International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
A SURVEY OF AUTOMATED TECHNIQUES
IDENTIFICATION IN DIABETIC RETINOPATHY.
199
FOR
RETINAL
DISEASE
Kade Mahesh k
BMIT,Solapur,India
maheshkade@gmail.com
Abstract :Computational methodologies have become a
significant part of the real time applications. One specific
application which highly depends on the computing
techniques is the medical field. Ophthalmology is a
significant branch of biomedical field which requires
computer-aided automated techniques for pathology
identification in human eyes. These automated techniques
must be highly accurate and also converge in quick time
period. The focus of this paper is on understanding the
procedure of automated techniques involved in retinal
disease identification in diabetic retinopathy and why it is
significant .This paper
focus on
importance of
understanding the motivation behind the retinopathy , its
intricacies and the role of technology involved in it . It also
gives information of available public data set on which
algorithms are tested and trained .This paper also briefly
touches the imaging method involved for acquiring the
fundus ,the gold standard rule for data set and imaging,This
paper also provides the performance parameter used for
analysis of algorithm and this paper will also highlight the
work done by various groups of various university , lastly
it gives the list of chronological work done in this field for
detecting
microanerusym, exudates ,optic disc ,blood
vessels. All this will help the aspirant researcher immensely.
Key words: fundus,lesion,
1 Introduction :
the outer retina it is by the choroidal circulation, a thick
vascular layer lying outside of the retinal pigment
epithelium (the outer blood-retinal barrier). The retinal
pigment epithelium and the choroid are critical to retinal
function. If you imagine that your eye is like a camera, then
the retina is the film. When rays of light enter the eye and
are focused on the retina by the cornea and lens, the retina
reacts. The light receptive nerve cells that comprise the
retina generate a nerve impulse whenever they are exposed
to light. The retina then sends these nerve impulses along
the optic nerve to the brain which interprets them as a
picture. It's rather like the film in the camera being
developed so that pictures can be viewed. However, just like
a picture cannot be developed if the camera has defective
film, vision is not possible in an eye with a defective retina.
IJOART
Diabetic retinopathy and retinal structure : Diabetic
retinopathy (DR) is the commonest complication of diabetes
and is one of the leading causes of blindness . Recent
research has given a better understanding of the disease
processes and is opening up new avenues for prevention and
treatment. Very effective treatments are available and are
optimally used when retinopathy is detected early, well
before the patient is aware of symptoms. For this reason
screening programs for early detection of retinopathy are an
essential part of diabetes care and are in widespread use.
There is potential for automated image analysis to manage
the large volume of images generated from screening the
ever-increasing number of people with diabetes. The retina
is a transparent layer of vascularized neural tissue lining the
inner layer of the back wall of the eye, between the retinal
pigment epithelium on the outer and the vitreous on the
inner side. The retina captures photons and converts these to
photochemical and electrical energy, integrates the signals,
and transmits the resultant signal to the visual cortex of the
brain via the optic nerve, tracts, and radiations. The retinal
architecture is lamellar. Within this there are major cell
types performing sensory, nutritional, regulatory,
immunomodulatory, and structural functions. The retina is
uniquely partitioned from the vascular system by the bloodretinal barrier and blood-aqueous barrier. The blood supply
is dual; to the inner retina it is by the retinal circulation lying
within the inner retina (the inner blood-retinal barrier) and to
Copyright © 2013 SciResPub.
Fig(a) Retinal structure
Fig(b) Retinal Structure Detail
1.1 Charaterestics of lesion in Diabetic
Retinopathy
Diabetic retinopathy is one of the complications
caused by diabetes. As indicated by the name, diabetic
retinopathy appears in the retina, which is the tissue
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
responsible for vision in the eye. Since diabetic retinopathy
causes changes in the eye, the disease may affect the vision.
In Diabetic retinopathy the blood vessel becomes weak and
due to this vessel leaks blood and fluid of lipoproteins this
creates abnormalaties in retina.There may exist different
kinds of abnormal lesions caused by diabetic retinopathy in
a diabetic’s eye.
The lesion types are briefly
described:microaneurysm, hard exudate, soft exudate,
hemorrhage, and neovascularization. Microaneurysms are
the earliest clinically detectable lesions , and thus it is
important to have a method for detecting microaneurysms so
that incipient retinopathy may be noticed in its early stages.
Microaneurysms are local distensions of retina
capillaryThese outpouchings of capillary walls can be seen
as tiny, round, red dots]. .The image is magnified,but it is
still difficult to distinguish the microaneurysm due to its
small size. When microaneurysms are coated with blood,
they may be indistinguishable from dot hemorrhages].Thus,
the term “red small dot” is used to cover both
microaneurysms
and
small
dot
hemorrhages.
Microaneurysms are not necessarily permanent changes, but
they may first appear and then disappear during some period
of time Intra-retinal hemorrhages appear when capillaries or
microaneurysms rupture and some blood leaks out of the
vessels. Hemorrhages can be seen as red, dot-blot or
flameshaped regions A fundus with multiple hemorrhages is
shown in F.Hard exudates are accumulated lipid formations
leaked from weakened vessels Hard exudate lesions are
waxy and yellow with relatively clear edges Hard exudates
often appear in clusters or rings represents a fundus where
multiple hard exudates appear as bright lesions. Soft
exudates, also called cotton wool spots or micro-infarctions,
appear when terminal retinal arterioles are obstructed. Soft
exudates are small, whitish lesions with blurry edges shows
a magnified fundus image where a soft exudates lesion is
marked with an arrow. As seen in the figure, soft exudates
are not usually as visible as hard exudates. Extensive lack of
oxygen caused by obstructions may lead to development of
new blood vessels that are weak and can therefore easily
tear . The disease where new blood vessels appear is called
neovascularization. Neovascularization is the most serious
abnormality type in diabetic retinopathy since profuse
bleeding may produce loss of vision shows a fundus with
neovascularization. New small vessels have appeared on the
optic disk that is marked with an arrow in the image. The
other arrow shows a hemorrhage lesion where blood has
leaked from a new blood vessel.
200
Fig(e) Hard exudates
Fig(f)soft exudates
Fig(g) Neovascularisatio
IJOART
Fig(h) Macula
Fig(i) Fovea
Fig(c)
Microanerusym
Fig(j)Retinal detachment
Fig(d)
Hemorrhages
Copyright © 2013 SciResPub.
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
Fig(l)Thickness
201
Fig(m)
Fig(k)choroid
The image processing algorithms is used to detect the lesions in retina and various lesions are microanerusym,.hard exudates,soft
,exudates,neovascularisation, retinal nerve fiber, macula degeneration, retinal thicknening ,fovea location , fliuid under retina,
retina detachment, blood vessel structure. The fig from c to l shows sample images with different lesion.
2 Imaging methods The size and facts of our eye and retina are described in next section to understand intricacy of
Imaging and the resolution of camera required to acquire the image. This will help us to visualize how the automated system
developing is complicated and why so research goes on ,This will also allow us to understand how actually the retina construction
it is and how the pathology development can affect the change in the size ,shape,thickness,and dimension of retina . Thus we
therefore focus how the imaging systems of fundus ,evolves to capture the various sections of retina and how it is complicated to
develop to see all the parameters of retina in one image itself . Retinal images are acquired by a specialized camera called fundus
camera. Mydriatic and non-mydriatic fundus cameras are used for retinal photography. The imaging modalaties are fundus film based
photography, color fundus digital photography,stereo photography,hperspectral photography, Scanning laser opthalmoscopy, color
Doppler imaging, computed tomography , ophthalmic ultrasound, retinal thickness analyzer and scanning laser polarimetry ,2D To
3D(by composing several axial scans and several OCT images) and now 4D and 5D just to see more and more details of retinaand
understand the pathology more clearly. No single method of imaging captures all the features of retina .Table I shows the
summary of imaging methods employed and their suitability for various abnormalities.
IJOART
TableI
Sr
No
1
Imaging Methods
Suitable for Abnormalaties in diabetic retinopathy
Fundus photography
To detect and evaluate symptoms of retinal detachment or eye diseases such as glaucoma, agerelated maculopathy ,optic disk.
Is used to record the condition of macula, tiny fovea, measuring only 500 microns across, the retinal
vasculature, and the optic nerve head (optic disc is 1.5mm in diameter.) and to record color images
of the condition of the interior surface of the eye, in order to document the presence of disorders and
monitor their change over time.
For macular edema,optic disk (glaucoma) Stereo fundus photography permits clinical examination
of the patient's pathology beyond the ordinary two-dimensional view of a conventional photograph.
viewer's brain recreates the three-dimensional view
2
Color
photography
fundus
3
Stereo
photography
fundus
4
Scanning
laser
ophthalmoscopy
(SLO
Hyperspectral
imaging
In the diagnosis of glaucoma, macular degeneration, and other retinal disorders . it gives sharp
images of of Retinal Pigment Epithelium (RPE) cells.
6
Adaptive optics SLO
Greater degree of accuracy for eye tracking.
7
Fluorescein
angiography
Optical
Coherence
Tomography Imaging
To detect vessel inflammation.
5
8
Intra-retinal microvascular abnormalities (IRMA) and neovascularization,retinal oxymetry , facilitate
localization of retinal and disc pathology.
High-resolution, cross-sectional images of the retina, retinal nerve fiber layer ,the optic nerve head.
Copyright © 2013 SciResPub.
