You are on page 1of 6

The Use of Fourier Phase Symmetry for Thin

Vessel Detection in Retinal Fundus Images


1st Mohammad A. U. Khan 2nd Tariq M. Khan 3rd Kahtan I. Aziz
Biometric and Sensor Lab Electrical and Computer Engineering Dept. College of Engineering and Computing
Effat University COMSATS University Al Ghurair University
Jeddah 22322, Saudi Arabia Islamabad, Pakistan Dubai Academic City, UAE 37374
mohammad a khan@yahoo.com tariq045@gmail.com kahtan@agu.ac.ae

4th Sayed S. Ahmad 5th Nighat Mir 6th Emhemed Elbakush


College of Engineering and Computing Department of Computer Science College of Engineering and Computing
Al Ghurair University Effat University Al Ghurair University
Dubai Academic City, UAE 37374 Jeddah 22322, Saudi Arabia Dubai Academic City, UAE 37374
sahmed@agu.ac.ae nmir@effatuniversity.edu.sa e.albakush@agu.ac.ae

Abstract—Fundus examination is a non-invasive procedure extent the stress on the clinicians. A CAD systems process
of observing changes in retinal vasculature linked with the retinal fundus images for an earlier warning systems and
identification and progression of certain ocular diseases. Seg- can emulate in some way the inspection by ophthalmologists
menting vessels from the rest of the structure is found helpful
in analyzing and later tracking the changes. Manual vessel for the diagnosis of several abnormalities, such as diabetic
segmentation requires clinical expertise, and with large scale retinopathy (DR), retinopathy, and maculopathy [1].
screening certainly puts a burden on already scarce clinical Blood vessel detection is a foremost step in constructing
resources. A computer-aided diagnosis (CAD) recently emerged of a computer-aided diagnostic (CAD) systems for retinal
to alleviate this burden. A variety of computerized methods have fundus images as they provide one of the most stable fea-
emerged recently with the primary aim of providing accurate
vessel segmentation. One particularly interesting approach is tures. Furthermore, vessel detection can help in finding other
multi-scale line filtering. However, its response diminishes in anatomical landmarks such as the optic nerve head (ONH)
low-contrast areas of the image causing certain vessels to be and the macula in the retinal image. The landmark feature
missed. In this paper, we investigate the use of phase symmetry spatial characteristics can be helpful in detection of other
detector to get help with low-contrast vessel detection. This abnormalities. It has been established that blood vessels appear
specific detector does not make any assumptions about the
luminance profile of the vessel but then has major drawback with Gaussian cross-sectional profile in fundus images and
of being sensitive to background noise. To reduce the noise spread as piecewise linear segments with a look of a tree
sensitivity, we adopted the multi-scale line filtering with an structure. Majority of the vasculature is found to converge at
improved vessel uniformity function as an input to the phase one specific location referred to as optic disc [2], [3].
symmetry detector. The low-contrast vessel information thus The blood vessel segmentation with computerized methods
made available helps in providing an improved accuracy for
automated vessel segmentation algorithms. The quantitative tests can be challenging for several reasons. Some of the noise
are conducted for two publicly available databases (DRIVE, sources are due to the acquisition process, while others are
STARE) of fundus images that shows promise of improvements intrinsic to retinal images. The two most dominating factors
in all three performance categories called accuracy, sensitivity, making segmentation hard are the nonuniform intensity for
and specificity. a vessel across its entire length and the uneven background.
Index Terms—segmentation, fundus image, multi-scale line
filtering, color-to-gray conversion, weighted linear square filter The uneven background intensity is due to the acquisition
conditions, while the unstable contrast is for the reason that
contrast agent injected may not reach to all parts of the retinal
I. I NTRODUCTION
vessels. Generally speaking, arteries have higher contrast than
Majority of ocular diseases undergo asymptomatic progres- veins. Thick vessels are found to possess a higher contrast
sion and the patient become aware only at the last stage than most of the thin ones. Additionally, presence of noise,
where treatment options become limited in preventing vision fovea and optical disk, vessels having width variation, effects
impairment. With regular eye screenings we can detect the of lesions, and pathological changes are probable concerns to
onsight of such diseases and intervention can be devised in be addressed. Therefore, development of an automatic blood
earlier stages. However, an increasing number of eye screening vessel detection system that extracts blood vessels of a retina
procedures would put a large burden on the already scarce with adequate accuracy is highly desired.
clinical resources. Computer-Aided-Diagnosis (CAD) systems A vessel extraction algorithm must take advantage of the
automate the detection process and thus can reduce to a larger geometry of the vessels; in particular, the characteristic cross-
lution SAR imagery [8]. A different application is its use
for discriminating lung disease in chest radiograph [9]. The
use of the phase congruency for image and video sharpness
improvement and blur reduction is discussed in [10]. The
phase congruency concept is also been tested for subsurface
