3194 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO.

11, NOVEMBER 2013

An Automated Method for Retinal Arteriovenous
Nicking Quantication From Color Fundus Images
Uyen T. V. Nguyen, Alauddin Bhuiyan, Laurence A. F. Park, Ryo Kawasaki, Tien Y. Wong, Jie Jin Wang,
Paul Mitchell, and Kotagiri Ramamohanarao

AbstractRetinal arteriovenous (AV) nicking is one of the

prominent and signicant microvascular abnormalities. It is char-
acterized by the decrease in the venular caliber at both sides of
an artery-vein crossing. Recent research suggests that retinal AV
nicking is a strong predictor of eye diseases such as branch retinal
vein occlusion and cardiovascular diseases such as stroke. In this
study, we present a novel method for objective and quantitative AV
nicking assessment. Fromthe input retinal image, the vascular net-
work is rst extracted using the multiscale line detection method.
The crossover point detection method is then performed to local-
ize all AV crossing locations. At each detected crossover point, the
four vessel segments, two associated with the artery and two as-
sociated with the vein, are identied and two venular segments
are then recognized through the artery-vein classication method.
The vessel widths along the two venular segments are measured
and analyzed to compute the AV nicking severity of that crossover.
The proposed method was validated on 47 high-resolution retinal
images obtained from two population-based studies. The experi-
mental results indicate a strong correlation between the computed
AV nicking values and the expert grading with a Spearman corre-
lation coefcient of 0.70. Sensitivity was 77% and specicity was
92% (Kappa = 0.70) when comparing AV nicking detected us-
ing the proposed method to that detected using a manual grading
method, performed by trained photographic graders.
Index TermsArteriovenous (AV) nicking, blood vessel seg-
mentation, crossover point detection, retinal image, vessel width
measurement.
Manuscript received March 7, 2013; revised May 20, 2013; accepted June
16, 2013. Date of publication June 25, 2013; date of current version October
16, 2013. Asterisk indicates corresponding author.
U. T. V. Nguyen is with the Department of Computing and Information Sys-
tems, The University of Melbourne, Parkville, VIC 3010, Australia (e-mail:
thivun@student.unimelb.edu.au).
A. Bhuiyan is with the ICT Centre, Australian E-Health Research Centre,
Commonwealth Scientic and Industrial Research Organization (CSIRO), Perth
WA 6014, Australia (e-mail: alauddin.bhuiyan@csiro.au).
L. A. F. Park is with the School of Computing, Engineering and Math-
ematics, The University of Western Sydney, NSW 2751, Australia (e-mail:
lapark@scm.uws.edu.au).
R. Kawasaki is with the Department of Public Health, Yamagata Univer-
sity Faculty of Medicine, Yamagata 990-9585, Japan (e-mail: ryok@med.id.
yamagata-u.ac.jp).
T. Y. Wong is with the Singapore Eye Research Institute, National University
of Singapore, Singapore (e-mail: tien_yin_wong@nuhs.edu.sg).
J. J. Wang and P. Mitchell are with the Centre for Vision Research, Depart-
ment of Ophthalmology, Westmead Millennium Institute, University of Syd-
ney (C24), Sydney, NSW 2006, Australia (e-mail: jiejin.wang@sydney.edu.au;
paul.mitchell@sydney.edu.au).

