IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 8, NO.

3, SEPTEMBER 1989 263

Detection of Blood Vessels in Retinal Images Using

Two-Dimensional Matched Filters
SUBHASIS CHAUDHURI, STUDENT MEMBER, IEEE, SHANKAR CHATTERJEE, MEMBER, IEEE,
NORMAN KATZ. MARK NELSON, AND MICHAEL GOLDBAUM

Abstract-Although current literature abounds in a variety of edge of noise in the image degrades the performance of gra-
detection algorithms, they do not always lead to acceptable results in dient based algorithms appreciably. The use of gray-scale
extracting various features in an image. In this paper, we address the
problem of detecting blood vessels in retinal images. Blood vessels nsu- morphology has been suggested in 131. It involves simple
ally have poor local contrast and the application of existing edge de- erosion and dilation operators and can be implemented for
tection algorithms yield results which are not satisfactory. We intro- real time processing by bsing suitable hardware. This al-
duce an operator for feature extraction based on the optical and spatial gorithm appears to work better in extracting edges from
properties of objects to be recognized. The gray-level profile of the
images with salt-and-pepper type of noise.
cross section of a blood vessel is approximated by a Gaussian shaped
curve. The concept of matched filter detection of signals is used to de- The Sobel operator involves the computation of local
tect piecewise linear segments of blood vessels in these images. We con- intensity gradients and the responses due to nonideal step
struct 12 different templates that are used to search for vessel segments edges are not good. A modification of that is the detection
along all possible directions. We discuss various issues related to the of second-order zero-crossings, and the corresponding
implementation of these matched filters. The results are compared to edge operator is called the Laplacian [ 2 ] . In order to re-
those obtained with other methods. The automatic detection of blood
vessels in the retina could help physicians in diagnosing ocular dis- duce the effect of high frequency noise, a bandwidth re-
eases. striction can be imposed by convolving the original data
by a two-dimensional Gaussian window before applying
the Laplacian operator [4]-[7]. It has been shown in [4]
I. INTRODUCTION that the above scheme is a limiting process of the differ-

E DGE detection plays an important role in a number of

image processing applications such as scene analysis
and object recognition. The edges in an,image provide
ence of Gaussian (DOG) filter. The advantage of using a
Gaussian filter is that it regularizes the ill-posed problem
of differentiation [SI.
useful structural information about object boundaries, as Algorithms belonging to the second category, the edge
the edges are caused by changes in some physical prop- fitting method, minimize the distance between the origi-
erties of surfaces being photographed, such as illumina- nal noisy data and a predefined edge model in a finite-
tion, geometry, and reflectance. An edge is, therefore, dimensional space. Thus, the results of these algorithms
defined as a local change or discontinuity in image lumi- are optimal for the chosen edge model and are less sen-
nance. sitive to noise. However, they usually involve more com-
There are two basic approaches to edge detection: the putations. One such algorithm that recognizes edges and
enhancement/thresholding method and the edge fitting lines (edges of finite thickness) was introduced by Hueckel
method [ 13. In the former method, the discontinuities in [9], [lo]. Here both the edge template and the picture
the image gray tone are enhanced by neighborhood oper- function (within the specified window) are parameterized
ators. The simplest type of differential operator, the Rob- and are represented by corresponding vectors of coeffi-
ert’s operator, was introduced more than two decades ago. cients. The Euclidean distance between these two vectors
This operator consists of 2 X 2 convolutional kernels. is minimized. Although this algorithm performs very well,
Prewitt and Sobel operators [2] are improvements on Rob- a significant savings in computation can be achieved by
ert’s operator, where the masks have sizes 3 x 3 pixels. using the scheme proposed in [ 111. Hueckel used the Fou-
A quantitative comparison of the performances of these rier series in polar coordinates as basis functions to obtain
operators have been reported in [ 11. All of these operators the orthogonal components. These basis functions are re-
yield good results when the edges are sharp. The presence placed by a seemingly reasonable choice of Walsh func-
tions which do not require any floating point multiplica-
Manuscript received September 12, 1988; revised March 31, 1989. This tions or divisions. Attempts have also been made to detect
work was supported by the United States Public Health Service under Re-
search Grant EY 05996-02. object boundaries by projecting the data onto a complete
S . Chaudhuri and S . Chatterjee are with the Department of Electrical set of basis functions in a Hilbert space, which are mu-
Engineering, University of California, San Diego, La Jolla, CA 92093. tually orthogonal [12]. An optimal frequency domain fil-
N. Katz, M. Nelson, and M . Goldbaum are with the Department of
Opthalmology, University of California, San Diego, La Jolla, CA 92093. ter for edge detection has been suggested in [13], where
IEEE Log Number 8928436. the filter transfer function is specified in terms of prolate

