J. Biomedical Science and Engineering, 2010, 3, 101-107
doi:10.4236/jbise.2010.31015 Published Online January 2010 (http://www.SciRP.org/journal/jbise/
JBiSE
).
Published Online January 2010 in SciRes. http://www.scirp.org/journal/jbise
Automatic detection of multiple oriented blood vessels in
retinal images
P. C. Siddalingaswamy1, K. Gopalakrishna Prabhu2
1Department of Computer Science & Engineering, Manipal Institute of Technology, Manipal, India;
2Department of Biomedical Engineering, Manipal Institute of Technology, Manipal, India.
Received 9 September 2009; revised 16 October 2009; accepted 19 October 2009.
ABSTRACT
Automatic segmentation of the vasculature in retinal
images is important in the detection of diabetic reti-
nopathy that affects the morphology of the blood ves-
sel tree. In this paper, a hybrid method for efficient
segmentation of multiple oriented blood vessels in
colour retinal images is proposed. Initially, the ap-
pearance of the blood vessels are enhanced and back-
ground noise is suppressed with the set of real compo-
nent of a complex Gabor filters. Then the vessel pixels
are detected in the vessel enhanced image using en-
tropic thresholding based on gray level co-occurrence
matrix as it takes into account the spatial distribution
of gray levels and preserving the spatial structures.
The performance of the method is illustrated on two
sets of retinal images from publicly available DRIVE
(Digital Retinal Images for Vessel Extraction) and
Hoover’s databases. For DRIVE database, the blood
vessels are detected with sensitivity of 86.47±3.6
(Mean±SD) and specificity of 96±1.01.
Keywords: Blood Vessel Segmentation; Gabor Filter;
Co-Occurence Matrix; Diabetic Retinopathy.
1. INTRODUCTION
In clinical ophthalmology colour retinal images acquired
from digital fundus camera are widely used for detection
and diagnosis of diseases like diabetic retinopathy, hyper-
tension and various vascular disorders. Retinal images
provide information about the blood supply system of the
retina. Diabetic retinopathy is a disorder of the retinal
vasculature that eventually develops to some degree in
nearly all patients with long-standing diabetes mellitus [1].
The timely diagnosis and referral for management of dia-
betic retinopathy can prevent 98% of severe visual loss,
for that, the patient has to undergo regular screening of
eye for retinopathy. The process involves dilating the eyes
with mydriatic drops and capturing the retinal image using
standard digital colour fundus camera. Screening pro-
gram results in large number of retinal images needed to
be examined by ophthalmologists. Manual diagnosis is
usually performed by analyzing the images from a pa-
tient, as not all images show signs of diabetic retinopathy,
it increases the time and decreases the efficiency of
ophthalmologists. Therefore, an automatic segmentation
of the vasculature could save workload of the ophthal-
mologists and may assist in characterizing the detected
lesions and to identify false positives [2]. Another im-
portant application of automatic retinal vessel segmenta-
tion is in the registration of retinal images of the same
patient taken over period of time [3]. The registered im-
ages are useful in monitoring the progression of disease
and to observe the effect of treatment.
Different techniques of segmentation of retinal images
have been investigated so far. They are filter based meth-
ods, vessel tracking methods, classifier based methods and
morphological methods. The techniques utilize the prior
knowledge such as, contrast that exists between the blood
vessels and surrounding background, origin of vasculature
from the same point that is the optic disc and connectivity
of the vessels. Filter based methods [4,5] and [6] employ a
two dimensional linear structural element that has a Gaus-
sian cross-profile section to identify the blood vessels,
which typically has a Gaussian profile. The gaussian ker-
nel is rotated into different orientations to fit into vessels
of different configuration to obtain a vessel enhanced im-
age. The image is then thresholded to extract the vessel
part from the background. In Hoover et al., 2000 [4]
threshold is computed using piece wise threshold probing
of matched filter response image. This works well on
images of healthy retina, but in diseased states such as
diabetic retinopathy it results in detection of many false
positives. These methods suffer from problems associ-
ated with detecting smaller and tortuous vessels that are
prone to changes in background intensity. Vessel track-
ing methods [7] and [8] use a model to track the vessels
starting at given points. Here each vessel segment is
defined by three attributes which are direction, width,
and center point. Individual segments are identified us-
ing a search procedure which keeps track of the center of
the vessel and makes some decisions about the future

