IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 27, NO.

2, FEBRUARY 2008 237

A Novel Vessel Segmentation Algorithm for

Pathological Retina Images Based on the
Divergence of Vector Fields
Benson Shu Yan Lam* and Hong Yan

Abstract—In this paper, a method is proposed for detecting blood to 140 are grouped into another level. In each level, candidate
vessels in pathological retina images. In the proposed method, vessels are obtained by thresholding. In a supervised method,
blood vessel-like objects are extracted using the Laplacian oper- the criteria are determined by the ground truth data based on
ator and noisy objects are pruned according to the centerlines,
which are detected using the normalized gradient vector field. The given features. However, a prerequisite for a supervised method
method has been tested with all the pathological retina images in is the availability of the ground truth data that are already
the publicly available STARE database. Experiment results show classified, which may not be available in real life applications.
that the method can avoid detecting false vessels in pathological An average of 2 h is needed to label a single retinal image [3].
regions and can produce reliable results for healthy regions. Staal et al. employ more than 10 features, including width of
Index Terms—Blood vessel segmentation, gradient vector field, the vessel, intensity, and edge strength [3]. Soares et al. make
image segmentation, retina image analysis. use of the Gabor wavelet transform [4]. As supervised methods
are designed based on preclassified data, their performance is
usually better than that of unsupervised ones and can produce
I. INTRODUCTION
very good results for healthy retinal images.
Although existing methods are robust for many retinal im-
N retinal images, blood vessels are landmarks for localizing
I the optic nerve, the fovea and lesions, which are useful for
medical diagnosis. However, in these images, many vessels are
ages, there is still room for further improvement, especially for
pathological retina images. A pathological retina may suffer
from a certain disease and there may contain some spots (light
narrow and close to each other, forming a network-like struc- or dark). Existing methods may recognize those spots as part of
ture. Also, due to the reflection on the tiny uneven surface of the the vessels. Due to the unknown characteristics of a patholog-
soft tissue in the image, the low contrast between the vessel and ical region, widely used features such as intensity are not effec-
background, and the pathological variations, detecting blood tive for solving the problem. The supervised method of Soares
vessels automatically from a retinal image is a challenging et al. has the same limitation. In their paper, the authors stated
problem. A number of techniques have been proposed to solve “Though very good ROC results are presented, visual inspec-
this problem. They can be classified into unsupervised and tion shows some typical difficulties of the method that must be
supervised methods. In an unsupervised method, a pixel is solved by future work. The major errors are in false detection of
assigned to a candidate vessel according to several predefined noise and other artifacts. False detection occurs in some images
criteria. Chaudhuri et al. propose a matched filter response for the border of the optic disc, haemorrhages, and other types
(MFR) method [1], which applies rotated Gaussian filters to of pathologies that present strong contrast” [4].
the image. If the pixel has a large filtered value, it is a part of Researchers have made many proposals to analyze patho-
a vessel. Jiang and Mojon propose an adaptive thresholding logical retina images. Chanwimaluang et al. suggest that more
technique for vessel segmentation [2]. The detection is con- constraints should be added in order to remove the spots [5].
ducted in different levels of image intensities. For example, the However, there is no discussion on how we can select the con-
pixels with intensity values from 80 to 100 are grouped into straints. One of the widely used constraints for noise removal is
one level while the pixels with the intensity values from 110 the split-and-merge system [6]. If the size of an object is small
enough, it will be treated as noise. An implicit assumption for
this pruning operation is that the size of the vessel should be
Manuscript received June 4, 2007; revised September 14, 2007. This work larger than that of noise. However, many blood vessels after
was supported by the City University of Hong Kong under Project 9610034.
Asterisk indicates corresponding author.
splitting are very short and can be removed easily. Staal et al.
*B. S. Y. Lam is with the Department of Electronic Engineering, City Uni- suggest to solve the problem by removing a pathological region
versity of Hong Kong, Kowloon, Hong Kong (e-mail: [email protected]. in a preprocessing step or by selecting more training data sets
edu.hk).
H. Yan is with the Department of Electronic Engineering, City University
that include pathological features in a supervised approach [3].
of Hong Kong, Kowloon, Hong Kong and with the School of Electrical and The removal of a pathological region in a preprocessing step can
Information Engineering, University of Sydney, Sydney, NSW 2006, Australia be difficult. Actually, as we do not know where a pathological
(e-mail: [email protected]). region will be, it is not easy to remove it in advance. One way
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. is to use an adaptive thresholding technique proposed by Jiang
Digital Object Identifier 10.1109/TMI.2007.909827 and Mojon [2]. They separate the image into several levels by
0278-0062/$25.00 © 2007 IEEE
238 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 27, NO. 2, FEBRUARY 2008

