SEGMENTATION AND ANALYSIS OF RETINAL VASCULAR TREE

FROM FUNDUS IMAGES PROCESSING

Sandra Morales

, Valery Naranjo

, Jes

us Angulo

, Juan Jos

e Fuertes

and Mariano Alca

niz

1,3

Instituto Interuniversitario de Investigaci

on en Bioingenier

ıa y Tecnolog

ıa Orientada al Ser Humano

Universitat Polit

ecnica de Val

encia, I3BH/LabHuman, Camino de Vera s/n, 46022 Valencia, Spain

Centre de Morphologie Math

ematique, Math

ematiques et Syst

emes, MINES Paristech, Paris, France

Ciber, Fisiopatolog

ıa de Obesidad y Nutrici

on (CB06/03), Instituto de Salud Carlos III, Madrid, Spain

Keywords:

Retinal vessel segmentation, Retinal vascular tree, Vessel caliber, Bifurcation angle.

Abstract:

From a fundus image, the system proposed in this paper automatically detects retinal vessels and measures

some geometrical properties on them such as caliber and bifurcation angles. Its goal is to establish objective

relations between different vessels, thus being able to determine cardiovascular risk or other diseases, as well

as to monitor their progression and response to different treatments. The proposed approach is mainly based on

mathematical morphology although also incorporates curvature evaluation for the detection of retinal vascular

tree. This method has been validated on a public database improving the results of previous published methods.

1 INTRODUCTION

Retinal vasculature is able to indicate the status of

other vessels of the human body. Classically, its study

is included in the standard screening of any patients

with cardiovascular risk and other diseases in which

the vessels may be altered inasmuch as it is a non-

invasive or minimally invasive procedure.

Nowadays, due to high resolution of digital fun-

dus images, they can be automatically processed pro-

viding invaluable help to clinicians in diagnosis and

disease prevention. Speciﬁcally, a system capable of

detecting the retinal vessels and measuring some geo-

metrical properties has been developed from a fundus

image. Vascular changes produced in systemic dis-

eases usually induce particular modiﬁcations in the

vessels, such as changes in the angle of intersection

between arteries and veins, and changes in the vessel

calibers. Based on these facts, the goal of the pro-

posed system is to establish objective relations be-

tween the different vessels, to determine cardiovas-

cular risk or other diseases, as well as to monitor

progress and response to different treatments. This

tool has been initially applied in a study carried out

by the Department of Pediatrics of General Hospital

of Valencia about the retinal microvascular architec-

ture of children with low birthweight, analyzing the

relationship between the measurements obtained in

these children and diseases such as hypertension and

cardiovascular problems in adult life, thus using their

results as a prognostic marker for this type of pathol-

ogy.

At ﬁrst, a segmentation process to characterize

any retinal morphological changes is necessary. Af-

terwards, detected vessels must be labeled as a means

to be able to perform desired measures on them and

to quantify these changes.

The mainly methodology proposed in this paper

is, above all, focused on mathematical morphology

although also incorporates curvature evaluation to get

a correct detection of retinal vascular tree.

2 SEGMENTATION METHOD

Although fundus images are RGB images, in the

present work we draw on monochrome images ob-

tained from the green band which provides a im-

proved visibility of retinal blood vessels.

The segmentation method presented in this paper

is based on mathematical morphology and curvature

evaluation for the detection of retinal vascular tree.

Mathematical morphology (Soille, 2003) is a non-

linear image processing methodology based on min-

imum and maximum operations whose aim is to ex-

tract relevant structures of an image. This is achieved

by probing the image with another set of known shape

321

Morales S., Naranjo V., Angulo J., José Fuertes J. and Alcañiz M..

SEGMENTATION AND ANALYSIS OF RETINAL VASCULAR TREE FROM FUNDUS IMAGES PROCESSING.

DOI: 10.5220/0003704603210324

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2012), pages 321-324

ISBN: 978-989-8425-89-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

called structuring element (SE).

First, a small opening, using a disc of radius 1 as

SE, is performed on the enhanced green component

image to ﬁll in any gaps of the vessels which could

provoke errors in segmentation. To enhance the image

a local shade correction (Walter and Klein, 2002) is

applied.

