A Retinal Image Enhancement Technique for Blood

Vessel Segmentation Algorithm

A. M. R. R. Bandara P. W. G. R. M. P. B. Giragama
University of Moratuwa, Katubedda, Base Hospital Kebithigollawa
Moratuwa, Sri Lanka. Kebithigollawa, Sri Lanka.
[email protected] [email protected]

© 2017 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other
uses, in any current or future media, including reprinting/republishing this material for advertising or
promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse
of any copyrighted component of this work in other works.

Original Paper:
A. M. R. R. Bandara and P. W. G. R. M. P. B. Giragama, "A retinal image enhancement technique for blood
vessel segmentation algorithm," 2017 IEEE International Conference on Industrial and Information Systems
(ICIIS), Peradeniya, Sri Lanka, 2017, pp. 1-5.
doi: 10.1109/ICIINFS.2017.8300426
keywords: {Biomedical imaging;Blood vessels;Heuristic algorithms;Image
segmentation;Lighting;Retina;Wavelet transforms;SUACE;adaptive contrast enhancement;blood vessel
segmentation;illumination normalization;retinal fundus image},
URL: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8300426&isnumber=8300323
 A Retinal Image Enhancement Technique for Blood
Vessel Segmentation Algorithm
A. M. R. R. Bandara P. W. G. R. M. P. B. Giragama
University of Moratuwa, Katubedda, Base Hospital Kebithigollawa
Moratuwa, Sri Lanka. Kebithigollawa, Sri Lanka.
[email protected] [email protected]

Abstract— The morphology of blood vessels in retinal fundus histogram equalization and also can regulate the amplification
images is an important indicator of diseases like glaucoma, of the details. However, it introduces a box-shaped artifact
hypertension and diabetic retinopathy. The accuracy of retinal which may cause to suppress some details and also it amplifies
blood vessels segmentation affects the quality of retinal image some undesirable details [10].
analysis which is used in diagnosis methods in modern
ophthalmology. Contrast enhancement is one of the crucial steps Speeded up adaptive contrast enhancement (SUACE)
in any of retinal blood vessel segmentation approaches. The algorithm is an outcome of one of our previous studies, in
reliability of the segmentation depends on the consistency of the which the goal is about enhancing superficial vein images that
contrast over the image. This paper presents an assessment of the have been captured with infrared radiation in real-time [10].
suitability of a recently invented spatially adaptive contrast Although SUACE was invented for the purpose of real-time
enhancement technique for enhancing retinal fundus images for execution on resource constrained devices, it also exhibited
blood vessel segmentation. The enhancement technique was several other superior properties such as self-illumination
integrated with a variant of Tyler Coye algorithm, which has balancing, ability to enhance a particular type of detail and the
been improved with Hough line transformation based vessel ability to regulate the detail amplification [10]. In this study,
reconstruction method. The proposed approach was evaluated on we assess the suitability of SUACE in retinal image
two public datasets STARE and DRIVE. The assessment was enhancement for segmenting the blood vessels.