a thresholding method based on pixel intensities. However, if

there are large intensity variations within the spots, this method
may not work well. To produce more training data, one possible
solution is to trace the vessels from the optic nerve, a user-de-
fined point or a given labeled vessel on the image [7]–[15]. How-
ever, if the spots are close to the blood vessels, which are not
connected to the optic nerve or the given information, it may
not be possible to remove them.
Based on the above discussions, there is a clear need for a new
approach to detect vessels in pathological retina images. In this
paper, we propose a new method to solve this problem, espe-
cially for those pathological retina images containing bright ab-
normality. In our previous study, we proposed a partial differen-
tial equation based approach [16]. Partial differential equations
have been widely used in many image processing problems in-
cluding segmentation and inpainting [17]–[21]. This method is
robust for healthy retinal images, but it is not so for pathological
retina images. The method presented in this paper is an exten-
sion of our preliminary work in [16]. In the proposed method, Fig. 1. Illustrations of the divergence of vector fields. (a) Expanding vector
blood vessel-like objects are extracted using the Laplacian op- field. Each vector has the same magnitude. (b) Expanding vector field. (c) Con-
stant vector field. (d) Contracting vector field.
erator and noisy objects are pruned according to the centerlines,
which are detected using the normalized gradient vector field.
Experiment results show that the proposed method is able to
avoid detecting false vessels in pathological regions and can pro-
duce reliable results for healthy regions.
The organization of the paper is as follows. In Section II, we
review the divergence operator. The concept of the proposed
method is presented in Section III. In Section IV, we show the
implementation details for each step. The robustness of the pro-
posed method is shown by experiments in Section V. Finally,
discussions and conclusions are given in Section VI.

II. REVIEW OF DIVERGENCE OPERATOR

The divergence of the vector field has been widely used in
fluid dynamics and electrodynamics [22], [23]. The definition
of divergence of a vector field at point ( ) is given below
[24] Fig. 2. Flow chart of the proposed method.

(1)
III. ALGORITHM DESIGN

where is the volume of the sphere , is the surface, and A. Outline of the Proposed Method
is the outward unit normal vector of the surface of the sphere In the proposed method, blood vessel-like objects are ex-
with radius and center at ( ). The divergence of a vector tracted using the Laplacian operator and noisy objects are
field at point ( ) is equal to the flux across the point divided pruned according to the centerlines, which are detected using
by the volume of the sphere as approaches 0. the normalized gradient vector field. The method consists of
The physical meaning of this divergence operator is as fol- three main steps illustrated in the flow chart shown in Fig. 2.
lows. Since the divergence measures the flux across a point, it The implementation details of the method will be given in
will be positive if the vector field at the point is expanding. Oth- Section IV.
erwise, the divergence will be negative. If there is no change in
the vector field at the point, the divergence will be zero. Fig. 1(a) B. Locating the Centerlines Using the Normalized Gradient
shows an expanding vector field and each vector has the same Vector Field
magnitude but different directions. Fig. 1(b) shows another ex- We use the normalized gradient vector field to locate the cen-
ample of an expanding vector field. Fig. 1(c) shows an example terlines. As each pixel in the vector field has the same unit mag-
of a constant vector field, where each vector has the same mag- nitude, the sign of the divergence depends only on the direc-
nitude and the same direction. Fig. 1(d) shows an example of tion. A blood vessel is always darker than its neighborhood and
a contracting vector field, where the magnitudes of the vectors thus, has a concave shape. Its normalized gradient vector field
are decreasing. is expanding. This is similar to the example shown in Fig. 1(a).
LAM AND YAN: A NOVEL VESSEL SEGMENTATION ALGORITHM FOR PATHOLOGICAL RETINA IMAGES 239

Fig. 3. Pathological retina image (Im0001).

