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

7, JULY 2013 1851

Simultaneously Identifying All True Vessels From
Segmented Retinal Images
Qiangfeng Peter Lau

, Mong Li Lee, Wynne Hsu, and Tien Yin Wong

AbstractMeasurements of retinal blood vessel morphology
have been shown to be related to the risk of cardiovascular dis-
eases. The wrong identication of vessels may result in a large
variation of these measurements, leading to a wrong clinical di-
agnosis. In this paper, we address the problem of automatically
identifying true vessels as a postprocessing step to vascular struc-
ture segmentation. We model the segmented vascular structure as
a vessel segment graph and formulate the problem of identifying
vessels as one of nding the optimal forest in the graph given a
set of constraints. We design a method to solve this optimization
problem and evaluate it on a large real-world dataset of 2446 reti-
nal images. Experiment results are analyzed with respect to actual
measurements of vessel morphology. The results showthat the pro-
posed approach is able to achieve 98.9% pixel precision and 98.7%
recall of the true vessels for clean segmented retinal images, and
remains robust even when the segmented image is noisy.
Index TermsOphthalmology, optimal vessel forest, retinal im-
age analysis, simultaneous vessel identication, vascular structure.
I. INTRODUCTION
A
RETINAL image provides a snapshot of what is happen-
ing inside the human body. In particular, the state of the
retinal vessels has been shown to reect the cardiovascular con-
dition of the body. Measurements to quantify retinal vascular
structure and properties have shown to provide good diagnostic
capabilities for the risk of cardiovascular diseases. For example,
the central retinal artery equivalent (CRAE) and the central reti-
nal vein equivalent (CRVE) are measurements of the diameters
of the six largest arteries and veins in the retinal image, respec-
tively. These measurements are found to have good correlation
with hypertension, coronary heart disease, and stroke [1][3].
However, they require the accurate extraction of distinct ves-
sels from a retinal image. This is a challenging problem due to
ambiguities caused by vessel bifurcations and crossovers.
Fig. 1(a) shows an example retinal image where vessels I and
II cross each other at two places (indicated by circles). These
Manuscript received August 31, 2012; revised December 7, 2012; accepted
January 21, 2013. Date of publication January 29, 2013; date of current version
June 24, 2013. This work was supported by A*STAR Exploit Flagship Grant
ETPL/10-FS0001-NUS0. Asterisk indicates corresponding author.

