bovenkamp et al., 2004; Mahdjoub et al.,

2006; Richard et al., 2004). This association has

been revealed as a research expanded area. As far as

known, multi-agent approaches have never been

applied to retinal images. In this study, an approach

based on Mahdjoub et al. (2006) previous study is

applied to fundus images for the blood vessels

segmentation. This new approach uses some image

processing algorithms as concrete perception and

action tools to define autonomous agents which

interact among themselves and with environment

(image). Then blood vessel segmentation emerges as

a global behavior.

2 METHODOLOGY

The proposed approach uses a MAS model to

improve retinal blood vessels edges detection

resulting from a preprocessing phase. This

preprocessing phase consists of a conventional

image processing algorithms group providing

information (environment) for the MAS model.

2.1 Image Pre-processing

In this first step, we first use the preprocessing phase

from Niemeijer et al. (2005) developed study. In

order to remove noise from fundus image while

preserving the edges, Kuwahara filter (Kuwahara et

al., 1976) was then applied. Finally, a modified

Kirsch filter was employed to the image resulting

290

Pereira C., Mahdjoub J., Guessoum Z., Gonçalves L. and Ferreira M..

Retinal Blood Vessel Segmentation by a MAS Approach.

DOI: 10.5220/0004200602900293

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 290-293

ISBN: 978-989-8565-36-5

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

from the last step. The Kirsch filtering is an image

edges enhancing method using a basic convolution

filter eight times rotated. The improved Kirsch filter

(Mahdjoub et al., 2006) enables edges detection with

a two pixels thickness whose external edge is

represented by a positive or negative value, whereas

the internal edge has an opposite value (Fig. 1 a).

This enables the MAS model detection process since

the blood vessels gradient reveals a specific pattern

(Fig. 1 b), as they can be represented by two parallel

linear segments series. Thus, the agents search for

blood vessels edges by looking for this gradient

specific pattern.

Figure 1: a) Modified Kirsch filter resultant image where

the blue and white pixels represent negative and positive

gradient values, respectively. b) One section image

expanded version of a).

2.2 Multi-Agent System Model

MAS is composed by an agent set and their

environment. The environment contains the green

plane image in which each pixel contains the

intensity grey level and a Boolean value defining if

the pixel has already been explored by an agent.

Moreover, when located in the environment the

agents perceive the modified Kirsch gradient which

defines a right visible edge. Agents are of several

kinds with different behaviors according to their

current state and perception: search agents (SA),

following agents (FA), node agents (NA), and region

agents (RA).

MAS is initialized with a SA launched on one of

the white points from Fig. 1 a), randomly chosen.

This SA in the “operating” state has to find edges

belonging to blood vessel regions. It evolves in the

environment by following positive gradient points.

When it finds an edge, it initializes a new contour

and launches two NAs belonging to this contour:

one in the “active” state and another in the

“inactive” one (Fig. 2 a). Furthermore, the SA

changes its state to “suspended”. Then the “active”

NA has to allow contour extension and closure.

Therefore, it determines the possible directions to

follow the contour, creates FA and becomes

“inactive” (Fig. 2 b). FA follows the detected edge

until there is no direction to follow or until the

contour reaches a specific length (Fig. 2 c). FA

launches then an “inactive” NA on its position and

an “active” NA on the perpendicular direction where

it was moving (Fig. 2 d). Moreover, FA gives to the

“active” NA information about this direction

allowing it to launch another FA on the opposite

direction (Fig. 2 e). When this FA reaches the

initially launched “inactive” NA (Fig. 2 f) it

launches a RA (Fig. 2 g) which will be responsible

for the contour delimited region. RA sends a

message to SA to change its state to “operating” and

repeats all the process until all the blood vessels

contours are founded by MAS. There so, MAS

detects one contour each time avoiding regions

intersection at this phase. Afterwards SA sends a

message to all RAs to change their state to “filling”.

RA fills all the contours taking into account the

image grey levels (Fig. 3). Finally, RAs attempt

fusions with each other.

At the end of the process MAS has to reconstruct

the vessels by representing them with a succession

of regions initially represented by contours.

Figure 2: Contour formation graphical representation to

which a RA is assigned.

Figure 3: RA “filling” state graphical representation. It

analyses the pixels located between each pair of two

points belonging to its contour by determining the line

linking these two points a) – c); d) all the points that are

inside the contour, with grey level value similar to the

contour average grey level value, are added to the region.

3 RESULTS AND DISCUSSION

The proposed MAS model was implemented with

MadKit (Gutknecht and Ferber, 2000) and tested

RetinalBloodVesselSegmentationbyaMASApproach

291

with the publicly available DRIVE dataset (Staal et

al. 2004).

To measure the overall approach performance, it

is important to compare the resulting image with the

information detected by the Kirsch filter. Moreover,

the differences between the binary resultant image

with blood vessel segmentation and the ground truth

vessel map should be evaluated. In that way, three

common measurements, namely, sensitivity,

specificity and predictive value (Lalkhen and

McCluskey, 2008) were used for testing the

proposed algorithm. Fig. 4 and Fig. 5 show the

quantitative results obtained with this approach in

the DRIVE dataset. The proposed approach results

applied to normal retinal images are shown in Fig. 6.

These illustrate the original and resulting images

where the proposed approach had the best and the

worst performance respectively, with 73.6% and

51.9% sensitivity values and 97.4% and 99.7%

specificity values.

Figure 4: Sensitivity, specificity and predictive values

obtained for the 40 images of the DRIVE database with

the proposed approach.

Figure 5: Sensitivity values obtained for the 20 images of

the DRIVE database test set, using both hand labeled

databases.