Thus, we can locate a vessel by checking whether the pixel’s di-

vergence is positive. Fig. 3 shows a retinal image and Fig. 4(a)
shows an example of blood vessels in a healthy region located
in the upper right box in Fig. 3. The corresponding vector fields
are shown in Fig. 4(b) and (c). The vectors point outwards in-
side the vessel and thus the corresponding pixels have positive
divergence.
Fig. 5(a) shows a pathological region with bright abnormality
in the center box in Fig. 3, while Fig. 5(b) shows the boundary
of this region. Fig. 5(c) is the corresponding normalized gra-
dient vector field. As the pathological region is brighter than its
neighborhood, it has a convex shape and a negative divergence.
Although positive divergence may be found in a group of noisy
pixels in the vector field outside the vessel, they can be elimi-
nated easily by the artifact removal scheme based on the con-
nectivity of the vessels. Fig. 6 shows the detection result. We
can see that no vessels are detected in the pathological region.

C. Detection of Blood Vessel-Like Objects Using Gradient

Vector Field
The goal of this step is to detect the blood vessel-like objects
using the gradient vector field. As the vector field does not have
the same magnitude of change in every direction, the sign of the
divergence does not only depend on the orientation but also on
Fig. 4. Blood vessels in a healthy region and their normalized gradient vector
the magnitude. This is similar to the situation shown in Fig. 1(b). field. (a) Blood vessels in a healthy region, the upper right box in Fig. 3. (b) Nor-
Fig. 7 shows a healthy region and its corresponding gradient malized gradient vector field inside the vessels. (c) Normalized gradient vector
vector field. We can see that the vessels in the healthy region field outside the vessels.
are expanding. By locating the pixel with positive divergence,
the detection result can be obtained. The result after this step
is shown in Fig. 8. We can see that almost all the vessel-like IV. IMPLEMENTATION DETAILS
objects are detected successfully.
In this section, we explain the implementation details for the
proposed method.
D. Pruning Operation
A. Computation of Divergence Values and Artifact Removal
In this step, a pruning operation is carried out. The spurious
detected blood vessel-like objects, as illustrated in Fig. 8, are Due to the presence of reflection on the tiny uneven surface
pruned according to the detected centerlines. The key idea is of the soft tissue, some vessels have a large divergence value
to remove a pixel from the vessel-like objects if this pixel is in some directions while the values in other directions are very
far away from the centerline. Fig. 9 shows the result after this small. These vessels can be detected more effectively by consid-
step for the image in Fig. 8. We can see that the falsely detected ering the divergence values under different smoothness and ori-
vessels near the pathological region are removed successfully. entations. The procedure is given below. For a given input image
240 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 27, NO. 2, FEBRUARY 2008

Fig. 5. Pathological region and its normalized gradient vector field. (a) Patho-
logical region, the center box in Fig. 3. (b) Boundary of the pathological region.
(c) Normalized gradient vector field of the region.
Fig. 7. Illustration of the divergence of the gradient vector for a blood vessel
image. (a) Example of blood vessels. (b) Gradient vector field.

Fig. 6. Detected centerlines of the image in Fig. 3.

Fig. 8. Detected vessel-like objects.

, we apply a Gaussian filter with fixed window size 41 and vari-
ance to the image. Then, we apply a horizontal and vertical
edge detector to obtain the vector field . A rotation matrix with
angle is applied to this vector field. After that, the divergence
is computed by the summation of the two components, and the
horizontal edge detector is applied to the -component and ver-
tical edge detector is applied to the -component of the vector
field followed by the rotation matrix. This is the computa-
tion for the divergence value of the image using the gradient
vector field and we denote it as . Here, is also
known as the oriented Laplacian. The same procedure (shown Fig. 9. Detection result after the pruning operation applied to the image in Fig.
above) is used for the normalized gradient vector field, except 8.
that normalization is applied to the vector field so that each
pixel shares the same unit magnitude. We denote this divergence
value as . edness. In removal of artifacts, each connected region with a
In order to remove artifacts in each processing step, we mea- skeleton’s length smaller than is removed. Here, is a
sure the length (in pixels) of the skeleton in each connected re- user-defined threshold constraining the length of the skeleton.
gion. Skeletonization is conducted using the morphological op- A similar technique is adopted by Vermeer et al. [26]. They use
erator [25]. The connected region is formed by four-connect- to for artifact removal. In our experiments,
LAM AND YAN: A NOVEL VESSEL SEGMENTATION ALGORITHM FOR PATHOLOGICAL RETINA IMAGES 241

Fig. 11. Low-contrast centerlines. (a) Detected low-contrast centerlines. (b)

Extra number of detected low-contrast pixels.

