Available online at www.sciencedirect.

com

ScienceDirect
Future Computing and Informatics Journal 2 (2017) 103e109
http://www.journals.elsevier.com/future-computing-and-informatics-journal/

A thresholding based technique to extract retinal blood vessels from fundus

images
Jyotiprava Dash*, Nilamani Bhoi
Veer Surendra Sai University of Technology, Burla, Odisha, India
Received 2 June 2017; accepted 11 October 2017
Available online 28 October 2017

Abstract

Retinal imaging has become the significant tool among all the medical imaging technology, due to its capability to extract many data which is
linked to various eye diseases. So, the accurate extraction of blood vessel is necessary that helps the eye care specialists and ophthalmologist to
identify the diseases at the early stages. In this paper, we have proposed a computerized technique for extraction of blood vessels from fundus
images. The process is conducted in three phases: (i) pre-processing where the image is enhanced using contrast limited adaptive histogram
equalization and median filter, (ii) segmentation using mean-C thresholding to extract retinal blood vessels, (iii) post-processing where
morphological cleaning operation is used to remove isolated pixels. The performance of the proposed method is tested on and experimental
results show that our method achieve an accuracies of 0.955 and 0.954 on Digital retinal images for vessel extraction (DRIVE) and Child heart
and health study in England (CHASE_DB1) databases respectively.
© 2017 Faculty of Computers and Information Technology, Future University in Egypt. Production and hosting by Elsevier B.V. This is an open
access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Keywords: Retinal imaging; Blood vessels; Mean-C; DRIVE and CHASE_DB1 databases; Medical imaging; Ophthalmologist

1. Introduction glaucoma and cardiovascular diseases [4]. In fundus image

analysis the automatic extraction of object from background is
The exponential growth of the quantity of digital images an essential task. But there are certain difficulties for this such
has covered the direction towards the study of computer vision as the variability in vessel width and low resolution databases
[1]. Exploration in medical informatics focuses predominantly that includes nose and changes in brightness [5].
on assimilating computer technology into the practice of Many methods have already proposed for blood vessel
medicine in order to advance in the area of analysis and segmentation by taking different databases which can be
treatment [2]. The retinal vasculature offers facts which is mainly categorized as: machine learning method, filtering
used to localize the optic disc and fovea, and act as a main part based method and model based method. The supervised
for recognition of pathological changes in automated diag- methods are coming under machine-learning methods where
nostic system [3]. Fundus imaging in ophthalmology is a the pixels are marked either as vessels or non-vessels. In su-
method for medical verdict and exploration of numerous dis- pervised learning, classifiers are accomplished with data from
eases like hypertension, diabetic retinopathy, cataract, hand-labeled images [6]. In Ref. [6], the author proposed a
supervised method where the image is initially enhanced and
* Corresponding author. Department of Electronic & Tele-communication, then features are extracted using 7D feature vector and to this
Veer Surendra Sai University of Technology, Burla, 768018, India. features undergoes neural network classification for labeling
E-mail addresses: [email protected] (J. Dash), nilamanib@ the pixels as vessels or nonvessel. Finally, a post-processing
gmail.com (N. Bhoi). step is applied for gap filling and removal of insulated pixels.
Peer review under responsibility of the Faculty of Computers and
Information, Future University in Egypt
Fraz et al. [7], suggested a supervised technique that utilizes a

https://doi.org/10.1016/j.fcij.2017.10.001
2314-7288/© 2017 Faculty of Computers and Information Technology, Future University in Egypt. Production and hosting by Elsevier B.V. This is an open access
article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
104 J. Dash, N. Bhoi / Future Computing and Informatics Journal 2 (2017) 103e109