0278-0062/89/0900-0263$01.OO 0 1989 IEEE

264 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL 8, NO 3, SEPTEMBER 1989

spheroidal wave functions. This filter is optimal in the produces dilated tortuous veins, arteriosclerosis can cause
sense that it produces maximum energy for a specified the arteries to acquire a copper or silver color, and dia-
resolution of edge localization in the vicinity of the edge. betes can generate new blood vessels (neovasculariza-
A stochastic model of the edge structure in the image has tion). Thus, a reliable method of vessel detection is
been proposed in [14], where a two-dimensional digital needed, which preserves various vessel measurements.
recursive filter has been used to minimize the least mean An edge detection algorithm, similar to the morpho-
square (LMS) estimation errors. Characterization of in- logical edge detector, has been used for automatic rec-
dividual edge elements in an image by using generalized ognition of arterio-venous intersections [ 191. A line find-
splines and discrete cosine transform (DCT), in relation ing algorithm along with a probabilistic relaxation scheme
to its neighborhood and in terms of the topographic prop- has been used for the extraction and subsequent descrip-
erties of the image intensity surface, has been reported in tion of blood vessel patterns in retinal images [20], [2 11.
~51. However, these algorithms do not result in a continuous
An important motivation for using algorithms based on map of the vasculature of the retinal surface. An individ-
template matching is the ease of implementation on spe- ual blood vessel may be broken into several segments due
cialized high speed hardware as well as the possibility for to branching of vessels or arteriovenous intersections.
parallel computations. In all of these approaches, com- Most of the algorithms mentioned above and in the pre-
putations could be performed in parallel on individual vious section may be considered as generalized edge de-
pixels. Such kernel-based methods for edge detection tectors, and do not take account of the special properties
have, thus, become increasingly popular. As the size of of the objects being detected. Blood vessels have a special
the convolution mask is increased, a better operator (in property that the two edges of a vessel always run parallel
the sense that it is less sensitive to noise) can be designed,to each other. Such objects may be represented by piece-
but the computation increases significantly. Although wise linear directed segments of finite width. The gradient
these methods may reduce the spatial resolution signifi- directions for these two edge elements are 180” apart and
cantly [ 111, they may be used very efficiently not just to hence, they are sometimes referred to as “anti-parallels”
detect individual edge elements but to extract certain spe- [22]. Thus, one method of extraction of such objects in
cial features in the image. Our goal is to design an oper- an image consists of identifying the anti-parallels and
ator which is nearly optimal for the recognition of certain tracking them along their gradient directions [22]. This
special objects in the image, particularly for the detection algorithm has been used for locating roads, rivers, etc.,
of blood vessels in retinal fundus images. in geological images. In our experience, the contrast be-
The organization of this paper is as follows. The im- tween vessels and the background was too low for these
portance of edge detection in image processing applica- methods to reliably define blood vessels without fragmen-
tions and the currently available different types of edge tation [23].
detectors have been reviewed in this section. Section I1 Methods of extraction or enhancement of directed line
describes the need for the detection of blood vessels in segments in an image have been proposed in [24], [25].
retinal images and a brief survey of the existing literature. These methods employ sector filters in the frequency do-
Some of the special properties of blood vessels in fundus main and fail to yield good results in the presence of blob
images are reported in Section 111. The concept of matched type objects in the image [26]. It has recently been brought
filter detection of 4gnals is briefly reviewed here, and then to our attention that a method of enhancement of finger-
extended to the detection problem for 2-D signals. Sec- prints has been proposed in [27], [28]. The intensity pro-
tion IV describes the implementation of the proposed al- files for individual lines are modeled by cosine shaped
gorithm. Some of the results of application of this scheme curves, the dominant angle of orientation in a local neigh-
to fundus images are given in Section V. The performance borhood is estimated, and then a proper template is used
of this algorithm is compared to that of some other com- to enhance the line segments. However, the process of
monly known algorithms. Finally, conclusions and rec- estimating the local orientation consumes about 85 per-
ommendations are presented in Section VI. cent of the computation time. The method that we propose
here is quite similar to the above approach, but it does not
11. DETECTION OF BLOODVESSELS require this pre-processing step. The proposed method is
Information about blood vessels can be used in grading discussed in the next section.
disease severity or as part of the process of automated
diagnosis of diseases with ocular manifestations [ 161- 111. DESIGNOF MATCHED FILTER
[ 181. Blood vessels can act as landmarks for localizing the We observe three interesting properties of the blood
optic nerve, the fovea (central vision area), and lesions. vessels in retinal images.
As a result of systemic or local ocular disease, the blood 1) Since the blood vessels usually have small curva-
vessels can have measurable abnormalities in diameter, tures, the anti-parallel pairs may be approximated by
color, and tortuosity. For example, central retinal artery piecewise linear segments.
occlusion usually causes generalized constriction of reti- 2) Since the vessels have lower reflectance compared
nal arteries, hypertension may result in focal constriction to other retinal surfaces, they appear darker relative to the
of retinal arteries, central retinal vein occlusion typically background. A few representative samples of blood vessel
CHAUDHURI et u l . : DETECTION OF BLOOD VESSELS I N RETINAL IMAGES 265