of all these results. Fig. 10(a)–(d) shows the results using dif-
ferent pairs of and Fig. 10(e) shows the high-contrast
centerlines. We can see that vessels are detected successfully.
For detecting the low-contrast centerlines, no rotation is ap-
plied and the parameter is set to zero. Two ranges for smooth-
ness parameter are taken. In the first range, is 3 to 4 with
interval 0.2, while in the second range, is 4 to 5 with in-
terval 0.2. We use large smoothness parameters here because
the contrast in this kind of centerlines is low and small smooth-
ness parameters may produce many artifacts. The rule, which
determines a candidate of the centerline, is given below

(2)

where represents the Gaussian filter with smoothness pa-

rameter , and is the convolution operator. is a user-de-
Fig. 10. Detected high-contrast centerlines. (a) Detected centerlines using  = fined parameter and will be defined in Section V. In each range,
45 and  = 0:2. (b) Detected centerlines using  = 145 and  = 0:2. (c) a pixel satisfying (2) passing the artifact removal scheme is a
Detected centerlines using  = 45 and  = 3. (d) Detected centerlines using candidate for a centerline. By combining the results from these
 = 145 and  = 3. (e) Combining all the detected results.
two ranges, the low-contrast centerlines (Fig. 11) are obtained.
We can see that the vessels placed near the edge of the FOV can
is taken and a perturbation to this value yields almost be detected successfully.
the same result. The centerlines of the retinal image are obtained by com-
bining the low-contrast centerlines and the high-contrast center-
B. Locating the Centerlines Using the Normalized Gradient lines. In the low-contrast centerlines, long line-shaped artifacts
Vector Field can be produced using large smoothing parameters. We use the
The normalized gradient vector field is used to de- following strategy to prune these artifacts. If a segment of any
tect the centerlines. This step consists of two different phases: low-contrast centerline is not close to any high-contrast center-
detecting the high-contrast and the low-contrast centerlines. The line, it will be pruned. Here, the low-contrast centerlines are first
high-contrast centerlines refer to the vessels in the region with separated into four-connected regions. If the region has a 40%
high foreground and background contrast, such as those placed intersection with the high-contrast centerlines, it is merged with
at the middle of the image. The low-contrast centerlines refer to the high-contrast centerlines. Fig. 11(b) shows the extra number
the vessels in the region with low contrast, such as those placed of detected pixels compared with the high-contrast centerlines
near the boundary of the field-of-view (FOV) of the image. after this step.
In each of these phases, the whole image domain is consid-
C. Detection of Blood Vessel-Like Objects Using the Gradient
ered as an input. For detecting the high-contrast centerlines,
Vector Field
the smoothness parameter from 0 to 4 with interval 0.2 and
the rotation parameter from 0 to with interval are The gradient vector field is used in this step to detect
used. For each pair , a pixel having a divergence value the blood vessel-like objects. The range for is from 0 to 4
of which passes the artifact removal scheme with interval 0.2. Two rotation parameters are 0 and 45
is a candidate for the centerline. Here, is a user-defined pa-
rameter. Its selection will be discussed in Section V. Based on (3)
these parameter settings, the algorithm produces 100 centerline
results in total in Fig. 10. By combining all the results, we ob- For each pair , a pixel satisfying (3) and passing the arti-
tain the high-contrast centerlines. The combination is the union fact removal scheme is a candidate for the vessel-like objects.
242 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 27, NO. 2, FEBRUARY 2008

