2019 IEEE International Conference on Intelligent Systems and Green Technology (ICISGT)

Retinal Blood Vessel Segmentation by Employing Various Upgraded Median Filters

Sonali Dash Gupteswar Sahu

Department of Electronics and Communication Engineering Department of Electronics and Communication Engineering
Raghu Institute of Technology (A), Visakhapatnam-531162 Raghu Engineering College (A), Visakhapatnam-531162
Andhra Pradesh, India Andhra Pradesh, India
[email protected] [email protected]

Abstract— Retinal image analysis plays a vital part in diagnosis of

Diabetic Retinopathy. It is a progressive disease and leads to II. RIVIEW OF VARIOUS MEDIAN FILTERING TECHNIQUES
blindness which can be prevented if detected at the earlier stage. In
this paper, a robust method of blood tree segmentation is proposed
A. Standard Median filter
by using some advanced median filters such as improved median, The standard median filter is a simple rank selection filter,
hybrid median and weighted median filters. The algorithm for which efforts to eliminate impulse noise by varying the values of
vessel extraction is carried out in three different stages: luminance of the center pixel of the filtering window with the
preprocessing, extraction of blood vessel, and post processing. The median of the luminance values of the pixels restricted inside the
results are compared with the median filter segmentation approach. window. The filtered image T={T(i, j)} from the standard
The approach is validated on DRIVE database. The suggested median filter can be defined as follows.
method delivers more accurate outputs than the median filter in
terms of accuracy, sensitivity, and specificity .
(, )= ( , )∈ , { ( + , + )} (1)
Keywords—Retinal image; blood vessel; vessel extraction; advanced
median filter. Where is a sliding window of size mun pixels centred at
coordinates (i, j).
I. INTRODUCTION
B. Improved median filter
In the field of biomedical presentations, automated analysis
of retinal image helps in the diagnosis of retina pathologies easy An improved median filter is a combination of both
for ophthalmologists, while traditional approaches like dilating averaging filter and a median filter. Filtering of random noises
the eye pupil takes much time [1]. In the literature, many are performed better through average filter, thus to assured the
different approaches are suggested to extract the retinal blood size of the filtering mask median filter is integrated with average
vessels automatically [2-6]. Since last decades, there has been a filter [7]. Supposing {T(i, j), iI} is the gray value of each pixel
great deal of attention in the development of retinal vessel in the image, the filter mask is NuN, Tij’ is the median of Tij in
extraction techniques. Maximum works are revolves around an the mask, then the equation is
illustration named as a “median filter”, that is adaptable, c
= ( − )( − ) , … , , . . ( + )( + ) (2)
convenient and effectual for the utilization. In the literature, it is
c
reported that the upgraded versions of median filters are ( +1) = ( − )( − +1) , … , ,.. ( + )( + +1) (3)
successfully applied to the image processing, speech processing c
= , =0
and delivered improved outcomes than the standard median
filter. Where Tll is the low frequency noise, , , = 1, 2, …n, and
In this paper, we focus on the improvement of performance = − ,…, ; − ,…,
measures such as specificity, sensitivity, and accuracy by
employing different approaches of upgraded median filters. The C. Hybrid median filter
recommended approach is tested on DRIVE dataset. The results There is a necessity to remove this noise in maximum amount
are compared with the standard median filter. The experimental by preserving the main image details. One such filter which can
results demonstrate that the suggested approaches give better remove noise efficiently without destroying the fine details of the
performance measures than standard median filter. image is hybrid filter, which is an improved version of median
The paper is arranged as follows. Brief review of different filter. This is a three step process where a window of size 5×5 is
median filters is given in section II. The suggested approach is selected and two sub-neighbourhoods are formed. The medians
given in Section III. Section IV explains the experiments and of both the neighbourhoods are calculated and the centre pixel
results. Conclusions are given in Section V. value is compared with both the obtained median value. Now the
median of all the three values is computed and the centre pixel is
replaced by that value [8]. The hybrid version of median filter