gray-level profiles along directions perpendicular to their 100 1

length are plotted in Fig. 1. It was observed that these

vessels almost never have ideal step edges. Although the
80
intensity profile varies by a small amount from vessel to
h
vessel, it may be approximated by a Gaussian curvef(x,
y) = A { 1 - k exp ( - d 2 / 2 a 2 ) ) , where d is the perpen-
dicular distance between the point ( x , y ) and the straight
line passing through the center of the blood vessel in a
direction along its length, U defines the spread of the in-
tensity profile, A is the gray-level intensity of the local * vesscll
background, and k is a measure of reflectance of the blood -D vessel 2

vessel relative to its neighborhood.

3 ) Although the width of a vessel decreases as it travels
20
-
*
+
vessel3
vessel4
vessel5

radially outward from the optic disk, such a change in 0 1 . I . I ' I ' I ' I ' I ' I ' I ' I ' I

vessel caliber is a gradual one. The widths of the vessels -10 -8 -6 -4 -2 0 2 4 6 8 10

are found to lie within a range of 2-10 pixels (36-180 Distance in pixels from center of blood vessel
pm). For our initial calculations, however, we shall as- Fig. I . The gray-level (brightness) profiles of the cross section of several
blood vessels in the image of Fig. 7(a). The shapes of the curves are
sume that all the blood vessels in the image are of equal similar although the vessels have different widths and intensities. The
width 20. trough i n each curve represents the center of the corresponding blood
Now let us consider the detection of an arbitrary 1-D vessel.
signal s ( t ) in an additive Gaussian white noise. If the
signal is passed through a filter with transfer function
H ( f),the output signal s o ( t )is given by sponses are to be compared, and for each pixel only the
maximum response is to be retained.
Consider the response of this filter for a pixel belonging
to the background retina. Assuming the background to
have constant intensity with zero mean additive Gaussian
where S( f ) is the Fourier transform of s ( t ) , and r ( f ) white noise, the expected value of the filter output should
is the noise spectrum. Using Schwartz's inequality, it can ideally be zero. The convolution kernel is, therefore,
be proved [29], [30] that the filter H ( f ) that maximizes modified by subtracting the mean value of s ( t ) from the
the output signal-to-noise ratio is given by Hop,( f ) = function itself.
S * ( f ). Since the input signal s ( t ) is real valued, hopt (t) It has already been mentioned that the vessels may be
= s ( - t ) . This optimal filter with the impulse response considered as piecewise linear segments. Instead of
h ( t )is commonly known as the matched filter for the sig- matching a single intensity profile of the cross section of
nal s ( t ) . In a typical communication system, if there are a vessel, a significant improvement can be achieved by
n different signals s , ( t ) , i = 1, 2 , * . * , n , the received matching a number of cross sections (of identical profile)
signal is passed through a stack of n matched filters. If along its length simultaneously. Such a kernel may be
the response due to the j t h filter is the maximum, it is mathematically expressed as
concluded that the signal s, ( t ) was transmitted.
~ ( xy ,) = -exp ( - x 2 / 2 a 2 ) for lyl I~ / 2( 2 )
Under the present context, it may be noted that the in-
tensity profile can be assumed to be symmetrical about the where L is the length of the segment for which the vessel
straight line passing through the center of the vessel. is assumed to have a fixed orientation. Here the direction
Hence, s ( - t ) = s ( t ) . The optimal filter must have the of the vessel is assumed to be aligned along the y-axis.
same shape as the intensity profile itself. In other words, For the vessels at different orientations, the kernel has to
the optimal filter is given by h,,,(d) = -exp ( - d 2 / 2 a 2 ) . be rotated accordingly. L was determined experimentally
The negative sign indicates that the vessels are darker than by analyzing vessels in both normal and abnormal retinas.
the background. Also, note further that instead of n dif- The above procedure reduces the possibility of false de-
ferent types of objects having to be identified, the prob- tection of blood vessels in a nonideal environment. Also,
lem reduces to deciding whether or not a particular pixel it suppresses the response due to the noise significantly in
belongs to a blood vessel. If the magnitude of the filtered the background retina where no blood vessel is present.
output at a given pixel location exceeds a certain thresh- Such a scheme may be compared to Canny's edge detector
old, the pixel is labeled as a part of a vessel. [7], where the maximum of directional derivatives at every
When the concept of matched filter is extended to two- pixel is calculated and is averaged along its perpendicular
dimensional images, it must be appreciated that a vessel direction.
may be oriented at any angle 0 (0 I0 5 a).The matched
filter s ( t ) will have its peak response only when it is IV. IMPLEMENTATION OF THE ALGORITHM

aligned at an angle 0 f a / 2 . Thus, the filter needs to be The two-dimensional matched filter kernel in a discrete
rotated for all possible angles, the corresponding re- grid is designed as follows. Let = [x y ] be a discrete
266 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 8, NO 3, SEPTEMBER 1989

point in the kernel and 8, be the orientation of the ith ker- 0 0 0 0 0 0 4 0 0 0 0 0 0 0 0 0

0 0 0 0 0 4 4 3 0 0 0 0 0 0 0 0
nel matched to a vessel at an angle 8,. In order to compute 0 0 0 0 4 4 3 2 0 0 0 0 0 0 0 0
0 0 0 4 4 3 2 0 - 2 - 4 0 0 0 0 0 0
the weighting coefficients for the kernel, it is assumed to 0 0 4 4 3 2 0 - 2 - 4 - 5 - 6 0 0 0 0 0
be centered about the origin [ 0 01. The rotation matrix is 0 4 4 3 2 0 - 2 - 4 - 5 - 6 - 5 - 4 0 0 0 0
4 4 3 2 0 - 2 - 4 - 5 - 6 - 5 - 4 - 2 0 0 0 0
given by 0 3 2 0 - 2 - 4 - 5 - 6 - 5 - 4 - 2 0 2 3 0 0
0 0 0 - 2 - 4 - 5 - 6 - 5 - 4 - 2 0 2 3 4 4 0
-
7, = [ cos 8, -sin 8,
sin 0, cos 0, 1
0
0
0
0
0