K. Ramamohanarao is with the Department of Computing and Information

Systems, The University of Melbourne,Parkville, VIC 3010, Australia (e-mail:
kotagiri@unimelb.edu.au).
Color versions of one or more of the gures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identier 10.1109/TBME.2013.2271035
Fig. 1. Examples of AV crossings: (a) normal AV crossing and (b) AV nicking
crossing. (Please refer to the color version of the paper for a clear view of the
images.)
I. INTRODUCTION
R
ETINAL arteriovenous nicking (AV nicking or AVN) is
the phenomenon where the vein is compressed by a stiff
artery at their crossing location in response to a rise in blood
pressure (i.e., hypertension) [1], [2]. In retinal photographs, AV
nicking exhibits itself as the decrease in the venular caliber at
both sides of an artery-vein (AV) crossing (see Fig. 1). The
prevalence of moderate to severe AV nicking is reported as
2.2% in the Beaver Dam Eye population [3] while this number
is found higher (7.7%) in the older population of the Cardio-
vascular Health Study [4]. It was claimed that AV nicking is
associated with not only current blood pressure but also past
blood pressure, implying that it is a persistent and long-term
marker of hypertension [2]. Strong and consistent association
between AV nicking and systemic diseases has also been found
by recent population-based studies [5][13]. The Atheroscle-
rosis Risk in Communities (ARIC) study reported that people
with AV nicking are two times more likely to develop an inci-
dent stroke than those without (relative risk 2.21, 95% con-
dence interval [CI] 1.443.38) [6]. In addition, AV nicking was
also found strongly associated to retinal vein occlusion, a com-
mon sight-threatening ocular disorder (odds ratio 16.75, 95%
CI 7.3338.24) [14]. Therefore, the assessment of AV nicking
is extremely important in order to identify people at high risk of
cardiovascular diseases for early and in-time treatment.
Currently, the assessment of AV nicking is done by human
graders in a subjective and qualitative manner. In this process,
a grader examines all AV crossing points in a retinal image
and compares each of them with standard photographs to assess
the presence and severity of AV nicking for that image [15].
This grading process, however, depends heavily on the graders
expertise and, thus, its accuracy and reproducibility are of con-
cerns. The ARIC study reported that only a fair to moderate
0018-9294 2013 IEEE
NGUYEN et al.: AN AUTOMATED METHOD FOR RETINAL ARTERIOVENOUS NICKING QUANTIFICATION FROM COLOR FUNDUS IMAGES 3195
agreement is achieved with the manual AV nicking assessment
(inter- and intragrader Kappa value = 0.40 to 0.61 [15] and
= 0.56 to 0.57 in a more dedicated study [16]). In addition,
the current manual grading process is very time consuming,
which makes it infeasible for large-scale screening and clinical
applications. In [17], we have proposed a computerized method
for AV nicking measurement which helps to reduce the grading
time substantially. However, this method also requires human
intervention as the user needs to select the crossover positions
in each retinal image. Motivated by this, we propose in this
paper a fully automated method, which is to the best of our
knowledge, the rst computerized method proposed for auto-
mated AV nicking assessment. The proposed method produces
a continuous value, implying the AV nicking severity level, for
each AV crossing point in a retinal image. The advantages of
the proposed method include:
1) The measurements produced are objective, reproducible,
and repeatable.
2) The computed measurements reveal more details on AV
nicking severity, which are not available in the human
grading system. Such information may help to strengthen
the relationship between AV nicking and known diseases
such as hypertension and stroke.
3) The computed measurements provide an important basis
toward the development of a computer system for au-
tomatic AV nicking detection for large-scale screening
systems.
The rest of this paper is organized as follows. In Section II, the
details of the proposed method are described. The experimental
results to demonstrate the performance of the proposed method
are presented in Section III. Finally, the paper is concluded with
Section IV.
II. METHODOLOGY
The overall framework of the proposed AV nicking quanti-
cation method is illustrated in Fig. 2. The system takes as input
the retinal image and returns a real number quantifying the sever-
ity level for each AV crossing point detected from that image.
From the input retinal image, the vessel segmentation method is
applied to extract the blood vessels from the image background
for further analysis. A crossover point detection method is then
performed to detect AV crossing locations within the retinal im-
age. At each detected crossing position, the four major vessel
segments constituting the two vessels (i.e., the artery and the
vein) are localized and the two venular segments are identied
through an artery-vein classication process. The vessel widths
of each venular segment are then measured and analyzed for
AV nicking measurement. The following sections describe each
step of the system in details.
A. Vessel Segmentation
Blood vessel segmentation from high-resolution retinal color
images is complicated by retinal vessel central light reex,
choroidal vascularization, and image artifacts such as back-
ground homogenization and other impulse noises. Although
many methods have been proposed for retinal vessel extraction,
they are not effective for detecting blood vessels in our image
set that includes high-resolution images with the presence of
vessel central light reex (i.e., the bright strip along the middle
of a vessel). Motivated by this, we have proposed a new method
for effective retinal vascular network extraction using multi-
scale line detection technique, which is described in [18]. The
proposed segmentation method combines line detectors at vary-
ing scales to produce an enhanced retinal image and the nal
segmentation is obtained by a thresholding operator. The pro-
posed method was proven to be effective in dealing with central
light reex problem while providing accurate vessel boundary
detection.
Fig. 3 shows a cropped retinal image with the presence of
central light reex and the segmented image obtained using our
method. It can be seen that the proposed method can correctly
segment all blood vessels, even on vessels with the presence
of central light reex. Hence, it was used in our system for
automatic blood vessel extraction.
B. Artery-Vein Crossover Point Detection
The vascular network is constructed of three important land-
mark points: branching, bifurcation, and crossover. A crossover
is the place where two vessels (i.e., a vein and an artery) cross
each other while a branching or a bifurcation is the place where
one vessel splits into two vessels. Since AV nicking happens
at the AV crossing locations, the method rst needs to detect
all crossover points within a retinal image. For this task, an
efcient method for automated crossover point detection has
been proposed and described in details in [19]. This section pro-
vides a brief summary of the proposed method for the sake of
completeness.
To detect AVcrossover points, the vascular skeleton and edge
images are rst extracted from the segmented image using bi-
nary morphological operators. The vessel skeleton is extracted
using an iterative thinning process [20] which successively
erodes away pixels on the boundary of the objects while pre-
serving its connectivity until no more thinning is possible. In
contrast, the vessel edge image is achieved by a morphological
operator that removes interior pixels and retains only pixels on
the vessel boundaries [20]. The crossover locations are then de-
tected by an extensive analysis on the extracted vascular skeleton
and edge images as follows.
At each skeleton pixel P, we compute the cross-point number
(cpn) [21] as follows:
cpn(P) =
1
2
8

i=1
|N
i
(P) N
i+1
(P)| (1)
where N
i
(P) are the neighbor pixels of P (in a 3 3 neighbor-
hood) named in an anticlockwise order and N
9
(P) = N
1
(P).
The cpn number computed for each skeleton pixel represents
the number of vessel segments connected to that pixel in the
skeleton image. Hence, all skeleton pixels with cpn equal to 3
are initially marked as bifurcation candidates while those with
cpn equal to 4 are labeled as crossover candidates. Then, two bi-
furcation candidates are further grouped as one crossover if they
are connected by at most T pixels. The segment that connects
3196 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 11, NOVEMBER 2013
Fig. 2. Main steps of the proposed method for automated AV nicking quantication.
Fig. 3. Example showing the vessel segmentation result: (a) cropped retinal
image with the presence of center light reex (pointed by black arrows) and
(b) segmented image.
Fig. 4. Geometrical features of two vessels constituting a crossover: (a)
1
and