P. C. Siddalingaswamy et al. / J. Biomedical Science and Engineering 3 (2010) 101-107
Published Online January 2010 in SciRes. http://www.scirp.org/journal/jbise
102
(a) (b)
(c) (d)
Figure 1. (a) Colour retinal image; (b-d) Red, Green and Blue component images.
path of the vessel based on certain vessel properties.
These methods require that beginning and ending search
points are manually selected using cursor or by using
simple thresholding techniques. Vessel tracking methods
provide very accurate measurements of vessel widths but
tracking methods often tend to terminate at branch points.
Classifier-based method employs two-step approach [9].
They start with a segmentation step often by employing
one of the mentioned matched filter-based methods and
then the regions are classified according to many fea-
tures. In the next step neural networks classifier is con-
structed using selected features by the sequential for-
ward selection method with the training data to detect
vessel pixels. Mathematical Morphology is employed for
segmentation of blood vessels as reported in [10,11] and
[12]. These methods exploits features of the vasculature
shape that are known prior, such as it being piecewise
linear and connected. They work well on normal retinal
images with uniform contrast but suffer when there is a
noise due to pathologies within the retina of eye. Many
papers have reported work on segmentation of vessels,
but still there is scope for improvement as these methods
detect vessels along with artifacts. Also detection proc-
ess becomes much more complicated in presence of le-
sions and other pathological changes affect the retinal
images.
The proposed retinal vessel detection method is com-
prised of two steps that is the retinal vessel enhancement
followed by entropic thresholding. A set of Gabor filters
tuned to particular frequency and orientation are used to
enhance the blood vessels suppressing the background.
Entropy based thresholding based on gray level co-oc-
currence matrix is employed for the segmentation of the
vessels. The following sections elucidate materials and
methods for vessel segmentation method, results and
discussion.
2. MATERIALS AND METHODS
To develop a retinal vessel segmentation system, the first
important thing is to obtain an effective database. To
realize this and also for facilitating comparison with the
existing methods, two sets of publicly available data-
bases are used.
2.1. Retinal Image Database
The DRIVE database provides forty images [19]. The
images are acquired using a Canon CR5 non-mydriatic
3CCD camera with a 45 degree field of view (FOV).
Each image was captured using 8 bits per colour plane at
768×584 pixels. The FOV of each image is circular with
a diameter of approximately 540 pixels. The Hoover’s
database [4] consists of twenty digitized slides captured
by a TopCon TRV-50 fundus camera at 35 FOV. The
slides were digitized to 700× 605 pixels with eight bits
per colour channel. Both the databases provide the hand
labeled blood vessels identified by experts as gold stan-
dard and binary mask image identifying the boundary of
the effective portion of each image. Apart from these

P. C. Siddalingaswamy et al. / J. Biomedical Science and Engineering 3 (2010) 101-107
SciRes Copyright © 2010 JBiSE
103
(a) (b)
Figure 2. (a) Surface representation of Gabor filter; (b) Real part of Gabor filter.
two databases the images are also acquired from the
Department of Ophthalmology, Kasturba medical col-
lege (KMC), Manipal using Sony FF450IR digital colour
fundus camera with 24 bit colour depth and 768×576
pixel resolution. These images are of very large variabil-
ity in terms of fundus disease and image quality and
were used to test the robustness of proposed retinal ves-
sel detection method.
2.2. Preprocessing
In the colour retinal images, blood vessels appear darker
than the background similar to the colour of lesions like
microaneurysms and hemorrhages. So it becomes essen-
tial to exempt the vessel area during the detection of
lesions to avoid false positives. Only one step is in-
volved in the preprocessing of retinal images for seg-
mentation of vessels. It can be seen in the Figure 1 that
the blood vessels appear most contrasted in the green
channel compared to red and blue channels in RGB im-
age. Only the green channel image is used for further
processing suppressing the other two colour compo-
nents.
2.3. Vessel Enhancement
The Gabor filters are widely applied to image processing
and computer vision application problems such as face
recognition and texture segmentation, strokes in charac-
ter recognition and roads in satellite image analysis
[13,14] and [15]. Since, the vessels in the retinal image
are connected and piecewise linear, for their segmenta-
tion gabor filters are better suited as they are capable of
detecting oriented features and can be fine tuned to spe-
cific frequencies. Because of their frequency sensitive-
ness it is possible to filter out the background noise of
retinal images. The spatial Gabor filter kernels are sinu-
soids modulated by a Gaussian window, the real part of
which is expressed by
 