collaboration of bagged and boosted decision trees utilizes a stage image fusion is applied to obtain a final segmented image.
feature vector to analyze both healthy and pathological images. In Ref. [15], the author introduced a new vessel segmentation
In Ref. [8], the author introduced a method where in the earlier algorithm where the image is pre-processed in order to get an
stage two binary images are created from green plane using enhanced and smoothed image. The blood vessels are then
high pass filtering and morphological reconstruction method. extracted using level set. In Ref. [16], Cinsdikici et al. intro-
Next, the two binary image common areas are extracted as main duced a method that extract the blood vessels using improved
vessels. All the remaining pixels in the two binary images are ant colony method.
categorized using a Gaussian mixture model (GMM) with a set For study of retinal imaging, detection of blood vessels is a
of eight features that are extracted based on pixel neighborhood basic step. Although several methods are introduced for seg-
and first and second-order gradient images. In the final post menting retinal blood vessel still accurate blood vessels sep-
processing stage, the major vessels are joined with the classified aration is a challenging task due to vessel width disparity and
vessel pixels. In the filtered based approach, retinal vessel low quality retinal images. Many thresholding technique have
segmentation is performed using morphological operators. In been proposed for segmenting retinal images but we have used
morphological image processing, the form is known priory here the local adaptive thresholding for segmenting retinal
which generates a structuring element that is used for filtering blood vessels as it can gives better segmentation performance
the objects from the background [5]. In Ref. [5] the author with a lower execution time as compared to other conventional
developed a new method for identification of blood vessels. thresholding approaches. This paper presents an automated
This method segments the large blood vessels as a solid struc- segmentation algorithm which is carried out in three stages:
ture without artifacts. A methodology that includes vessel pre-processing, vessel extraction and post-processing. In the
skeletal recognition with morphological bit planes for seg- pre-processing stage the image is enhanced using contrast
mentation of retinal blood vessels is offered by Fraz et al. [9]. limited adaptive histogram equalization (CLAHE) [17] and
An automated novel segmentation method has been introduced then the image is thresholded using mean-C thresholding in
by Azzopardi et al. [10]. It is based on the combination of order to detect the required blood vessels from the back-
receptive fields (CORF) raking model of a simple cell in the ground. A post-processing phase is carried out in order to
visual cortex and its employment called combination of shifted obtain a final segmented image. Many authors have proposed a
filter responses (COSFIRE). For the extraction of retinal blood variety of algorithm to segment the blood vessels of the retinal
vessels a rod-shaped COSFIRE non-linear filter is used. In Ref. images but we have presents a very simple, time efficient
[11], the author introduced a system that utilizes the conception thresholding method that gives a good accuracy as compared
of matched filter with first order derivative of Gaussian to many of the existing blood vessel segmentation approaches.
considering that the vessel cross section is a symmetric
Gaussian function. To extract the blood vessels a pair of zero 2. Proposed method
mean Gaussian filter and the first order derivative of Gaussian is
used. Roychowdhury et al. [12], presents a first-hand method The proposed method for extraction of retinal blood vessels
where the new pixels generated iteratively by adaptive global comprises of three phases: pre-processing, segmentation and
thresholding for vessel approximations. A stopping condition is post-processing. For the purpose of vessel extraction the green
used to dismiss the iterative vessel addition process and hence channel of RGB image undergo different phases. The green
reduce the false positives. In model based approach, a vessel channel is chosen because red and blue channel suffers from
model is useful to identify the retinal blood vessels. These poor illuminance [18]. Fig. 1 (a) and (b) shows the original RGB
approaches are sensitive to their parameterization. Here, the and green channel image respectively. All the three phases
trouble is that the parameters should be selected sensibly to responsible for blood vessel extraction are explained below.
extract thin and large vessels at once [13]. In Ref. [14], a
structure is used where initially the Hessian method is used to 2.1. Pre-processing
segment the blood vessels from the rescaled image and after this
again it is back sampled to original size. Subsequently blood The fundus images may suffers from nonuniform illumi-
vessels are take-out using Hysteresis thresholding. At the last nation, so before extraction of blood vessels pre-processing is

Fig. 1. (a) Original retinal image, (b) Extracted green channel image, (c) Contrast limited adaptive histogram equalized image and (d) Median filtered image.
 J. Dash, N. Bhoi / Future Computing and Informatics Journal 2 (2017) 103e109 105

the basic step for getting better segmentation performance.