structures delineation within migrated seismic volumes [11].
A strong impediment in its use is its high sensitivity to
pick background noise. We proposed a variant of fourier
(a) (b)
phase characteristic known as phase symmetry. At points of
Fig. 1. Multiscale Line filtering limitation: Image on the left depicts Image symmetry in natural images we find readily identifiable phase
30 of DRIVE database having vasculature (thick and thin vessels) marked patterns. The bilateral symmetry identification is found helpful
as yellow color; while image on the right displays Multi-scale line filtering
output superimposed as green color with undetected thin vessels appearing as
in detecting low-contrast thin vessels. Measures are discussed
yellow. to reduce the sensitivity of the phase-symmetry detector for
noise such as instead of applying directly on acquired images,
it works on output of the multi-scale line filtering rather
vessel profile is Gaussian in nature together with constant than directly on acquired images. A specialized vessel linking
intensity along the flow direction. A large majority of vessel algorithm is then enacted to provide response to breakage
enhancement filters reported earlier make use of eigen-value in low-contrast vessel network. The processing flow of the
analysis of the second-order spatial derivatives matrix (Hessian proposed vessel segmentation method is as shown in Fig. 1.
matrix ) of the image intensity in neighborhood of a pixel in The rest of the paper is organized as follows. The details of the
order to detect elongated linear structures. One most popular proposed method are described in Section 2. The experimental
example is the Frangi’s filter [4]. The second most popular results and performance evaluation are provided in Section 3.
approach in this genre is the employment of a specially- Final Section 4 presents conclusions.
designed linear filter. It is more like an oriented line, which is
used to scan the neighborhood with rotation through a number II. P ROPOSED M ETHODOLOGY
of angles while convolving with the image at the same time,
called multi-scale line filtering [5], which was later improved Our proposed methodology for retinal blood vessel seg-
in [6] resulting in acceptable accuracy scores. The simulation mentation comprises of four consecutive stages: 1) Weighted-
tests revealed that line filtering technique is not adequate, as Least-Square Filtering with SUACE algorithm for background
it responds more strongly to bright distracters than to those homogenization and the removal of non-uniform illumination;
of actual vessels. This happens due to the method’s direct 2) vessel enhancement using multi-scale line filter and its
proportionality of the magnitude of the eigenvalues. However, improvement for uniformity of vessels; 3) Morphological
in reality vessels carry eigenvalue heterogeneity and variation Vessel Enhancement with Phase Congruency detector; and
in eigenvalue magnitude across background, this proportion- 4) Hysteresis threshold operation with post-processing for
ality dependence results in nonuniform enhancement. In [7], removing artifacts. For the illustration of the flow process, Fig.
authors suggested a modification to mitigate the vessel contrast 1 displays the original color image 12 of the DRIVE database
variation and provide filters that depends on eigenvalue ratio and its ground-truth image.
rather than magnitude. This results in a regularised filter to
show robustness to small changes in magnitude due to back- A. Grayscale Conversion
ground noise. The modified vessel enhancement filter detects Fundus image is an RGB color image, which consists of
large percentage of the vasculature. however, certain thin low- three channels (red, green and blue). It has been observed
contrast vessels still remains illusive as they are not raised to that the green channel provides the highest vessel/background
adequate level due to the structural limitations of the vessel contrast; whereas the red shows vasculature with saturation,
enhancement filters. Certain fundus images are quite noisy whereas blue channels is largely noisy [12]. Therefore, the
and it happens so that fine vessels ranging down to sub-pixel blue channel is avoided altogether, while the green and red
level shows their presence with low contrast and therefore channels are linearly combined to provide a grayscale image,
hard to be identified even by a human eye. Segmenting these according to the relationship as described below:
very fine vessels is the focus of the research presented here.
The paper proposes the use of contrast- and illumination- I = k · Ig + (k − 1) · Ir , (1)
invariant phase-symmetry detector to recover many of the
thin low-contrast vessels. There has been a renewed interest where Ig and Ir represent the green and red color components
recently in employing phase congruency characteristics such of a retinal image. Parameter k is the weight and is set at 0.75
as phase symmetry detector for the purpose of line and blob here. The purpose of assigning more weight to Ig is to keep
identification in diverse applications. A relevant application the image variation across the background suppressed and get
is to develop a reliable and robust detector for extracting better contrast for the retinal vessels which are usually much
the lake outlines of glacial masses in high-temporal reso- stronger in Ig as compared to in Ir .
Fig. 2. Block diagram showing the flow from an input color fundus image to the final binary vessel image.