22
,c
pp

os2
p
yx
xy
g
x yexpfx 



  





(1)
where,
p
x
=xcos+ ysin


p
yxsin+ycos




The Gabor function is defined by five parameters as
follows.

is the orientation of the filter, an angle of
zero gives a filter that responds to vertical features.
f
is the central frequency of pass band.
x

is the stan-
dard deviation of Gaussian in x direction along the filter
that determines the bandwidth of the filter. y

is the
standard deviation of Gaussian across the filter that con-
trol the orientation selectivity of the filter. The parame-
ters are to be derived by taking into account the size of
the lines or curvilinear structures to be detected. In reti-
nal images the width of the vessels varies along the
length of the vessel and in manual segmentation it is
found that majority of the vessel diameters are of 4 pix-
els wide. Therefore to accommodate the all vessels in the
detection the thickness parameter t is set to four. The
other parameters are derived using the procedure in [15]
as 0.5t
f, 0.75
xt



 and 0.85
yx

  and
2ln2 π
 .
The surface representation and real part of resulting
Gabor kernel is shown in Figure 2. It can be seen that it
is suited for the orientation of directional features to
provide good response for pixels associated with retinal
blood vessels.
To obtain good response of the vessels oriented along

P. C. Siddalingaswamy et al. / J. Biomedical Science and Engineering 3 (2010) 101-107
SciRes Copyright © 2010 JBiSE
104
Figure 3. Enhanced vessels in Gabor response.

Figure 4. (a) Gray level distribution in GLCM; (b) Represen-
tation of GLCM of 4 quadrants
different directions, the

of filter is rotated from 0o to
170o in the steps of ten degrees to produce a single peak
response on the center of a vessel segment. At each pixel
only the maximum response is retained. Figure 3 shows
the result of convolving the image in Figure 1(c) with
the vessels oriented along
the set of gabor filters. It can be seen that the vessels are
enhanced and background is suppressed considerably.
To obtain good response of
different directions, the

of filter is rotated from 0o to
170o in the steps of ten degrees to produce a single peak
response on the center of a vessel segment. At each pixel
only the maximum response is retained. Figure 3 shows
the result of convolving the image in Figure 1(c) with
the set of gabor filters. It can be seen that the vessels are
enhanced and background is suppressed considerably.
2.
.
contains information on
the distribution of gray level frequency and edge infor-
4. Entropic Thresholding
Thresholding is used to segment the blood vessels from
the background. An entropy based thresholding method
based on gray level co-occurrence matrix (GLCM) is
used to find optimal threshold as it takes into account the
spatial distribution of gray levels and preserves the spa-
tial structures in thresholded image [16,17] and [18]
Gray level co-occurrence matrix
mation, as it is very useful in finding the threshold value.
The gray level co-occurrence matrix T = [ti,j ] of the
image I with M×N dimensional matrix gives an idea
about the transition of intensities between adjacent pix-
els, indicating spatial structural information of an image.
Depending upon the ways in which the gray level i fol-
lows gray level j, different definitions of co-occurrence
matrix are possible. The GLCM is obtained as follows:
,
11
MN
ij
lk
t




 (2)
where,
(, )(,1)
(, )(1, )
1{
f
lkiand flkj
or
f
lkiand flkj
if 



0otherwi se


The probability of co-occurre
and j is written as
nce Pij of gray levels i
,
,
,
ij
ij
ijij
t
Pt  (3)
Let Th be the threshold within tnge h Oh TL
w
e 4. Quadrant A represents gray le
transition within the object while qu
gray level transition within the ba
level transition between the object and the background
or adran
an
he ra-1,
vel
here L is the number of gray levels. Threshold Th parti-
tions the GLCM into four quadrants, namely A, B, C,
and D as in Figur
adrant D represents
ckground. The gray
across the object’s boundary is placed in qut B
d quadrant C.
The probabilities of object class and background class
are defined as
,
00
hh
TT
Aij
ij
PP



 (4)
Table 1. Performance of retinal blood vessels segmentation
method.
Database No. of Sensitivity Specificity
images (%) (%)
DRIVE 40 86.47±3.6 96±1.01
Table 2. Comparison of vessel segmentation results on Hoo-
Method Sensitivity range
(%)
Specificity range
ver’s database.
(%)
Proposed method79-91 94-98
Hoover 2000 80-90 92-93 et al.