done by comparing the segmentation performance with five
widely used contrast enhancement techniques based on wavelet The evaluation of the suitability of the enhancement
transform, contrast limited histogram equalization, local technique which is done by human experts is subjective.
normalization, linear un-sharp masking and contourlet Therefore we used an existing blood vessel segmentation
transform. The results revealed that the assessed enhancement method combining with SUACE to obtain the accuracy of
technique is well suited for the application and also outperforms segmentation, which then used as a measurement to quantify
all compared techniques. the suitability of the contrast enhancement technique for the
blood vessel segmentation task. We used Tyler Coye
Keywords— adaptive contrast enhancement; blood vessel
algorithm [11] which has been used in numerous related
segmentation; illumination normalization; retinal fundus image;
studies published recently[1], [3], [6], [12], with a slight
SUACE
modification and combining with SUACE. The modification
I. INTRODUCTION has been done for better dealing with the noises caused by the
contrast enhancement techniques without losing parts of
Early diagnosis is crucial in many sight-threatening vessels which are detached from the main vessel network due
diseases like glaucoma, hypertension and diabetic retinopathy to foreground misclassification. Precisely a connected
which cause blindness among working age people [1], [2]. component analysis based noise removal technique and vessel
Therefore retinal image analysis has become one major reconstruction with probabilistic Hough line
diagnosis method in modern ophthalmology. Retinal image transformation[13] are used instead of using only connected
analysis typically involves in blood vessel segmentation, component based technique which is specified in the original
optical disc segmentation and fovea segmentation for implementation of the Tyler Coye algorithm. In the
detecting and analyzing any abnormalities [3], [4]. The experiment, the effect of SUACE in blood vessel segmentation
contrast enhancement is one mandatory step in any of the is compared by applying the modified Tyler Coye algorithm
related image analysis approaches [1]–[6]. The main challenge on images from two public retinal image datasets STARE[14]
of any contrast enhancement algorithm is finding a method to and DRIVE[15], which were enhanced by different contrast
regulate the amplification according to the illumination enhancement techniques based on wavelet transform[16],
variations over the image [7]. A typical solution is applying a CLAHE, LN, linear unsharp masking (LUM) and contourlet
homomorphic filter to normalize the illumination. However, transform[5]. The details in-depth of SUACE algorithm is
some contrast enhancement techniques such as contrast specified in [10]. However, the algorithm has been briefly
limited adaptive histogram equalization (CLAHE) [8] and described in the next section in order to self-contain the
local normalization (LN)[9] have the capability of analyzing article.
the local illumination and regulate the amplification to bring
the final outcome up to an acceptable level of quality. CLAHE
is able to handle the illumination variation by doing local