B. SUACE algorithm value of each pixel to be a linear combination of its neighbors


The retinal background shows considerable intensity vari- N (i). The non-linear weights wij provides the edge-preserving
ation, and the contrast of the thin vessels quite low. If properties for the smooth image, and are
P set as follows:
background is left untreated will produce artifacts, especially wij ∝ exp α||g(i) − g(j)||2 , where j∈N (i) wij = 1,
around light objects such as exudates, reflection artifacts and α is a parameter to emphasize the edge-preserving power.
the optic nerve head. Recently, a Speeded-Up Adaptive Con- The smooth image approach has three parameters: choice
trast Enhancement (SUACE) has been proposed in [14], which of neighborhood N (i), λ to trade off smoothing and edge-
provides background homogenization. The SUACE algorithm preservation, and α for computing weights.
uses a fixed distance for enhancing the contrast of the image. We seek a new image u, which, on the one hand, is as close
However, the distance is moved up or down based on the as possible to g, and, while on the other hand, is as smooth as
calculated illumination map corresponding to local intensity possible everywhere, except across some significant gradients
mean, i.e. the two boundary values of the range are varied in g, which happens to be at the vessels boundaries [15] in
from point to point. The two boundary values are calculated our case.
as follows: The enhanced image can be obtained by using the condi-
d tional relation as described below:
a(x, y) = g(x, y) − , 
2 (2) 0,
 I(x, y) < a(x, y)
d Ienh (x, y) = 1, I(x, y) ≥ b(x, y) (4)
b(x, y) = g(x, y) + ,
2  I(x,y)−a

d × k, otherwise
where a(x, y) and b(x, y) are the lower and upper endpoints
of the fixed distance d for the scaling, respectively. where Ienh (x, y) is the contrast enhanced image and k is the
The image g(x, y) is e low-frequency response of the image new dynamic range.
which was obtained by filtering the original image with the
Gaussian kernel. However, in our application, vessels are C. Multi-scale Line Filtering
involved, that are regarded as linear features, so a better Though the SUACE algorithm provided background homog-
alternative is to preserve the edges of the vessels, while enized image, the vessel contrast has to be enhanced with
smoothing the rest of the image. For this purpose, an edge- background noise suppressed. For this purpose, Nguyen et al.
preserving smoothing is utilized which provides a compromise [5] proposed a multi-scale line detector, which uses a range
between two possibly contradictory goals: smoothness and of lengths assigned to lines aligned to the underlying linear
edge-preservation. Given an input image g with n pixels,the features. The generalized multi-scale line detector is defined
image u is an edge-preserving smooth result, which minimizes as follows:
the following energy expression: L L L
RW = Imax − Iavg , (5)
 
n L
X X where, 1 ≤ L ≤ W , and Iavg is the average intensity of
λ||u(i) − g(i)||2 + ||u(i) − wij u(j)||2  , (3) all pixels lying along the line of length L. By changing L,
i=0 j∈N (i)
a multi-scale line detector is constructed. The rationale for
where i denotes the index of the pixel. The data term combining a range of various length filters is understood
||u(i)−g(i)||2 guarantees that the smooth image approximates by observing individual line filter responses. The line filter
the original image, while the smoothing term constraints the having longer length lines can successfully recognize central
The input image is convolved with each of the quadrature pairs
of wavelets. If we let I denote the image and Mne and Mno
denote the even-symmetric and odd-symmetric wavelet at scale
n, respectively. This results in vectors en (x) = I(x) ∗ Mne and
on (x) = I(x) ∗ Mno , referred to as real and imaginary parts of
(a) (b) (c) (d) complex valued frequency component. The phase is given by:

φn (x) = atan2(en (x), on (x)).