0
0
0 - 4 - 5 - 6
0 0 0 - 6 -
0
0
0
0
0 -4
0 0
5
- 5 -

-2
0
-
4
4
0
2
-
-
2
3
2
2
0

3
4
0
2

4
4
2 3
3

4
0
4

0
0
4
4 4 0
0
0
0

0
0
0
0
0
0
0
0 0 0 0 0 0 0 3 4 4 0 0 0 0 0 0
and the corresponding
- point in therotated coordinate sys- 0 0 0 0 0 0 0 0 4 0 0 0 0 0 0 0

tem is given by p , = [ U U ] = p.7. Assuming an angular (a)

resolution of 15", we need 12 different kernels to span all
possible orientations. A set of 12 such kernels is applied 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
to a fundus image and at each pixel only the maximum of 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 4 3
2 1 - 2 - 5 - 6 - 5 - 2 1 2 3 4 0 0
their responses is retained. 0 4 3 1 2 - 2 - 5 - 6 - 5 - 2 1 2 3 4 0 0
A Gaussian curve has infinitely long double sided trails. 0 4 3 1 2 - 2 - 5 - 6 - 5 - 2 1 2 3 4 0 0
0 4 3 2 1 - 2 - 5 - 6 - 5 - 2 1 2 3 4 0 0
We truncate the trail at U = +3a. A neighborhood N is 0 4 3 2 1 - 2 - 5 - 6 - 5 - 2 1 2 3 4 0 0
defined such that N = { ( U , U ) I 1 U I 5 3a, I U 1 I L / 2 } . 0
0
4 3 2 1
4 3 2
-
1
2 - 5
- 2 - 5
- 6 -
- 6 - 5
5
- 2
- 2 1
1 2 3
2 3
4
3 0 0
0 0
The corresponding weights in the ith kernel are given by 0 4 3 2 1 - 2 - 5 - 6 - 5 - 2 1 2 3 4 0 0
0 4 3 2 1 - 2 - 5 - 6 - 5 - 2 1 2 3 4 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
K,(x,y ) = -exp ( - u 2 / 2 a 2 ) VP, E N . (3) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
If A denotes the number of points in N , the mean value of
(b)
the kernel is determined as
Fig 2 Two of the 12 different kernels that have been used to detect vessel
m, = K,(x,
y)/A. segments along different directions (a) segments along the 45" direc-
FEN tion, and (b) segments along the vertical direction. We have used U =
2.0 and L = 9 to generate diflerent kernels.
Thus, the convolutional mask used in this algorithm is
given by K:(x,y ) = K,(x,y ) - m, V p , E N .
Because of the hardware design of the image processing
system, the weighting coefficients in the kernel need to be
integers in the range (-128, 127). The coefficients are
each multiplied by a scale factor of 10 and truncated to
their nearest integer. A higher scale factor would increase
the accuracy, but might lead to buffer overflow. Due to
truncation errors, the mean value of the kernel may not
be exactly equal to zero. Kernels, for which the mean
value is positive, are forced to have slightly negative mean
values in order to reduce the effect of background noise
where no blood vessel is present. Two such kernels for
two different angles are given in Fig. 2 .
Our system consists of a special image processing board
from Matrox, Inc., known as the MVP-AT/NP, which is
Fig 3 A typical retinal image obtained from a fundus camera The mon-
capable of computing the convolution of sizes up to 32 x ochrome image represents the information contained in the green color
32 for an image of size 512 x 480 in a few hundred mil- plane only since this plane provides the highest vessel contrast
liseconds. The software runs on a Compaq 386 and the
execution time for this algorithm is about 50 s, including
the rotation of the kernel for 12 different directions. The Figs. 5 and 6, respectively. Comparing these three fil-
images are digitized from 35 mm transparencies taken tered images, we may conclude that the given algorithm
from a TOPCON TRV-50 fundus camera at 35" field of achieves significant improvement compared to other al-
view. gorithms in detecting blood vessels. For the purpose of
comparison, the original gray-scale output images were
V . RESULTSAND DISCUSSIONS used.
An original fundus image is given in Fig. 3, and the Fig. 7(a) represents another fundus image, and Fig.
result of the application of this filter is given in Fig. 4. 7(b), (c), and (d) are the results of applying the Sobel
For this case, we choose a = 2 and L = 9. The original operator, the morphological edge detector, and the pro-
image was smoothed by a 5 X 5 mean filter to reduce the posed algorithm, respectively. It may be noticed here that,
effect of spurious noise. The Sobel operator and the mor- unlike Sobel or morphological operators where only the
phological edge detector [3] were also applied to the same edges are detected, the proposed algorithm extracts the
image, and the corresponding filtered images are given in blood vessel as a whole. The "anti-parallels" are de-
C H A U D H U R I rf U/.: DETECTION OF BLOOD VESSELS I N RETINAL IMAGES 267