The pre-processing step includes: contrast limited adaptive
histogram equalization and median filtering.

2.1.1. Contrast limited adaptive histogram equalization

Contrast limited adaptive histogram equalization
(CLAHE), splits the entire space into a number of tiny sections
of identical size and operates on each region where the
contrast of each small region is increased so that the histogram
of the output image corresponds to the histogram indicated by
the distribution parameter. The small neighboring regions are
then combined using bilinear interpolation which suppress the
artificially induced limits. Over noise amplification can be
avoided by limiting the contrast of the individual homoge-
neous region. Fig. 1 (c) represents the CLAHE enhanced
image.

2.1.2. Median filtering

The median filter is a non-linear digital filtering technique, Fig. 2. Variation of accuracy with respect to constant (C).
often used to eliminate noise. This noise reduction is a distinct
pretreatment footstep to upturn the properties of future treat-
ments. The median filter is cast-off to eradicate salt and pepper observed that the accuracy value increases till the value of
noise, to smooth the image. So, it's employ in the segmentation C ¼ 0.042 and then begins to decrease. Therefore, in order to
process [19]. The median filtered image is shown in Fig. 1 (d). get a better result with maximum accuracy, the optimum value
of C is taken as 0.042. Fig. 4 depicts the output images with
2.2. Segmentation using mean-C thresholding different values of C. From Fig. 3 (a), it can observed that at
C ¼ 0.03 the output image includes vessels with unwanted
The black-and-white image is generated from grayscale spurious noise. At C ¼ 0.042 the output image extract vessels
image by setting the pixels to white whose value is beyond a with maximum accuracies 0.9629 as shown in Fig. 3 (b). For
certain threshold, setting the other pixels to black. Here, higher value of C ¼ 0.05 less vessels are identified which is
mean-C thresholding technique is taken where a threshold is shown in Fig. 3 (c).
calculated for each pixel in the image based on some local
statistics such as mean, median of the neighborhood pixels 2.2.2. Experiment 2: To determine the value of window size
and threshold is updated each time. The main advantage of In the second experiment, by increasing the window sizes,
this thresholding method is that it is applicable in case of the unnecessary black pixels are removed from the image
badly illuminated images [20]. In mean-C thresholding background in a better way with filling out characters and vice
(where C is a constant), for extraction of blood vessels the versa. For selection of window size different sizes from 3
3
pixels lies in a uniform neighborhood are set to background to 19
19 are taken and calculated the accuracies. The ac-
value. The mean-C thresholding can be carried out as curacy starts increasing from window size 3
3 to 13
13
follows: and after 13
13 it starts decreasing. So, as at window size
13
13 maximum accuracy is obtained, the optimum window
a. A mean filter of window size N
N is selected. size is considered as 13
13. The variation of accuracy with
b. The enhanced image is convolved with the mean. respect to different window sizes is shown in Fig. 4.
c. A difference image is obtain by subtracting the convolved
image from enhanced image. 2.3. Post-processing
d. The difference image is thresholded with the constant
value C. The image obtained from the segmentation process con-
e. The complement of the thresholded image is calculated. tains some non-vessel which are removed with the help of a
morphological cleaning operation. In this operation six
Two experiments are conducted to determine the values of neighbors containing pixels are characterized as vessel [21].
constant C and window size. The original image and the segmentation result of DRIVE and
CHASE_DB1 are shown in Fig. 5.
2.2.1. Experiment 1: To determine the value of C
In the first experiment, for selection of C, a series of values 2.4. Algorithm for proposed method
from 0.03 to 0.05 with a variation of 0.002 are taken. For every
value of C, accuracy is calculated and plotted against different i. Read the original RGB image.
values of C. This is shown in Fig. 2. From the graph, it can ii. Extract the green plane of the image.
106 J. Dash, N. Bhoi / Future Computing and Informatics Journal 2 (2017) 103e109

Fig. 3. Segmented image at different values of C (a) C ¼ 0.03, (b) C ¼ 0.042 and (c) C ¼ 0.05.

iii. Intensify the image using CLAHE and remove noise

using median filtering.
iv. Extract blood vessels using mean-C thresholding.
v. Remove the isolated pixels using morphological cleaning
operation.