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
202
9
Time-domain OCT,
or time-of-flight OCT
Age-related Macular Degeneration (AMD), central serous chorioretinopathy,macular hole, vitreomacular interface syndrome and diabetic maculopathy.
10
Spectral-domain
OCT
Three-Dimensional
OCT Imaging
Choroidal thickness in , age-related macular degeneration
Longer Wavelength
OCT Imaging
More image depth in retinal layers
11
12
Retinal detachment , intrar etinal layers in macula odema ,Fovea ,,r etinal thickness and thicknesses
of the nerve fiber layer
Fig(n) normal and Disease Retina
IJOART
The list of available Public data set is STARE, DRIVE,
DIARETDB1, DIARETDB0, MESSIDOR, Retinopathy
online Challenge, CMIF, REVIEWdb, JRD, ONHSEG,
DIARETSPECDB1, RIM-ONE, ARIA, INSPIRE. The
dataset can be downloaded
2.1 Facts of Retina .
is 1.5 mm (Polyak, 1941) ie .1.2-1.5 mm (Ahnelt and Kolb,
unpublished data).Cross diameter of central rod free area
is.400-600 µm (Polyak, 1941) ,750 µm (Hendrickson and
Youdelis, 1984) ,570 µm (Yamada, 1969) ,250 µm (Ahnelt
et al., 1987).Vertical thickness of the fovea from ILM to
ELM.In the foveal pit is 150 µm (Yamada, 1969) ,foveal
rim is 300 µm.Length of foveal axons (Henle fibers) is 150300 µm (Ahnelt and Pflug, 1986).Highest density of cones
at center of the fovea (counted in a 50 x 50 µm square)
is147,000/mm2 (Osterberg, 1935) 96,900-281,000/mm2
mean161,900/mm2 (Curcio et al., 1987).Total number of
cones in fovea.Approximately is 200,000. There are 17,500
cones/degree2. Rod free area is approximately 1o thus there
are 17,500 cones in the central rod-free fovea , 178,000238,000/mm2 (Ahnelt et al., 1987).Total number of cones in
the retinais 6,400,000 (Osterberg, 1935).. Total number of
rods in the retina is110,000,000 to 125,000,000 (Osterberg,
1935).Rod distribution ie Rods peak in density is 18o or
5mm out from the center of the fovea in a ring around the
fovea at 160,000 rods/mm2. No rods in central 200 µm.
Average 80-100,000 rods/mm2 are present .Rod acuity peak
is at 5.2o or 1.5 mm from foveal center where there are
100,000 rods/mm2 (Mariani et al.,1984)..Number of axons in
the optic nerve is .564776-1,140,030 (Bruesch and Arey,
1942) ,800,000-1,000,000 (Polyak, 1941) ,1,200,000
(Quigley et al., 1982; Balaszi et al., 1984).Number of cones
to ganglion cells in the fovea is1 cone to 2 ganglion cells out
to about 2.2o (Schein, 1988).. Number of cones/retinal
pigment epithelial cell (RPE) is 30 cones/RPE in fovea
(Rapaport et al., 1995).Number of rods/retinal epithelial cell
(RPE).In periphery is 22 rods/RPE cell. The above data is
useful to understand and decide the resolution of camera.
The resolution of camera should be high to capture the detail
of fundus. , for eg if resolution is 10 micometer per pixel
and if the size of retina thickness is 100 micrometer then the
number of pixel for the thickness is 10 pixel , so the
algorithm should correctly identify this 10 pixel. Therefore
the resolution of fundus camera is important and depending
on the which parameter is to be detected the minimum
resolution of camera is decided.
The eyeball is about 1 inch (2.5 cm) in diameter. The adult
human eye weighs approximately 7.5 grams and measures
approximately 24.5 millimeters in its anterior-to-posterior
diameter. The entrance pupil is typically about 4 mm in
diameter . Size of the retina30-40 mm from ora to ora along
the horizontal meridian . Area of the human retina is 1094
square mm (Bernstein, personal communication), calculated
from the expectations that the average dimension of the
human eye is 22 mm from anterior to posterior poles, and
that 72% of the inside of the globe is retina . Size of optic
nerve head or disc.1.86 x 1.75 mm. One degree of visual
angle is equal to 288 µm on the retina without correction for
shrinkage . The fovea (arrow) is the center most part of the
macula. This tiny area is responsible for our central, sharpest
vision . Unlike the peripheral retina, it has no blood vessels.
Instead, it has a very high concentration of cones
(photoreceptors responsible for color vision), allowing us to
appreciate color.The fovea is the absolute center of vision.
The macula is actually an area where the central vision is
and the fovea is a point . One degree of visual angle is equal
to 288 µm on the retina without correction for shrinkage
(Drasdo and Fowler (1974).Foveal position is 11.8o or or 3.4
mm temporal to the optic disk edge.Cross diameter of the
macula is .3 mm of intense pigmentation, surrounded by 1
mm wide zone of less pigmentation (Polyak, 1941).Cross
diameter of the central fovea from foveal rim to foveal rim
Copyright © 2013 SciResPub.
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
.
Fig(m)
2.3 Image and performance analysis standard.
The image is to taken as per standards defined by Early
Treatment Diabetic Retinopathy Study ETDRS research
group or gold standard rules The gold standard data is
generated by a group of experts working in the field of
retinal image analysis and clinicians from the cooperated
ophthalmology clinics . The ground truths are recorded by
multi experts and then we need to develop the algorithm to
detect the the lesion on train images and the algorithm is
tested on the test images. The results are compared whether
the algorithms correctly identifies the disease pixels or not
and how many correct pixel are identified and how many
wrong pixels are detected are analyzed in terms of
sensitivity and selectivity and ROC plot.
The internationally recognized gold standard test for DR is
seven-field stereo 30° fundus photography on color
transparency film, performed by certified photographers (the
original work was done using Zeiss Oberkochen FF-3 and
FF-4 fundus cameras giving images with 2.5:1
magnification on Kodachrome film ,certain other cameras
and films are approved by the Fundus Photograph Reading
Center in Madison, Wisconsin with seven-field stereo
photograph grading performed by certified graders using the
set of Standard Photographs (SPs) obtainable from the
Fundus Photograph Reading Center in adison,Wisconsin
(http://eyephoto.ophth.wiscfudu/Photography/Protocols/Mo
d7&FA-ver1.pdf) and special grading forms; for further
details
see theETDRS manual of operations.Retinal
photography is a widely utilized method of screening. The
United Kingdom Screening Committee has put digital
retinal photography as the preferred modality for newly
established screening programs .Seven standard field
stereoscopic color photography as defined by the ETDRS is
the gold standard for detecting and classifying diabetic
retinopathy . It serves asthe benchmark to which other
screening techniques can be compared. In the United
Kingdom a Minimum Camera Specification for screening
has been proposed. Nonmydriatic digital cameras are
preferred even when mydriasis is routinely undertaken. The
recommendations are that Image file formats should not
result in the loss of any clinically significant information.
Camera resolution of the original images, as output by the
camera, should be a minimum of 20 pixels per degree of
retinal image, both horizontally and vertically.. The field of
view, as permanently recorded, should be a minimum of 45°
horizontally and 40° vertically.Images must be reviewed by
a skilled reader. For viewing the image The reference
standard remains Kodachrome color transparencies, viewed
203
with optimum sharpness and contrast with the right color
temperature of retro illumination. Today, images are usually
graded using office-type LCD monitors that are not colorcalibrated and whose color fidelity is dependent on angle of
view. One study showed an LCD monitor to be inferior to a
CRT monitor for the purpose of DR diagnosis . Fundus
camera such as Canon CR5 nonmydriatic, Topcon
TRCNW6, , Zeiss FF450 are employed to capture images.
A valid automated analysis system will need to perform
above the 80% sensitivity and 95% specificity set by the
World Health Organization for diabetic retinopathy.. An
image must be of sufficient quality,in focus, and have
adequate contrast. Sensitivity is the proportion of disease
positive individuals who are detected by the test, also known
as the true positive fraction (TPF). An ideal test has a
sensitivity of 100% (TPF = 1). Specificity is the proportion
of disease free people correctly labeled as disease free. An
ideal test has a specificity of 100% meaning that the false
positive fraction (FPF) is zero and (1-FPF) = 1.The “Exeter
Standards” for sensitivity and specificity were set in 1984 at
a meeting organized by Diabetes UK (then the British
Diabetic Association) and later accepted by the UK National
Screening Framework A perfect test will predict disease
perfectly, with PPV = 1 and NPV = 1.( positive predictive
value (PPV) is the probability that the subject is diseased if
the test is positive. The predictive value of a negative test,
negative predictive value (NPV), is the probability that the
subject is nondiseased when the test is negative. The
receiver operating characteristic (ROC) curve is a useful
way of describing the performance of medical tests whose
results are not simply positive or negative,but are measured
on continuous or ordinal scales. The area under the ROC
curve (AUC) is frequently reported. A perfect test has the
value AUC = 1:0 and an uninformative test has AUC = 0:5,
while most tests have values falling in between. The AUC is
interpreted by as an average TPF, averaged uniformly over
the whole range of false positive fractions.
IJOART
Copyright © 2013 SciResPub.
3 Automated system
The literature on
automatic retinal image analysis
algorithms are classified according to the following broad
categories: 1) Pre-processing (illumination correction and
contrast enhancement) 2) Segmentation of different features
(blood vessels and OD)
3) Detection of lesions
(microaneurysms, exudates, haemorrhages and cotton wool
spots) Pre-processing is an essential step in retinal image
analysis which attenuates image variation by normalizing
the original image with a reference model. It helps in
reducing the intra image as well as inter image variability.