Q. P. Lau is with the Department of Computer Science, National University

of Singapore, 117417 Singapore (e-mail: plau@comp.nus.edu.sg).
M. L. Lee and W. Hsu are with the Department of Computer Science, National
University of Singapore, 117417 Singapore (e-mail: leeml@comp.nus.edu.sg;
whsu@comp.nus.edu.sg).
T. Y. Wong is with the Singapore Eye Research Institute, Singapore National
Eye Centre, 168751 Singapore (e-mail: tien_yin_wong@nuhs.edu.sg).
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.2013.2243447
Fig. 1. Vessels, I and II, crossing each other twice at white circles. (a) Wrong
Identication of I and II. (b) Correct Identication of I and II.
Fig. 2. (a) Vessel III wrongly connected to a segment that should belong to IV.
(b) Vessel IV correctly identied.
crossovers are often mistaken as vessel bifurcations, leading to
I and II being regarded as a single vessel. Fig. 1(b) shows the
correctly identied vessel structures for vessels I and II marked
in blue and red, respectively. Note that the line segment at the
second crossing (larger circle) is shared by vessels I and II.
In order to disambiguate between vessels at bifurcations and
crossovers, we need to gure out if linking a vessel segment
to one vessel will lead to an adjacent vessel being wrongly
identied. For example, in Fig. 2(a), if we identify vessel III
rst without any knowledge of vessel IV, the junction indicated
0018-9294/$31.00 2013 IEEE
1852 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 7, JULY 2013
by the white arrow may be mistaken as a bifurcation instead of
a crossover. Consequently, vessels III and IV will be incorrectly
identied, leading to a large difference in vessel measurements.
However, if both vessels were constructed and considered at the
same time, it becomes obvious that one of the branches of vessel
III should be an extension of vessel IV, as shown in Fig. 2(b). By
considering multiple vessels simultaneously, information from
other vessels can be used to better decide on the linking of vessel
segments.
In this paper, we describe a novel technique that utilizes the
global information of the segmented vascular structure to cor-
rectly identify true vessels in a retinal image. We model the
segmented vascular structure as a vessel segment graph and
transform the problem of identifying true vessels to that of
nding an optimal forest in the graph. An objective function
to score forests is designed based on directional information.
Our proposed solution employs candidate generation and expert
knowledge to prune the search space. We demonstrate the effec-
tiveness of our approach on a large real-world dataset of 2446
retinal images. The proposed technique has been incorporated
as part of the semiautomated Singapore Eye Vessel Assessment
(SIVA) system that has been used in real-world studies in both
the community and hospital-based patient populations [3], [4].
II. RELATED WORK
Retinal vessel extraction involves segmentation of vascular
structure and identication of distinct vessels by linking up
segments in the vascular structure to give complete vessels.
One branch of works, termed vessel tracking, performs vessel
segmentation and identication at the same time [5][8]. These
methods require the start points of vessels to be predetermined.
Each vessel is tracked individually by repeatedly nding the
next vessel point with a scoring function that considers the
pixel intensity and orientation in the vicinity of the current
point in the image. Bifurcations and crossovers are detected
using some intensity prole. Tracking for the same vessel then
continues along the most likely path. This approach of tracking
vessels one-at-a-time does not provide sufcient information for
disambiguating vessels at bifurcations and crossovers.
Another branch of works treat vessel identication as a post-
processing step to segmentation [9][11]. The work in [9] re-
quired the user to resolve the connectivity of bifurcation and
crossover points before vessels were individually identied.
For [10], a graph formulation was used with Dijkstras shortest-
path algorithm to identify the central vein. Similarly, Joshi
et al. [11] used Dijkstras algorithm to identify vessels one-
at-a-time and evaluated their method on a set of 15 images.
However, these methods may lead to incorrect vessel identica-
tion because choosing the correct vessel segment to connect at a
bifurcation or crossover requires information from other nearby
vessels. Al-Diri et al. [12] used expert rules to resolve vessel
crossovers and locally linked up segments at these crossovers
to give a vascular network. However, they did not identify com-
plete vessels.
Our work is focused on vessel identication as a post-
processing step to segmentation. Our approach differs from ex-
Fig. 3. (a) Zone of interest and vessel structures. (b) Line image of vessel
segmentation.
Fig. 4. Example of junctions. Pixels belonging to a junction are shaded.
isting works in that we identify multiple vessels simultaneously
and use global structural information to gure out if linking
a vessel segment to one vessel will lead to an overlapping or
adjacent vessel being wrongly identied.
III. PRELIMINARIES
We rst dene the zone of interest in the retinal image. This
is a circular ring bounded by two concentric circles of radii 2r
and 5r [see Fig. 3(a)], where r is the radius of the optic disc
(OD). Measurements from this zone are used in a number of
clinical studies [3], [4]. Each vessel starts from a pixel near the
circle of radius 2r. These pixels are called root pixels and are
denoted in yellow in Fig. 3(a).
We utilize existing vessel segmentation methods and apply
any skeletonization procedure [9], [13] to obtain the line image
in the zones of interest [see Fig. 3(b)]. The lines in the line image
depict the topological connectivity of the vessel structures.
Let P be the set of all white pixels in a line image. Two
pixels p
i
, p
j
P are adjacent, i.e., adj(p
i
, p
j
), if and only if
p
j
neigh8(p
i
), where neigh8(p) = p1, p2, . . . , p8 is the
eight-neighborhood of p [see Fig. 4(a)].
Denition 1 (Connected Pixels): Pixels p
i
, p
j
P are con-
nected, i.e., conn(p
i
, p
j
), if adj(p
i
, p
j
) or p
c
P p
i
, p
j