Then, a dual top-hat, with a SE larger than the

biggest vessel, is applied with the goal of extracting

all of them and eliminating structures with high cur-

vature that are not vessels, as occurs in the optic disc.

After that, to highlight the vessels on the back-

ground, the next steps are followed. Principal curva-

ture is calculated as the maximum eigenvalue of the

Hessian matrix (Martinez-Perez et al., 2007). Subse-

quently, a linear opening by reconstruction is applied

to reconstruct the principal curvature from a supre-

mum of openings which were calculated using a line

of size 31 as SE every 15

◦

. This operation removes

any structure smaller than this SE in any orientation

obtaining a cleaner background.

Finally, to binarize the resulting image a k-means

clustering (MacQueen, 1967) is used with a k value

equals to 3. Two of the three obtained clusters are

deﬁned as vessel. Three classes are requiered because

thick and thin vessels can be very different.

Figure 1 shows the resulting images of the men-

tioned method.

(a) (b) (c)

(d) (e) (f)

Figure 1: Segmentation steps: (a) Original fundus image,

(b) Green component enhancement, (c) Dual top-hat ﬁlter-

ing, (d) Principal curvature, (e) Opening by reconstruction

and (f) k-means clustering (segmented image).

3 RETINAL TREE LABELING

Retinal vascular tree labeling is focused on obtain-

ing the skeleton of vascular tree, detecting signiﬁcant

points and a tracking process.

The skeleton of the vascular tree is obtained by

a thinning process from the segmented binary image

(Soille, 2003). After, a pruning process is applied to

eliminate possible spurs and to avoid multiple paths

that are inherent in a 8-connected boundary.

In a vascular tree skeleton there are three types of

signiﬁcant points and all of them must be detected:

terminal, bifurcation and crossing points (Martinez-

Perez et al., 2002). To locate the terminal and bifur-

cation points the hit-or-miss transform is utilised. The

hit-or-miss transform is a binary morphological oper-

ation that can be used to look for particular patterns

in an image (Soille, 2003). Figure 2 shows the SEs

used to detect these points on a skeleton. 1’s deﬁne

the set of pixels that should match the foreground and

0’s the background. Notice that all of them must be

used in all its orientations, one every 90

. On the other

hand, to detect the crossing points it will be necessary

a manual intervention due to there are some of them

that are incorrectly deﬁned as bifurcation points.

0 1

1 0 1

(a)

1 0 1

(b)

1 1

(c)

0 1 0

0 0 0

(d)

Figure 2: Structuring elements used to detect signiﬁcant

points: (a-c) bifurcation points and (d) terminal points.

The tracking purpose is to analyze the direction of

every pixel of the skeleton and to identify which pix-

els belong to each branch. The method used is pro-

posed in (Martinez-Perez et al., 2002). The algorithm

is able to specify the direction of any skeleton point

from a starting point. In order to keep the relation in-

formation between the branches, a determinated num-

bering scheme is employed.

4 GEOMETRIC MEASURES

Certain geometric measurements of blood vessels can

help to establish whether they have undergone mor-

phological changes over time and facilitate diagnose

illness. We focus on next parameters due to the fact

that they have particular interest for the early hyper-

tension detection:

Vessel Caliber. Is equal to two times the average of the

geodesic distance calculated from the skeleton points

of the branch to the edge of the corresponding vessel.

(3a).

Bifurcation Angle. Angle formed by the daughter

branches for each bifurcation point. The branches are

ﬁtted for straight lines by least-squares into a circular

window centered on these points (3b).

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

322

(a) (b)

Figure 3: Measures: (a) Caliber and (b) Bifurcation angles.

5 RESULTS

To validate the segmentation, our method results

have been compared with a public database of hand-

segmented images. In DRIVE database (Staal et al.,

2004) original images and the images manually seg-

mented by two different specialists are included. The

ﬁrst-observer images are taken as reference (gold

standard) to calculate similarity degree between them

and our segmentation. The segmented images by the

second observer also has been compared with the gold

standard in order to note that there are also differences

between both experts.

The performance of the method has been evalu-

ated based on three concepts: accuracy (Ac) and true

positive (TPF) and false positive (FPF) fractions. In

table 1 these results can be observed. In summary, the

average values obtanained by the proposed method

are: Ac=0.9417, TPF=0.6570 and FPF=0.0166.