Non uniform illumination results shading artifact and
degrades the efficiency of image analysis. Adaptive contrast
enhancement overcomes this problem. Wang et al., uses
point transformation for intensity correction. Histogram
equalization reduces inter image variability segmentation
of different features A number of algorithms are available
for the detection of different features such as blood vessels,
OD and macula. Detection of lesions microaneurysms,
exudates,haemorrhages.This
paper
focus
on
Microaneruysm,Exudates, optic disk and blood vessel
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
3.1 Introduction of Microanerusym
Microaneurysms (MA) are small saccular bulges in the
walls of the retinal capillary vessels.As the retinal capillaries
are too small to be resolved by normal fundus photography,
microaneurysms are classically described as distinct small
round objects, ranging from 10 mm to 100 mm in diameter.
In color fundus images they appear red and in red-free
images they appear dark. With fluorescein angiography
microaneurysms are hyperfluorescent, thus appear as bright
dots in positive images. Microaneurysms are better
visualized with fluorescein angiography with more MAs
visible than in color fundus images But As fluorescein
angiography poses an unacceptable health risk for eyescreening and color digital camera technology has advanced
sufficiently, there is now a move to perform automated MA
detection on color fundus images as an adjunct to automated
eye screening. Accurate microaneurysm counts are of
considerable practical importance at present because of their
proven prognostic value—this work,however, is tedious;
automated detection should be welcomed if reliable The
earliest clinically recognizable characteristics of DR are
microaneurysms. Retinal microaneurysms are focal
dilatations of retinal capillaries. They are about 10 to 100
microns in diameter and appear as small, round, dark red
dots on the retinal surface.
3.1.1 Survey on microanerusym detection .
204
vessels. Further improvements to the Aberdeen
automatedMAdetector were made by Cree et al.. The
region-growing algorithm was made computationally
efficient, and the classifier was redesigned. Extra features
using intensities of candidate MAs normalized intra- and
inter-image were added as was the peak response of the
matched filter that marked an object as a candidate . Cree et
al. also reported an automated method to detect the fovea
based on Fourier-space correlation of a model of the gross
shading of the fovea against a low-passed filtered version of
the retinal image. Thus the whole process of MA detection
and counting, including locating the region-ofinterest
centered on the fovea, was automated. the research group
established by ,Hipwell et al. of Aberdeen University
modified the automated MA detector of Cree et al. to work
on red-free images of 1024_1024 pixel resolution of a 50°
FOV retina either centered on the macula or nasal of the
optic disc. Against a testing set of 62 images, containing a
total of 88 MAs that broadly reflected the level of
retinopathy expected in an eye-screening population, the
automated MA detector achieved 43% sensitivity with 0.11
false-positives per image for detecting and correctly locating
MAs Acha and Serrano] used the preprocessing of
Mendonc¸a et al and extended the classifier to include
features that are based on the correlation of various shape
filters with the candidates. On five angiographic images
with a mean of 100 MAs per image they achieve 82%
sensitivity with 3.3 false positives per image. Serrano et al.
further enhance the image processing strategy by using the
errors of linear prediction coefficients as a means of
adaptively thresholding the matched-filtered image to
generate seed-points for region-growing. The classifier
appears to be essentially the same as Acha and Serrano .On
11 angiographic images containing a total of 711 MAs they
report a sensitivity of 91% for detecting MAs at 7 falsepositives per image. Yang et al. report on the detection of
MAs in low-resolution color retinal images. The green-plane
of the color images is used for preprocessing and detection
of candidates A preliminary test on 46 images (of healthy
and diseased retinae) achieved 90% sensitivity at 80%
specificity for detecting presence of MAs in the diseased
retinal images. Lalonde et al. [24] use this MA detector in
their RetsoftPlus retinal analysis software. Niemeijer et al.]
find that the standard approach for segmenting candidates
fails to segment all true MAs, thus they develop a new
method to generate candidates that is based on pixel
classification. After shade-correcting the image, the pixel
intensities and first- and second-order Gaussian derivatives
at various scales are provided to a k-nearest neighbor
classifier that produces a probability map for each pixel to
be a small red lesion. This is thresholded and filtered to
remove vessel segments and objects that are too large,
leaving the candidates. They test the standard approach and
their new pixel-based approach for generating candidates
and find that both methods fail to detect 14% of true
microaneurysms, but when combined together into a hybrid
system that figure is reduced to 10%. When coupled with
the usual region-growing and final classification to detect
MAs, the hybrid system achieved 100% sensitivityat 87%
specificity for detecting the presence of retinopathy in the 50
training images (27 DR; 23 NDR) of 768_576 pixel
resolution. Fleming et al. showed that the use of their local
contrast normalization and the detection of candidates lying
IJOART
The first attempts at automated microaneurysm detection
date back to the early 1980s, and were on fluorescein
angiographic images. La¨y et al. described a morphological
approach to MA detection. Image resolution and dynamic
range were low (256_256 pixels with 100 gray levels),
however, by using an imaging microscope to digitize film
negatives, the digital image covered a small field-of-view of
the macula so that MAs were well resolved. As negatives
were used, the MAs appeared as dark dots, thus as localized
minima in the intensity image. A tophat transform was
applied to identify the local minima. The automated MA
detector achieved 58% sensitivity with 4.2 false detections
per image for detecting and locating MAs. This process was
extended by Baudoin et al. [9] with extra morphological
processing to remove false detections such as objects too
small to be MAs. The automated MA detector achieved 70%
sensitivity for detecting and locating MAs with 4.4 false
detections per image. Spencer et al. used matched-filtering
to detect MAs and to remove falsedetections on vessels in
angiographic images of the retina. Film negatives of the
macula were projected onto a screen and digitized with a
monochrome CCD camera at 512_512 pixel resolution .
Preprocessing included radiometric correction for the
illumination of the negatives, and subtractive shadecorrection to remove choroidal fluorescence. MAs were
identified by matched-filtering with a circularly symmetric
two-dimensional Gaussian (s = 1 pixel) and a mask of size
5_5 pixels . It was concluded that the digitized image
resolution was insufficient for reliable MA detection
researchers at Aberdeen University, Scotland, developed the
image processing strategy for detection of MAs into what
we shall call the “standard approach” of automated MA
detection In the standard approach, a matched-filter with a
circularly symmetric 2-D Gaussian as the kernel is used to
highlight MAs, but this process tends to false-detect on
Copyright © 2013 SciResPub.
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
on vessel improves sensitivity by about 6% and specificity
by 8% for detecting retinal images containing MAs on a
database of 720 good quality retinal images, of which 216
contained MAs. Fleming et al. [28] compared a number of
local contrast normalization schemes and found that their
proposed scheme performed best and gave a useful
improvement to microaneurysm detection. The system
achieved a sensitivity of 85.4% at 83.1% specificity for
detecting images that containMAs The problem that global
shade-correction does not normalize the contrast of
microaneurysms sufficiently has also been described by
Huang and Yan [30]. They improve local contrast by first
applying shade-correction coupled with a nonlinear contrast
enhancing transform to local 120_120 pixel blocks of the
image (that slightly overlap with one another). Candidate
microaneurysms (small circular objects) are found with a
tophat transform, and the area and maximum intensity of the
candidates must be greater than locally determined
thresholds for the candidate to be accepted as a
microaneurysm. Gardner et al. detected lesions and features
other than microaneurysms, their work is notable in that
they break the image down into small windows and use the
pixel data of the window (after preprocessing with median
filter noise suppression) as well as the result of the Sobel
edge operator as inputs to a back propagation neural
network. This approach is based on the philosophy that the
neural network should be able to recognize the
interrelationships between pixels that enable lesions to be
detected
The output of image processing operators
that aid segmentation, such as matchedfilters and
morphological operators, can be provided as input to the
neural network, thus potentially giving the neural network
better information to work with. This can be taken to the
extreme in which the standard approach can be used to
segment candidate microaneurysms, and then use a neural
network to classify the candidates as microaneurysms or
spurious objects. This is the approach taken by Frame et al.
final classification stage of the microaneurysm detector was
tested with various classifiers including a neural network.
The manually constructed rule-based classifier, as used by
Cree et al. Quellec et al. [38] use a wavelet approach to
detect microaneurysms. Out of three wavelets tested they
find the Haar wavelet performed best with 88% sensitivity at
96% specificity for detecting microaneurysms in retinal
images. Their database included 995 images (1280_1008
pixel resolution) including both angiographic images and
photographs.
Most
publications
on
automated
microaneurysm detection use monochromatic data, even if
full color information is available. The green part of the
spectrum contains the most useful information for
discriminating microaneurysms and vascular structure, but
whether there is any useful extra information in the red and
blue parts of the spectrum is an open question. It is
interesting to note that Olson et al. [18] reported improved
sensitivity and specificity for manual screening for DR with
color slides compared to manual screening with red-free
images, It is normally said that in retinal images, the green
plane contains the best detail; the red plane, while brighter
and sometimes saturated, has poorer contrast; and the blue
plane is dark and typically not much better than noise .