978-1-7281-1423-1/19/$31.00 ©2019 IEEE 35

DOI 10.1109/ICISGT44072.2019.00023

Authorized licensed use limited to: Newcastle University. Downloaded on October 16,2020 at 13:08:57 UTC from IEEE Xplore. Restrictions apply.
works as “apply median filter with cross-mask, apply median TABLE 1: PERFORMANCE MEASURES
filter with x-mask and take the median of got results and element
itself”. Measure Definition

D. Weighted median filter

TPR TP/vessel pixel count
Weighted median filter can preserve details, e.g., pulses, of
any preferred length by choosing appropriate weights [9]. FPR FP/non-vessel pixel count
However, in practical conditions, to select the appropriate
weighting coefficients is a challenging task. In addition, when Sensitivity TP/(TP+FN)
the weights are large the computational time is high. The filtered
image T={T(i, j)} from the weighted median filter can be defined Specificity TN/(TN+FP)
as follows.
T (i, j ) Median(k , l ) Wm,n {Wm,n (k , l ) D(i  k ), j  l} Precision TP/(TP+FP)

(4) Accuracy (TP+TN)/(TP+TN+FP+FN)

III. PROPOSED METHODOLOGY

Retinal vessel segmentation methodologies are frequently
based on visual and geometric features that distinguish arteries Green
and veins. However, in most of the cases to differentiate arteries Morphological
channel CLAHE
from veins is quite difficult because of low quality images and operation
extracted
the vessels in the outer regions of the image are very dark due to
shading effect resulted from inhomogeneous lighting of the
image. In these cases, arteries and veins look very similar that Background Upgraded
leads to wrong segmentation. For this reason denoising of retinal extraction by median
images are of utmost important before extraction of the retinal morphological filters
vessels. Many authors have utilized median filter for denoising
of retinal images. This paper recommends denoising of the Fig. 1. Procedural steps of the recommended approach.
retinal images by using different upgraded median filters that is
more robust to noise than the standard median filter. The
suggested technique for the retinal blood vessels extraction
contains of three phases: pre-processing, segmentation and
postprocessing. The green channel of the RGB retinal image is
selected because it has high intensity as compared to red and (a) (b) (c) (d)
blue channel. The procedures of the proposed method are shown
in Fig. 1.
Pre-processing is a step where the green channel is enhanced
by applying contrast limited adaptive histogram equalization
(CLAHE), which divides the whole space into a number of (e) (f) (g)
minute regions of equal size and performs on each region. Then Fig. 2. (a) Original image (b) Extracted green channel image (c) CLAHE image
by applying morphological structuring element, retinal blood (d) Median filtered image (e) Improved median filtered image (f) Hybrid median
filtered image (g) Weighted median filtered image
vessels are highlighted. Afterwards by utilizing the
morphological open function the blood vessels are thickened. IV. EXPERIMENTAL RESULTS, COMPARISION AND DISCUSSION
Fig. 2 (a), (b), and (c) display the original RGB, green channel,
and CLAHE enhanced images respectively. Then the standard This paper presents a method for vessel enhancement filtering
median filter and various upgraded median filters such as which is formulated on different upgraded median filters and
improved median filter (IMF), hybrid median filter (HMF), and used for retinal vessel detection. This paper presents a method
weighted median filter (WMF) are applied to eliminate the noise. for vessel enhancement filtering which is formulated on different
A simple thresholding function is used to extract the retinal upgraded median filters and used for retinal vessel detection.
blood vessel. The standard median filter image, improved This work utilizes the DRIVE database to verify the suggested
median filtered image, hybrid median filtered image, and methods. The numbers of retinal images are 40 in the DRIVE
weighted median filtered image in Fig. 2. (d), (e), (f), and (g) database out of which seven of them have various pathological
respectively. cases. The dimensions of the images are 768u584 pixels with 24
In the postprocessing stage, the image achieved through the bits. The images in the database are divided into two parts as
segmentation procedure carries certain non-vessels that are training and testing and each part contains 20 images. The testing
eliminated through the support of a cleaning process by using images are used for performance evaluation. The evaluation of
morphological operation. the suggested approaches are computed by using various
The details of performance measures are given below in table 1. performance measures such as accuracy, false positive rate
(FPR), true positive rate (TPR), and precision. TABLE 2 shows