It has been observed that at a point of symmetry the absolute


value of the even-symmetric filter outputs is large and the ab-
solute value of the odd-symmetric filter outputs is small. Thus,
(e) (f) (g) (h)
to emphasize symmetry a natural way is to take the absolute
Fig. 3. Various Stages of the Proposed Method: (a) GrayScale Image, value of the even-symmetric filter output and subtract from it
(b) SUACE algorithm output displaying background homogenization, (c) the absolute value of the odd-symmetric filter output. Then to
the Multi-scale Line filtering Output, (d) High-contrast vessels detected by
hysteresis threshold, (e) Phase Symmetry detector output for low-contrast aggregate responses from filter responses over multiple scales
vessels, (f) combined low and high-contrast vessel network, (g) morphological a weighted average is formed. For this purpose, difference of
cleaning and vessel linking output as green segments, (h) final vasculature the absolute values of the even and odd filter responses at each
detected superimposed on the original color image showing high-contrast
vessels with yellow shade and low-contrast vessels with red shade. scale is weighted by the corresponding magnitude of the filter
response vector at each scale. The quantity is then normalized
by summing the magnitude filter response over all scales. This
reflex pixels as vessel pixels, because the candidate line produces the following phase symmetry detection equation:
includes only a small number of central reflex pixels; hence P
the average intensity of pixels in the candidate line is not bAn (x)[| cos(φn (x))| − | sin(φn (x))|] − T c
Sym(x) = n P .
much affected, and the central reflex pixels have a high n An (x) + 
vessel-ness measure. The shorter-length line filters detect close (7)
vessels more effectively, but introduce background noise. The The term  is a small numerical quantity to prevent division by
variances of line responses decrease with an increase in length zero. The factor T is a noise compensation term representing
from smaller scalers to large ones, which demonstrates that the maximum response that could be generated from noise
the smaller length filters produce higher noise responses. In alone in the image, and the symbols bc represent the enclosed
order to remove biasses for large-length filters and eliminate quantity to be equal to itself iff its value is positive, and zero
the drawback of each individual line detector, line responses at otherwise. For the purpose simulations conducted in this paper,
all scales are linearly combined with their respective weights the local frequency information is obtained via a bank of log
to produce the segmentation for each retinal image. The overall Gabor wavelets tuned to different spatial frequencies, and the
response at each image pixel is defined as: phase symmetry of vessel detection image is then segmented
! using hysteresis threshold as shown in Fig. 3 (e). We observe
1 1 X L that many fine vessels are identified in the process.
Rcombined = L × RW + Ive , (6)
nL + 1 nL
L
L E. Hysteresis Thresholding
where nL is the number of scales used, RW is the response
of the line filter at scale L, and Ive is the enhanced image. The retinal vessel structure is like a tree with branches,
and does form a large connected component. However, some
D. Phase Congruency Detection parts of the tree are not high enough in terms of magnitude to
The use of fourier phase patterns for marking features be picked as vessel part. We resort to Hysteresis threshold for
has significant advantages over gradient-based methods. Since restoring the connectivity of the vessel tree during binarization
phase is a dimensionless quantity invariant to changes in process. Hysteresis is kind of inertia (lagging of an effect).
image brightness or contrast, it can be used more effectively This threshold process involves two threshold values- TL
for bringing out hidden edges and lines. The research work (a low threshold) and TH (a high threshold). Pixels below
presented here takes a look at an aspect of phase patterns TL constitute to background and pixels above TH form the
known as phase symmetry for linear feature detection. It has vessel part. The pixels between low and high threshold are
been observed that the axis of symmetry corresponds to the then checked for connectivity with the already designated
linear feature central points (like central-line of the vessel ) vessel part. If found connected, then they are also declared
where all the frequency components are at either the minimum vessels, otherwise background. Because of its effectiveness,
or maximum points in their respective periodicity cycles [17]. we adopted the hysteresis threshold approach for both the
Using two quadrature filters allows us to calculate the phase high- and low-contrast vessel parts in the research reported