is readily seen that the proposed algorithm preserves the

continuity of the vessels in the image. Also, we have
found the algorithm to perform effectively in detecting
blood vessels even when the local contrast is quite low.
Careful observation of results in Figs. 4 and 7(d) would
reveal that we, indeed, get some response due to edges of
other bright objects (such as lesions, optic disk, etc.)
which do not resemble any vessel segment. This is due to
the fact that the local contrast is very high and the edges
of the objects partially match the shape of the kernel. Such
edges also appear when other methods are used. A post-
processing step is needed to identify and subsequently
eliminate such false detection.
Some vessels may have a specular bright reflex running
Fig. 4. The result of application of the matched filter to the image given through their center [31]. This could be due to certain
in Fig. 3. The edges at thc image border due to the apcrture ofthe fundus
camera have been removed by fitting an ellipse around the image. The pathological conditions of the patient or due to the nature
output imagcs shown in this paper represent the gray-scale response of of illumination. However, we found such a “bump” at
the filters. the center of the intensity profile to be quite small com-
pared to the brightness of its neighborhood. The effect of
such a reflex on the results of the matched filter were,
therefore, negligible.
It was mentioned earlier that the parameter L was cho-
sen to be equal to 9 pixels which correspond to about 160
pm. Ideally, one would like to have a larger value of L
which would further reduce the noise. However, the re-
sulting template might not match well if the vessels were
quite tortuous in nature (as in central retinal vein occlu-
sions). L = 9 at the working level of magnification was
experimentally found to be a good choice as it performs
well even for detecting tortuous vessel segments.
Twelve different templates were used in the proposed
scheme to search for directional components along all
possible orientations, since vessel segments lying within
Fig. 5. Edge detection is pcrformcd on the image given in Fig. 3 u h i n g the k 7 . 5 ” of the direction of the chosen kernel were found
Sobel operator. to respond well. The performance of the proposed scheme
with 24 templates, using integer valued kernels, was vir-
tually indistinguishable from that of 12 templates. Higher
precision can be achieved by performing convolutions in
floating point hardware or software, with a sacrifice in
speed. A fewer number of templates may be used to ob-
tain faster results, however, at least 4-6 such templates
are needed to have a reasonably good enhancement of
vessel segments.
Applications of the above designed filters enhance in-
dividual vessel segments in the image. A proper thresh-
olding scheme can be used to distinguish between en-
hanced vessel segments and the background. An automatic
thresholding algorithm that maximizes the inter-class in-
tensity variance [32] was used to binarize the enhanced
images.
Fig. 6. The result of application of the morphological edge detector to the
Convolution with a Gaussian window does not displace
image in Fig. 3. the second-order zero-crossings of the original intensity
data. Thus, the detected blood vessels have good edge
localization. This convolution process also eliminates the
tected from the edge detected image in Fig. 7(b) and are noise outside the desired frequency band [4].
tracked along the gradient direction (refer to [22]) to ex- We initially assumed that the blood vessels are of uni-
tract the blood vessel segments. The result is given in Fig. form caliber. One way of accommodating the varying
8. When the results in Figs. 7(d) and 8 are compared, it vessel widths in the proposed scheme is to run the pro-
268 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 8, NO. 3, SEPTEMBER 1989

Fig. 7. Three different filters are applied to a different fundus image: (a)
the original image; (b) edge detection by using Sobel operator; (c) edge
detection by using morphological edge detector; and (d) detection of
blood vessels by using the proposed 2-D matched filter.