P. C. Siddalingaswamy et al. / J. Biomedical Science and Engineering 3 (2010) 101-107
SciRes Copyright © 2010 JBiSE
105
(a) (b) (c)
(d) (e) (f)
Figure 5. Retinal vessel segmentation. (a) Image from DRIVE database; (b) Corresponding manual segmentation; (c) Vessel seg-
mentation result; (d) Image from Hoover’s database; (g) Corresponding manual segmentation; (f) Vessel segmentation result.
(a) (b)
Figure 6. (a) Digital color retinal image from KMC database; (b) Segmented vessels.
Using (4) and (5) as normalization factors, the nor-
malized probabilities of the object class and background
class are functions of threshold v
fined as
11hh
LL
,
11
hh
Cij
iT jT
P
 
 P (5)
ector (Th, T
h) are de-
,
,
ij
A
ij
A
P
PP  (6)
(7)
the obct is given by
,
,
I
J C
ij
c
P
Pp 
The second-order entropy ofje
,2,
00
1
()
TT
lo g
2
hh
AA
Ahij ij
ij
H
ThP P

(8)
Similarly, the second-order
is given by
entropy of the background
11
,2,
lo g
CC
ij ij
T11
1
() 2
hh
hh
LL
C h
iT j
H
Th
 

nd
the background is given by
P P
 (9)
The total second-order local entropy of the object a

P. C. Siddalingaswamy et al. / J. Biomedical Science and Engineering 3 (2010) 101-107
SciRes Copyright © 2010 JBiSE
106
()() ()
ThAh Ch
HTHTHT
Finally TE the gray level corresponding to the maxi-
m
T (11)
3. RESULTS AND DISCUSSION
automatic
retinal vessels. Figure 5 shows the m
im
with 1.66 GHz CPU and
512MB memory using Matlab 7.0
about 20 seconds to detect the vesse
ng
comparison
(10)
um of HT(Th) gives the optimal threshold for vessel
and non vessel classification.
1
max
0
arg[( )]
h
ETLTh
TH