36

Authorized licensed use limited to: Newcastle University. Downloaded on October 16,2020 at 13:08:57 UTC from IEEE Xplore. Restrictions apply.
the performance measures of the original standard median filters. TABLE 3. EVALUATED PARAMETERS FOR THE UPGRADED MEDIAN FILTERS
TABLE II represents the performance measures of each image Upgrade TPR FPR Accuracy Precision
by using the improved median filter, hybrid median filter, and d MF
weighted median filter. Retina1
From the TABLE 3 it is observed that for all the retinal IMF 0.778701 0.019376 0.97062 0.7965
images by using the improved median filter gives the highest
percentage of accuracy than the traditional standard median HMF 0.69966 0.013639 0.96078 0.8340
filter. Similarly, by using the hybrid median filter all most all the WMF 0.69966 0.013639 0.96078 0.8340
retinal images achieve higher accuracy than the standard median Retina2
filter. However, the weighted median filter is unable to deliver
IMF 0.708701 0.008937 0.969347 0.9004
better performance measures than the standard median filter. The
reason of reducing of performance measure is that the selection HMF 0.665256 0.006469 0.969913 0.9214
of the weight is not appropriate. Figure 3 shows the segmented WMF 0.69195 0.011371 0.958246 0.8740
images obtained from various upgraded median filter Retina3
approaches.
IMF 0.742863 0.034130 0.958639 0.7067
HMF 0.679293 0.022267 0.950982 0.7715
WMF 0.730338 0.030235 0.949896 0.7278
(a) (b) (c) Retina4
IMF 0.58111 0.004339 0.969525 0.9313
HMF 0.545134 0.003394 0.958073 0.9420
WMF 0.603545 0.007937 0.946322 0.8851
(d) (e) (f)
Fig. 3. Segmentation images of various approaches (a) Original image (b)
Retina5
Ground truth image of the first observer (c) Standard median filter (d) Improved IMF 0.659711 0.011101 0.959059 0.8599
median filter(e) Hybrid median filter (f) Weighted median filter.
HMF 0.581166 0.006045 0.955282 0.9085
TABLE 2. EVALUATED PARAMETERS FOR TRADITIONAL MEDIAN FILTER WMF 0.637778 0.015291 0.952206 0.8117
Retina6
Median TPR FPR Accuracy Precision IMF 0.726616 0.037549 0.939496 0.6760
filter HMF 0.624144 0.011637 0.952912 0.8525
Retina1 0.711923 0.0193764 0.9622 0.7965 WMF 0.619722 0.012849 0.951388 0.8387
Retina2 0.678662 0.00689807 0.960901 0.918198
Retina7
Retina3 0.698203 0.0241932 0.948133 0.761649 IMF 0.718526 0.033614 0.6825
0.948736
Retina4 0.553469 0.00349125 0.955752 0.941387 HMF 0.646093 0.019705 0.949755 0.7673
Retina5 0.596079 0.00610939 0.956622 0.909791 WMF 0.73806 0.040882 0.938917 0.6448
Retina6 0.616204 0.00915916 0.934376 0.878852
Retina8
Retina7 0.645761 0.0181783 0.951112 0.781309 IMF 0.814153 0.092578 0.4529
0.949397
Retina8 0.711895 0.0345259 0.943657 0.659983 HMF 0.698158 0.03119 0.945524 0.6781
Retina9 0.676826 0.021064 0.954452 0.739157 WMF 0.84811 0.155529 0.904784 0.3392
Retina10 0.598652 0.00680308 0.960726 0.887536
Retina9
Retina11 0.62321 0.0165002 0.951246 0.787854 IMF 0.803485 0.073395 0.4912
0.959626
Retina12 0.725904 0.0187946 0.959162 0.784947 HMF 0.66011 0.019299 0.7510
0.954719
Retina13 0.565982 0.00540475 0.952691 0.919011 WMF 0.847164 0.124943 0.932797 0.3742
Retina14 0.749859 0.0230346 0.958604 0.741163
Retina10
Retina15 0.701194 0.0213223 0.958819 0.717107 IMF 0.680881 0.013810 0.8155
0.968062
Retina16 0.645497 0.0101676 0.958743 0.863028 HMF 0.581492 0.006165 0.8942
0.969898
Retina17 0.700883 0.0248322 0.952015 0.722385 WMF 0.487885 0.006793 0.951618 0.8656
Retina18 0.697139 0.0162598 0.961032 0.786756
Retina11
Retina19 0.75032 0.00894943 0.971081 0.883502 IMF 0.682284 0.02138 0.7583
0.957091
Retina20 0.712137 0.0176189 0.962508 0.762375 HMF 0.630285 0.015941 0.7954
0.955388
WMF 0.615356 0.018933 0.948327 0.7616
Retina12