achieved does not warrant such extra computations. We

have used u = 2, which matched well with a blood vessel
of medium caliber for the retinal images being considered
with our apparatus. Nonetheless, from Figs. 4 and 7(d),
it is clear that vessels from maximum diameter down to
nearly capillary size also responded well to the chosen
value of U .
Lastly, it may be commented here that the matched fil-
ter detection method is applicable only to stationary pro-
cesses. Usually, a pre-processing step is performed in
which the local mean is subtracted from the image and
then the pixel intensities are divided by the square root of
the local variances to approximate the image as a station-
ary process. This pre-processing step may be omitted as
Fig. 8. The edges from Fig. 7(b) are tracked as shown here by the contin-
uous lines. The “anti-parallels” are then detected and filled in to iden-
the qualitative improvement is almost insignificant.
t i I j blood vessel segments. The results show that the detected blood ves-
sels arc broken into several segments due to thc vessel branchings and VI. CONCLUSIONS
crossings.
We have introduced a simple operator for feature de-
tection based on the optical and spatial properties of ob-
gram with kernels having different values of U. Here the jects (namely blood vessels) to be recognized. The pro-
computation increases linearly with the number of differ- posed scheme retains the computational simplicity of the
ent widths being considered. However, practical experi- enhancement/thresholding type of edge operators, and at
mentations suggest that the amount of improvement the same time incorporates the advantages of using model-
CHAUDHURI rt al.: DETECTION OF BLOOD VESSELS IN RETINAL IMAGES 269