We use three different measures to evaluate the performance

of the methods. They are the overall accuracy, receiver operating
characteristic (ROC) curve and the accuracy in the pathological
region. Accuracy for a retinal image with a given set of param-
eters is defined as the number of true positive vessel pixels
plus the number of true positive background pixels divided by
the total number of pixels in the FOV. The FOV is the region of
the retina while the region not in the FOV is the background of
the retinal image, which is usually dark. As the dark background
outside the FOV can be easily detected, all the measurements
are carried out inside the FOV. In this paper, the accuracy for
the pathological retina images in the STARE database is taken
as the maximum of the average accuracy for all the images in
the database with respect to the parameter . In the ROC curve,
the two axes are the true positive fraction (tpf) for the vertical
axis, and the false positive fraction (fpf) for the horizontal axis
[27], [28]. The closer the curve approaches the top left hand
corner, the better the performance of the method. A common
Fig. 12. Pruning operation. (a) Skeleton of the vessel-like object. (b) Skeleton single measure to quantify the performance of the method is to
belonging to the second group. (c) Artifacts removed. (d) Detection result. compute the area under the curve . An area close to one
means a good performance. In this paper, the tpf and fpf values
are computed by varying the parameters of the algorithm. Then,
is a user-defined parameter and will be defined in Section V. By the is approximated by adding (0, 0) and (1,1). In this
combining all the results, the vessel-like objects are obtained. database, we employ the first observer’s manual segmentation
Fig. 8 shows the result of this step. as ground truth for evaluation. The second observer’s (tpf, fpf)
are plotted on the ROC graph. In addition to the classification
D. Pruning Operation accuracy and the ROC curve, we also propose a new measure to
evaluate the performance in the pathological region (PUR) of a
After detecting the blood vessel-like objects in the image, the
detection method. This measure is similar to accuracy, but it puts
spurious objects are pruned according to the detected centerlines
the emphasis on pathological regions that are especially impor-
so that the falsely detected vessels near the pathological region
tant but difficult to deal with. The PUR is defined as the number
can be removed. The procedure is given below. They are first
of true positive pixels plus the number of true positive back-
converted into its skeleton form using a morphological operator
ground pixels that are at least pixels away from the true pos-
[25]. The vessel-like objects are then classified into two groups.
itive blood vessels pixels, divided by the total number of pixels
The first group contains pixels closer, or having equal distance,
involved. As a pathological region is near the true positive back-
to the centerlines and the second group contains pixels closer to
ground pixels and somewhat away from the true positive vessels
the skeleton. The artifact removal scheme with eight connected
pixels, a large PUR means that the method is less sensitive to the
regions is applied to the second group and the pixels passing
presence of the pathological region with respect to the param-
this scheme are pruned. These pruned pixels are merged back
eter . In this paper, varies from 0 to 25 pixels. If is equal
to the centerlines to form the detection result. Fig. 12(a) shows
to zero, all the pixels in the whole retina image are taken into ac-
the skeleton of the vessel-like objects and Fig. 12(b) the group
count for computation. If is large, only the pixels, which are
far away from the centerlines. Fig. 12(c) shows the artifacts re-
far away from the true vessels, are considered in the true positive
moved, and Fig. 12(d) the detection result.
background pixels. It is noticed that the true positive fraction is
independent of the choice of the parameter . For convenience,
V. EXPERIMENT EVALUATION we use PUR( , ) to denote the PUR measure for different
values and parameter setting .
A. Databases
We use all the pathological retina images in the STARE data- B. Parameter Settings
base to evaluate the proposed method and compare its perfor-
The input image is obtained from the green channel of the
mances with other methods. The performances on the DRIVE
image, which contains the richest information on the blood ves-
database are given in the supplementary materials as reference.
sels [3], [8]. For the proposed method, tpf and fpf are set to
The reasons for using the STARE database for evaluation is that
the following values: , and
it is available publicly and the reader can download the database
for . Under this setting,
on the website. Also, several difficult pathological retina images
the proposed method requires about 25 min to produce the de-
can be found in this database.
tection results for a single retinal image. We use MATLAB
1http://www.parl.clemson.edu/stare/probing/ to implement our method on a computer with Intel Pentium
2http://www.isi.uu.nl/Research/Databases/DRIVE/ 2.66 GHz and 512 MB RAM. The algorithms can be optimized
LAM AND YAN: A NOVEL VESSEL SEGMENTATION ALGORITHM FOR PATHOLOGICAL RETINA IMAGES 243