37

Authorized licensed use limited to: Newcastle University. Downloaded on October 16,2020 at 13:08:57 UTC from IEEE Xplore. Restrictions apply.
 TABLE 4. COMPARATIVE STUDY
IMF 0.846859 0.078966 0.918629 0.5033
HMF 0.70888 0.016930 0.959395 0.7982 Approach Accuracy
WMF 0.715374 0.021869 0.955443 0.7555
Yin et al. [10] 0.9267
Retina13
Salazar-Gonzalez et al. [11] 0.9412
IMF 0.625717 0.011548 0.968988 0.8544 Roychowdhury et al. [12] 0.9519
HMF 0.565672 0.005324 0.955734 0.9200 Original standard median filter 0.9556
WMF 0.516352 0.006640 0.954672 0.8939 Proposed Improved median filter 0.9609
Retina14 Proposed Hybrid median filter 0.9590
IMF 0.839112 0.054984 0.960453 0.5730 Proposed Weighted median filter 0.9475
HMF 0.736027 0.020960 0.959392 0.7554 V. CONCLUSION
WMF 0.784833 0.038132 0.947554 0.6441 In this paper, different approaches based on various upgraded
Retina15 median filters for the retinal vessels segmentation is
IMF 0.79275 0.042611 0.969606 0.5891 recommended. Instead of utilizing the existing median filter, the
proposed techniques have considered various types of upgraded
HMF 0.700347 0.021563 0.961534 0.7145 median filters for the denoising of retinal images before vessel
WMF 0.796773 0.053508 0.935777 0.5344 extraction. The results are compared with the existing standard
Retina16 median filter. The results demonstrate that the suggested
approaches deliver better performance measure than the standard
IMF 0.746232 0.020884 0.968089 0.7800
median filter in retinal vessel segmentation.
HMF 0.64647 0.010277 0.968731 0.8619 However, the weighted median filter does not perform well in
WMF 0.726327 0.028240 0.9496 0.7185 delivering good results because of the selection of the weights. In
Retina17 future, the weighted median filter with appropriate selection of
IMF 0.790607 0.061597 0.968927 0.5419 the weights can be used for achieving the better results.
HMF 0.6783 0.022733 0.952031 0.7333 REFERENCES
WMF 0.696754 0.026325 0.9503 0.7093 [1] A. Sopharak, B. Uyyanonvara, S. Barman, “Automatic detection of diabetic
retinopathy exudate from nondilated retinal images using mathematical
Retina18 morphology methods,” Comput Med Imaging Graph, vol. 32(8), pp. 720-
IMF 0.796053 0.032529 0.967888 0.6780 27, 2008.
HMF 0.70192 0.017536 0.962235 0.7749 [2] D. Marin, A. Aquino, M. Gegundez-Arias, J. Bravo, “A new supervised
WMF 0.77295 0.037970 0.947048 0.6365 method for blood vessel segmentation in retinal images using gray-level
and moment invarianta-based features.” IEEE Transc on Med Imag, vol.
Retina19 30(1), pp. 146-58, 2011.
IMF 0.822988 0.015017 0.978545 0.8321 [3] U. R. Acharya, C. M. Lim, E. Y. Ng, C. Chee, T. Tamura, “Computer-
HMF 0.749004 0.007888 0.975945 0.8957 based detection of diabetes retinopathy stages using digital fundus images,”
Proc Inst Mech Eng H. vol. 223(5), pp. 545-553, 2009.