s.t. conn(p
i
, p
c
) conn(p
c
, p
j
).
Denition 2 (Pixel Crossing Number): Let p1, . . . , p8 be a
clockwise sequence of the eight neighbor pixels of pixel p.
Then, xnum(p) is the number of black to nonblack transitions
in this sequence of neighbor pixels of p.
Denition 3 (Junction): Let white8(p) neigh8(p) be the
set of white pixels that are neighbors of p. The set of junction pix-
els in P is }
P
= p P[xnum(p) > 2 [white8(p)[ > 3. A
junction is a set of connected junction pixels, i.e., J }
P
such
LAU et al.: SIMULTANEOUSLY IDENTIFYING ALL TRUE VESSELS FROM SEGMENTED RETINAL IMAGES 1853
Fig. 5. Example of segment pixels (in white) with their end pixels (in red),
and junction pixels (in blue).
that p
i
, p
j,=i
J, conn(p
i
, p
j
), where conn is restricted to the
set }
P
. Then, the set of all junctions in P is
P
.
Fig. 4 depicts examples of junction pixels. In Fig. 4(a), we
have white8(p) = p2, p4, p6, and xnum(p) = 3 due to the
transitions (p1, p2), (p3, p4), and (p5, p6). This is the straight-
forward case where the shaded pixel p is a junction pixel with
xnum(p) > 2. In Fig. 4(b), the top and left neighbors of the
shaded pixel have crossing numbers of 2 and hence are not
junction pixels. In Fig. 4(c), all four shaded pixels are junction
pixels since they each have more than 3 white pixels in their
eight-neighborhood.
Denition 4 (Segment): A segment s is a sequence of unique
white pixels p
1
, . . . , p
n
) in P such that all of the following
conditions are true:
1) n > 0 and i [1, n], p
i
/
P
2) n > 1 i [1, n 1], adj(p
i
, p
i+1
)
3) i 1, n, [white8(p
i
)[ = 1
p
j

P
s.t. adj(p
i
, p
j
)
4) n > 2 i [2, n 1], xnum(p
i
) = 2.
We call p
1
and p
n
the end pixels of s. Let S
P
be the set of all
segments in P and ^
P
= P }
P
, i.e., ^
P
contains nonjunc-
tions pixels that are part of segments. Then, s S
P
is adjacent
to a junction J, i.e., adj(s, J), if p
j
J s.t. adj(p
j
, p
1
)
adj(p
j
, p
n
). Consequently, two segments s
a
, s
b
S
P
are adja-
cent, adj(s
a
, s
b
) if J
P
s.t. adj(s
a
, J) adj(s
b
, J).
Fig. 5 shows the examples of segments, end pixels, and junc-
tion pixels according to Denitions 3 and 4 for a region from a
line image. Each segment is indicated by the connected white
pixels.
IV. GRAPH TRACER
Our proposed method aims to identify vessels and represent
them in the form of binary trees for subsequent vessel mea-
surements. It has two main steps: 1) identify crossovers, and 2)
search for the optimal forest (set of vessel trees). We describe
the details in the following sections.
A. Identify Crossover Locations
Vessels in a retinal image frequently cross each other, at a
point or over a shared segment. We call the former crossover
points and the latter crossover segments.
Denition 5 (Crossover Point): Given the set of white pixels
P in a line image, a junction J
P
is a crossover point if and
only if the number of segments that are adjacent to J is greater
Fig. 6. Example crossover segment, point, and possible ambiguity. (a) Ex-
ample of a crossover segment. (b) Example of a short segment between two
junctions (white arrow).
Fig. 7. Examples of directional change between segments.
than or equal to 4, i.e., cross(J) is true iff [s S
P
[adj(s, J)[
4.
For example, the lower junction in Fig. 5 is a crossover point
as it has four segments adjacent to it.
A crossover segment occurs when two different vessels
share a segment as shown in Fig. 6(a). Given the set of
white pixels P of a line image, a segment s S
P
is a can-
didate crossover segment if [s[ < L and J
1
, J
2

P
s.t.
adj(s, J
1
) adj(s, J
2
) cross(J
1
) cross(J
2
). Lis a pa-
rameter to limit candidates to short segments.
Note that short segments between two junctions are not nec-
essary true crossover segments, as shown in Fig. 6(b). Hence,
we propose to use the directional change between adjacent seg-
ments and their pixel intensity values to differentiate crossover
segments.
Denition 6 (Directional Change Between Segments): Given
two segments s
a
and s
b
that are adjacent to a common junction,
let p
a
and p
b
be the end points of s
a
and s
b
that are nearest to
each other. Let v
a
be a vector that starts on s
a
and ends at p
a
,
and v
b
be a vector that starts from p
b
and ends on s
b
. Then, the
directional change between s
a
and s
b
is given by
D(s
a
, s
b
) = cos
1

v
a
v
b
[v
a
|v
b
[

where D(s
a
, s
b
) [0

, 180

].
Intuitively, D(s
a
, s
b
) measures the magnitude of a change
in direction if we were to go from s
a
to s
b
. Fig. 7 shows ex-
ample angles representing directional change between various
segments using Denition 6.
Denition 7 (Crossover Segment): Given a candidate seg-
ment seg between two junctions J
1
and J
2
, let S
i
= s
a