TABLE I
PARAMETERS FOR JIANG AND MOJON’S METHOD

TABLE II
PARAMETERS FOR VERMEER et al.’S METHOD

TABLE III
RESULTS FOR THE STARE DATABASE USING DIFFERENT METHODS

and speeded up using other programming languages, such as

C. We compare the performance of the proposed method with
two unsupervised learning methods and two different results
of a supervised learning method. We implement the two un-
supervised methods, which are the adaptive thresholding tech-
nique by Jiang and Mojon [2] and the model-based technique
by Vermeer et al. [26]. The parameters for these two methods
are shown in the Tables I and II. The values of tpf and fpf are Fig. 13. ROC curves for the pathological retina images in the STARE database.
obtained by experiments so that the optimal is reached. Comparing the second observer to the first observer, true and false positive frac-
Also, one step of Vermeer et al. method conducts pruning with tions of (0.8215, 0.0307) are found. (a) ROC curves of the proposed method and
the two unsupervised methods. (b) ROC curves of the proposed method and the
the aid of human visual assessment [26, Sec. III-B ]. As this step two results by the supervised method.
is very subjective, we have only carried out a major clean up of
the detection results along the boundary of the FOV so that max-
imum accuracy can be reached. This procedure is not applied to C. Results
the proposed method or the method of Jiang and Mojon. For the
The evaluation measures using the overall accuracy and
supervised method, we use the method proposed by Soares et
values are given in Table III. In this table, we can see that the
al. Their results are available on the website. We use the two
proposed method has a better performance than both supervised
results from the classifier with 20 Gaussian models trained by
and unsupervised methods in accuracy.
the STARE and the DRIVE databases, which produce the best
Fig. 13(a) shows the ROC curves of the proposed method and
results in their paper. The reason for comparing the results ob-
the two unsupervised methods. We can see that the curve for the
tained from training the DRIVE database to detect vessels in
proposed method is completely above the two other curves and
the STARE database is that the ground truth labels for a data-
yields the largest value. Fig. 13(b) shows the ROC curves
base may not be available in real applications. We can only use
of the proposed method and the results using the supervised
existing databases to train the classifier for blood vessel segmen-
learning method. In this figure, we can see that the ROC curve
tation.
from the proposed method is not completely above the curve
3http://www.retina.iv.fapesp.br from the two results of Soares et al. except in the initial part of
244 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 27, NO. 2, FEBRUARY 2008

Fig. 14. Performance evaluation measure PUR for different methods using all
the pathological images.

the curve. The main reason for the proposed method having a
smaller and part of its initial ROC curve below the super-
vised method is given as follows. As discussed in the introduc-
tory section, the supervised method learns from human-labeled
data and it is able to yield accurate results near the edge of the
ground truth region. Thus, in the ROC curves, an increase in
the true positive fraction does not lead to a large increase in the
false positive fraction and the supervised method is able to yield
a larger value. In contrast to the supervised method, the
proposed method is more robust to the pathological regions and
can avoid detecting false vessels in these regions. It is able to
yield a solution very similar to the ground truth. Thus, it pro-
duces the highest accuracy among others, including the super-
vised method, and yields a ROC curve closest to the pair (fpf,tpf)
from the second observer for the pathological images.
We also use the third evaluation measure to verify the robust-
ness of the methods for pathological regions. Fig. 14 shows the
PUR of all the pathological images using different methods. The
axis represents the number of pixels used in computing
PUR and the -axis refers to the PUR. The way to select the
parameter set is as follows. For a given pixel distance pa-
rameter , the optimal value of , denoted as , is found by
maximizing the average PUR( , ) for all the images. The en-
tire range of for the two unsupervised methods and the su-
pervised method are used as the search space. For the proposed
method, a smaller range, to 5 is used. The reason for Fig. 15. Results for a pathological retina image with a dark abnormality
the restriction is to show the robustness of the proposed method obtained using different methods. (a) Retinal image with dark abnormality
(Im0139). (b) Hoover et. al. (tpf ; fpf ) = (0:6660; 0:0499). (c) Jiang and
compared to other methods under different parameter settings. Mojon (tpf ; fpf) = (0:9009; 0:1245). (d) Vermeer et al.; (tpf ; fpf) =
Fig. 14 shows that the proposed method has the highest PUR (0:8626; 0:0949). (e) Soares et al. (DRIVE) (tpf ; fpf) = (0:8966; 0:0868).
from to . When is near zero, PUR( , ), (f) Soares et al. (STARE) (tpf ; fpf) = (0:8981; 0:0640). (g) Proposed
method. (tpf ; fpf ) = (0:8935; 0:0839). (h) Ground truth 1. (i) Ground truth
using different methods, shares almost the same value. How- 2.
ever, when increases, the proposed method yields a higher
accuracy than others and their difference becomes larger. As a
pathological region is near the true positive background pixels
VI. DISCUSSION AND CONCLUSION
and somewhat away from the true positive blood vessel pixels,
the PUR value for changing shows the accuracy of a method Detecting blood vessels in a pathological retina image is a
towards the pathological region. This means that the proposed challenging problem. In our proposed method, we solve this
method is more robust than other methods in pathological re- problem by detecting the blood vessel-like objects in the image
gions. using the Laplacian operator and the noisy objects are pruned
LAM AND YAN: A NOVEL VESSEL SEGMENTATION ALGORITHM FOR PATHOLOGICAL RETINA IMAGES 245