Retinal images from the DRIVE database and Hoover’s
database are used for the segmentation of
anual segmented
age and the final output obtained by the proposed
method on the images from both the image databases.
On Windows XP, Intel PC
, the method takes
ls in retinal image.
The performance of the method is evaluated usi
sensitivity and specificity at pixel level in
with manually segmented vessels by an expert.. Sensi-
tivity gives the percentage of pixels correctly classified
as vessels by the method and specificity gives the per-
centage of non vessels pixels classified as non vessels by
the method. given by
p
p n
T
Sensitivity TF  (12)
n
nP
T
Specificity TF  (13)
where Tp is true positive, Tn is true negative, Fp is false
positive and Fn is false negative at each pixel. Table 1
show the performance of the proposed method on forty
images from DRIVE database.
The results of the proposed method are also comp
with those of Hoover et al. 2000
from the Hoover’s database. The
mal and ten abnormal retinal images and the resu
depicted in Table 2. It can be
method performs better with lower specificity even in
of reti-
na
er segmentation. The
e used to obtain the control
registration techniques. It is
. and Ali, E. (2002) A
the diagnosis of dia-
[3] Laliberté, F., Gagnon, L. and Sheng, Y. (2003) Registra-
inal images: an evaluation study.
cal. Imaging, 22, 661-673.
sh Journal of Ophthalmology, 83, 902-910.
ared [6] Thitiporn, C. and Fan, G.L. (2003) An efficient Algo-
rithm for extraction of anatomical structures in retinal
images. Proc. Of Intl. Conf. on Image Processing, 1,
[4] on twenty images
database has ten nor-
lt is 1093-1096.
[7] Wu, D., Zhang, M., Liu, J.C. and Wendall, B. (2006) On
the adaptive detection of blood vessels in retinal images.
IEEE Transactions on Biomedical Engineering, 53,
seen that the proposed
the presence of lesions in the abnormal images.
Apart from two standard databases the method is also
tested on retinal images obtained from Ophthalmology
department, KMC. These images are of large variability
in terms of presence of lesions and image quality. These
are considered to evaluate the robustness of the method.
The result for one of the image is shown in Figure 6 and
is validated by ophthalmologists.
4. CONCLUSIONS
An efficient method for automatic segmentation
l blood vessels has been presented. Images from three
different datasets are used to evaluate the robustness and
accuracy of the method, demonstrating that it may be
useful in a wide range of retinal images. Based on a brief
comparison with some other vessel segmentation algo-
rithms, we can conclude that the Gabor filter and en-
tropic threshold provides a bett
segmented vessels can b
points used in the retinal
hoped that automated segmentation of vessel technique
can detect the signs of diabetic retinopathy in the early
stage, monitor the progression of disease, minimize the
examination time and assist the ophthalmologist for a
better treatment plan.
5. ACKNOWLEDGEMENT
Our sincere thanks to ophthalmologists of the Department of Oph-
thalmology, Kasturba Medical College, Manipal for providing the
necessary images and clinical details.
REFERENCES
[1] Emily, Y.C. (2003) Diabetic retinopathy. Preferred Prac-
tice Patterns, American Academy of Ophthalmology,
USA.
[2] Thomas, W., Klein, J.C., Pascale, M
contribution of image processing to
betic retinopathy-detection of exudates in color fundus
images of the human retina. IEEE Trans. Medical. Imag-
ing, 21, 1236-1243.
tion and fusion of ret
IEEE Trans. Medi
[4] Hoover, A., Kouznetsoza, V., Goldbaum, M. (2000) Lo-
cating blood vessels in retinal images by piecewise
threshold probing of a matched filter response. IEEE
Transactions on Medical Imaging, 19, 203-210.
[5] Chaudhuri, S., Chatterjee, S., Katz, N., Nelson, M. and
Goldbaum, M. (1989) Detection of blood vessels in reti-
nal images using two dimensional matched filters. IEEE
transactions on Medical Imaging, 8, 263-269.
341-343.
[8] Pinz, A., Bernogger, S., Datlinger, P. and Kruger, A.
(1998) Mapping the human retina. IEEE Transactions on
Medical imaging, 17, 606-619.
[9] Sinthanayothin, C., Boyce, J.F., Cook, H.L. and Wil-
liamson H.T. (1999) Automated location of the optic disc,
fovea, and retinal blood vessels from digital color fundus
images. Briti
[10] Zana, F. and Klein, J. (2001) Segmentation of vessel-like
patterns using mathematical morphology and curvature
evaluation. IEEE Transactions on Image Processing, 10,
1010-1019.
[11] Siddalingaswamy, P.C., Prabhu, G.K. and Mithun, D.
(2006) Feature extraction of retinal image. Proc. Of the
National Conference for PG and Research scholars,
24-27.

P. C. Siddalingaswamy et al. / J. Biomedical Science and Engineering 3 (2010) 101-107
SciRes Copyright © 2010
107
to, Y. and Tamura, S. (2000) Orientation
, Cai, J. and Buse, R. (2003) Hand-writing rec-
996) Hierar-
nal entropic thresholding algo-
d Chang, C. (1995) Target de-
d S.A.R and Bakar A. (2007) Adaptive
(2004) Ridge based vessel seg-
JBiSE
[12] Yang, Y., Huang, S.Y. and Rao, N. (2008) An automatic
hybrid method for retinal blood vessel extraction. Jour-
nal of Appl. Math. Comp. Sci., 18, 399-407.
[13] Lee, T.S. (1996) Image representation using 2D Gabor
wavelets. IEEE Transactions of Pattern Analysis and
Machine Intelligence. 18, 959-971.
[14] Chen, J., Sa
space filtering for multiple orientation line segmentation.
IEEE Transactions of Pattern Analysis and Machine In-
telligence, 22, 417-429.
[15] Liu, Z.Q.
ognition: soft computing and probablistic approaches.
Springer Verlag. Berlin.
[16] Yang, C.W., Chung, P.C. and Chang, C.I. (1
chical fast two dimensio
rithm using a histogram pyramid. Optical Engineering,
35, 3227-3241.
[17] Mark, L. Althouse, G. an
tection in multispectral images using the spectral co- oc-
currence matrix and entropy thresholding. Optical Engi-
neering, 34, 2135-2148.
[18] Mokji, M.M. an
thresholding based on co occurrence matrix edge infor-
mation. Journal of Computers, 2, 44-52.
[19] Staal, J.J., Abràmoff, M.D., Niemeijer, M., Viergever, M.
A. and van Ginneken, B.
mentation in color images of the retina. IEEE Trans. Med.
Imag., 23, 501-509.