Hemoglobin has an absorption peak in the green part of the
spectrum, thus features containing hemoglobin (e.g.,
microaneurysms) absorb more green light than surrounding
205
tissues, and appear dark and well contrasted in the green part
of the spectrum. The red light penetrates deeper into the
layers of the inner eye and is primarily reflected in the
choroid, explaining the reddish appearance of fundus
images. Because red light has a lower absorption coefficient
in the tissues of the eye than green, the red part of the
spectrum is less contrasted. It also depends on the melanin
concentration in the choroid. A comparison between gray
world normalization, color histogram equalization and color
histogram specification to a standard image was undertaken
by Goatman et al. Interestingly they found histogram
specification performed best in the sense that the class
separateness between exudates, cotton wool spots,
hemorrhage, and blot hemorrhages in the chromaticity
components was best with this method. The automated
microaneurysm detector, developed by the authors and coworkers at the University of Waikato], is a modification of
the standard approach to work on color retinal images. The
processing of the images for microaneurysms is as follows:
The green plane only of the retinal image is used to segment
candidate microaneurysms. It is divisive shade-corrected
and then lightly median filtered for noise-suppression. The
tophat transform is followed by a morphological
reconstruction to improve the removal of vessels from the
image. A circularly symmetric 2-D Gaussian is used as the
template for matched-filtering to highlight microaneurysms.
An adaptive threshold then isolates candidates from the
match-filtered image, and region-growing on the shadecorrected green plane at the positions of
candidates is used to isolate the morphology of the
underlying candidate. The microaneurysm detector was
trained on 80 retinal images (60 DR; 20 NDR), captured at
50_ field-of-view with a Topcon fundus camera and a Nikon
D1X 6 megapixel digital camera,.moreever the issue of
image quality and compression also affects the detection
rate.
IJOART
Copyright © 2013 SciResPub.
Table II
Sr Year
No
1
2013
Summary of Microa nerusym detection
Author/researcher Method
Lazar i
2
3
2012
2012
Agurto C
Lochan
4
2012
Antal B
5
2012
Lahmiri S
6
2012
Lahmiri S
7
2012
Narsimha K
8
2012
Hatnaka
Local rotating cross
section profile
Multiscale
Feature set of color size,
edge strenghth texture
Adaptive base ensemble
approach
DWTand
Emperical
mode
decomposition
with SVM and Radial
basis function kernel
(polynomial )
Statistical features (T
statistic
,entropy,bhattacharya
statistic)
PCA + SVM with
polynomials + radial
Basis Function
SVM
+
Bayseian
classifier
Double ring and Feature
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
9
2012
San G
10
11
12
2011
2011
2011
Vidya sari
Wen Xua
Qu quan
13
14
15
16
17
18
2011
2011
2011
2010
2010
2010
quellec
Purvita
Giancardo Luca
Esmaelli
Istavan l
Antal b
19
20
2010
2010
Agurto c
Paripurna s
21
2010
Pouerreza
22
2009
Martins
23
2008
Pradhan S
24
2008
Quellec
25
2008
A Bhalerao
26
2007
Grisan
28
29
2006
2005
Quellec
Pallawala
30
31
2004
2004
Raman B
Kahai
32
33
2003
2003
Goatman
Usher
34
36
2001
2000
Kamel
EgeA
37
2000
Hipwell
38
39
1999
1998
Mendoca
Hipwell
40
1996
Cree J
41
42
1996
1992
Gardner
Lay
Analysis
Constrained Maximally
stable external region
method
Filter based
SVM
Extraction of connection
components
Optimal filter
Math morphology
Radon transform basedf
New curvelet Transform
Local maxima map
Multi level ensemble
based
Multiscale AM FM
Fractal dimension in
spatil frequency domain
Value I of each pixel is
computed
Radon + morphological
reconstruction
Us
of
vessel
segmentation
for
posterior analysis of MA
Seed
generation
&
hybrid Classification
Optimal
wavelet
transform
Complex valued circular
symmetric filter
Local thresholding pixel
density
W Avelet transform
Generalized
Eigen
Vectors
Spatial resolution
Decision support
Binary
hypothesis
testing problem
ADDR
ANN
Recursive growing &
adaptive thresholing
NN
Mahalanobis classifier (
statistical measure of
similarity )
Shade
correcting
image+top
hat
transformation
as
distracters
+Gaussian
filter
Local intensity ,Contrast
Time
to
maximum
Intensity of fluroscence
Region
growing
&
classification
Back propagation NN
Thresholding
206
3.2 Introduction for Detection of Exudates
Among lesions caused by DR, exudates are one of the most
commonly occurring lesions. They are associated with
patches of vascular damage with leakage.The presence of
hard exudates within one disc diameter of the center of the
fovea is believed to be sight-threatening; this is called
macular involvement or diabetic maculopathy, but
confusingly sometimes referred to as “macular edema.” In
the first stage of the work, data goes through two
preprocessing steps.
3.2.1Survery of Exudates detection
Ege et al. located the optic disc, fovea, and four red and
yellow abnormalities (microaneurysms, hemorrhages,
exudates, and cotton wool spots) in 38 color fundus images,
which were previously graded by an ophthalmologist. The
abnormalities were detected using a combination of template
matching, region growing, and thresholding techniques.
87% of exudates were detected during the first stage while
the detection rate for the cotton wool spots was up to 95% in
terms of the lesion-based criterion. Following preliminary
detection of the abnormalities, a Bayesian classifier was
engaged to classify the yellow lesions into exudates, cotton
wool spots, and noise. The classification performance for
this stage was only 62% for exudates and 52% for the cotton
wool spots Wang et al. addressed the same problem by
using a minimum distance discriminant classifier to
categorize each pixel into yellow lesion (exudates, cotton
wool spots) or nonlesion (vessels and background) class.
The objective was to distinguish yellow lesions from red
lesions, therefore other yellowish lesions (e.g. cotton wool
spots) were incorrectly classified at the same time. The
image-based diagnostic accuracy of this approach was
reported as 100% sensitivity and 70% specificity. Gardner et
al. broke down the retinal images into small squares and
then presented them to a back propagation neural network.
After median smoothing, the photographed red-free images
with a field-of-view of 60_ were fed directly into a large
neural network (using 20_20 patches, with 400 inputs). This
technique recognized the blood vessels, exudates, and
hemorrhages. The neural network was trained for 5 days and
the lesion-based sensitivity of the exudate detection method
was 93.1%. Walter et al. identified exudates from the green
channel of retinal images, according to their gray-level
variation. After initial localization, the exudate contours
were subsequently determined by mathematical morphology
techniques. This approach had three parameters, the size of
the local window, which was used for calculation of the
pixel local variation, and two other threshold values. The
table III gives the list of work done in exudates detection
,list is not complete but is exhaustive.
IJOART
Copyright © 2013 SciResPub.
Table III Exudates detection summary
Sr
No
Year
Author/
Method used
Researcher
1
2013
Tamalirasi
Genetic based fuzzy
region growing
seeded
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
2
2012
3
Harngi b
Naives Bayesian classifier.
Hangi b
Active contour & regionwise
classification
207
24
2009
Xu lili
SVM and Wavelet, gray level
co occurance matrix, Gaussian
radial basis function
25
2008
Kande G
Specially weighted fuzzy C
means clusterring
26
2008
Bin m
Fuzzy morphology
27
2007
Sagar A
Dynamic
detection
28
2007
Garcia M
Feature Extraction ,multilayer
perceptron
Gaussian mixture model for
classification
29
2007
Gnsan E
Local thresholding & pixel
density
4
2012
Esmaeli M
Curvlet Transform
5
2012
Chen
Xiang
Histogram
equalization
,morphological
reconstrucor,SVM
6
2012
Zelkovich
Classification algorithm
7
2012
Aftab u
Filter bank for detection
thresholding
Edge
8
2012
Agurto
Multiscale approach
31
2004
Li huiqi
Model based approach
9
2012
Rochha
Point of interest and visual
dictionaries
32
2004
Zhiang X
For bright lesion fuzzy c mean
SVM
10
2012
Edgahi
Morphology
33
2001
Hsu W
Domain knowledge based
11
2011
Poostch H
Preprocessing,
enhancement
contrast
IJOART
3.3 Survey of optic disc detection
12
2011
Nagy E
Ensemble based
13
2011
Soares
14
2010
Kumara V
15
2010
Sae tang
Nonuniform
subtraction n
16
2010
Sanchez C
Combination of loca
contextual information
Scale
space
thresholding
curvature
Feature extractors
illumination
and
Improves detection
17
2010
YOUSEEF
New feature based
18
2010
Shivram J
Knowledge based frame work
19
2010
Jaffar h
Pure splitting technique
20
2010
Langaourdi
Morphology
21
2010
Agurto c
Multiscale AM FM
22
2010
Fang G
Boosted soft segmentation
23
2009
Wang h
Median
filtering
clustering
Copyright © 2013 SciResPub.
dynamic
The optic disc is the entrance and exit region of blood
vessels and optic nerves to the retina, and its localization
and segmentation is an important task in an automated
retinal image analysis system. Indeed, optic disc localization
is required as a prerequisite for the subsequent stages in
most algorithms applied for identification of the anatomical
structures and pathologies in retinal images Optic disc
localization is required as a prerequisite for the subsequent
stages in many algorithms applied for identification of the
anatomical-pathological structures in retinal images
Reliable optic disc localization is surprisingly difficult, due
to its highly variable appearance in retinal images. One
method is edge detection is followed by a circular Hough
transform to locate the optic disc. The algorithm
commenced with locating the optic disc candidate area..