978-1-5386-1676-5/17/$31.00 ©2017 IEEE

 II. RELATED WORKS
Input Image
A. SUACE Algorithm
SUACE is based on the linear contrast stretching to the
image by selecting the reference dynamic range based on the Convert to Grayscale
illumination response around the point which is to be
enhanced. The illumination response is obtained by using a
low-pass filter with a controllable blur radius. SUACE uses
Contrast Enhancement
Gaussian smooth filter to obtain the lower frequency band.
The two boundaries of the reference range can be calculated
by
Background Exclusion
d
a x,y =g x,y ‐
2 (1)
d
b x,y =g x,y + Binarization
2
where a(x,y) and b(x,y) are the lower and upper boundaries of
the reference range for the scaling, respectively and d is the
Remove Smaller Components
width of reference dynamic range which is an application
dependent variable. The effect of the d has been explained in
[10].
Network of Blood Vessels
The g(x,y) is the low-frequency response of the image
which is obtained by filtering the original signal I(x,y) with the
Fig. 1. Flow of Tyler Coye algorithm
Gaussian kernel as shown in
g x,y =I(x,y)* (x,y)
2 2 (2) grayscale image which is obtained by PCA based Lab color to
1 –x +y2 grayscale conversion.
x,y = e 2σ
2πσ2
It is common that any contrast enhancement algorithm
amplifies both the signal and noises simultaneously. In this
The details about parameters selection have been application, the amplified noises appeared as several
described in [10]. The enhanced image can be obtained by connected components which are detached from the network
using of blood vessels. The simple connected component analysis
0 I x,y <a x,y method which is used in the original implementation of Tyler
l I x,y ≥b x,y (3) Coye tends to remove parts of some smaller vessels with the
I x,y = width of a single pixel, due to the discontinuities over the
I x,y –a x,y
×k, otherwise vessel network as shown in Fig. 2. We extend this original
d noise removal technique by fixing the discontinuities in the
vessel network. The specific procedure is given below.
where Ȋ(x,y) is the contrast enhanced image and k is the new
dynamic range. Then the enhanced image can be processed 1) Remove the connected components with the area less
with Tyler Coye algorithm. than a1. The value of a1 is selected to be a smaller value
B. Tyler Coye Algorithm somewhere between 5-15 pixels.
2) Obtain the probabilistic Hough line transformation of
Fig. 1 shows the flow chart of the Tyler Coye algorithm
for blood vessel segmentation in retinal images. The principle every local window with the size h and fill the disconnected
component analysis (PCA) of weighted Lab color model is gaps in the vessel network by drawing lines for the each pair
used for converting the image into grayscale. The contrast of endpoints having the line votes more than the given
enhancement is done by adaptive histogram equalization. threshold v.
Following the contrast enhancement step, it excludes the 3) Remove the connected components with the area less
background by subtracting the average filtered image. than a2 where a2>>a1.
Isodata[17] is used for extracting a fair threshold level for the
binarization process and finally, the smaller components are The a1 was selected as it does not remove any of the
removed by considering the size of each connected vessels in the image hence the first step can remove some of
components. The method of combining the SUACE with Tyler the noises safely. The window size h was selected based on
Coye algorithm and the modification done for the original the average distance of the discontinued spaces in the vessel
algorithm is described in the next section. networks. Since the window is small, the threshold for the
III. INTEGRATION OF SUACE Hough line votes should be kept smaller. We empirically
select a value for v throughout the experiments. After fixing
The SUACE algorithm is injected into the process at the the discontinuities, only the noises are remained without
contrast enhancement step. The input for the SUACE is the
 for the one based on the LN have amplified the signal from
fovea and the optical disc. But the LN based method has
introduced a kind of random noise all over the image hence
the enhancement become less usable for the vessel
segmentation. Both the contourlet and wavelet based contrast
enhancement algorithm have introduced black patches to the
image in different scales which can be hard to remove in the
segmentation process. Besides these newly introduced noises,
the contourlet based enhancement technique suffers from the
(a) (b) (c) imbalance illumination problem compared to the uniform
Fig. 2. Hogh line transformation based reconstruction. (a) Image with a illumination distribution of the output of SUACE.
sample region marked; (b) Binarized image of the sample region; (c) Image