S
P
[adj(s
a
, J
i
) s
a
,= seg for i 1, 2. Each S
i
contains
two segments sharing the same junction as one end pixel of
seg. Let A = seg S
1
S
2
and = s
a
, seg, s
b
[s
a

S
1
, s
b
S
2
. Then seg is a crossover segment, i.e., cross(seg)
is true, if all of the following conditions are true:
1854 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 7, JULY 2013
Fig. 8. Segment 3 is a crossover segment, while segment 4 is not a crossover
segment because it does not satisfy condition 1 in Denition 7.
1) s, s
/
S
i
, i 1, 2, D(s, s
/
) > 30

2) [seg[ L

[s
a
, s
b
S
1
, s
c
, s
d
S
2
,
s.t. D(s
a
, s
c
) < 30

D(s
b
, s
d
) < 30

]
min

[sd(M()) +sd(M(A))] < sd(M(A))

3) [seg[ > L

[s S
1
S
2
, D(seg, s) <
low
]
[s S
1
S
2
, D(seg, s) <
high
min

[sd(M()) +sd(M(A))] < sd(M(A))]

where (s) is the mean intensity of the pixels in segment s, the
bag M(S) = (s)[s S for a set of segments S, and sd is
the standard deviation of the numbers in M(S).
Condition 1 of Denition 7 handles the case when seg is at a
bifurcation. For example, segment 4 in Fig. 8 is not a crossover
segment due to the small directional change between segments
1 and 5.
Condition 2 in Denition 7 handles the case when the length
of seg is too short to determine the directional change. In this
case, we check if the adjacent segments of seg forms a rea-
sonable cross pattern, i.e., if there exists some pairing of the
segments in S
1
with those in S
2
such that their directional
change are less than 30

. Otherwise, we partition A into two

such that the sum of the sd of both partitions is minimum. If
this minimum is less than the sd of all the segments in A, then
seg is a crossover segment.
Condition 3 of Denition 7 states that if the length of seg
is long enough and the directional change between seg and
each of its adjacent segment is less than
low
, then seg is a
crossover segment. Otherwise, if directional change is less than

high
, we compare the sds of the segments intensity values as
in Condition 2.
Note that
low
and
high
can be determined empirically. For
our experiments, we set
low
= 65

and
high
= 85

. Fig. 9
shows the crossover segments identied for the retinal image in
Fig. 3(a).
B. Find the Optimal Forest
Next, we model the segments as a segment graph and use
constraint optimization to search for the best set of vessel trees
(forest) from the graph.
Denition 8 (Segment graph): Given the set of white pixels P
in a line image, a segment graph G
P
= (S
P
, E
P
), where each
vertex in S
P
is a segment and an edge e
i,j
= (s
i
, s
j
) E
P
exists if adj(s
i
, s
j
), s
i
, s
j
S
P
, i ,= j.
Fig. 9. Identied crossover segments for Fig. 3(a), highlighted in red as indi-
cated by the white arrows.
Fig. 10. (a) Segment graph corresponding to the segments in Fig. 8. (b) Ex-
ample forest of two binary trees (gray and black) corresponding to two vessels
rooted at segments 1 and 2 in Fig. 8.
Typically, G
P
consists of disconnected subgraphs that are
independent and can be processed in parallel. Without loss of
generality, we refer to each of these subgraphs as the segment
graph G
P
. The goal is to obtain a set of binary trees from the
segment graph such that each binary tree corresponds to a vessel
in the retinal image.
Denition 9 (Vessel): Given a segment graph G
P
=
(S
P
, E
P
), a vessel is a binary tree, T = (s
root
, V
T
, E
T
) such
that s
root
is the root node, root(T) = s
root
, V
T
S
P
, and
E
T
E
P
. A set of such binary trees is called a forest.
A binary tree is a natural representation of an actual blood
vessel as it only bifurcates. Segment end points near the inner
circle of the zone of interest are automatically identied as root
pixels. The root of each tree corresponds to the root segment
that contains a unique root pixel, i.e., the yellow dots in Figs.
1 and 2. Fig. 10 shows the segment graph and two binary trees
corresponding to the two vessels in Fig. 8. We formulate the
goal of simultaneous identication as a constraint optimization
problem (COP).
Given a segment graph G
P
= (S
P
, E
P
), and a set of root
segments S
root
, let T
P
be the set of all possible forests fromG
P
for each root segment in S
root
. The optimal forest, F