3. Results and discussion

By using DRIVE (Digital retinal images for vessel

extraction) and CHASE_DB1 (Child heart and health study in
England) the presented method's performance is appraised.
The DRIVE database [22] is divided into two set: training set
and test set and each data set contain twenty images. The
ophthalmologists provide a single training set and two testing
set of manual segmentation of the retinal images respectively.
The manual segmentation suggested by the first observer of
the test set is considered as ground truth. The CHASE_DB1
[23] database encompasses total of 28 images of both the left
and right eyes of each child. Two sets of manual segmentation
Fig. 4. Variation of accuracy with respect to window size (w).
results is provided by two observers.

Fig. 5. Segmentation results for DRIVE and CHASE_DB1 databases.

 J. Dash, N. Bhoi / Future Computing and Informatics Journal 2 (2017) 103e109 107

Table 1 Table 4
Mathematical parameters for performance evaluation. Performance comparison (DRIVE database).
Measure Expression Sr. No. Methods TPR TNR ACC
TPR TP=TP þ FN 1 AI-Rawi et al. [3] e e 0.942
TNR TN= 2 Odstrcilik et al. [5] 0.780 0.971 0.947
TN þ FP 3 Marin et al. [6] 0.706 0.980 0.945
PPV TP=
TP þ FP 4 Fraz et al. [7] 0.740 0.980 0.948
NPV TN=TN þ FN 5 Roychowdhury et al. [8] 0.725 0.983 0.952
FDR FP=FP þ TN 6 Fraz et al. [9] 0.730 0.974 0.942