On the other hand, in (Niemeijer et al., 2004) a

study about the accuracy of different vessel segmen-

tation methods was carried out on the same database.

This study compared several approaches: matched ﬁl-

ter (Chaudhuri et al., 1989), scale-space analysis and

region growing (Martinez-Perez et al., 1999), math-

ematical morphology and curvature estimation (Zana

and Klein, 2001), veriﬁcation-based local threshold-

ing (Jiang and Mojon, 2003) and pixel classiﬁcation

(Niemeijer et al., 2004). Table 2 shows a comparison

between the aforementioned methods and ours.

We can conclude that our method has achieved a

higher accuracy ratio and its false positive fraction is,

even, lower than the second observer fraction.

Table 1: Comparison between the results of the proposed

method and the 2

observer (average values and standard

deviations) regarding the gold standard.

Proposed method 2

observer

Ac 0.9417 (0.0076) 0.9473 (0.0048)

TPF 0.6570 (0.0668) 0.7757 (0.0596)

FPF 0.0166 (0.0093) 0.0275 (0.0083)

Table 2: Accuracy (average and standard desviation) of sev-

eral methods on the same public database.

Proposed method 0.9417 (0.0076)

observer 0.9473 (0.0048)

(Niemeijer et al., 2004) 0.9416 (0.0065)

(Zana and Klein, 2001) 0.9377 (0.0077)

(Jiang and Mojon, 2003) 0.9212 (0.0076)

(Martinez-Perez et al., 1999) 0.9181 (0.0240)

(Chaudhuri et al., 1989) 0.8773 (0.0232)

6 COMPARISON WITH OTHER

METHODS

In the literature there are numerous techniques for

the extraction of blood vessels, but not too focused

on the extraction of retinal vessels. Most edge de-

tection algorithms assess changes between pixels val-

ues by calculating image gradient magnitude and

then it is thresholded to create a binary edges im-

age (Martinez-Perez et al., 2007; Jiang and Mojon,

2003). Matched ﬁlters are ﬁlters rotated in differ-

ent directions in order to identify the cross section

of blood vessels (Hoover et al., 2000).Artiﬁcial neu-

ral networks use a ”weight” to determine probability

that some input data belonging to a particular output;

this system must be adjusted using known output data

(Sinthanayothin et al., 1999). Morphological pro-

cessing exploits vessels characteristics known a priori

(line connected segments) and combines morpholog-

ical operators to achieve the segmentation (Zana and

Klein, 2001; Walter and Klein, 2002).

Techniques based on edge detectors lack of

strength in distinguishing which are desired edges and

which are not, in our case, vessels and optic disc.

Matched ﬁlters have difﬁculty adapting to changes in

width and orientation of vessels. Techniques focused

on neural network are difﬁcult to automate for any

image and need a lot of training images. And mor-

phological processing runs the risk of fragmenting the

vessels.

7 CONCLUSIONS AND FUTURE

WORK

From a fundus image, the implemented system auto-

matically detects blood vessels of a speciﬁc region of

the image. Moreover, it allows to measure bifurca-

tion angles found and to select branches to know their

caliber after a manual intervention, in case it was nec-

essary. These data facilitate expert medical diagnosis

SEGMENTATION AND ANALYSIS OF RETINAL VASCULAR TREE FROM FUNDUS IMAGES PROCESSING

323

and study of the progression of a disease.

The method proposed for the segmentation takes

advantage of some of the techniques mentioned in

the previous section, besides new contributions, to

achieve the correct segmentation and to avoid the dis-

advantages of using them separately. Speciﬁcally, it

is based mainly on morphological operators but also

used principal curvature information, allowing an en-

hanced detection of the vessels on to the background

of the image. This algorithm has been compared with

several methods and, apart from improving the accu-

racy ratio, it must be stood out that the optic disc edge

is not detected as vessel unlike the most edge detec-

tion methods. Figure 4 shows our method along with

the proposed in (Martinez-Perez et al., 2007). The

non-detection of this feature is crucial to avoid multi-

ple errors in the tracking process.

(a) (b) (c)

Figure 4: Segmentation comparison: (a) Proposed method,

(b) (Martinez-Perez et al., 2007) method and (c) Manually

segmented image belonging to the DRIVE database.