T
P
,
that corresponds to vessels in G
P
is given by
F

= argmin
F T
P
cost(F)
LAU et al.: SIMULTANEOUSLY IDENTIFYING ALL TRUE VESSELS FROM SEGMENTED RETINAL IMAGES 1855
subject to the following constraints:
1) Roots are unique to each tree
T
1
F, T
2
F T
1
, root(T
2
) , V
T
1
.
2) Directional change between parent and child segments are
within the threshold
T F, (s
p
, s
c
) E
T
,
[s
p
> L

s
c
> L

] D(s
p
, s
c
) < 135

.
3) Any segment appearing in more than one tree must be a
crossover segment
s o
P
, [T F[s V
T
[ > 1 cross(s).
4) A parent segment at crossover junction must connect to
the child with minimum directional change
T F, (s
p
, s
c
) E
T
,
J
P
s.t.cross(J) adj(s
p
, J) adj(s
c
, J)
child(T, s
p
) = 1 s
c
= argmin
s/
D(s
p
, s)
where / = s S
P
s
p
[adj(s, J).
5) Crossover segment is the only child and have only one
child that has the minimum directional change
T F, (s
p
, s
x
), (s
x
, s
c
) E
T
,
cross(s
x
) child(T, s
p
) = 1 child(T, s
x
) = 1
s
c
= argmin
so
D(s
p
, s)
where o = s[(s
x
, s) E
P
s ,= s
p
adj(s, s
p
).
6) Leaf segments cannot be crossovers segments
T F, s V
T
, leaf(T, s) cross(s).
In addition to the above constraints, each binary tree T F
should not contain cycles. Consequently, a segment may not
appear twice in the same vessel.
Our cost function makes use of change in direction between
segments. For a vessel T, let the set of bifurcations be
Y
T
= (s
y
, s
1
, s
2
)[s
y
, s
1
, s
2
V
T
(s
y
, s
1
), (s
y
, s
2
) E
T
.
Further, let the set of single parentchild nodes in T be
I
T
= (s
p
, s
c
)[s
p
, s
c
V
T
child(s
p
) = 1
[(cross(s
p
) cross(s
c
) (s
p
, s
c
) E
T
)
((s
p
, s
m
), (s
m
, s
c
) E
T
s.t. cross(s
m
) child(s
m
) = 1)].
Then, we dene the following functions with Y
T
and I
T
:

Y
(T) =

(s
y
,s
1
,s
2
)Y
T
0.5[D(s
y
, s
2
) + D(s
y
, s
1
)]

I
(T) =

(s
p
,s
c
)I
T
D(s
p
, s
c
).

Y
(T) sums the average of the parentchild directional changes
at bifurcations in T; hence, smaller D are preferred as child
segments seldom branch off at obtuse angles to the parent seg-
ment.
I
(T) sums the change in direction between parents in
the tree with only one child segment. This favors smaller direc-
tional changes when choosing between segments to connect at
junctions. Finally, the cost function on forests is dened as
cost(F) =

T F
[
I
(T) +
Y
(T)].
This COP differs from other graph problems in several ways.
First, instead of dening the cost on edges as in minimum span-
ning tree (MST) and minimum spanning forest (MSF) prob-
lems [14], [15], we dene the cost on forests to allow weighting
and fusing of multiple cost criteria at the forest level. Second,
we nd a forest from a connected graph, while MSF nds a tree
in each graph [16]. Third, our denition of vessel trees does not
allow us to use the weighted-SAT formulation in [13] as it may
produce broken vessels.
To solve the COP, we use a candidate enumeration algorithm
that utilizes the lower bound of the cost function to prune the
search space. This lower bound LB
cost
(F) is based on the
following theorem.
Theorem 1 (Lower bound of cost): Given a set of binary trees
F and any vessel T F, we construct the vessel T
/
by growing
one leaf node of T such that it has either one or two chil-
dren. Let F
/
= F T T
/
. Then, cost(F) cost(F
/
),
i.e., cost(F) is the lower bound cost of any F
/
resulting from
growing the vessels in F.
Proof: By adding new children to a leaf node, we increase the
size of I
T
by one, Y
T
by one, or neither, but not both. As Dhas
the codomain [0