of the signal for a particular scale at a given spatial location. here..
F. Vessel Linking and Morphological Cleaning
The low-contrast vessel map shows gaps at some places. To
fill it we adopted the following algorithm. First,we compute
the endpoints of vessel segments, then label the segments and
get a list of all the points in the segment. Then a list of the
distances is created from each endpoint to all other endpoints (a) (b) (c) (d)
that are not in the same segment, and pick the endpoint with
the smallest distance. The smallest distance is compared with
some specified allowable distance (in our case it is 15 pixels),
a line is erected into the binary image connecting those two
endpoints. We need to repeat for every endpoint. We also
added additional criteria that the angle of the ends needs to be
about the same angle as the two endpoints. After the linking (e) (f) (g) (h)
process, we implemented the morphological cleaning process, Fig. 4. Comparison of proposed method with Nguyen et al. [ [5]] and
which removes all small unconnected components less than a Modified Frangi. First column show input image, second shows the output
specified area. of Nguyen et al. [ [5]], third shows the output of modified Frangi and the last
one shows the output of the proposed method.
III. E XPERIMENTAL R ESULTS
A. Evaluation Criterion
The sensitivity of the proposed method is higher than both
Vessel segmentation algorithms are binary classifiers that Zhao et al. [ [18]] and Mendonca et al. [ [19]] on DRIVE
should distinguish between vessels and non-vessels. Their dataset. It is to be noted that the proposed method’s sensitivity
performance is evaluated by comparing their results with the on STARE database is greater than all existing state-of the art
“ground truth”. The ground truth is a set of images that are methods. Specificity(Sp)of the proposed method is at par with
annotated manually by human observers. The following four to the state-of-art methods on these two datasets. On STARE
parameters can be extracted by comparing the algorithm output dataset, proposed method’s specificity is higher than other
with the ground truth: existing methods while on DRIVE database it is slightly lower
1) True Positive (TP): when vessels are rightfully predicted then some of the state-of-the-art methods, showing promise of
to be vessels improvement for better noise suppression in DRIVE images.
2) False Negative (FN): when vessels are predicted to be Accuracy (AC) is the mean measurement more closer to the
non-vessels interpreted standards. The accuracy of the proposed method is
3) True Negative (TN): When non-vessels are correctly found comparable to the state-of-art methods on two datasets.
identified as non-vessels & On STARE dataset, proposed method’s accuracy is higher than
4) False Positive (FP): When non-vessels are denoted to be other methods while on STARE dataset it is slightly lower then
vessels that of Zhao et al.(IPACHI Method) [ [18]].
These parameters can be used to evaluate the quality of
the segmentation algorithm for the pixels inside the FOV as IV. C ONCLUSION
follows above:
In this paper, we investigate the use of phase symmetry de-
Se = T PT+F
P
N,
TN tector to detect low-contrast thin vessels. The detector does not
Sp = T N +F P ,
make any assumptions regarding vessel’s luminance profile.
Acc = T P +FT N
P +T N
+T N +F P , We have used the multi-scale filtering to suppress non-linear
Here, Se is the sensitivity, indicating to what extent the objects and then enhanced features of vessel as an input to the
classifier can identify original vessel pixels. The Sp is the phase symmetry detector stage. This led to a better sensitivity
specificity, which represents classifier discrimination power to of automated vessel segmentation algorithms, especially for
mark non-vessel pixels. The Acc is the overall accuracy score thin vessel data. Two public databases (DRIVE, STARE) of
of the segmentation algorithm, reflecting the ratio between all fundus images are used to quantify proposed algorithm in all
correctly classified pixels (vessels or non-vessels) to all the three dimensions called accuracy, sensitivity and specificity.
pixels present in an image. The proposed method’s sensitivity and accuracy is promising
in comparison to the existing state-of-the-art methods.
B. Comparison of performance measures with other algo-
rithms
ACKNOWLEDGMENT
Different approaches are selected for comparison, both
supervised and unsupervised. The results on the two datasets The authors would like to thank Effat University in Jeddah,
are shown in Table I. On DRIVE datebase, proposed method’s Saudi Arabia, for funding the research reported in this paper
sensitivity is 0.761 while on STARE its sensitivity is 0.757. through the Research and Consultancy Institute.
TABLE I
P ERFORMANCE A NALYSIS OF S EGMENTATION M ODEL