based edge detectors. Because of the large size of the con- Trans. Pattern Anal. Machine Intell., vol. PAMI-I, pp. 37-49, Jan.
1979.
volutional kernel, it may take a long time to run this al- 141 J. W. Modestino and R. W. Fries, “Edge detection i n noisy images
gorithm on an ordinary computer. However, these oper- using recursive digital filtering,” Comput. Graphics Image Process-
ators can be efficiently implemented in many machine ing, vol. 6 , pp. 409-433, 1977.
151 T. H. Laffey, L. T . Watson, and R. M. Haralick, “Topographic clas-
vision systems having special hardware support for real sification of digital image intensity surfaces using generalized splines
time convolution operations. and the discrete cosine transformation,” Compur. Vision, Graphics,
This method performs well in analyzing fluorescein an- Image Processing, vol. 29, pp. 143-167, 1985.
161 L. D. Andrews and S . Sinclair, “Macro-vascular topography in dia-
giogram images of the retina as well. With minor modi- betic retinopathy: Analysis of fundus pictures using high resolution
fications, this method could very well be extended to the digital image processing,” in ARVO Abstract, p. 262, 1988.
extraction of geological features from satellite images and [I71 M. Tanaka and K. Tanaka, “An automatic technique for fundus-pho-
tograph mosaic and vascular net reconstruction,” in MEDlNFO ’80.
the enhancement of fingerprint images. Amsterdam, The Netherlands: North-Holland, 1980. pp, I 16-120.
[I81 Y. Okamoto, S . Tamura, and K. Yanashima, “Zero-crossing interval
ACKNOWLEDGMENT correction in tracing eye-fundus blood vessels,” Pattern Recognition,
vol. 21, pp. 227-233, 1988.
The authors would like to thank T. Clark for his help 1191 S . Yamamoto and H. Yokouchi, “Automatic recognition of color
fundus photographs,” in Digital Processing o j Biomedical Images.
in obtaining the images used in this research. The sug- New York: Plenum, 1976, pp. 385-397.
gestions and comments by anonymous reviewers are also [20] K. Akita and H. Kuga, “Digital processing of color ocular fundus
gratefully acknowledged. images,’’ in MEDINFO ’80. Amsterdam, The Netherlands: North-
Holland, 1980, pp. 80-84.
[21] K. Akita and H. Kuga, “Pattern recognition of blood vessel nctworks
REFERENCES in ocular fundus images,’’ in IEEE 1nt. Workshop Phys. and Eng. in
Med. Imaging, Mar. 15-18, 1982, pp. 436-441.
[ I ] I. E. Abdou and W. K . Pratt, “Quantitative design and evaluation of [22] R. Nevatia and K. R. Babu, “Linear feature extraction and descrip-
enhancementlthresholding edge detectors,” Proc. IEEE, vol. 67, pp. tion,” Comput. Graphics Image ProcessinR, vol. 13, pp. 257-269.
753-763, May 1979. 1980.
[2] W. K. Pratt, Digital Image Processing. New York: Wiley, 1978. [23] M. Nelson N. Katz, M . Goldbaum, and S . Chaudhuri, “An image
[3] R. M. Haralick, J . S . J. Lee, and L. G. Shapiro, “Morphological processing system for automatic retina diagnosis,” in Proc. SPlE