B. Future Work
Although the proposed method shows a good performance for
pathological images, there is still room for further improvement.
In Section VI-A, it is shown that the proposed method shows
a better performance than other methods for the pathological
retina image having dark abnormality. However, several spots
are falsely detected as vessels. In the future, we may be able to
use other features to resolve this problem. One possible solution
is to prune the vessel-like objects having spherical shape, as the
false detected vessels appear in this form.
Another improvement that can be made with the proposed
method is the performance near blood vessel edges. Although
the supervised method has limitations for pathological regions,
it produces good results near blood vessel edges. In future, we
can combine the proposed method and the supervised method
Fig. 16. Performance evaluation in the pathological region of the retinal image
(Im0139).
to yield overall better results.

ACKNOWLEDGMENT
according to the detected centerlines, which are extracted using
the normalized gradient vector field. Experiment results show The authors would like to thank J. J. Staal and A. Hoover
that our method is able to yield an accurate result especially for for making their databases publicly available. Also, the sug-
pathological retina images. gestions and comments of anonymous reviewers, which have
greatly helped to improve the quality of this paper, are acknowl-
edged.
A. Limitation of the Proposed Method
In this paper, we focus our discussion on bright abnormality. REFERENCES
Dark abnormality is also present in a pathological retina image.
Fig. 15(a) is an example. Fig. 15(b)–(g) shows the results using [1] S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and M. Goldbaum,
“Detection of blood vessels in retinal images using two-dimensional
different methods. Fig. 15(h) and (i) shows the ground truths matched filters,” IEEE Trans. Med. Imag., no. 3, pp. 263–269, Sep.
obtained by two different human observers. There is a remark- 1989.
able difference between the two ground truths. The reason is [2] X. Jiang and D. Mojon, “Adaptive local thresholding by verification-
based multithreshold probing with application to vessel detection in
that the pixel values are small compared with other retinal im- retinal images,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 25, no. 1,
ages. For human vision, a different conclusion can be made as pp. 131–137, Jan. 2003.
to whether a pixel belongs to a vessel or a nonvessel, in a dark [3] J. J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, and B.
van Ginneken, “Ridge based vessel segmentation in color images of
region with a tiny uneven surface. Fig. 15(c) shows the result the retina,” IEEE Trans. Med. Imag., vol. 23, no. 4, pp. 501–509, Apr.
using the method of Jiang and Mojon and several dots are ap- 2004.
parent. The reason for the spots is given below. The detection [4] J. V. B. Soares, J. J. G. Leandro, R. M. Cesar, Jr., H. F. Jelinek, and M.
J. Cree, “Retinal vessel segmentation using the 2-D gabor wavelet and
is conducted under different intensity levels. If an abnormal re- supervised classification,” IEEE Trans. Med. Imag., vol. 25, no. 9, pp.
gion of large size exists in one of these levels, then this region 1214–1222, Sep. 2006.
cannot be pruned by the algorithm and thus becomes part of the [5] T. Chanwimaluang and F. Guoliang, “An efficient blood vessel detec-
tion algorithm for retinal images using local entropy thresholding,” in
detection results. Fig. 15(e) and (f) show the results obtained Proc. 2003 Int. Symp. Circuits Syst., 2003, vol. 5, pp. 21–24.
using the method of Soares et al. There are more spots detected [6] A. Hoover, V. Kouznetsova, and M. Goldbaum, “Locating blood ves-
near the vessels using Soares et al. (DRIVE). This is because sels in retinal images by piece-wise threshold probing of a matched
filter response,” IEEE Trans. Med. Imag., vol. 19, no. 3, pp. 203–210,
a different training image set is used for the results of Soares Mar. 2000.
et al. (DRIVE), which can degrade the performance. The result [7] A. Pinz, S. Bernogger, P. Datlinger, and A. Kruger, “Mapping the
obtained using the proposed method is shown in Fig. 15(g). We human retina,” IEEE Trans. Med. Imag., vol. 17, no. 4, pp. 606–619,
Apr. 1998.
can see that it is able to detect many short vessels without pro- [8] Y. A. Tolias and S. M. Panas, “A fuzzy vessel tracking algorithm for
ducing many spots near the pathological region. However, if a retinal images based on fuzzy clustering,” IEEE Trans. Med. Imag., vol.
dark abnormality is very close to a vessel, an artifact can still be 17, no. 2, pp. 263–273, Apr. 1998.
[9] A. K. Klein, L. Forester, and A. A. Amini, “Quantitative coronary an-
formed. giography with deformable spline models,” IEEE Trans. Med. Imag.,
Now, we compare the performance of the proposed method vol. 16, no. 5, pp. 468–482, Oct. 1997.
and the method of Soares et al. near the pathological region [10] S. R. Fleagle, M. R. Johnson, C. J. Wilbricht, D. J. Skorton, R. F.
Wilson, C. W. White, M. L. Marcus, and S. M. Collins, “Automated
using the PUR measure. This is shown in Fig. 16. We can see analysis of coronary arterial morphology in cineangiograms: Geo-
that the proposed method generally has a better performance metric and physiological validation in humans,” IEEE Trans. Med.
than Soares et al. This implies that the performance of the pro- Imag., vol. 8, no. 4, pp. 387–400, Dec. 1989.
[11] S. Tamura, Y. Okamoto, and K. Yanashima, “Zero-crossing interval
posed method is still better than the supervised method in the correction in tracing eye-fundus blood vessels,” Pattern Recognit., vol.
pathological region with a dark abnormality. 21, pp. 227–233, 1988.
246 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 27, NO. 2, FEBRUARY 2008