, 180

],
I
(T)
I
(T
/
)
Y
(T)
Y
(T
/
).
Thus, cost(F) cost(F
/
).
Fig. 11 shows the details of our tracing algorithm,
GraphTracer. The input is the segment graph G
P
with n root
segments given in S
root
. Lines 16 initialize the global variables
and call the recursive procedure Trace. F[1. .n] corresponds to
the initial forest of n vessels. R[1. .n] denotes a fringe stack for
each vessel. F
min
and c
min
record the minimum cost forest and
its corresponding cost.
In Trace, if F satises all constraints and cannot be grown
further, we update F
min
if cost(F) < c
min
(Line 7). Otherwise,
we may prune descendant forests grown from F with the lower
bound LB
cost
(F) (at Line 9). The outer loop at Line 10 orders
each vessel T [1, n] for growth. T ranges from the current
index i to n, ensuring that Trace does not enumerate duplicate
forests. Each vessels fringe stack, R[T], stores its current leaf
nodes to be grown. R[T] is used in conjunction with the loop at
Line 11 to enumerate vessels in a depth-rst traversal order.
A subprocedure FindChildren returns pairs (s
l
, s
r
) of pos-
sible children for the current fringe node s
T
. If only one child
is to be added, we set s
r
= . FindChildren employs forward
checking to eliminate children pairs that violate constraints (1)
(6) in the COP formulation.
The time complexity of GraphTracer is exponential to the
number of edges in G
P
and is independent of the size of the
retinal image as it only deals with the connectivity of entire
1856 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 7, JULY 2013
Fig. 11. Details of Algorithm GraphTracer.
segments without measuring pixel properties such as intensity.
However, in practice, FindChildren eliminates many combi-
nations using the constraints presented.
V. EXPERIMENT RESULTS
We evaluate our proposed method on 2446 retinal images
of patients from the Singapore Malay eye study [17]. For each
image, the line image of the retinal vessels is obtained using
the semiautomated retinal image analysis tool, SIVA. Trained
human graders then follow a protocol to verify the correctness
of the vascular structure obtained, e.g., arteries, veins, crossover
locations, and branch points. We use these veried vascular
structures as the gold standard and call the corresponding vessel
center lines as clean line images.
We implement the Graph tracer and a Solo tracer that traces
vessels individually without regard for other vessels. The Solo
tracer works as follows: it starts from one root pixel and follows
the adjacent pixels in the line image. When a split is encountered,
a local lookahead is done to inspect the directional change of
the segments. If they t the crossover prole, the split is treated
as a crossover; otherwise, it is a bifurcation and the tracer will
follow both paths. It is greedy because unless a crossover is
identied, it will add all the connected pixels to the same vessel.
All tracers are given the same OD, line image, artery/vein
labeling, and use the same method to compute the vessel diam-
eters. We evaluate their performance on both clean and noisy
line images. Noisy line images are obtained using an existing
vessel segmentation algorithm [18] and is representative of the
real-world situation where segmentation is often imperfect. We
Fig. 12. Results of tracers on clean and noisy line images.
use the following evaluation metrics based on the pixels in the
entire vessels. Let Big6 refer to the six largest arteries and veins
ranked by the average width of the rst segment of each vessel.
Further, if a pixel of a traced vessel exists in the gold standard,
it is called a matched pixel.
1) Pixel precision: Total number of matched pixels divided
by total number of traced vessel pixels.
2) Pixel recall: Total number of matched pixels divided by
total number of gold standard pixels.
3) Big6 precision: Total number of matched pixels divided
by total number of traced pixels of Big6.
4) Big6 recall: Total number of matched pixels divided by
total number of gold standard pixels of Big6.
In our rst set of experiments, we use both clean and noisy
line images as inputs to the tracers. Fig. 12 shows the results.
For the clean line images, both Solo and Graph tracers display
good performance. In particular, the Graph tracer is able to
achieve near perfect pixel precision (98.9%) and pixel recall
(98.7%). The performance of both tracers decrease for noisy
line images. We observe that the difference between the Solo
tracer and Graph tracer is more pronounced, indicating that the
Graph tracer is more robust. Fromthese results, we conclude that
tracing all vessels simultaneously is better than tracing vessels
individually without current knowledge of other vessels.
For the second set of experiments, we analyze the impact of
our methods on measurement quality by computing the Pearson
correlation coefcient (PCC) between the measurements of the
traced vessels and the gold standard. These measurements are
automatically computed from the traced and gold standard vas-
cular structures, respectively. The vessel measurements CRAE,
CRVE, and average curvature tortuosity of arteries (CT
a
) and
veins (CT
v
) have been found to be correlated with risks factors
of cardiovascular diseases and are positive real numbers.
CRAE and CRVE are computed by iteratively combining the
mean widths of consecutive pairs of vessels in the Big6 arteries
and veins [19], respectively, as follows:
Arteries: w = 0.88 (w
2
1
+w
2
2
)
1
2
Veins: w = 0.95 (w
2
1
+w
2
2
)
1
2
where w
1
, w
2
is a pair of width values and wis the newcombined
width value for the next iteration. Iteration stops when one width
value remains.
LAU et al.: SIMULTANEOUSLY IDENTIFYING ALL TRUE VESSELS FROM SEGMENTED RETINAL IMAGES 1857
Fig. 13. PCCwith the gold standard for clean and noisy line images. Subscripts
indicate veins (v), or arteries (a).
Fig. 14. Example of Solo versus Graph tracer on a clean line image. (a) Solo
Tracer. (b) Graph Tracer.
The average curvature tortuosity measures how tortuous the
center lines of arteries or veins are. The curvature tortuosity
of a vessel is the average of its segments curvature measures
weighted by length [3]
ct(T) =