MCC TP
TN  FP
FN pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðTP þ FPÞðTP þ FNÞðTN þ FPÞðTN þ FNÞ
7 Azzopardi et al. [10] 0.765 0.970 0.944
8 Zhang et al. [11] 0.712 0.972 0.938
ACC TP þ TN=TP þ FN þ TN þ FP 9 Roychowdhury et al. [12] 0.739 0.978 0.949
10 You et al. [13] 0.741 0.975 0.943
11 Zhao et al. [15] 0.735 0.978 0.947
Table 2 12 Cinsdikici et al. [16] e e 0.929
Performance evaluation on the DRIVE database. 13 Fraz et al. [21] 0.715 0.976 0.943
14 Proposed method 0.719 0.976 0.955
Image TPR TNR PPV NPV FDR MCC ACC
1 0.776 0.979 0.789 0.978 0.160 0.762 0.962
2 0.731 0.983 0.831 0.969 0.112 0.757 0.957
3 0.700 0.972 0.736 0.964 0.183 0.676 0.944
4 0.692 0.977 0.751 0.967 0.205 0.683 0.949 Table 5
5 0.694 0.984 0.799 0.963 0.167 0.686 0.951 Performance comparison (CHASE_DB1 database).
6 0.676 0.983 0.793 0.959 0.175 0.671 0.957 Sr. No. Methods TPR TNR ACC
7 0.695 0.969 0.687 0.966 0.245 0.694 0.952
8 0.694 0.982 0.768 0.963 0.157 0.655 0.960 1 Fraz et al. [7] 0.722 0.971 0.946
9 0.729 0.982 0.759 0.967 0.210 0.667 0.954 2 Roychowdhury et al. [8] 0.720 0.982 0.953
10 0.701 0.979 0.748 0.971 0.212 0.684 0.955 3 Azzopardi et al. [10] 0.758 0.958 0.938
11 0.709 0.965 0.660 0.969 0.273 0.642 0.940 4 Roychowdhury et al. [12] 0.761 0.957 0.946
12 0.700 0.975 0.727 0.972 0.228 0.687 0.952 5 Proposed method 0.704 0.976 0.954
13 0.695 0.977 0.758 0.963 0.198 0.676 0.956
14 0.735 0.972 0.702 0.976 0.225 0.693 0.963
15 0.740 0.960 0.688 0.979 0.211 0.631 0.944 Table 6
16 0.717 0.976 0.755 0.972 0.205 0.710 0.953 Comparative analysis of the methods in terms of time required to
17 0.680 0.983 0.792 0.970 0.148 0.712 0.967 process image from the DRIVE database.
18 0.765 0.973 0.710 0.979 0.250 0.713 0.965 Method Processing time
19 0.820 0.978 0.771 0.983 0.183 0.776 0.965
20 0.736 0.976 0.714 0.979 0.225 0.703 0.959 Marin et al. [6] 15 min
Average 0.719 0.976 0.746 0.970 0.198 0.694 0.955 Fraz et al. [7] 2 min
Roychowdhury et al. [8] 3.115 s
Azzopardi et al. [10] 10 s
Roychowdhury et al. [12] 2.45 s
Table 3
Zhao et al. [15] 2 min
Performance evaluation on the CHASE_DB1 database.
Cinsdikici et al. [16] 35 s
Image TPR TNR PPV NPV FDR MCC ACC Proposed method 1.66 s
1 0.771 0.971 0.703 0.978 0.283 0.711 0.953
2 0.698 0.981 0.768 0.967 0.231 0.674 0.952
3 0.693 0.967 0.643 0.972 0.274 0.633 0.944 Table 7
4 0.684 0.982 0.793 0.966 0.206 0.684 0.952 Comparative analysis of the methods in terms of time required to
5 0.784 0.970 0.699 0.979 0.273 0.715 0.953 process image from the CHASE_DB1 database.
6 0.701 0.981 0.769 0.974 0.230 0.706 0.958 Method Processing time
7 0.643 0.980 0.784 0.965 0.215 0.683 0.956
8 0.694 0.976 0.756 0.958 0.243 0.649 0.948 Roychowdhury et al. [8] 7.913 s
9 0.718 0.970 0.690 0.980 0.209 0.713 0.954 Roychowdhury et al. [12] 11.711 s
10 0.659 0.973 0.732 0.958 0.267 0.644 0.963 Proposed method 1.98 s
11 0.747 0.961 0.646 0.984 0.245 0.713 0.958
12 0.681 0.986 0.761 0.959 0.238 0.702 0.951 (FDR), Matthews's correlation coefficient (MCC) and accu-
13 0.650 0.983 0.769 0.966 0.230 0.654 0.953
14 0.741 0.981 0.726 0.972 0.273 0.759 0.957
racy (ACC) [24]. Table 1 represents the vessel classification
Average 0.704 0.976 0.731 0.969 0.244 0.689 0.954 and performance matrices respectively. True positive ratio is
the degree of proportion of positives that are appropriately
recognized as such. True negative rate measures the ratios of
The performance of the proposed method is evaluated by negatives that are truly marked as such. Positive and negative
utilizing different performance measures like true positive rate predictive values are the quantity of positives and negatives
(TPR), true negative rate (TNR), positive predictive value results in the diagnosis procedure that are true positive and
(PPV), negative predictive value (NPV), false discovery rate true negative results respectively. False discovery rate is the
108 J. Dash, N. Bhoi / Future Computing and Informatics Journal 2 (2017) 103e109

Fig. 6. Segmentation results of first retinal image from the DRIVE database using different methods (a) Original image, (b) Ground truth image of the first observer,
(c) AI Rawi et al. [3], (d) Cinsdikici et al. [16], (e) Fraz et al. [9] and (f) Proposed method.