On the other hand, the included measures are ac-

curate and reliable but also dependent on a correct

image analysis and rectiﬁcation of some signiﬁcant

points by the user.

About future work lines, a clinical validation will

be carried out to determine the speciﬁcity of the sys-

tem to distinguish between healthy and ill patients.

Afterwards, the method will be applied to analyze

the retinal microvascular architecture of children with

low birthweight and to use it as a prognostic marker

of cardiovascular risk.

ACKNOWLEDGEMENTS

This work has been funded by the project IMIDTA/

2010/47 and partially by projects Consolider-C

(SEJ2006-14301/PSIC), ”CIBER of Physiopathology

of Obesity and Nutrition, an initiative of ISCIII” and

Excellence Research Program PROMETEO (Gener-

alitat Valenciana. Conselleria de Educaci

on, 2008-

157). We would like to express our deep gratitude

to Imex Clinic S.L., the Department of Pediatrics of

General Hospital of Valencia and the Fundaci

on Of-

talmol

ogica del Mediterr

aneo for its participation as

well as the DRIVE database.

REFERENCES

Chaudhuri, S., Chatterjee, S., Katz, N., Nelson, M., and

Goldbaum, M. (1989). Detection of blood vessels in

retinal images using two-dimensional matched ﬁlters.

Medical Imaging, IEEE Transactions, 8(3):263 –269.

Hoover, A., Kouznetsova, V., and Goldbaum, M. (2000).

Locating blood vessels in retinal images by piecewise

threshold probing of a matched ﬁlter response. IEEE

Transactions on Medical Imaging, 19:203–210.

Jiang, X. and Mojon, D. (2003). Adaptive local threshold-

ing by veriﬁcation-based multithreshold probing with

application to vessel detection in retinal images. IEEE

Trans. Pattern Anal. Mach. Intell., 25:131–137.

MacQueen, J. B. (1967). Some methods for classiﬁcation

and analysis of multivariate observations. In Proc. of

the ﬁfth Berkeley Symposium on Mathematical Statis-

tics and Probability, volume 1, pages 281–297.

Martinez-Perez, M., Hughes, A., Stanton, A., Thom, S.,

Bharath, A., and Parker, K. (1999). Retinal blood

vessel segmentation by means of scale-space analy-

sis and region growing. In Medical Image Comput-

ing and Computer-Assisted Intervention MICCAI 99,

volume 1679 of Lecture Notes in Computer Science,

pages 90–97.

Martinez-Perez, M. E., Hughes, A. D., Stanton, A. V.,

Thorn, S. A., Chapman, N., Bharath, A. A., and

Parker, K. H. (2002). Retinal vascular tree morphol-

ogy: A semi-automatic quantiﬁcation. In Biomedical

Engineering, volume 49.

Martinez-Perez, M. E., Hughes, A. D., Thom, S. A.,

Bharath, A. A., and Parker, K. H. (2007). Segmen-

tation of blood vessels from red-free and ﬂuorescein

retinal images. Medical Image Analysis, 11:47–61.

Niemeijer, M., Staal, J., van Ginneken, B., Loog, M., and

Abramoff, M. (2004). Comparative study of retinal

vessel segmentation methods on a new publicly avail-

able database. In SPIE Medical Imaging, volume

5370, pages 648–656.

Sinthanayothin, C., Boyce, J. F., Cook, H. L., and

Williamson, T. H. (1999). Automated localisation of

the optic disc, fovea, and retinal blood vessels from

digital colour fundus images. British Journal of Oph-

thalmology, 83:902–910.

Soille, P. (2003). Morphological Image Analysis: Prin-

ciples and Applications. Springer-Verlag New York,

Inc., 2nd edition.

Staal, J., Abrmoff, M. D., Niemeijer, M., Viergever, M. A.,

and Ginneken, B. V. (2004). Ridge-based vessel seg-

mentation in color images of the retina. IEEE Trans-

actions on Medical Imaging, 23:501–509.

Walter, T. and Klein, J. (2002). A computational approach

to diagnosis of diabetic retinopathy. In 6th Conference

on Systemics, Cybernetics and Informatics (SCI).

Zana, F. and Klein, J. C. (2001). Segmentation of vessel-

like patterns using mathematical morphology and cur-

vature evaluation. IEEE Transactions on Image Pro-

cessing, 10:1010–1019.

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

324