edge detection,” IEEE J. Robotics Automat., vol. RA-3, pp. 142- Conj: Elec. Imaging, Devices and Syst., Los Angeles, CA, Jan. 10-
155, Apr. 1987. 14, 1988, vol. 902.
[4] D. Marr and E. Hildreth, “Theory of edge detection.” Proc. R r y . 1241 E. Peli, “Adaptive enhancement based on visual model,” Opt. E n g . ,
Soc. London, Ser. B, vol. 207, pp. 187-217, 1980. vol. 26, pp. 655-660, July 1987.
[5] D. Talton, “Implementation of a Gaussian-smoothed gradient-edge [25] R. M. Rangayyan, S . Walsh, S . Chaudhuri, H. Nguyen, and C. 5 .
operator,” Dep. Elec. Eng., Univ. Pennsylvania, Tech. Rep. Frank, “A Fourier domain directional filtering method for analysis of
MS-CIS-85-12, 1985. collagen alignment in ligaments,” IEEE Trans. Biomc,d. En,?. , vol.
161 T . Elfving, J.-0. Eklundh, and S . Nyberg, “Edge detection using the BME-34, pp. 509-518, July 1987.
Marr-Hildreth operator with different sizes,” in Int. Con$ on Pattern 1261 S. Chaudhuri, “Digital image processing techniques for quantitative
Recognition, 1982, pp. 1109-1 112. analysis of collagen fibril alignment in ligaments,” M.S. thesis, Univ.
171 J. Canny, “A computational approach to edge detection,” IEEE Calgary, 1987.
Trans. Pattern Anal. Machine Intell., vol. PAMI-8, pp. 679-697, [27] L. O’Gorman and J. V . Nickerson, “Matched filter design for fin-
Nov. 1986. gerprint image enhancement,” i n IEEE Inr. Con$ 4~ ouyt., Sprech,
[SI V . Torre and T . A . Poggio, “On edge detection,” IEEE Trans. Pat- and Signal Processing, New York, 1988, pp. 916-919.
tern Anal. Machine Intell., vol. PAMI-8, pp. 147-163, Mar. 1986. 1281 L. O’Gorman and J. V. Nickerson, “An approach to fingerprint filter
L 91 M . H. Hueckel, “An operator which locates edges in digitized pic- design,” Pattern Recognition, vol. 2 2 , pp. 29-38, Jan. 1989.
tures,” J . Ass. Comput. Mach., vol. 18, pp. 113-125, Jan. 1971. [29] S . Haykin, Communication Systems. New Delhi: Wiley Eastern,
[IO] -, “A local operator which recognizes edges and lines,” J . Ass. 1979.
Comput. Mach., vol. 20, pp. 634-647, Oct. 1973. [30] D. H. Friedman, Detection of Signals hy Ternplate Mrrtching. Bal-
[ I 11 F. O’Gorman, “Edge detection using Walsh functions,” Art@cial timore, MD: Johns Hopkins University Press, 1969.
Intell., vol. 10, pp. 215-223, 1978. [31] F. C. Delori, “Noninvasive technique for oximetry of blood in retinal
[ I21 W. Frei and C. C. Chen, “Fast boundary detection: A generalization vessels,” Appl. Opt.. vol. 27, pp. I 1 13-1 125, Mar. 1988.
and a new algorithm,” IEEE Trims. Comput., Oct. 1977. [32] N. Otsu, “A threshold selection method from gray-level histo-
[I31 F. M. Dickey, K. S. Shanmugam, and J. A. Green, “An optimal grams,” IEEE Trans. Syst., Man, Cybem. vol. SMC-9, pp. 62-66,
~

frequency domain filter for edge detection in digital pictures,” IEEE Jan. 1979.