The, the Sobel operator is used to detect the edge points of
the located candidate area. The contours were then detected
by means of the circular Hough transform, i.e., the gradient
of the image was calculated, and the best fitting circle was
then determined. These approaches are quite time
consuming and rely on certain conditions regarding the
shape of the optic disc that are not always met. Moreover,
edge detection algorithms often fail to provide an acceptable
solution due to the fuzzy boundaries, inconsistent image
contrast, or missing edge features.Li and Chutatape
proposed a method for locating the optic disc using a
combination of pixel clustering and principal component
analysis techniques. They first determined the optic disc
candidate regions by clustering the brightest pixels in
graylevel retinal images. This strategy can only work when
there is no abnormality in the image.Sinthanayothin used an
80_80 sub-image to evaluate the intensity variance of
adjacent pixels, and marking the point with the largest
variance as the optic disc location. This technique has been
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
shown to work well with 99.1% sensitivity and specificity
when there are no or only few small exudate pathologies
that also appear very bright and are also well contrasted. In
fact, the authors made an assumption to locate the optic disc
from those retinal images with no visible symptoms. This
work has also mainly focused on locating the optic disc
center. Lalonde et al. localized the optic disc using a
combination of two procedures including a Hausdorff-based
template matching technique on the edge map, guided by a
pyramidal decomposition technique. A priori information of
the image characteristics, e.g., right/left eye image and
whether the input image is centered on the macula or an
optic disc have to be provided. The edge maps were
calculated using Canny edge detection and therefore the low
and high hysteresis thresholds must be defined properly. The
authors reported an average error of 7.0% in locatig the disc
center and 80% area overlap between their ground truth
optic disc and the localized one against a database
comprising 40 retinal images. The optic disc boundary was
found in Mendels et al. using a twofold method. The optic
disc boundary is highly influenced by the strong blood
vessel edges crossing the optic disc region. Therefore, a
preprocessing step was first applied based on local minima
detection and morphological filtering to remove the
interfering blood vessels and locate the optic disc boundary
accurately. To do that, a gray-level dilation operator was
applied based on a 5_5 pixel structuring element. This was
followed by erosion using the same structuring element to
re-establish the optic disc contour to its original place. A
morphological reconstruction operator was then applied by
maintaining the maximum of the dilated/eroded image and
the original one. Then, the optic disc contour was located
using the deformable active contours (snakes) technique.
The snake was based on an external image field called
Gradient Vector Flow (GVF). The technique was tested
against a set of 9 retinal images and the author reported an
accurate optic disc boundary localization in all the test
images. To locate the optic disc in this study, two different
techniques are investigated including template matching and
active contours. The first strategy provides an approximate
location of the optic disc center. The optic disc can
sometimes have an elliptical shape, and therefore the
baseline plotting of its exact boundaries is necessary. To
achieve this goal a snake technique is applied to obtain a
more accurate discboundary. The Table IV Gives the list of
work done in detecting optic .
208
7
2012
Yin
8
9
2012
2012
Damon
Malik J
10
2012
Zhang
11
2012
Sekar G
12
13
2012
2012
Qureshi
Noronha K
14
2011
Cheng J
15
16
17
2011
2011
2011
LU ,S
Koukaunos D
Giachetti A
18
2011
Canas
19
2010
Vahabi
Sector based with intensity
& blood vessel priori
Using vessel kinking
Region based active contour
model In variational Level
Set Formulation
Projection
with
vessel
distribution and appearance
characteristic
Clustering & Histogram
Technique
Using multiple features
Review on fundus image
analysis
Peripallary
atrophy
detection,
Edgedetectionn + constraint
Elliptical Hough Transfporm
Circular Transformation
Multiscale products
Use
inpaintedbackground
and vessel information
Use of shape based
intensity & Information
regarding vascular Structure
Wavelet as pre processing,
Sobel edge detector +
Texture analysis +Intensity
+ template matching
Adaptive
morphological
Operation
Back ground Estimation
Edge detection
Identification of round shape
Use of regional information
,
Vessel bend based cup
Modified
gradient vector
flow
Morphological ,edge +
feature extraction
Mixture model bases
Priori knowledge shape
Level set segmentation
K NN classifier
Using projection of image
features
Level set method
Genetic algorithm
IJOART
TableIV Summary of optic disk detection
Sr Year Author
Method
no
1
2013 Ching J
Super pixel classification
2
2013 Lajao I
Local rotating cross section
profile analysis (For MA)
3
2012 M muthu
Intuitioitic Fuzzy histon
4
2012 Hemanth D
Hybrid clustering model
5
2012 Agurtus C
Am-fm
208epresentation
granulometry used (contrast
enhancing method ,partial
least square used for final
classification
6
2012 Yu H
Directional Matched filter &
Level Sets
Copyright © 2013 SciResPub.
20
2010
Daniel
21
22
2010
2010
LU,Shijan
Chaichana
23
2010
Joshi G
24
2010
Huiyu
25
2010
Aquino
26
27
28
29
30
2010
2010
2009
2009
2009
Tan
Lee
Echigaray
Meindert
Mahfouz
31
32
2008
2008
Liu J
Enrique
33
34
2008
2008
Eswaran
Kande
35
2008
Niemeijier
36
2008
Youssif
Marker controlled watershed
Local variance,geometrical
active
contour,spatially
weigted fuzzy c mean
clusterring
Integration of local vessel
geometry & Image intensity
feature (fusion).
Use ofK NN classifier
Vessels direction matched
filter
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
37
38
2007
2007
Virance
Deformable model based
approach using gradient
vector floe Snake
K
nearest
Neighbors
Classifier.
Pixel based classification,
Intensity,edge,Gaussian
filter,gabor filter
Abramoff
39
2007
Juan
40
2007
Anantha
41
2007
Juan
42
2007
Sun
43
2007
Santhi
44
2007
Huajun
45
2007
Niemeijerm
46
2006
Kim
47
48
2006
2006
Xu.Juan
Yang
49
2004
50
51
2004
2004
Li
Chutatape
Lowell
Foracchi
52
53
2003
2003
Hoover
Thitiporn
54
2003
Huiqui
55
56
2002
2002
Osarech
Radim
57
2001
Lalonde
58
1999
sinthanyothin
Deformable
model
technique modified active
contour model.
Knowledge based clustering
Eigen vector based(eigen
vector
calculated
from
covariance matrix & image
constructed from eigen
vector
Modified active contour
Update equation of snake
model
Warping
technique
&
RANSAC
Multilevel thresholding
Morphological for object
roundness & For Circle
detection
Circle
fitting
Method
Histogram based
Concept of fractal analysis
Fitting a single point
distribution model
Hybrid parabolic & markow
Model
Modified snake technique
Multiscal
region
&
boundary hybrid snake
method
Model based (use of PCA)
209
59
1999
Francois
60
1999
Chaudhari
61
62
63
64
65
1999
1999
1995
1988
1982
Lee
Mendib
Kalvianian
Tamura
Akita kuga
vessel
Active contour Gradient
Vector Flow method
Cluster of high intensity
Pixel
Active contour model
Hough Transform
Hough Transform
Variance
of
intensity
between OD &adjacent
vessel
3.4 Importance of vascular detection.
There is considerable interest in the relationship between
retinal vascular morphology and metabolic and
inflammatory diseases . Several researchers have measured
blood vessel diameter, for the purposes of assessing retinal
blood flow or variations of vascular diameter (venous
beading). There have been attempts to detect venous loops,
intraretinal microvascular abnormalities (IRMA) that predict
the onset of proliferative DR, and small tufts of new blood
vessels (proliferations) .Parr, Hubbard, and their associates
developed formulae to generate summarized measures of
retinal arteriolar (central retinal arteriolar equivalent
[CRAE]) and venular (central retinal venular equivalent
[CRVE]) diameters, as well as their dimensionless quotient
(arteriovenous ratio [AVR]) The association of retinal
vascular changes with blood pressure, is strong, graded, and
consistently seen in both adult . In the Beaver Dam Eye
Study, each 10 mmHg increase in mean arterial blood
pressure was associated with a 6 mm (or 3%) decrease in
retinal arteriolar diameter, even after adjusting for age,
gender, diabetes, smoking, and other vascular risk factors
.The retinal and cerebral vasculature share similar
embryological origin, anatomical features, and physiological
properties This concept provides strong biological rationale
for the use of retinal image analysis to indirectly study the
cerebral microvasculature and related diseases. ( ref :Beaver
Dam, Rotterdam, and the Blue Mountains Eye studies)
.Retinopathy and nephropathy are two well known
microvascular complications of Diabetes An earlier study of
individuals with hypertension demonstrated a strong
relationship of retinopathy grades with microalbuminuria, a
preclinical marker of renal Dysfunction Atherosclerosis, the
key pathophysiological disease process of CHD and stroke,
has been the focus of interest in studies of retinal vascular
measurements . Retinal vascular changes in diabetes,
hypertension,and other systemic disease processes may be
related to inflammation and or endothelial dysfunction..Thus
it becomes clear from various study that the vascular
structure detection is important in diabetic retinopathy and
the measurement of the parameter of vessel such as the
width ,thickness,vurtosis,junction of vessel are of most
important in early diagonosis of cardiovascular disease.
3.4.1 Survey on vascular detection
IJOART
&
Template matching based
Geometrical
parameter
model(directional
pattern
converge on OD)
Fuzzy convergence
Snake active contour ,
Local
entropy
Threshold(VD)
Active shape model
(region growing &edge
detection
Modified ASM(use of PCA)
Improved gradient vector
flow
Active contour model
Nonlinear filtering (filter +
Edge detector)
Pyramidal decomposition &
hausdroff based template
matching
Variance
of
intensity
between OD &adjacent
Copyright © 2013 SciResPub.