sT

4
(s) [s[

sT
[s[
where
4
is the curvature of a line segment given in [20].
Fig. 13 shows the PCCof various measures for both clean and
noisy line images. It is clear that the Graph tracer performs better
than the Solo tracer, particularly for the CT measures. This is
because Graph tracer is able to more accurately trace the small
vessels. The PCC for noisy line images reafrms that Graph
tracer is more robust than the Solo tracer in the presence of
noise. We observe that measurements of veins are consistently
more correlated than those of arteries, indicating that arteries
are more difcult to segment than veins.
Visual inspection on the results reveals that the Solo tracer
often traces overlapping vessel segments or wrongly connected
bifurcations that lead to poor measurements. An example mis-
take made by the Solo tracer on a clean line image is shown
in Fig. 14. The arrows indicate the location where two vessels
cross near bifurcations that causes the Solo tracer to erroneously
link segments belonging to other vessels. In contrast, the Graph
tracer takes both overlapping vessels into account when jointly
identifying the vessels resulting in better identication.
VI. CONCLUSION
We have presented a novel technique to identify true vessels
fromretinal images. The accurate identication of vessels is key
to obtaining reliable vascular morphology measurements for
clinical studies. The proposed method is a postprocessing step
to vessel segmentation. The problem is modeled as nding the
optimal vessel forest from a graph with constraints on the vessel
trees. All vessel trees are taken into account when nding the
optimal forest; therefore, this global approach is acutely aware
of the mislinking of vessels. Experiment results on a large real-
world population study show that the proposed approach leads
to accurate identication of vessels and is scalable.
REFERENCES
[1] T. Y. Wong, F. M. A. Islam, R. Klein, B. E. K. Klein, M. F. Cotch, C. Castro,
A. R. Sharrett, and E. Shahar, Retinal vascular caliber, cardiovascular
risk factors, and inammation: The multi-ethnic study of atherosclerosis
(MESA), Invest Ophthalm. Vis. Sci., vol. 47, no. 6, pp. 23412350, 2006.
[2] K. McGeechan, G. Liew, P. Macaskill, L. Irwig, R. Klein, B. E. K. Klein,
J. J. Wang, P. Mitchell, J. R. Vingerling, P. T. V. M. Dejong, J. C. M.
Witteman, M. M. B. Breteler, J. Shaw, P. Zimmet, and T. Y. Wong, Meta-
analysis: Retinal vessel caliber and risk for coronary heart disease, Ann.
Intern. Med., vol. 151, no. 6, pp. 404413, 2009.
[3] C. Y.-L. Cheung, Y. Zheng, W. Hsu, M. L. Lee, Q. P. Lau, P. Mitchell, J. J.
Wang, R. Klein, and T. Y. Wong, Retinal vascular tortuosity, blood pres-
sure, and cardiovascular risk factors, Ophthalmology, vol. 118, pp. 812
818, 2011.
[4] C. Y.-L. Cheung, W. T. Tay, P. Mitchell, J. J. Wang, W. Hsu, M. L. Lee,
Q. P. Lau, A. L. Zhu, R. Klein, S. M. Saw, and T. Y. Wong, Quantitative
and qualitative retinal microvascular characteristics and blood pressure,
J. Hypertens, vol. 29, no. 7, pp. 13801391, 2011.
[5] Y. Tolias and S. Panas, A fuzzy vessel tracking algorithm for retinal