2
represent intersection angles formed by the two vessels; (b) representation
of
1
,
2
in the skeleton image; and (c)
1
and
2
represent the curvature of
each vessel at the crossing.
the two bifurcation points is then identied and its middle point
is located and served as the true position of that crossover. This
step helps to detect crossover points that are represented by two
bifurcation points in the skeleton image when two vessels inter-
sect at an acute angle. However, this process also falsely detects
those points that have the same conguration in the skeleton im-
age as crossovers. To distinguish true crossovers from spurious
results, two geometrical features of each detected crossover are
analyzed:
1) The intersection angles formed by the two vessels at the
crossing (
1
and
2
in Fig. 4(a) and (b)).
2) The two angles representing the curvature of each vessel
at the crossing (
1
and
2
in Fig. 4(c)).
A detected crossover is valid if it satises:
(max(
1
,
2
) <
max
) and (min(
1
,
2
) >
min
) (2)
where
max
and
min
are two parameters to be set. We used
ve retinal images for parameter tuning and the experimen-
tal results show that with our image set (see Section III-A),
{T = 75,
max
= 110,
min
= 140} is a suitable setting. The
performance of the proposed method was assessed on a subset
Fig. 5. Performance of the proposed crossover point detection method in terms
of precision, recall, and F1-measure on 15 test images when T changes from 0
to 100 pixels.
of 15 test images randomly selected from our image set (differ-
ent from ve training images). The results show that out of 41
crossover points detected, 38 of them are correct, which indi-
cates an accuracy (or precision) of 93%. In addition, our method
has missed four crossovers, which means that the recall of our
method is 90%. This has demonstrated the effectiveness of our
method for crossover point detection. The detected crossing lo-
cations are then fed directly to the AV nicking quantication
method so that their AV nicking severity can be assessed.
It should be noted that among the three parameters (T,
max
,
and
min
), T is the only parameter that is dependent on the image
set (i.e., image resolution). Hence, the same values of
max
and
min
can be used for a new image set. To investigate the
dependence of the system on T, we examined the performance
achieved in terms of precision, recall, and F1-measure (F
1
=
2precisionrecall
precision+recall
) on 15 test images when T changes from 0 to
100 pixels (see Fig. 5). This gure shows that similar results are
obtained for a long range values of T (65 < T < 100), which
implies that the performance of the proposed method is not very
sensitive to the setting of T.
NGUYEN et al.: AN AUTOMATED METHOD FOR RETINAL ARTERIOVENOUS NICKING QUANTIFICATION FROM COLOR FUNDUS IMAGES 3197
C. AV Nicking Quantication
At each detected crossover, a subimage centered on the cross-
ing location is identied and used for all subsequent analysis
for computational efciency. Since AV nicking is characterized
by the decrease in the venular width at the crossing location, the
two vessel segments constituting the vein are extracted and their
vessel widths are measured and analyzed for AV nicking com-
putation. To achieve this, the segmented image at each crossover
is applied to a sequence of the following steps:
1) Major vessel identication
2) Venular segment identication
3) Vessel width measurement
4) AV nicking measurement
The details of each step are described in the following
sections:
1) Major Vessel Identication: This step aims to identify the
main vessels constituting the crossover and isolate them from
unnecessary structures such as small branches or noisy artifacts.
To achieve this, the system rst examines both vessel skeleton
and edge images to detect and classify all feature points as
crossover (C), bifurcation (Bi) or branching point (Br). We have
described the method for crossover point detection in Section II-
B. Branching point is the place where one small vessel comes
out from a main vessel while a bifurcation point is the place
where one vessel splits into two similar vessels. To distinguish
branching point from the bifurcation point, the vessel widths of
the three vessel segments associated with that point are mea-
sured and recorded as W
c
(the vessel width of the trunk vessel),
W
a
(the vessel width of the smaller branch), and W
b
(the vessel
width of the larger branch). A feature point is classied as a
bifurcation if
W
a
W
b
f
b
(0 < f
b
< 1) and as a branching point
otherwise. For example, f
b
= 0.5 means that at a branching lo-
cation, if the vessel width of the smaller branch is equal to or
greater than half of the vessel width of the bigger branch, that
feature point is considered as a bifurcation. This setting guaran-
tees that two branches of a bifurcation are of similar diameter.
To identify a suitable setting for f
b
, we randomly selected ten
bifurcation and ten branching points in the working image set
and examined f
b
values (f
b
=
W
a
W b
) of these selected landmark
points. For this, at each selected landmark point, we manually
measured the vessel width of its smaller branch (W
a
) and the
bigger branch (W
b
). f
b
is then computed as the ratio of
W
a
W
b
.
Fig. 6 shows the distribution of f
b
values of these 20 landmark
points and it is shown that there is a clear separation in f
b
values
of bifurcation and branching points: all bifurcation points have
f
b
0.8 while all branching points have f
b
0.73. Hence, f
b
is
set as 0.8 in our system for bifurcation/branching classication.
The process of major vessel identication is then performed as
follows. From the extracted vascular network, the system traces
from the crossing position through its four connected vessel
segments and performs appropriate action when a branching, a
crossover, or a bifurcation is encountered. If a crossover or a bi-
furcation is found, the system stops its traversal and terminates
the vessel along that direction. Otherwise, if a branching point
is found, the smaller branch is trimmed off and the system con-
tinues its traversal along the main vessel until a bifurcation or a
Fig. 6. Distribution of f
b
values over 20 randomly selected landmark points
(ten branching and ten bifurcation points).
Fig. 7. Example showing the cutoffs at: (a) crossover point C; (b) bifurcation
point Bi; and (c) branching point Br.
crossover point is encountered. To realize this, at each crossover
(except the crossover to be examined) or bifurcation point, cut-
offs are performed to separate its constituting segments while
at a branching point, the cutoff is to separate the smaller branch
from the main vessel. This process is demonstrated in Fig. 7.
Once the cutoffs are performed at all detected feature points, the
simplied segmentation is obtained by a morphological recon-
struction starting at the crossing position, following the direction