after the noise removed in vessel reconstructed image. Fig. 3 also shows the results after segmenting the images
enhanced by the 6 different contrast enhancement algorithms
connecting to the vessel network. Therefore we can set a
with their ground truth segmentations. Both the CLAHE and
larger value to a2 in the third step to remove all the noises
LN based methods show that it detects most part of the vessels
safely without false classifying the vessels as the background.
and also misclassified most of the background as well. The
The Hough line transformation votes for potential lines having
LUM, wavelet and contourlet based methods show a less
similar angles hence a slight discontinuity in a vessel does not
misclassification of background yet misclassified several
affect the voting process. Therefore the discontinuities can be
vessel areas. It is clearly observable that the segmentation
fixed computationally by drawing lines with the highest votes.
performance over the image which has been enhanced by
IV. EXPERIMENTAL SETUP SUACE has correctly segmented most of the blood vessels
including the smaller vessels.
The suitability of SUACE is evaluated by comparing the
segmentation accuracy of blood vessels with other state-of- The result of quantitative analysis of the segmentation is
the-art contrast enhancement techniques. The σ and d given in Table I. The enhancement of the SUACE outperforms
parameters of the SUACE were set to 7 and 16 respectively the other methods by means of TPR and accuracy. The reason
throughout the experiments done in this study. These values might be that there is no threshold value which can be used to
were selected empirically where the best average performance separate the foreground and background all over the image
is given in the experiments. Further we set the first connected due to the imbalance illumination produced by the other
component area threshold a1 to 10, the window size h to 5, the enhancement methods. The higher threshold provided by
Hough line vote threshold v to 3 and the second connected isodata algorithm might remove most of the noises as well as
component area threshold a2 to 50 throughout the experiments. the part of the vessels in the network.
We repeated the same procedure for the alternative contrast
The lowest FPR is shown by the wavelet-based method
enhancement algorithms used in this study. The segmentation
where SUACE shows the second best value. However, the
performance is measured by using the accuracy (ACC) which
TPR of the wavelet-based method is extremely lower. The fact
is defined by the sum of the true positives (pixels correctly
proves that the SUACE still can compete with the wavelet-
classified as on a vessel) and the true negatives (correctly
based method. The TPR is gained by the contribution of both
classified as on non-vessel region), divided by the total
the Hough line transformation based vessel reconstruction and
number of pixels in the images. Further, we obtain the true
the performance of the contrast enhancement algorithm.
positive rate (TPR) and false positive rate (FPR) to measure
However, due to the vessel reconstruction process is done with
the correctness of the segmentation. TPR is defined as the total
all the compared contrast enhancement techniques; we can
number of true positives, divided by the number of true blood
conclude that the best TPR is achieved due to the relative
vessel pixels. The FPR is calculated as the total number of
contribution from the SUACE algorithm. Table I also showed
false positives divided by the number of non-vessel pixels
that the integration of the SUACE with the modified Tyler
marked in the ground truth image. We report the average of
Coye algorithm achieved the best accuracy compared to the
TPR, FPR, and accuracy for two publicly available datasets
other integrated approaches.
namely DRIVE and STARE.
TABLE I. PERFORMANCE COMPARISON
V. RESULTS AND DISCUSSION
STARE DRIVE
Fig. 3 shows the output of a sample image which has been Method
TPR FPR ACC TPR FPR ACC
enhanced by each of the contrast enhancement methods. In CLAHE 0.7765 0.0520 0.9245 0.7414 0.0498 0.9122
addition to the blood vessels in the image, there are several
LUM 0.7789 0.0326 0.9112 0.7698 0.0331 0.9058
other components such as the fovea which is the darker region
at the middle of the image and the optical disc which is at the Wavelet 0.6017 0.0311 0.8954 0.5958 0.0315 0.8902
convergence point of all the blood vessels. SUACE has Contorulet 0.6724 0.0325 0.9012 0.6733 0.0327 0.9101
enhanced the blood vessels yet suppressed the fovea and most LN 0.7456 0.0942 0.8735 0.7389 0.0955 0.8598
part of the optical disk. However, all other algorithms except SUACE 0.0324 0.0334
0.7885 0.9489 0.7432 0.9411
 (a) (b)