Fig. 7 illustrates a superimposed image of the hand

labeled image with the hand labeled image after

morphological opening and with MAS result.

In this figure, the white pixels represent the pixels

common to the three images; the yellow and green

pixels represent the pixels that belong to blood

vessels but are not detected by MAS, that is, the

Figure 6: Images resulting where the proposed approach

had the best (above) and worst (below) performance in the

DRIVE database. From left to right: original color fundus

image; hand labeled image; blood vessel segmentation

using MAS approach.

false negative pixels; and the false positive pixels

are represented in blue. As it can be seen, the most

part of false positive pixels are located at the

manually segmented vessels border and therefore,

they should not be considered as false positive.

Actually, manual blood vessels segmentation from

retinal images is a very arduous and difficult task,

leading two people to segment the same image in

different ways. This can be observed in Fig. 5 where

two different hand labeled images for the same color

fundus images resulted in different sensitivity values

with the same approach.

Analyzing Fig. 6 (right) and Fig. 7 it can be

observed that MAS reconstructed the most part of

the vessels, especially the thickest ones. Some of the

thinnest vessels were also segmented but not all,

affecting the sensitivity values. In fact, after

removing the thinnest vessels from the hand labeled

image (green pixels of Fig. 7) the sensitivity values

of the proposed approach increased to values higher

than 80%. So, improvements have to be made in

MAS model to deal with the small vessels.

Moreover, there are some thickest vessels

portions not detected by MAS model mostly near the

FOV border and the optic disc contour. This last

problem may be related with the preprocessing

phase, mainly with Kuwahara filter since this often

produces clearly noticeable artifacts. (Papari, et al.,

2007).

Therefore, MAS is efficient in segmenting the

blood vessels from where the edges were already

detected in the preprocessing phase and in excluding

the detected pixels that did not belong to vessels.

Furthermore, MAS is able to close and aggregate

regions that were delimited by interrupted edges in

the kirsch resultant image, as a RA fusion process

result.

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

292

Figure 7: Hand labeled image superimposition with the

hand labeled image after the morphological opening and

with MAS result.

4 CONCLUSIONS

In this study, a MAS approach is proposed where

agents enrich a traditional edge detector algorithm.

The experiments show that the use of a MAS model

in the micro level could be an effective way to

segment structures in complex images such as retinal

images. In fact, through environment perception and

local interactions, a simple agent organization can

have as global behavior the most part of retinal

vasculature detection. The use of an improved

version of agent society with some knowledge a

priori about the retina proprieties, complemented

with some other traditional image processing

algorithms, could have the potential to develop a

system to detect and differentiate all the anatomic

and pathological structures of the fundus images.

Such an approach will overcome the classic image

processing algorithms that are limited to macro

results which cannot take into account the local

characteristics of a complex image. Therefore, it

could be a fundamental tool responsible for a very

efficient system development to be used in screening

programs concerning DR diagnosis.

ACKNOWLEDGEMENTS

Work supported by FEDER funds through the

"Programa Operacional Factores de Competitividade

– COMPETE" and by national funds by FCT-

Fundação para a Ciência e a Tecnologia. C. Pereira

thanks the FCT for the SFRH/BD/61829/2009 grant.

REFERENCES

Bovenkamp, E.G.P., Dijkstra, J., Bosch, J.G., and Reiber,

J. H. C. (2004). Multi-agent segmentation of IVUS

images. Pattern Recognition, 37:647-663.

Gutknecht, O., Ferber, J., (2000). Madkit: a generic multi-

agent platform. In: Proceedings of the 4th

international conference on Autonomous agents,

Barcelona, Spain, 78-79.

Kuwahara, M. , Hachimura, K., Eiho, S., and Kinoshita,

M. (1976) Digital Processing of Biomedical Images

(pp. 187-203). Plenum Press: New York.

Lalkhen, A. G., McCluskey, A. (2008). Statistics VI:

Clinical tests: sensitivity and specificity. Contin Educ

Anaesth Crit Care Pain, 8(6):221-223.

Mahdjoub, J., Guessoum, Z., Michel, F., and Herbin, M.

(2006). A multi-agent approach for the edge detection

in image processings. in 4th European Workshop on

Multi-Agent Systems, Lisbon, Portugal.

Mendonça, A., and Campilho, A. (2006). Segmentation of

retinal blood vessel by combining the detection of

centerlines and morphological reconstruction. IEEE

Trans. Med. Imag., 25(9):1200-1213.

Niemeijer, M., Ginneken, B. van, Staal, J., Suttorp-

Schulten, M., and Abràmoff, M.D.(2005). Automatic

detection of red lesions in digital color fundus

photographs. IEEE Trans. Med. Imag., 24(5): 584-

592.

Papari, G., Petkov, N., Campisi, P., (2007). Artistic edge

and corner enhancing smoothing. IEEE Transactions

on Medical Imaging 16 (10):2449 - 2462.

Richard, N., Dojat, M., and Garbay, C. (2004). Distributed

Markovian segmentation: Application to MR brain

scans. Pattern Recognition, 40:3467 – 3480.

Staal, J. J., Abrámoff, M. D., Niemeijer, M., Viergever, M.

A., and Ginneken, B. van (2004). Ridge-based vessel

segmentation in color images of the retina. IEEE

Trans. Med. Imag., 23(4): 501-509.

Zhang, B., Zhang, L., Zhang, L., and Karray, F. (2010).

Retinal vessel extraction by matched filter with first-

order derivative of Gaussian. Computers in Biology

and Medicine, 40:438-445.

RetinalBloodVesselSegmentationbyaMASApproach

293