expected proportion of false positives among all significant accurately. The method introduced by Zhang et al. [11] detect
hypotheses. The Matthews's correlation coefficient is castoff as some unwanted structures along with the vessels. In Ref. [13],
a portion of the value of binary classifications. Accuracy is the the limitation of the method using radial projection and semi-
measure of capability to classify the amount of conventionality supervised method is overestimation of thin vessels due to
of the resulted image to the manually segmented image. noise and it results in false vessel detection. In the method
Tables 2 and 3 shows the performance matrices for every elaborated by Zhao et al. [15] leads to an inaccurate seg-
images of DRIVE and CHASE_DB1 databases. It is perceived mentation for some pathological images.
that the average accuracy, true positive rate and true negative This paper presented a thresholding based method for
rate of each of the databases is approximately equal to 0.955, vessel identification. The proposed method gives high values
0.695 and 0.976 respectively. The high value of positive pre- of different performance matrices as compared to other
dictive value and low value of false discovery rate indicates existing blood vessel segmentation methods. The average ac-
that the vessel and nonvessel pels are marked with more curacy for the proposed method is 0.955 with TPR, TNR, PPV,
exactness. Higher the value of Matthews's correlation coeffi- NPV, FDR, and MCC of 0.719, 0.976, 0.746, 0.970, 0.198 and
cient better is the performance. The presentation of the 0.694 respectively for DRIVE database. Similarly in CHA-
anticipated technique is compared with the existing method- SE_DB1 database the TPR, TNR, PPV, NPV, FDR, MCC and
ologies in respect of TPR, TNR and ACC, which are given in ACC are 0.704, 0.976, 0.731, 0.969, 0.244, 0.689, and 0.954
Tables 4 and 5 for DRIVE and CHASE_DB1 databases respectively. Fig. 6 represents the comparison of different
respectively. Tables 6 and 7 shows the comparison of proposed images of proposed method with existing methods.
method with existing methods in terms of execution time for
DRIVE and CHASE_DB1 databases respectively. 4. Conclusion
The performance matrices for the existing methodologies
are achieved from their corresponding work. The segmentation In this paper, we have proposed a method based on local
performances of Zhang et al. [11], You et al. [13] are obtained adaptive thresholding for extraction of retinal vessels. The
from Zhao et al. [15] and Fraz et al. [7] and Roychowdhury main contribution of the paper is the application of mean-C
et al. [12] are obtained from Azzopardi et al. [10]. The results thresholding for retinal blood vessels extraction. The method
of Al-Rawi et al. [3] is obtained from Cinsdikici et al. [16]. is evaluated using two different benchmark databases. The
The outcomes of Budai et al. [14] is obtained from its original technique attains a mean accuracies of 0.955 and 0.954 for
literature. The genetic algorithm matched filter optimization DRIVE and CHASE_DB1 databases respectively which are
method [3] is not able to accurately detect the thin vessels better than numerous existing approaches. This automated
 J. Dash, N. Bhoi / Future Computing and Informatics Journal 2 (2017) 103e109 109