of its four segments and stopping when a cutoff is met.
This major vessel identication process helps to retain im-
portant vessel structure while removing unnecessary structures
to simplify the subsequent analysis. Fig. 8 shows the example
results obtained through this process on the two crossover points
in Fig. 1. We can see that this process has effectively removed
small branches and noisy segments from the vascular network,
retaining only important vessels constituting the crossover for
accurate AV nicking assessment.
To illustrate the effect of f
b
value on the resulting segmenta-
tion, Fig. 9 shows the results obtained with two different settings
of f
b
on a crossover. On this example crossover, there are two
branching points (Br
1
and Br
2
) on the vein vessel close to
the crossing region. The results show that if f
b
is set too small
(i.e., f
b
= 0.1, Fig. 9(b)), the two small branches of the vein are
incorrectly classied as bifurcation and this leads to the early
termination of the systemnear the crossing location. This results
in an inadequacy in the vascular network obtained since only a
small portion of the two venular segments are retained for further
3198 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 11, NOVEMBER 2013
Fig. 8. Example results of the major vessel identication process: (a) ini-
tial segmentation; (b) skeleton image with detected branching points marked
with cross signs and detected bifurcation points marked with circles; (c) seg-
mented image with cutoffs at branching and bifurcation points; (d) simplied
segmentation.
Fig. 9. Example results obtained after the major vessel identication process
with two different settings of f
b
: (a) cropped retinal image at a crossover
point (top) and its initial segmentation (bottom), (b) and (c): (top) segmentation
with cutoffs at the detected bifurcation and branching points and (bottom) its
simplied segmentation when (b) f
b
= 0.1 and (c) f
b
= 0.8.
analysis. On the other hand, when f
b
is set as 0.8 (see Fig. 9(c)),
the two small branches of the vein are correctly identied and
they are successfully trimmed off from the main vessel. This
helps to retain the most important parts of the vein and make
them available for the subsequent AV nicking measurement.
2) Venular Segment Identication: The simplied segmen-
tation provides us with a vascular network containing four seg-
ments; two associated with the vein and two associated with
the artery. Our analysis is performed on the vein, therefore we
need to determine which pair is associated to the vein. This is
done by an analysis on the vessel skeleton and edge images ex-
tracted from the simplied segmentation. The vascular skeleton
divides the boundary into four sections. For each section, the
edge point closest to the crossing position is identied, resulting
in four edge points: E
1
, E
2
, E
3
, E
4
(see Fig. 10(a)). Connect-
ing these four points in a convex hull order will divide the whole
segmentation into ve parts: the four vessel segments and the in-
tersection region. The four segments are labeled in a clockwise
order from S
1
to S
4
. Then, two opposite segments are paired to
represent each vessel, i.e., VSS
1
= (S
1
, S
3
), VSS
2
= (S
2
, S
4
).
Fig. 10 depicts this identication process.
Fig. 10. Example showing the individual vessel identication process:
(a) cutoff at the crossing region to separate the vascular network into four vessel
segments, labeled in a clockwise order (from S
1
to S
4
); (b) and (c) two oppo-
site segments are paired to represent each vessel: VSS
1
= (S
1
, S
3
), VSS
2
=
(S
2
, S
4
).
Fig. 11. Example results of the venular segment identication process: (a)
simplied segmentation with the cut-off at the crossover point; (b) AV classied
image (artery is in red, vein is in green); (c) extracted venular segments; (d)
extracted venular skeleton and edge superimposed on the original image.
The AVclassication process is then performed to distinguish
the vein from the artery. For the majority of crossover points, it
is observed that the artery appears brighter than its vein counter-
part. This is due to the fact that the artery contains oxygenated
blood which is pumped from the heart, making it red, while
the vein carries deoxygenated blood back to the heart, which
makes it darker. Hence, the color information of the two vessels
is computed and the artery is assigned to the vessel with higher
intensity value. Different color spaces, RGB and HSV as well
as the gray level, were used to identify the most discriminative
feature. The median intensity value of each vessel is computed
as
f
VSS
i
= median(X(VSS
i
)) (3)
where i {1, 2}, X {R, G, B, H, S, V, Gray} representing
different color components. For example, R(VSS
1
) contains
the intensity values of the red channel of the rst vessel. Then,
the vein is identied as the vessel with lower intensity level.
Once the vein is correctly recognized, its two segments are
extracted for subsequent analysis. The results achieved through
this process are shown in Fig. 11(a)(d).
The artery-vein classication accuracy (ACC) is used to eval-
uate the performance of our system at this stage. ACC is dened
as the fraction of the number of crossover points where the
artery and vein were correctly classied over the total number
of crossovers that were considered. Table I presents the ACC on
90 working crossovers (described in Section III-A) using differ-
ent color features. It is shown that the green channel provides
NGUYEN et al.: AN AUTOMATED METHOD FOR RETINAL ARTERIOVENOUS NICKING QUANTIFICATION FROM COLOR FUNDUS IMAGES 3199
TABLE I
ARTERY-VEIN CLASSIFICATION ACCURACY (ACC) OF DIFFERENT COLOR
FEATURES ON 90 WORKING CROSSOVERS
Fig. 12. Method for identifying vessel boundary pixels representing the vessel
width at a skeleton pixel.
highest accuracy, followed by the gray level. The red channel
and the value component (in HSV space) yield similar good
performance while the blue channel gives lowest accuracy. This
result is in accordance with previous studies ( [22][24]) where
the green channel is found most effective for blood vessel ex-
traction as it provides best contrast between the vessel and the
background. This result further conrms that the green channel
provides best contrast between artery and vein and being the
most discriminative feature for AV classication.
3) Vessel Width Measurement: In this step, the vessel widths
of the two venular segments are measured using the vessel skele-
ton and edge images extracted in the previous step. For each
skeleton pixel C
i
, a set of pairs of edge pixels whose connected
line going through C
i
is identied (see Fig. 12)
E = {(E
1j
, E
2k
) : C
i
E
1j
+ (1 )E
2k
} (4)
where E
xi
E
x
, [0, 1], E
1
, and E
2
are the set of edge
pixels at two sides of the vessel skeleton [see Fig. 12(a)]. Let
C
iK
and C
i+K
be the two skeleton pixels at K pixel distance
to either side of C
i
, the gradient at C
i
can be estimated by the
vector