(c) (e) (g)

(d) (f) (h)

(i) (k) (m)

(j) (l) (n)

Fig. 3. Outputs of different enhancement methods applied to a sample image from DRIVE dataset. (a) & (b) Original image and its ground truth segmentation
respectively; (c) & (d) output of CLAHE and its segmentation; (e) & (f) output of LUM and its segmentation; (g) & (h) output of Wavelet and its segmentation;
(i) & (j) output of Contourlet and its segmentation; (k) & (l) output of LN and its segmentation; (m) & (n) output of SUACE and its segmentation.
 [6] M. Saranya and A. G. Selvarani, “Fundus Image Screening for Diabetic
VI. CONCLUSION Retinopathy,” Indian J. Sci. Technol., vol. 9, no. 25, 2016.
[7] J. H. Yoon and Y. M. Ro, “Enhancement of the contrast in
In this paper, we presented an assessment of the suitability of mammographic images using the homomorphic filter method,” IEICE
a recently invented contrast enhancement algorithm namely Trans. Inf. Syst., vol. 85, no. 1, pp. 298–303, 2002.
SUACE for retinal blood vessel segmentation in fundus [8] A. W. Setiawan, T. R. Mengko, O. S. Santoso, and A. B. Suksmono,
“Color retinal image enhancement using CLAHE,” in ICT for Smart
images. The results of both the qualitative and quantitative Society (ICISS), 2013 International Conference on, 2013, pp. 1–3.
analysis showed the superiority of SUACE in enhancing [9] J. Staal, M. D. Abràmoff, M. Niemeijer, M. A. Viergever, and B. Van
retinal images for blood vessel segmentation, hence we can Ginneken, “Ridge-based vessel segmentation in color images of the
conclude that SUACE is not only suitable for this application retina,” IEEE Trans. Med. Imaging, vol. 23, no. 4, pp. 501–509, 2004.
[10] R. Bandara, H. Kulathilake, M. P. Giragama, “Super-Efficient Spatially
but also outperforms all other enhancement techniques which Adaptive Contrast Enhancement Algorithm for Superficial Vein
are used in the comparison. Further works can be directed to Imaging,” in 2017 IEEE 12th International Conference on Industrial
tune up the parameters of SUACE to detect the other possible and Information Systems (ICIIS) (ICIIS’2017), Peradeniya, Sri Lanka,
components in retinal images such as fovea, optical disc and 2017.
[11] C. Tyler, “A Novel Retinal Blood Vessel Segmentation Algorithm for
retinal lesions. Fundus Images,” MATLAB Cent. File Exch., 2016.
[12] P. Saponaro et al., “Three-Dimensional Segmentation of Vesicular
REFERENCES Networks of Fungal Hyphae in Macroscopic Microscopy Image Stacks,”
[1] P. Furtado, C. Travassos, R. Monteiro, S. Oliveira, C. Baptista, and F. ArXiv Prepr. ArXiv170402356, 2017.
Carrilho, “Segmentation of Eye Fundus Images by density clustering in [13] J. Matas, C. Galambos, and J. Kittler, “Robust detection of lines using
diabetic retinopathy,” in Biomedical & Health Informatics (BHI), 2017 the progressive probabilistic hough transform,” Comput. Vis. Image
IEEE EMBS International Conference on, 2017, pp. 25–28. Underst., vol. 78, no. 1, pp. 119–137, 2000.
[2] S. W. Franklin and S. E. Rajan, “Computerized screening of diabetic [14] “STARE project website,” The STARE Project, Jul-2003. [Online].
retinopathy employing blood vessel segmentation in retinal images,” Available:
Biocybern. Biomed. Eng., vol. 34, no. 2, pp. 117–124, 2014. http://www.ces.clemson.edu/~ahoover/stare.
[3] T. Intaramanee, S. Rasmequan, K. Chinnasarn, B. Jantarakongkul, and [15] M. Niemeijer, J. Staal, B. van Ginneken, M. Loog, M. D. Abramoff, and
A. Rodtook, “Optic disc detection via blood vessels origin using others, “Comparative study of retinal vessel segmentation methods on a
Morphological end point,” in Advanced Informatics: Concepts, Theory new publicly available database,” in SPIE medical imaging, 2004, vol.
And Application (ICAICTA), 2016 International Conference On, 2016, 5370, pp. 648–656.
pp. 1–6. [16] J. Fu, J. Chai, and S. T. Wong, “Wavelet-based enhancement for
[4] A. Salazar-Gonzalez, D. Kaba, Y. Li, and X. Liu, “Segmentation of the detection of left ventricular myocardial boundaries in magnetic
blood vessels and optic disk in retinal images,” IEEE J. Biomed. Health resonance images,” Magn. Reson. Imaging, vol. 18, no. 9, pp. 1135–
Inform., vol. 18, no. 6, pp. 1874–1886, 2014. 1141, 2000.
[5] P. Feng, Y. Pan, B. Wei, W. Jin, and D. Mi, “Enhancing retinal image [17] A. El-Zaart, “Images thresholding using ISODATA technique with
by the Contourlet transform,” Pattern Recognit. Lett., vol. 28, no. 4, pp. gamma distribution,” Pattern Recognit. Image Anal., vol. 20, no. 1, pp.
516–522, 2007. 29–41, 2010.