images based on fuzzy clustering, IEEE Trans. Med. Imag., vol. 17,
no. 2, pp. 263273, Apr. 1998.
[6] H. Li, W. Hsu, M. L. Lee, and T. Y. Wong, Automatic grading of retinal
vessel caliber, IEEE Trans. Biomed. Eng., vol. 52, no. 7, pp. 13521355,
Jul. 2005.
[7] E. Grisan, A. Pesce, A. Giani, M. Foracchia, and A. Ruggeri, A new
tracking system for the robust extraction of retinal vessel structure, in
Proc. IEEE Eng. Med. Biol. Soc., Sep. 2004, vol. 1, pp. 16201623.
[8] Y. Yin, M. Adel, M. Guillaume, and S. Bourennane, Aprobabilistic based
method for tracking vessels in retinal images, in Proc. IEEE Int. Conf.
Image Process., Sep. 2010, pp. 40814084.
[9] M. Martinez-Perez, A. Highes, A. Stanton, S. Thorn, N. Chapman, A.
Bharath, and K. Parker, Retinal vascular tree morphology: A semi-
automatic quantication, IEEE Trans. Biomed. Eng., vol. 49, no. 8,
pp. 912917, Aug. 2002.
1858 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 7, JULY 2013
[10] H. Azegrouz and E. Trucco, Max-min central vein detection in retinal
fundus images, in Proc. IEEE Int. Conf. Image Process., Oct. 2006,
pp. 19251928.
[11] V. S. Joshi, M. K. Garvin, J. M. Reinhardt, and M. D. Abramoff, Auto-
mated method for the identication and analysis of vascular tree structures
in retinal vessel network, in Proc. SPIEConf. Med. Imag., 2011, vol. 7963,
no. 1, pp. 111.
[12] B. Al-Diri, A. Hunter, D. Steel, and M. Habib, Automated analysis of reti-
nal vascular network connectivity, Comput. Med. Imag. Graph., vol. 34,
no. 6, pp. 462470, 2010.
[13] K. Rothaus, X. Jiang, and P. Rhiem, X. Jiang, and P. Rhiem, Separation
of the retinal vascular graph in arteries and veins based upon structural
knowledge, Imag. Vis. Comput., vol. 27, pp. 864875, 2009.
[14] R. L. Graham and P. Hell, On the history of the minimum spanning tree
problem, IEEE Ann. Hist. Comput., vol. 7, no. 1, pp. 4357, Jan.-Mar.
1985.
[15] C. All` ene, J.-Y. Audibert, M. Couprie, and R. Keriven, Some links be-
tween extremum spanning forests, watersheds and min-cuts, Imag. Vis.
Comput., vol. 28, pp. 14601471, 2010.
[16] J. Cousty, G. Bertrand, L. Najman, and M. Couprie, Watershed cuts:
Minimum spanning forests and the drop of water principle, IEEE Trans.
Pattern Anal. Mach. Intell., vol. 31, no. 8, pp. 13621374, Aug. 2009.
[17] 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 methodology
for a population-based study of eye diseases in Malay people: SiMES,
Ophthalm. Epidemiol., vol. 14, no. 1, pp. 2535, 2007.
[18] S. Garg, J. Sivaswamy, and S. Chandra, Unsupervised curvature-based
retinal vessel segmentation, in Proc. IEEE Int. Symp. Biomed. Imaging,
Apr. 2007, pp. 344347.
[19] C. Y.-L. Cheung, W. Hsu, M. L. Lee, J. J. Wang, P. Mitchell, Q. P. Lau, H.
Hamzah, M. Ho, and T. Y. Wong, A new method to measure peripheral
retinal vascular caliber over an extended area, Microcirculation, vol. 17,
pp. 19, 2010.
[20] W. E. Hart, M. Goldbaum, P. Kube, and M. R. Nelson, Automated mea-
surement of retinal vascular tortuosity, in Proc. AMIA Fall Conf., 1997,
pp. 459463.
Authors photographs and biographies not available at the time of publication.