C
iK
C
i+K
. For each pair of edge points {E
1j
, E
2k
}, the
angle of intersection with the skeleton is given as
cos (
jk
) =
|

C
iK
C
i+K


E
1j
E
2k
|

C
iK
C
i+K

E
1j
E
2k

. (5)
The two edge points {E
1m
, E
2n
}, where
{m, n} = argmin
j,k
cos (
jk
) (6)
construct the line that is most perpendicular to the skeleton
gradient. We use the distance between these two edge points

E
1m
E
2n

as the vessel width at C

i
.
However, the vessel width at C
i
is considered as valid only
if
mn
is close to 90
o
(i.e., |90
mn
| ). Otherwise, it is
marked as invalid and C
i
is removed from the result to ex-
Fig. 13. Example showing the effect of on vessel width measurements
achieved: (a) cropped retinal image with superimposed vessel edge (white)
and vessel skeleton (black); (b) vessel width measurements achieved without
constraint on
mn
( = +); and (c) vessel width measurements achieved
if is set as 5
o
. Invalid measurements are observed if no constraint is put on

mn
while the setting of a small value of has effectively removed incorrect
measurements.
Fig. 14. Example results of vessel width measurements produced by the pro-
posed method on two example crossovers shown in Fig. 1: imaginary lines
connecting two edge points represent the vessel width at each skeleton pixel.
clude from further analysis. A small value of ensures that the
vessel width achieved is perpendicular to the vessel axis. If no
constraint is put on
mn
, the vessel width obtained might be
incorrect due to the insufcient number of edge points to match
the vessel width, which often happens at the skeleton pixels
close to the crossing region, an example of which is shown in
Fig. 13.
Fig. 13(a) shows the vessel skeleton and vessel edge super-
imposed on a venular segment while the vessel widths obtained
if no constraint is put on
mn
(i.e., = +) is shown in
Fig. 13(b). Invalid measurements are found at the skeleton pix-
els close to the crossover since they do not reect the true width
of the vessel (they are not perpendicular to the vessel axis).
These invalid measurements have been removed when is set
as 5

(see Fig. 13(c)). Hence, it is set as 5

o
in our system. How-
ever, it should be noted that similar results were observed for
a range values of (5 20). The vessel widths achieved
on two example crossover points are shown in Fig. 14. We can
see that the straight lines representing the vessel width at each
skeleton pixel are perpendicular to the vessel axis while its two
edge points are tted well on the true vessel boundary. This has
indicated that vessel widths were accurately measured.
4) AV Nicking Measurement: As each vein is composed of
two segments at two sides of the crossing, the severity of AV
nicking is assessed separately for each segment and the nal AV
nicking index of the crossover point is obtained by a combination
3200 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 11, NOVEMBER 2013
of these two individual measures:
AVN(C) = f
A
(AVN(V
1
), AVN(V
2
)) (7)
where f
A
is a coalescing function, AVN(V
1
) and AVN(V
2
) are
the AV nicking severity levels of the two venular segments,
V
1
and V
2
, respectively. The method for computing the AV
nicking severity level of each venular segment (at one side of
the crossing) is as follows.
For each venular segment, four important measurements
(W, W
C
, S, and S
missing
) are automatically derived from the
vessel widths of that segment. W is the measurement represent-
ing the normal vessel width of the segment while W
C
represents
the vessel width at the region close to the crossing location. S
represents the total number vessel pixels at region close to the
crossover if there is no decrease in the vessel width (or the ab-
sence of AV nicking), while S
missing
represents the number of
vessel pixels that is missing due to the presence of AV nicking.
These measurements are inspired based on the fact that there is a
decrease in vessel width at the crossing region in the presence of
AV nicking, while this phenomenon is absent in normal cases.
Hence, in the presence of AV nicking, W
C
is much lower than
the normal vessel width W while the difference is not much
in the normal cases. Similarly, the number of missing pixels
S
missing
is much higher in AV nicking crossovers than in nor-
mal ones. The method for computing these measurements from
the vessel widths is as follows.
Suppose that the segment is made up of n vessel width mea-
surements, w = {w
i
|i {1, . . . , n}} (the vessel widths w
i
are
indexed and ordered so that w
1
is the width closest to the
crossover point, while w
n
is the furthest). W is approximated
as the median value of its measurements, or W = median(w).
Suppose that N is the number of successive cross sections,
starting from the rst measurement w
1
, with the measure-
ments lower than normal vessel width W, W
C
is approxi-
mated as the minimum value of the rst N vessel widths, or
W
C
= min{w
i
|i {1, . . . , N}}. The number of vessel pix-
els at the crossing region in the normal case, S, is estimated
as N W. Let S
1
and S
2
represent the number of missing
pixels at both side of the vessel at the crossing region (see
Fig. 15(a)), while S
3
represents the number of vessel pixels
composed by the rst N cross sections (S
3
=

N
i=1
w
i
), the
total number of missing pixels S
missing
can be computed as
S
missing
= S
1
+S
2
= S S
3
= N W