Fredric et al proposed the morphological operations based
vessel segmentation technique. This technique also
incorporated the advantages of differential operators.
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
Subhasis et al have used the matched filters to detect the
vascular network. The properties of optical and spatial
properties of the retinal images are used in this work. But
the requirement for huge convergence time period is the
major drawback of this technique. Riccardo et al have used
the Gaussian kernels based filtering approach for retinal
vessel segmentation. Hysteresis combined with thresholding
is used in this approach after the filtering technique. The
convergence rate of the proposed approach is very high but
the accuracy reported in the paper is very low. Chutatape et
al have proposed a hybrid approach for vascular network
detection. The advantages of tracking and filters are
combined in this technique. Selection of seed pixel for the
tracking algorithm is the major drawback of this method.
Edge thinning combined with Sobel operators are used for
vessel detection by Yiming et al (1998). The proposed
method also incorporated the concept of local thresholding.
Experimental results revealed the high convergence rate of
the proposed approach. The application of vessel detection
for image registration is explored by Zana et al (1999). The
Hough transform with Bayesian approach is used for vessel
segmentation in this approach. Ali et al (1999) have used the
tracing method to detect the vascular network of the retinal
images. This approach is based on the recursive procedure
to determine even the finest blood vessel. This technique is
also applicable for images with discontinued blood vessels.
But the selection of seed pixel for the tracing procedure is
the practical difficulty of this approach. Cristian et al (2000)
have used the watershed technique for blood vessel
segmentation. The selection of initial seed points for
watershed segmentation is the major disadvantage of this
approach. Huiqi et al (2000) have used the edge detection
technique to extract the blood vessels. A threshold is further
used to enhance the blood vessels. This technique is
implemented on green channel of the input RGB image.
Jorge et al (2001) have used the concept of wavelet
transform to extract the blood vessels Morphological
operations are also used in this work to extract the thin
blood vessels. Experimental results suggested the superior
nature of the proposed approach. Gaussian filter approach is
used for retinal vessel detection by Luo et al (2002). The
vessel width measurement is incorporated in this technique
which yields superior results than the matched filter
approach. Entropy thresholding based vessel detection
algorithm is demonstrated by Thitiporn et al (2003). The
concept of filtering and thresholding is used in this work.
The usage of non-adaptive thresholding technique is the
major drawback of this approach. Thitiporn et al (2003)
have used the local entropy thresholding technique for
efficient blood vessel detection. The concept of matched
filtering and length filtering is also involved in this
approach. But the proper selection of threshold value is the
major drawback of this approach. Model based vessel
segmentation technique is illustrated by Li et al (2003). The
Gaussian model is used in this work for feature extraction.
The proposed approach is inexpensive and highly robust.
Ridge based vessel segmentation technique is proposed by
Joes et al (2004). The image ridges coincide with the vessel
centerlines and this concept is used for segmentation in this
work. A comparative analysis with the conventional
methods is also reported in this work. Sequential forward
feature selection is used in this work. A spatial referencing
algorithm for vasculature extraction is reported by Gang et
210
al (2004). This technique is based on tracing methodology to
extract the features from the retinal blood vessels. This
method is then compared with other techniques such as
random scheduling to show the superior nature of the
proposed algorithm. A hybrid approach with Laplacian
operators and thresholding for retinal blood vessel detection
is proposed by Vermeer et al (2004). Experimental results
are analyzed in terms of sensitivity and specificity. This
method is capable of detecting blood vessels even in images
with specular reflection. A comparative study of retinal
vessel segmentation techniques are reported by Niemeijer et
al (2004). Five different algorithms are analyzed in this
approach. The conclusion of this report is that the
performance of the pixel classification methods is better
than the other techniques. Cree et al (2005) have used the
model and tracking based approach for retinal vessel
segmentation. The model aided in accurate tracking of the
blood vessels. But the selection of initial seed point is the
drawback of this approach. Morphology and matched filters
are collectively used for vessel segmentation by Dietrich et
al (2005). This technique is used to detect the glaucoma eye
disease. Comparative analysis and future scope of the work
is also discussed in this work. Andrew et al (2005) have
used the tramline filtering technique for retinal vessel
segmentation. The proposed approach is a nonlinear
approach which is highly robust in nature. The usage of
blood vessel extraction technique for Diabetic Retinopathy
detection is demonstrated by Cornforth et al (2005). The
concept of wavelet transforms is used in this work for
segmentation. But this approach is not applicable for images
with noisy background. Bevilacqua et al (2005) have used
the combined morphology and filtering techniques for
retinal feature extraction. Emphasis is laid on cluster filter to
extract the blood vessels. Skeleton process and median
filtering is used to enhance the blood vessels. Experimental
results have suggested the superior nature of the proposed
algorithm. A survey of different retinal image segmentation
techniques are reported by Mai et al (2006). The merits and
demerits of various segmentation techniques are analyzed in
detail in this report. This work also suggested suitable
techniques for various applications. Matched filter approach
based retinal vessel extraction technique is proposed by
Michal et al (2006). The experimental results are analyzed
in terms of confidence and edge measures. But the low
contrast vessels are not detected by this approach. Morlet
wavelet transform based retinal vessel segmentation is
performed by Joao et al (2006). Features are constituted
from the wavelet transform and pixel based classification is
further performed to segment the blood vessels. But addition
of noise to the image degrades the quality of the output. The
hybrid approach of multiscale analysis and adaptive
thresholding is used by Qin et al (2006). Gabor filter
approach is used in this work for segmentation. Small and
thin blood vessels are also detected by this method. Ana et
al (2006) have used the morphological operations for the
segmentation of retinal blood vessels. Changhua et al (2007)
have used the probabilistic methodology for retinal vessel
segmentation. The advantage of this approach is that it is
independent of vessel constraints. This approach is also
capable of enhancing the vessel junctions. Elena et al (2007)
have used the multiscale feature extraction principle for
retinal vessel segmentation. The main advantage of this
approach is that it is able to detect the blood vessels with
IJOART
Copyright © 2013 SciResPub.
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
different widths, lengths and orientations. Support Vector
Machine (SVM) based retinal blood vessel segmentation is
proposed by Elisa et al (2007). The results are compared
with the line detector technique which employed the concept
of thresholding. Incapability of thin vessel detection is the
demerit of this approach. The lack of a proper performance
measure is another drawback of this approach. Joao et al
(2007) have used the wavelet features and the supervised
classifier to segment the blood vessels. The classifier used in
this work is the Gaussian model classifier. Experimental
results are also compared with the conventional K-NN
classifier. But the proposed approach outperform the K-NN
classifier in terms of computation time period and accuracy.
The advantages of using structural features are demonstrated
by Harihar et al (2007). The structural features used in this
work are extracted from the vascular network. The structural
features are then used as input for the SVM classifier to
distinguish the various types of blood vessels. But the
system failed in case of thin blood vessels. Saurabh et al
(2007) have used the unsupervised curvature based retinal
vessel segmentation technique. The region growing method
is used for extracting the blood vessels. The random
selection of initial seed point is the major drawback of this
approach. Salem et al (2007) used the nearest neighbor
clustering algorithm for blood vessel segmentation. The
features used for this clustering approach are based on
gradient magnitude and Eigen values. The results are
compared with the conventional K-NN classifier to show the
superior nature of the proposed method. A modified
matched filter approach is demonstrated by Mohammed et al
(2007). Modifications are performed in the fixation of filter
parameters which yielded enhanced results. A comparative
analysis with the conventional technique is also provided in
this approach. The application of Genetic Algorithm (GA)
for retinal blood vessel segmentation is explored by
Mohammed et al (2007). GA is used for optimizing the
various parameters of the matched filter. Experimental
results suggested the superior nature of the optimized filter
over the un-optimized filter. Unsupervised segmentation of
retinal blood vessels is proposed by Nancy et al (2007). The
unsupervised technique used in this work is the clustering
algorithm combined with the Hessian matrix. Experimental
results are compared with the piecewise threshold probing
method to show the superior nature of the proposed
approach. Alauddin et al (2007) have used the unsupervised
texture classification technique for blood vessel
segmentation. Color and contrast are used as textural
features in this work. FCM algorithm is then used for
segmenting the blood vessels. Filtering approach is
demonstrated for blood vessel segmentation by Changhua et
al (2007). The Eigen vectors are extracted from
these images and used for vessel segmentation.The proposed
approach is tested on DRIVE database to evaluate the
segmentation algorithm. Yong et al (2008) have used a
hybrid method for retinal blood vessel extraction. The
concept of morphology and fuzzy clustering algorithms are
used in this work. But few post-processing procedures are
required to preserve the weak edges. Divergence of the
vector fields based vessel segmentation algorithm is
implemented by Benson et al (2008). Laplacian operators
are used in this work for segmentation. But spherical shaped
vessels are not detected by this approach. Deformable
contour model based retinal vessel segmentation is proposed
211
by Espona et al (2008). The topological properties of the
blood vessels are incorporated in this approach. The
performance measure used in this work is arteriovenous
index which is essential for pathology classification.
Supervised linear classifier based retinal image
segmentation is performed by Alexandru et al (2009).