method can segment retinal blood vessels effectively. At the [10] Azzopardi G, Strisciuglio N, Vento M, Petkov N. Trainable COSFIRE
same time, the proposed method is very time efficient as it can filters for vessel delineation with application to retinal images. Med
Image Anal 2015;19(1):46e57.
complete the process with in an execution time of 1.66 and [11] Zhang B, Zhang L, Zhang L, Karray F. Retinal vessel extraction by
1.98 s for DRIVE and CHASE_DB1 databases respectively. matched filter with first-order derivative of Gaussian. Comput Biol Med
Together the efficacy and strength with its easiness and fast 2010;40(4):438e45.
employment make the offered method a pertinent implement [12] Roychowdhury S, Koozekanani DD, Parhi KK. Iterative vessel segmen-
for being unified into a comprehensive pre-screening scheme tation of fundus images. IEEE Trans Biomed Eng 2015;62(7):1738e49.
[13] You X, Peng Q, Yuan Y, Cheung YM, Lei J. Segmentation of retinal
for early identification of ophthalmological disorders. blood vessels using the radial projection and semi-supervised approach.
Pattern Recognit 2011;44(10):2314e24.
References [14] Budai A, Bock R, Maier A, Hornegger J, Michelson G. Robust vessel
segmentation in fundus images. Int J Biomed Imaging 2013;2013.
[1] Harini DN, Bhaskari DL. Automatic image segmentation using grid- [15] Zhao YQ, Wang XH, Wang XF, Shih FY. Retinal vessels segmentation
based Ncut and RAG. Int J Image Min 2015;1(4):279e96. based on level set and region growing. Pattern Recognit 2014;47(7):
[2] Elesaid SA. 3D medical image segmentation technique. Int J Biomed 2437e46.
Eng Technol 2015;17(3):232e51. [16] Cinsdikici MG, Aydın D. Detection of blood vessels in ophthalmoscope
[3] Al-Rawi M, Karajeh H. Genetic algorithm matched filter optimization for images using MF/ant (matched filter/ant colony) algorithm. Comput
automated detection of blood vessels from digital retinal images. Comput Methods Programs Biomed 2009;96(2):85e95.
Methods Programs BIOMED 2007;87(3):248e53. [17] Gonzales R, Woods R. Digital image processing. Addison-Wesley Pub-
[4] Mapayi T, Viriri S, Tapamo JR. Adaptive thresholding technique for lishing Company; 1992.
retinal vessel segmentation based on GLCM-energy information. Com- [18] Walter T, Massin P, Erginay A, Ordonez R, Jeulin C, Klein JC. Auto-
put Math Methods Med 2015;2015. matic detection of microaneurysms in color fundus images. Med image
[5] Odstrcilik J, Kolar R, Budai A, Hornegger J, Jan J, Gazarek J, et al. Anal 2007;11(6):555e66.
Retinal vessel segmentation by improved matched filtering: evaluation on [19] Aqeel AF, Ganesan S. Retinal image segmentation using texture,
a new high-resolution fundus image database. IET Image Process 2013; thresholding, and morphological operations. Electro/Information Technol
7(4):373e83. (EIT) 2011:1e6. IEEE International Conference.
[6] Marín D, Aquino A, Gegúndez-Arias ME, Bravo JM. A new supervised [20] Singh TR, Roy S, Singh OI, Sinam T, Singh K. A new local adaptive
method for blood vessel segmentation in retinal images by using gray- thresholding technique in binarization. Int J Comput Sci 2012;8(6):271e7.
level and moment invariants-based features. IEEE Trans Med Imaging [21] Fraz MM, Barman SA, Remagnino P, Hoppe A, Basit A, Uyyanonvara B,
2011;30(1):146e58. et al. An approach to localize the retinal blood vessels using bit planes
[7] Fraz MM, Remagnino P, Hoppe A, Uyyanonvara B, Rudnicka AR, and centerline detection. Comput Methods Programs Biomed 2012;
Owen CG, et al. An ensemble classification-based approach applied to 108(2):600e16.
retinal blood vessel segmentation. IEEE Trans Biomed Eng 2012;59(9): [22] Niemeijer M, Staal J, Ginneken V, Loog B, Abramoff MD. Comparative
2538e48. study of retinal vessel segmentation methods on a new publicly available
[8] Roychowdhury S, Koozekanani DD, Parhi KK. Blood vessel segmenta- database. Med Imaging 2004:648e56.
tion of fundus images by major vessel extraction and subimage classi- [23] K. U. Research. (2011, Jan.). Chase_db1 [Online]. Available: http://
fication. IEEE J Biomed Health Inf 2015;19(3):1118e28. blogs.kingston.ac.uk/retinal/chasedb1/.
[9] Fraz MM, Basit A, Barman SA. Application of morphological bit planes [24] Fawcett T. An introduction to ROC analysis. Pattern Recognit Lett 2006;
in retinal blood vessel extraction. J Digit Imaging 2013;26(2):274e86. 27(8):861e74.