WMF 0.781338 0.024703 0.959207 0.741 [4] Z. Xiaohoui, C. Opas, “A SVM approach for detection of haemorrhages in
Retina20 background diabetic retinopathy,” Int Neural Networks, 2435-40, 2005.
IMF 0.822543 0.037468 0.968237 0.6353 [5] J. Dash, N. Bhoi, “A thresholding based technique to extract retinal blood
vessels from fundus images.” Futur Compt and Inform J, vol. 2, pp. 107-
HMF 0.722316 0.019804 0.963232 0.7432 09, 2017.
WMF 0.686544 0.021956 0.956607 0.7128 [6] M. M. Fraz, A. R. Rudnicka, C. G. Owen, D. P. Strachan, S. A. Barman,
“Automated Arteriole and venule recognition in retinal images using
ensemble classification,” 2014 International Conference on Computer
The suggested approaches deliver higher values of accuracy Vision Theory and Applications (VISAPP), IEEE, vol. 3, pp. 194-202,
as compared to the original traditional median filter. The 2014.
standard median filter delivers average accuracy of 0.9556. The [7] Y. Zhu, C. Huang, “An Improved Median Filtering Algorithm for Impulse
Noise Reduction”, International Conference on Solid State Devices and
improved median filter and hybrid median filter give average Material Sciences, vol. 25, pp. 609-16, 2012.
accuracy of 0.9609 and 0.9590, which is higher than the existing [8] Zeinab Mustafa, Banazier A. Abrahim, Yasser M. Kadah, “Modified
median filter. However, the weighted median filter delivers Hybrid Filter for Image Denoising”, 29th National Radio Conference, IEEE
average accuracy of 0.9475, which is lower than the existing transactions, pp 705-712, 2012.
median filter. The performance of the suggested approach on [9] D. R. K. Brownrigg, “The weighted median filter,” Commun ACM, vol. 27,
DRIVE dataset is compared with other approaches in respect to pp. 807-18, 1984.
accuracy: Yin et al. [10], Salazar-Gonzalez et al. [11], and Roy- [10] Y. Yin, M. Adel, S. Bourennane, “Automatic segmentation and
chowdhury et al. [12]. TABLE 4 illustrates the performance of measurement of vasculature in retinal fundus images using probabilistic
formulation,” in Computational and Mathematical Methods in Medicine,
our approach against the above approaches on DRIVE database. Hindawi Publishing Corporation Computational and Mathematical Methods
in Medicines, vol. 2013 (1), 2013.
[11] A. Salazar-Gonzalez, D. Kaba, Y. Li, X. Liu, “Segmentation of the blood
vessels and optic disk in retinal images,” IEEE J of Biomed and Health
Inform ,vol. 18(6), pp. 1874-1886, 2014.

38

Authorized licensed use limited to: Newcastle University. Downloaded on October 16,2020 at 13:08:57 UTC from IEEE Xplore. Restrictions apply.
[12] S. Roychowdhury, D. D. Koozekanani, K. K. Parhi, Blood vessel classification. IEEE J Biomed Health Inf, vol. 19(3), pp. 1118-1128, 2015
segmentation of fundus images by major vessel extraction and subimage

39

Authorized licensed use limited to: Newcastle University. Downloaded on October 16,2020 at 13:08:57 UTC from IEEE Xplore. Restrictions apply.