Feature vectors obtained from the blood vessels are used as
inputs for the linear classifier. Experimental results
suggested the superior nature of the proposed approach in
terms of convergence rate. Muhammed et al (2009)
combined the Ant Colony (ACO) approach with the
matched filter for blood vessel detection. The parameters are
optimized using ACO and a comparative analysis is
performed with the un-optimized network to show the
superior nature of the proposed approach. Retinal vessel
graph based vascular network detection is implemented by
Bashir et al (2009). The algorithm used in this work is a
junction resolution algorithm which forms the complete
graph of the blood vessels. But the low accuracy due to
many training error is the major drawback of this work.
Multi scale quadrature filtering based retinal blood vessel
detection is proposed by Gunnar et al (2009). This approach
combined the concept of both line and edge detection. The
experimental results claim that the proposed approach is
highly robust in nature. Line tracking based retinal vessel
segmentation is implemented by Marios et al (2009). This
method is also used on images with Gaussian noise and salt
& pepper noise. The major drawback of the proposed
algorithm is the high misclassification rate of the optic disk.
Fourier cross-sectional profile is formed for vessel detection
by Tao (2010). An extensive analysis is performed by
changing the nature of the profiles such as Gaussian
profiles. This technique also preserves the information
available in the blood vessels. Bob et al (2010) have used
the concept of matched filter combined with the first order
derivative of Gaussian filter for retinal vessel extraction. A
threshold is also used for post processing in this work. This
technique is also implemented on pathological retinal
images. But this technique is not suitable for noisy images
The Gabor wavelet shows itself efficient in enhancing vessel
contrast while filtering out noise, giving better performance
than the matched filter of Chaudhuri et al. . Information
from wavelet responses at different scales is combined
through the supervised classification framework, allowing
proper segmentation of vessels of various widths. Of the
three classifiers tested, the LMSE gave the worst
performance, but with fast training and classification phases.
The kNN classifier showed good performance, but with a
slow classification phase, complicating its use in interactive
applications. Finally, the GMM classifier has a demanding
training process, but guarantees a fast classification phase
and good performance, similar to that of the kNN. Feature
generation consists of computing a series of wavelet
coefficients and can be implemented efficiently, resulting in
an approach that can be included in an interactive tool. An
open-source prototype of an interactive software for vessel
segmentation — based on a graphical user interface that
assists the process of classifier training and vessel
segmentation. Vessel Segmentation Methods and extraction
techniques can also be classified into Pixel Methods ,
Edge Detection Methods, Exploratory Algorithms, Ridge
Detection, Morphological Methods Wavelets, The Haar
Transform, The Daubechies Transform, Adaptive
IJOART
Copyright © 2013 SciResPub.
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
Thresholding, Genetic algorithm,ensemble
based
classification .table below give the entire list of methods
used in vessel detection etc.
212
Fundus )
23
2011
Yavuz
Gabor filter & top hat
tranform
24
2011
Fraz
Line
,multiscale,gaborr
,morphological
25
2011
Selvathi d
Gabor wavelet & kernel
classifiers
26
2011
Marin deigo
New supervised
invariants based
27
2010
Salazar
Graph cut
28
2010
Paripurna S
Fractal dimension in spatial
freq domain
29
2010
Akram M
Thresholding probing
TableV Blood Vessel detection summary
1
2013
Lau PQ
Vessel segment graph
2
2013
Songhuo X
GA + FCM clustering
3
2013
Shanmugam
Extreme learning machine
approach
4
2012
Tang y
Merging shape ,region edge
information
5
2-12
Rattanapod
Multilevel line detection
6
2012
Chien chenhg
Counterlet transform
7
2012
Rouchdy
Geodesic voting methods
8
2012
Fraz
Ensemble classification
9
2012
Sumathy b
Morphological structur &
entropy thresholding
10
2012
Vora R A
2012
Kundu a
moment
Wavelet
30
2010
Peng Q
Radial
projection
supervised classification
&
IJOART
Wavelet energy entropy
Kekre wavelet feature
Euclidian distance
11
strength
31
2010
Polomera P
Parallel multiscale feature
extraction & region growing
32
2010
Lu shijian
Iterative
polynomial
smoothing , bilateral filter
33
2010
Youj
2D matche d filter _ gabor
filter,Otsu
_P
tile
thresholding
+
Morphological angular Scale
space
12
2012
Holbura C
13
2012
Lazar Istavan
Directional height statistic
34
2010
Demir S
Multithresholding
14
2012
Gegundez
Quality evaluation o vessel
35
2010
Du Xiaojun
Mumford Shah model gabor
wavelet filter
15
2012
Cao
Patch based
36
2010
Adil Mouloud
16
2012
Ol vera w
Average of synthetic exact
filters & hessian matrix
Statistical
technique
37
2010
Pourezza R
Radon
transform
morphological
reconstruction
38
2010
Lupassa
Using ad boost
39
2010
Moju s
Gabor & local binary pattern
40
2010
Palomera p
Parallel multiscale feature
extraction & region growing
41
2010
Bhuiyan
Varying contrast & central
reflex properties
Supervised
decision fusion
classifiers
17
2012
Ahmed M B
Phase congruency
18
2012
Lin K
Anatomica realism
19
2012
Baishing
Geodesic time transform
20
2012
Vargaic
Committee of local expert
21
2012
Yu Honggang
Directional mathed filter &
level set
22
2012
Hu zhihang
Multimodal 9SD OCT +
Copyright © 2013 SciResPub.
base
tracking
&
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
42
2010
Shivram J
Knowledge based
43
2010
Moghimirad
Multiscale
function
_
213
67
2007
Li wang
Multiresolution
model
68
2007
Wu Chang
Hybrid filtering
Hermite
medialness
44
2010
Lam B
Multi concavity modeling
69
2007
Garg S C
Unsuoervised
based
45
2009
Chang Chin
Line operator
detector
70
2007
Zhang
Non linear projection
&
curvature
edge
46
2009
Rezatofighi
Polar run length features
71
2007
Mac gillivray
Fractal analysis
47
2009
Akram M
2D gabor wavelet
72
2006
Salem N
Scale space features & k
nearest neighbour
48
2009
Akram M
Wavelet Tx ,gabor wavelet
73
2004
Qing Z
49
2009
Hani a
ICA
50
2009
Parnama
Branches filtering
2D entropies of gray level
gradient
co
occurrence
matrix
51
2009
Amin m
Phase congruency
52
2008
Rezatofiglu
Contourlet
53
2008
Salem nancy
Single parmeter
,eigen
value of hessian matriix
54
2008
Kande G B
55
2008
Espana
56
2008
Kande G B
57
2008
Narasimha
Generalized dual Gaussian
model
58
2008
Yedidya T
Kalman Filter
59
2008
Fraz M
Center line & morphology
60
2008
Lam B
Divergence of vectors field
4 Few Global contribution on automated
Diabetic retinopathy system
Dr. Peli was an early developer of image processing of
retinal images in the early 1980s before the era of personal
computers he used a main frame computer to lead some of
the earliest research in this field. Dr. Peli applied image
enhancement techniques to improve visualization of the
retinal nerve fiber layer. Dr. Peli also developed techniques
to automatically detect Drusen in retinal images and by
aligning retinal images taken over time he demonstrated that
Drusen change over time, they may disappear, and new ones
may appear at other places. Dr. Peli also developed a
number of photographic techniques to improve the quality
of the retinal images to facilitate better image processing
and has designed and implemented digital image processing
techniques to improve the visibility of retinal images
through cataracts. Dr peli also worked on drusen
measurement from fundus photographs, Fast registration of
digital retinal images, Computer measurement of retinal
nerve fiber layer striations, Fundus image analysis using
mathematical morphology, Use of circularly polarized light
in
fundus
and
optic
disc
photograph.
Ref(http://www.eri.harvard.edu/faculty/peli/
Researcher
from uiowa have developed a set of quantitative analysis
tools for highly accurate assessment of retinal layers from
clinically available optical coherence tomography (OCT)
images . they have developed a comprehensive system for
detecting lesions including hemorrhages, infarcts, and
neovascularizations from these photographs.Faculty of
uiowa developed On-line Retinal Diagnosis a system for
telediagnosis of retinal disease using retinal cameras in
family care clinics across the Midwest.. This system is fully
based on open-source components, allowing family
physicians to diagnose patients with diabetes.. Development
of algorithms and systems for automated classification of
the optic nerve head from stereo photographs is being also
developed. They are developing a method for measuring
retinal responses to visual stimuli using infrared light
sources. uiowa have developed a novel method allowing
optimal identification of 2-D, 3-D, and 4-D surfaces.. They
IJOART
Histogram
Matching +
spatially weighted fuzzy C
menas
Deformable contour model
Local
relative
thresholding
entropy
Laplacian operators
62
2008
Basher al
Joining retinal vessel
63
2007
Martinez p
Insight segemnetation
registration tool kit
64
2007
Ricci E
Line operator &SVM
65
2007
Bhuiyan A
Unsupervised
classification
66
2007
Supot s
Fuzzy K median clustering
Copyright © 2013 SciResPub.
&
texture
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
also have INSPIRE data set - Iowa Normative Set for
Processing Images of the REtina - online set of 30 pairs of
de-identified stereo color images of the optic nerve head,
with reference standard, based on 3D OCT
http://www.biomed-imaging.uiowa.edu/research.html
Researcher
of Lincoln universit
have worked on
accurate methods for manually marking retinal vessel
widths, validating retinal fundus image analysis algorithms:
issues and a proposal, an active contour model for
segmenting and measuring retinal vessels, review - a
reference data set for retinal vessel profiles, joining retinal
vessel segments, a ribbon of twins for extracting vessel
boundaries, automated measurements of retinal bifurcations,
automated analysis of retinal vascular network connectivity,
manual measurement of retinal bifurcation features.Ref:
http://ulincoln.academia.edu/BashirAlDir .The work done
by
following
university
UCDAVIS,ABERDEEN
UNIVERSITY,
SIUE
,UCSB,,VARPA,CALTECH,
WATERLOO are also very important. For details reader
can refer there website.