N
i=1
w
i
. The il-
lustration of these measurements is shown in Fig. 15.
From the four critical measurements (W, W
C
, S, and
S
missing
), three measures are derived to compute the ANnicking
severity of each venular segment as follows:
M
MF
=
S
missing
S
(8)
M
R
= 1
W
C
W
(9)
M
S
= W W
C
. (10)
All of these measures yield high values for AV nicking
crossovers and low values for normal cases. The rst measure
M
MF
evaluates the proportion of the missing pixels due to the
Fig. 15. Illustration of the measurements used for AV nicking computation
on (a) synthetic vessels and (b) vessel width measurements. The denition of
the annotations in this gure is as follows: W represents the vessel width in the
normal case; W
C
represents the vessel width at the crossing region; N is the
number of successive cross sections, starting from the rst measurement w
1
,
with the measurements lower than normal vessel width W; S represents the
number of vessel pixels at the crossing region in the normal case (i.e., without
AV nicking); S
1
and S
2
represent the number of missing pixels at both side
of the vessel at the crossing region; S
3
represents the number of vessel pixels
composed by the rst N cross sections.
presence of AV nicking while M
R
measures the ratio of the ves-
sel width at the crossing region compared to the normal width.
M
S
measure computes the difference in the vessel widths at the
crossing region and the normal vessel width. For all of these
measures, the higher the value, the greater the severity of AV
nicking at that crossover.
III. EXPERIMENTAL RESULTS
A. Material
The proposed method was evaluated using 47 high-resolution
retinal images obtained from two population-based studies, the
Blue Mountain Eye Study (BMES) [25][27] and the Singapore
Malay Eye Study (SiMES) [28], [29]. BMES photos were taken
in 35-mm color lm, then digitized using CanoScan FS2710
(Canon, Tokyo, Japan) set to automatic exposure and focus. Im-
ages were converted to 24-bit (8-bit for each color space of red,
green, and blue, with resolution 2720 dpi without enhancement,
in Tagged Image File Format, image resolution of 3888 2592
pixels). Retinal images in SiMES set were captured using a dig-
ital nonmydriatic retinal camera (Canon CR-DGi, Japan) with
the image resolution of 3072 2048 pixels.
From these images, 90 detected crossover points were se-
lected to evaluate the performance of the proposed method for
AV nicking assessment. Each crossover point was manually
graded by two experts at the Centre for Eye Research Australia
(Melbourne, Australia) independently, using 4-scale grading
system(from0 =normal to 3 =most severe). Disagreement be-
tween two experts was then reassessed in a joint session, which
resulted in a single grading for each crossing point. We note that
this evaluation process is very time consuming and expensive
with respect to graders time. Table II summarizes the distribution
of AV nicking levels within 90 crossover points in this dataset.
A set of four example crossover points with varying nicking
levels are shown in Fig. 16. The dataset of these 90 crossover
points with the manual assessment has been made publicly avail-
able at: http://people.eng.unimelb.edu.au/thivun/projects/AV_
nicking_quantication/.
NGUYEN et al.: AN AUTOMATED METHOD FOR RETINAL ARTERIOVENOUS NICKING QUANTIFICATION FROM COLOR FUNDUS IMAGES 3201
TABLE II
DISTRIBUTION OF AV NICKING LEVELS ACROSS 90 WORKING CROSSOVERS
Fig. 16. Examples of crossover points with increasing AV nicking level, from
(a) = 0 (normal) to (d) = 3 (most severe).
TABLE III
PERFORMANCE IN TERMS OF SPEARMAN CORRELATION COEFFICIENT
ACHIEVED BY THREE PROPOSED MEASURES ON 90 WORKING CROSSOVERS
B. Results and Discussions
In the rst experiment, we examine the applicability of the
three proposed measures (M
MF
, M
S
, and M
R
) for AV nicking
assessment. For this, the correlation of the computed values
and the manual grading was computed using the Spearmans
rank correlation coefcient . The Spearman coefcient mea-
sures the strength of association between two variables and it
assesses how well the relationship between two variables can be
described using a monotonic function [30]. In addition, the signi-
cation test result (P value) was reported to indicate if the corre-
lation is signicantly different fromzero. The correlation scores
achieved by the three measures with different coalescing func-
tions f
A
= {MIN, MAX, MEAN} are presented in Table III. It
should be noted that f
A
is the coalescing function (7) that is used
to produce a single AV nicking level for each crossover using
the AV nicking measurements of its two venular segments. The
results show that the M
S
measure achieves highest correlation
with a correlation score of = 0.70 (P < 0.0001) and when
MAX is used as the coalescing function. This means that the
raw difference in the vessel width at crossing region and the
normal vessel width, which is captured by M
S
measure, is a
good measurement of AV nicking. In addition, the severity level
at a crossover point should be assigned as the most severe of
its two individual segments (i.e., f
A
= MAX). It is interesting
to see that M
R
measure is far less effective than M
S
measure.
This means that the ratio between the normal vessel width and
the width at the crossing region (M
R
measures) is not effective
for AV nicking measurement compared to the raw difference
Fig. 17. Box plot showing the linear relationship between the AV nicking
values computed using M
S
measure against manual grading.
(M
S
measure). Finally, the fraction of the missing pixels at the
crossing region seems to be a poor measure for AV nicking
assessment.
From the experiment above, we can see that M
S
is the most
effective measure for AV nicking assessment. To visually assess
the results, the box plot in Fig. 17 shows the AV nicking values
computed using M
S
measure against the manual grading values.
Each box in the box plot graphically summarizes the computed
AV nicking values of crossovers with the same manual grading
level using ve numbers: the red central mark is the median, the
edges of the box are the 25th and 75th percentiles, the lowest
and highest lines are the smallest, and largest computed values,