[12] M. Figueiredo and J. Leitao, “A nonsmoothing approach to the esti- [20] J. A. Sethian, Level Set Methods: Evolving Interfaces in Geometry,
mation of of vessel controus in angiograms,” IEEE Trans. Med. Imag., Fluid Mechanics, Computer Vision and Materials Sciences. Cam-
vol. 14, pp. 162–172, 1995. bridge, U.K.: Cambridge Univ. Press, 1996.
[13] A. Hoover and M. Goldbaum, “Locating the optic nerve in a retinal [21] S. Osher and R. Fedkiw, Level Sets Methods and Dynamic Implicit
image using the fuzzy convergence of the blood vessels,” IEEE Trans. Surfaces. New York: Springer, 2003.
Med. Imag., vol. 22, no. 8, pp. 951–958, Aug. 2003. [22] G. K. Batchelor, An Introduction to Fluid Dynamics, 1st ed. Cam-
[14] M. Sofka and C. V. Stewart, “Retinal vessel centerline extraction bridge, U.K.: Cambridge Univ. Press, 2000.
using multiscale matched filters, confidence and edge measures,” [23] D. J. Griffiths, Introduction to Electrodynamics, 3rd ed. Upper Sad-
IEEE Trans. Med. Imag., vol. 25, no. 12, pp. 1531–1546, Dec. 2006. dler River, NJ: Prentice Hall, 1999.
[15] M. Sofka and C. V. Stewart, “Erratum to “retinal vessel centerline ex- [24] W. Kaplan, Advanced Calculus, 4th ed. Reading, MA: Ad-
traction using multiscale matched filters, confidence and edge mea- dison-Wesley, 1991.
sures”,” IEEE Trans. Med. Imag., vol. 26, no. 1, pp. 133–133, Jan. 2007. [25] K. R. Castleman, Digital Image Processing. Englewood Cliffs, N.J.:
[16] B. S. Y. Lam and H. Yan, “Blood vessel extraction based on Mum- Prentice-Hall, 1979.
ford-Shah model and skeletonization,” in Int. Conf. Mach. Learn. Cy- [26] K. A. Vermeer, F. M. Vos, H. G. Lemij, and A. M. Vossepoel, “A
bernetics, 2006, pp. 4227–4230. model based method for retinal blood vessel detection,” Computers
[17] S. Esedoglu and J. Shen, “Digital inpainting based on the Mumford- Biol. Med., vol. 34, pp. 209–219, 2004.
Shah-Euler image model,” Eur. J. Appl. Math, vol. 13, pp. 353–370, [27] P. Armitage and T. Colton, “Receiver operating characteristic (ROC)
2002. curves,” in Encyclopedia of Biostatistics. New York: Wiley, 1998, pp.
[18] T. Chan and J. Shen, Image processing and analysis: variational, 3738–3745.
PDE, wavelet, and stochastic methods. Philadelphia, PA: Soc. [28] N. A. Obuchowski, “Receiver operating characteristic curves and their
Indust. Appl. Math., 2005. use in radiology,” Radiology, vol. 229, pp. 3–8, 2003.
[19] T. Chan, S. Osher, and J. Shen, “The digital TV filter and nonlinear
denoising,” IEEE Trans. Image Process., vol. 10, no. 2, pp. 231–241,
Feb. 2001.