DRIV E ST ARE
Methods Se Sp AC Se Sp AC
Supervised Methods
Steal et al. [ [20]] - - 0.946 - - 0.951
Soares et al. [ [21]] - - 0.946 - - 0.948
Lupascu et al. [ [22]] 0.720 - 0.959 - - -
Orlando et al. [ [23]] 0.785 0.967 - - - 0.951
Liskowski [ [24]] - - 0.949 - - 0.949
Unsupervised Methods
Palomera-Perez et al. [ [25]] 0.66 0.961 0.922 0.779 0.940 0.924
Nguyen et al. [ [5]] - - 0.940 - - 0.932
Xiaoxia Yin et al. [ [26]] - - 0.947 - - -
Zhao et al. (IPACHI Method) [ [18]] 0.742 0.982 0.954 0.780 0.978 0.956
Soomro et al. [ [27]] 0.713 0.968 0.941 0.711 0.965 0.942
Khan et al. [ [28]] 0.734 0.967 0.951 0.736 0.971 0.95
Proposed Method 0.757 0.966 0.923 0.761 0.972 0.9537

R EFERENCES in 2017 IEEE International Conference on Industrial and Information


Systems (ICIIS). IEEE, 2017, pp. 1–6.
[15] Z. Farbman, R. Fattal, D. Lischinski, and R. Szeliski, “Edge-preserving
[1] M. M. Fraz, P. Remagnino, A. Hoppe, B. Uyyanonvara, A. R. Rudnicka,
decompositions for multi-scale tone and detail manipulation,” in ACM
C. G. Owen, and S. A. Barman, “Blood vessel segmentation method-
Transactions on Graphics (TOG), vol. 27, no. 3. ACM, 2008, p. 67.
ologies in retinal images–a survey,” Computer methods and programs
[16] P. Kovesi, “Image features from phase congruency. videre: A journal of
in biomedicine, vol. 108, no. 1, pp. 407–433, 2012.
computer vision research (vol. 1),” 1999.
[2] H. R. Taylor and J. E. Keeffe, “World blindness: a 21st century [17] P. Kovesi et al., “Symmetry and asymmetry from local phase,” in Tenth
perspective,” British journal of ophthalmology, vol. 85, no. 3, pp. 261– Australian joint conference on artificial intelligence, vol. 190. Citeseer,
266, 2001. 1997, pp. 2–4.
[3] C. Sinthanayothin, J. F. Boyce, H. L. Cook, and T. H. Williamson, [18] Y. Zhao, L. Rada, K. Chen, S. P. Harding, and Y. Zheng, “Automated
“Automated localisation of the optic disc, fovea, and retinal blood vessels vessel segmentation using infinite perimeter active contour model with
from digital colour fundus images,” British journal of ophthalmology, hybrid region information with application to retinal images,” IEEE
vol. 83, no. 8, pp. 902–910, 1999. Transactions on Medical Imaging, vol. 34, no. 9, pp. 1797–1807., 2015.
[4] A. F. Frangi, W. J. Niessen, K. L. Vincken, and M. A. Viergever, [19] A. M. Mendonca and A. Campilho, “Segmentation of retinal blood
“Multiscale vessel enhancement filtering,” in International conference on vessels by combining the detection of centerlines and morphological
medical image computing and computer-assisted intervention. Springer, reconstruction,” IEEE Transactions on Medical Imaging, vol. 25, no. 9,
1998, pp. 130–137. 2006.
[5] U. T. Nguyen, A. Bhuiyan, L. A. Park, and K. Ramamohanarao, “An [20] J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, and B. V.
effective retinal blood vessel segmentation method using multi-scale line Ginneken, “Ridge based vessel segmentation in color images of the
detection,” Pattern recognition, vol. 46, no. 3, pp. 703–715, 2013. retina,” IEEE Transactions on Medical Imaging, vol. 23, no. 4, pp. 501–
[6] Y. Hou, “Automatic segmentation of retinal blood vessels based on 509, April 2004.
improved multiscale line detection,” Journal of Computing Science and [21] J. V. B. Soares, J. J. G. L. Roberto, M. Cesar, J. H. F. Jelinek, and
Engineering, vol. 8, no. 2, pp. 119–128, 2014. M. J. Cree, “Retinal vessel segmentation using the 2-d gabor wavelet