respectively. It is shown that there is a clear linear association
between the subjective grading and objective measurements.
In addition, it also indicates that the computed values differ
from subjective assessment mostly by one grading level, which
is reected by small overlap in the computed values of two
successive grades, while they are well separated for grades of
two level apart.
To roughly assess the performance of our system in terms of
classication accuracy, we divide all crossovers into two classes
(we refer to this as binary separation) by performing thresh-
olding on the manual grading. With different threshold values,
different separations are obtained. For example, the separation
([0] vs. [1, 2, 3]) means that crossover points manually marked
as 0 are classied as the rst class while the remaining points are
classied as the second class. The performance in terms of ROC
curve is then evaluated for each such separation. In addition, the
separations are classied as 1-level separation and 2-level sep-
aration if the difference in the upper bound value of the rst
class is one or two levels apart from the lower bound value
3202 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 11, NOVEMBER 2013
Fig. 18. ROC curves obtained on 1-level separations.
Fig. 19. ROC curves obtained on 2-level separations.
of the second class, respectively. For example, the separation
([0, 1] vs. [2, 3]) is considered as 1-level separation while ([0]
vs. [2, 3]) is considered as 2-level separation.
Figs. 18 and 19 show the ROC curves together with the area
under the ROCcurve (AUC) obtained on 1-level and 2-level sep-
arations, respectively. The results show that high-classication
performance are achieved for all separations (AUC 0.86). In
addition, the ROC curves in Fig. 18 show that the separations
(([0, 1] vs. [2, 3]) and ([0, 1, 2] vs. [3])) achieve higher per-
formance than the ([0] vs. [1, 2, 3]) separation, which reects
that the system is more effective in detecting moderate to severe
AV nicking cases than the mild cases. This is of clinical impor-
tance since more severe AV nicking cases need to be effectively
identied for further examination. The results also showthat the
performance achieved with 2-level separations are much higher
than with 1-level separations, which further conrms that the
computed AV nicking values differ from subjective assessment
mostly at one grading level.
TABLE IV
THE PERFORMANCE OBTAINED BY THE SYSTEM ON DIFFERENT BINARY
SEPARATIONS IN TERMS OF MA, SEN, SPEC, AND
Table IV presents the performance of the proposed method
in terms of maximum accuracy (MA), sensitivity (SEN), speci-
city (SPEC), and Kappa measure () on different binary sep-
arations. MA is dened as the maximum accuracy obtained in
classifying crossover points into two classes using the computed
AV nicking values when compared with the expert grading. The
SEN, SPEC, and Kappa value are reported at the threshold
value that yields highest classication accuracy (i.e., MA). SEN
measures the proportion of actual positives which are correctly
identied, while SPEC measures the proportion of negatives
which are correctly identied. Kappa value is used to measure
the agreement in the grading of AV nicking produced by the
system when compared with expert judgment [31].
The results show that the accuracy of our method at all sepa-
rations is higher than 80%. In addition, the system can separate
moderate to severe cases from normal or mild cases (i.e., at ([0,
1] vs. [2, 3]) separation) with an accuracy of 88% (SEN = 0.77,
SPEC = 0.92, = 0.70). Higher performance is achieved in
classifying moderate to severe cases from normal cases (i.e.,
at ([0] vs. [2, 3]) separation) with an accuracy of 89% (SEN
= 0.85, SPEC = 0.92, = 0.76). Moderate agreements are
achieved at two separations, ([0] vs. [1, 2, 3], = 0.59) and
([0, 1, 2] vs. [3], = 0.52), while substantial agreements are
observed for the remaining separations ( 0.70) (according
to the interpretation of Kappa value in [31]).
IV. CONCLUSION
We have proposed and validated a new method for automated
AV nicking assessment. The accuracy of the method was eval-
uated on 90 AV crossover points of 47 high-resolution retinal
images. The results show that the computed AV nicking mea-
surements are strongly correlated with human judgment with a
correlation score of 0.70. In addition, the box plot and the ROC
curves indicate that the predicted values differ from manual
grading mostly at one grading level. In terms of classication
accuracy, the proposed system can detect moderate and severe
AV nicking cases with an accuracy of 88% (SEN of 0.77, SPEC
of 0.92, Kappa value of 0.70). These have demonstrated the
reliability and accuracy of the proposed system for AV nick-
ing assessment. The measurements produced by the proposed
method provide the basis toward the development of a systemfor
automatic AV nicking detection for a large-scale screening sys-
tem. Moreover, since the quantied values provide more details
on the AV nicking severity level, the computed measurements
may help to enhance the relationship between AV nicking and
NGUYEN et al.: AN AUTOMATED METHOD FOR RETINAL ARTERIOVENOUS NICKING QUANTIFICATION FROM COLOR FUNDUS IMAGES 3203
known diseases such as hypertension and stroke. At present, we
are applying our method on a large number of images to deter-
mine AV nicking severity level associated with cardiovascular
diseases. The results will be reported later to the appropriate
clinical journal.
ACKNOWLEDGMENT
The authors wish to thank to K. Y. Lee and L. Hodgson at the
Centre for Eye Research Australia (Melbourne, Australia) for
kindly providing us with the arteriovenous nicking grading and
clinical advice.
REFERENCES
[1] T. Y. Wong and P. Mitchell, Hypertensive retinopathy, New Engl. J.
Med., vol. 351, no. 22, pp. 23102317, 2004.
[2] T. Wong and P. Mitchell, The eye in hypertension, Lancet, vol. 369,
no. 9559, pp. 425435, 2007.
[3] R. Klein, B. Klein, S. Moss, and Q. Wang, Hypertension and retinopathy,
arteriolar narrowing, and arteriovenous nicking in a population, Arch.
Ophthalmol., vol. 112, no. 1, pp. 9298, 1994.
[4] T. Y. Wong, R. Klein, A. R. Sharrett, T. A. Manolio, L. D. Hubbard,
E. K. Marino, L. Kuller, G. Burke, R. P. Tracy, and J. F. Polak, The
prevalence and risk factors of retinal microvascular abnormalities in older
persons: The cardiovascular health study, Ophthalmology, vol. 110, no. 4,
pp. 658666, 2003.
[5] P. Mitchell, J. Wang, T. Wong, W. Smith, R. Klein, and S. Leeder, Retinal
microvascular signs and risk of stroke and stroke mortality, Neurology,
vol. 65, no. 7, pp. 10051009, 2005.
[6] T. Wong, R. Klein, D. Couper, L. Cooper, E. Shahar, L. Hubbard,
M. Wofford, and A. Sharrett, Retinal microvascular abnormalities and
incident stroke: The atherosclerosis risk in communities study, Lancet,
vol. 358, no. 9288, pp. 11341140, 2001.
[7] T. Wong, R. Klein, B. Klein, J. Tielsch, L. Hubbard, and F. Nieto, Reti-
nal microvascular abnormalities and their relationship with hypertension,
cardiovascular disease, and mortality, Survey Ophthalmol., vol. 46, no. 1,
pp. 5980, 2001.
[8] N. Roper, S. R. Patel, and C. ODonnell, Signs of stroke in the retina,
Optometry Practice, vol. 13, no. 1, pp. 918, 2012.
[9] T. Wong, R. Klein, A. Sharrett, B. Duncan, D. Couper, B. Klein,
L. Hubbard, and F. Nieto, Retinal arteriolar diameter and risk for hy-
pertension, Ann. Internal Med., vol. 140, no. 4, pp. 248255, 2004.
[10] L. S. Cooper, T. Y. Wong, R. Klein, A. R. Sharrett, R. N. Bryan,
L. D. Hubbard, D. J. Couper, G. Heiss, and P. D. Sorlie, Retinal mi-
crovascular abnormalities and MRI-dened subclinical cerebral infarc-
tion the atherosclerosis risk in communities study, Stroke, vol. 37, no. 1,
pp. 8286, 2006.
[11] M. Baker, P. Hand, J. Wang, and T. Wong, Retinal signs and stroke,
Stroke, vol. 39, no. 4, pp. 13711379, 2008.
[12] T. Wong and R. McIntosh, Hypertensive retinopathy signs as risk indica-
tors of cardiovascular morbidity and mortality, Brit. Med. Bull., vol. 73,
no. 1, pp. 5770, 2005.
[13] H. Yatsuya, A. R. Folsom, T. Y. Wong, R. Klein, B. E. K. Klein, and
A. R. Sharrett, Retinal microvascular abnormalities and risk of lacunar
stroke atherosclerosis risk in communities study, Stroke, vol. 41, no. 7,
pp. 13491355, 2010.
[14] R. Klein, B. E. Klein, S. E. Moss, and S. M. Meuer, The epidemiol-
ogy of retinal vein occlusion: The BeaverDam Eye Study, Trans. Amer.
Ophthalmol. Soc., vol. 98, pp. 133143, 2000.
[15] L. D. Hubbard, R. J. Brothers, W. N. King, L. X. Clegg, R. Klein,
L. S. Cooper, A. R. Sharrett, M. D. Davis, and J. Cai, Methods for eval-
uation of retinal microvascular abnormalities associated with hyperten-
sion/sclerosis in the atherosclerosis risk in communities study, Ophthal-
mology, vol. 106, no. 12, pp. 22692280, 1999.
[16] D. J. Couper, R. Klein, L. D. Hubbard, T. Y. Wong, P. D. Sorlie,
L. S. Cooper, R. J. Brothers, and F. J. Nieto, Reliability of retinal pho-
tography in the assessment of retinal microvascular characteristics: The
atherosclerosis risk in communities study, Amer. J. Ophthalmol., vol. 133,
no. 1, pp. 7888, 2002.
[17] U. T. V. Nguyen, A. Bhuiyan, L. A. F. Park, R. Kawasaki, T. Y. Wong,
J. J. Wang, P. Mitchell, and K. Ramamohanarao, Automated quantica-
tion of retinal arteriovenous nicking from colour fundus images, in Proc.
IEEE 35th Annu. Int. Conf. Eng. Med. Biol. Soc., 2013, pp. 58655868.
[18] U. T. V. Nguyen, A. Bhuiyan, L. A. F. Park, and K. Ramamohanarao, An
effective retinal blood vessel segmentation method using multi-scale line
detection, Pattern Recognit., vol. 46, no. 3, pp. 703715, 2012.
[19] U. T. V. Nguyen, A. Bhuiyan, L. A. F. Park, R. Kawasaki, T. Y. Wong,
and K. Ramamohanarao, Automatic detection of retinal vascular land-
mark features for colour fundus image matching and patient longitudinal
study, presented at the IEEE Int. Conf. Image Process., Melbourne, VIC,
Australia, 2013.
[20] R. Gonzalez, R. Woods, and S. Eddins, Digital Image Processing Using
MATLAB. Noida, India: Pearson Education India, 2004.
[21] V. Bevilacqua, S. Camb, L. Cariello, and G. Mastronardi, A combined
method to detect retinal fundus features, in Proc. IEEE Eur. Conf. Emer-
gent Aspects Clin. Data Anal., 2005.
[22] E. Ricci and R. Perfetti, Retinal blood vessel segmentation using line
operators and support vector classication, IEEE Trans. Med. Imag.,
vol. 26, no. 10, pp. 13571365, Oct. 2007.
[23] S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and M. Goldbaum, De-
tection of blood vessels in retinal images using two-dimensional matched
lters, IEEE Trans. Med. Imag., vol. 8, no. 3, pp. 263269, Sep. 1989.
[24] X. You, Q. Peng, Y. Yuan, Y. Cheung, and J. Lei, Segmentation of retinal
blood vessels using the radial projection and semi-supervised approach,
Pattern Recognit., vol. 44, pp. 23142324, 2011.
[25] P. Mitchell, W. Smith, K. Attebo, and J. J. Wang, Prevalence of age-
related maculopathy inAustralia. the Blue Mountains Eye Study, Oph-
thalmology, vol. 102, no. 10, pp. 145060, 1995.
[26] J. J. Wang, P. Mitchell, R. G. Cumming, and R. Lim, Cataract and
age-related maculopathy: The Blue Mountains Eye study, Neuro-
Ophthalmol., vol. 6, no. 4, pp. 317326, 1999.
[27] S. Cugati, R. G. Cumming, W. Smith, G. Burlutsky, P. Mitchell,
and J. J. Wang, Visual impairment, age-related macular degeneration,
cataract, and long-term mortality: The BlueMountainsEye Study, Arch.
Ophthalmol., vol. 125, no. 7, pp. 917924, 2007.
[28] A. W. P. Foong, S. M. Saw, J. L. Loo, S. Shen, S. C. Loon, M. Rosman,
T. Aung, D. T. H. Tan, E. S. Tai, and T. Y. Wong, Rationale and method-
ology for a population-based study of eye diseases in Malay people: The
Singapore Malay eye study (SiMES), Ophthalm. Epidemiol., vol. 14,
no. 1, pp. 2535, 2007.
[29] C. Y. Cheung, W. T. Tay, P. Mitchell, J. J. Wang, W. Hsu, M. L. Lee,
Q. P. Lau, A. L. Zhu, R. Klein, and S. M. Saw, Quantitative and quali-
tative retinal microvascular characteristics and blood pressure, J. Hyper-
tension, vol. 29, no. 7, pp. 13801391, 2011.
[30] Wikipediaa free encylopedia, (2013, May 08). [Online]. Available:http:
//en.wikipedia.org/wiki/Spearmans rank correlation coefcient
[31] A. J. Viera and J. M. Garrett, Understanding interobserver agreement:
The kappa statistic, Fam. Med., vol. 37, no. 5, pp. 360363, 2005.
Authors photographs and biographies not available at the time of publication.