[7] T. Jerman, F. Pernuš, B. Likar, and Ž. Špiclin, “Enhancement of vascular and supervised classification,” IEEE Transactions on Medical Imaging,
structures in 3d and 2d angiographic images,” IEEE transactions on vol. 25, no. 9, pp. 1214–1222., 2006.
medical imaging, vol. 35, no. 9, pp. 2107–2118, 2016. [22] C. A. Lupas, D. Tegolo, and E. Trucco, “Fabc: Retinal vessel segmenta-
[8] M. Zhang, F. Chen, B. Tian, and D. Liang, “Using a phase-congruency- tion using adaboost,” IEEE Transactions on Information Technology In
based detector for glacial lake segmentation in high-temporal resolution Biomedicine, vol. 14, no. 5, pp. 1267–1274, 2010.
sentinel-1a/1b data,” IEEE Journal of Selected Topics in Applied Earth [23] J. I. Orlando and M. Blaschko, “Learning fully-connected crfs for blood
Observations and Remote Sensing, 2019. vessel segmentation in retinal images,” Medical Image Computing and
[9] O. M. Rijal, H. Ebrahimian, N. M. Noor, A. Hussin, A. Yunus, and A. A. Computer Assisted Intervention (MICCAI),, vol. 17, pp. 634–641, 2014.
Mahayiddin, “Application of phase congruency for discriminating some [24] P. Liskowski and K. Krawiec, “Segmenting retinal blood vessels with
lung diseases using chest radiograph,” Computational and mathematical deep neural networks,” IEEE Transactions on Medical Imaging, vol. 35,
methods in medicine, vol. 2015, 2015. no. 11, pp. 2369–2380, 2016.
[10] A. Shcherbinin, K. Kolchin, I. Glazistov, and M. Rychagov, “Sharp- [25] M. A. Palomera-Perez, M. E. Martinez-Perez, H. Benitez-Perez, and J. L.
ening image details using local phase congruency analysis,” Electronic Ortega-Arjona, “Parallel multiscale feature extraction and region grow-
Imaging, vol. 2018, no. 13, pp. 1–5, 2018. ing: Application in retinal blood vessel detection,” IEEE Transactions
[11] M. A. Shafiq, Y. Alaudah, G. AlRegib, and M. Deriche, “Phase congru- on Information Technology in Biomedicine, vol. 14, no. 2, pp. 500–506,
ency for image understanding with applications in computational seismic 2010.
interpretation,” in 2017 IEEE International Conference on Acoustics, [26] X. Yin, B. W.-H. Ng, J. He, Y. Zhang, and D. Abbott, “Accurate
Speech and Signal Processing (ICASSP). IEEE, 2017, pp. 1587–1591. image analysis of the retina using hessian matrix and binarisation of
[12] J. Meier, R. Bock, G. Michelson, L. G. Nyúl, and J. Hornegger, “Effects thresholded entropy with application of texture mapping,” PLOS ONE,
of preprocessing eye fundus images on appearance based glaucoma vol. 9, no. 4, pp. 1–17, 2014.
classification,” in International Conference on Computer Analysis of [27] T. A. Soomro, M. A. U. Khan, J. Gao, T. M. Khan, M. Paul, and N. Mir,
Images and Patterns. Springer, 2007, pp. 165–172. “Automatic retinal vessel extraction algorithm,” in 2016 International
Conference on Digital Image Computing: Techniques and Applications
[13] F. A. Hashim, N. M. Salem, and A. F. Seddik, “Preprocessing of color
(DICTA), Nov 2016, pp. 1–8.
retinal fundus images,” in 2013 Second International Japan-Egypt Con-
[28] M. A. U. Khan, T. A. Soomro, T. M. Khan, D. G. Bailey, J. Gao,
ference on Electronics, Communications and Computers (JEC-ECC).
and N. Mir, “Automatic retinal vessel extraction algorithm based on
IEEE, 2013, pp. 190–193.
contrast-sensitive schemes,” in 2016 International Conference on Image
[14] A. Bandara, K. K. Rajarata, and P. Giragama, “Super-efficient spatially
and Vision Computing New Zealand (IVCNZ), Nov 2016, pp. 1–5.
adaptive contrast enhancement algorithm for superficial vein imaging,”

You might also like