5 Conclusion : This paper
try to cover the
understanding of the diabetic tetinopathy and the vision
system which affects due to the Diabetic retinopathy .
Therefore the need of automated screening and detection
system of various lesion , which may be occurring at the
early stages of Diabetic retinopathy , so that preventive
measures can be taken to prevent blindness. The work
done by various groups and university is also covered in
brief to make aware what is being done globally in this
field. . For details the reader can refer the website of the
university. This paper also give the
the theoretical
background of retina and its structure which is important to
understand because without this it is difficult to Develop the
algotithm for detecting the various parts and layers of retina
and to identify the change in structure which occurs due to
various pathology . Lastly the summary of survey for
detection in Microanerusym, Exudates, optic disk, vascular
structure is given in table and in chronological order. This
list do not exactly covers all the work but definitely it will
make reader to understand the methods and technique
employed for detection in various lesion and this will benefit
the new aspirant research scholar.
214
using marker-controlled water shed transformation’, Journal
of Medical Systems.
[5]. Akara Sopharak, Bunyarit Uyyanonvara, Sarah Barman,
Thomas H.Williamson (2008), ‘Automatic detection of
diabetic retinopathy exudates from non-dilated retinal
images using mathematical morphology methods’,
Computerized Medical Imaging and Graphics, Vol. 32, pp.
720-727.
[6]. Akara Sopharak, Bunyarit Uyyanovara and Sarah
Barman (2009), ‘Automatic Exudate detection from nondilated diabetic retinopathy retinal images using Fuzzy
Cmeans clustering’, Sensors, Vol. 9, pp. 2148 – 2161.
[7]. Alan D.F., Sam Philip, Keith A.G., John A.O. and Peter
F.S. (2006), ‘Automatedmicroaneurysm detection using
local contrast normalization and local vessel detection’,
IEEE Transactions on Medical Imaging,Vol. 25, No. 9, pp.
1223-1232.
[8]. Alauddin Bhuiyan, Baikunth Nath, Joselito Chua and
Ramamohanarao
Kotagiri
(2007),
‘Blood
vessel
segmentation from color retinal images using unsupervised
texture classification’, Proceedings of the International
Conference on Image Processing, pp.521-524.
[9]. Alexandru Paul Condurache, Alfred Mertins and Til
Aach (2009), ‘Supervised, hysteresis-based segmentation of
retinal images using the linear-classifier percentile’, Lecture
notes in Informatics, Vol. P154.
[10]. Aliaa A., Youssif A., Atef Z.Ghalwash and Amr
S.Ghoneim (2006), ‘Comparative study of contrast
enhancement and illumination equalization methods for
retinal vasculature segmentation’, Proceedings of
International Biomedical Engineering conference, pp. 1-5.
[11]. Alireza Osareh , Mirmehdi M., Thomas B., Markham
R. (2002), ‘Comparative exudates classification using
support vector machines and neural networks’, Proceedings
of the fifth International Conference on Medical Image
Computing and Computer-Assisted Intervention- part II, pp.
413-420.
[12]. Alireza Osareh , Mirmehdi M., Thomas B., Markham
R. (2003), ‘Automated identification of diabetic retinal
exudates in digital colour images’, British Journal of
Ophthalmology, Vol. 87, pp. 1220-1223.
[13]. Ana Maria Mendonca and Aurelio Campilho (2006),
‘Segmentation of Retinal Blood Vessels by Combining the
Detection of centerlines and Morphological Reconstruction’,
IEEE Transaction on Medical Imaging, Vol. 25, No. 9, pp.
1200- 1213.
[14]. Anantha Vidya Sagar, Balsubramanian S.,
Chandrasekaran V. (2007), ‘Automatic detection of
anatomical structures in digital fundus retinal images’,
Conference on Machine Vision Applications, pp. 483-486
[15]. Andrea Anzalone, Federico Bizzari, Mauro Parodi,
Marco Storace (2008), ‘A modular supervised algorithm for
vessel segmentation in red-free retinal images’,
Computers in Biology and Medicine, Vol. 38, pp. 913-922.
[16]. Andrew Hunter, James Lowell and David Steel (2005),
‘Tram-line filtering for retinal vessel segmentation’,
Proceedings of the 3rdEuropean Medical and Biological
Engineering Conference.
[17]. April Khademi and Sridhar Krishnan (2007), ‘Shiftinvariant Discrete WaveletTransform Analysis for Retinal
Image Classification’, Journal of Medical and Biological
Engineering and Computing,Vol. 45, No. 12, pp. 12111222.
IJOART
6 References
[1]. Abramoff M.D., Alward W.L.M., Greenlee E.C.,
Lesya Shuba, Kim C.Y., Fingert J.H. and Kwon Y.H.
(2007), ‘Automated Segmentation of the Optic Disc from
Stereo color Photographs Using Physiologically Plausible
Features’, Investigative Ophthalmology and Visual Science,
Vol.48, No.4, pp.1665-1673.
[2]. Acharya R., Wong L.Y., Ng E.Y.K. and Suri J.S.
(2007), ‘Automatic identification of anterior segment eye
abnormality’, ITBMRBM, Vol. 28, No.1, pp. 35-41.
[3]. Agostino Accardo P. and Stefano Pensiero (2003),
‘Neural network based system for early keratoconus
detection from corneal topography’, Journal of Biomedical
Informatics, Vol. 35, pp. 151-159.
[4]. Ahmed Wasif Reza, Eswaran C., Kaharudin Dimyati
(2010), ‘Diagonosis of diabetic retinopathy: Automatic
extraction of optic disc and exudates from retinal images
Copyright © 2013 SciResPub.
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
[18]. Arturo Aquino, Manuel Emilio Gegundez,Diego
Martin (2009), ‘ Automated optic disc detection in retinal
images of patients with diabetic retinopathy and risk of
macular edema’, World Academy of Science, Engineering
and technology, Vol. 60, pp.85-90.
[19]. Balanco M., Penedo M.G., Barreira N.,Penas M.,
Carreira M.J. (2006), ‘Localisation and extraction of the
optic
disc
using
the
fuzzy
circular
hough
transform’,Springer Lecture Notes in Computer Science, pp.
712-721.
[20]. Bashir Al-Diri , Andrew Hunter, David Steel, Maged
Habib (2009), ‘Automated analysis of retinal vascular
network connectivity’, Computerized Medical Imaging and
Graphics.
[21]. Benson Shu Yan Lam and Hong Yan (2008), ‘A Novel
Vessel Segmentation Algorithm for Pathological Retinal
Images Based on the Divergence of Vector Fields’,
Transaction on Medical Imaging, Vol. 27, No. 2, pp. 237246.
[22]. Bob Zhang , Lin Zhang , Lei Zhang , Fakhri Karray
(2010), ‘Retinal vessel extraction by matched filter with
first-order derivative of Gaussian’, Computers in Biology
and Medicine, Vol. 40, Issue 4, pp. 438-445.
[23]. Carla Agurto, Victor murray, Eduardo Barriga, Sergio
Murillo, Marios Pattichis, Herbert Davis, Stephen Russel,
Michael Abramoff, Peter Soliz (2010), ‘Multiscale AM-FM
methods for diabetic retinopathy lesion detection’, IEEE
Transactions on Medical Imaging, Vol. 29, No. 2, pp. 502512.
[24]. Cemal Kose, Ugur Sevik, Okyay Gencalioglu (2008),
‘Automatic segmentation
of age-related macular
degeneration in retinal fundus images’, Computers in
Biology and Medicine,Vol.38, pp. 611-619.
[25]. Cemal Kose, Ugur Sevik, Okyay Gencalioglu, Cevat
Ikibas and Temel Kayikicioglu (2008), ‘A statistical
segmentation method for measuring agerelated macular
degeneration in retinal fundus images’, Journal on Medical
Systems, Vol. 3, pp. 1-13.
[26]. Changhua Wu, Gady Agam (2007), ‘Probablistic
retinal vessel segmentation”,SPIE Medical Imaging-Image
processing’, Proceedings of the SPIE, Vol. 6512, pp.
651213.
[27]. Changhua Wu, Gady Agam, Peter Stanchev, (2007),
‘A general framework for vessel segmentation in retinal
images’, International Symposium on Computational
Intelligence in Robotics and Automation,pp. 37-42.
[28]. Chrastek R., Wolf M., Donath K., Niemann H., Paulus
D., Hothorn T., Lausen B.,Lammer R., Mardin C.Y.,
Michelson G. (2005), ‘Automated Segmentation of the
Optic Nerve Head for Diagnosis of Glaucoma’, Medical
Image Analysis, Vol.9,pp. 297-314.
[29]. Chutatape.O (1988), ‘Retinal Blood Vessel Detection
and Tracking by Matched Gaussian and Kalman Filters’,
Proceedings of the 20th Annual International Conference of
the IEEE Engineering in Medicine and Biology society, Vol.
20, No. 6, pp. 3144-3149.
[30]. Clara I.S., Roberto Hornero, Maria I.L., Mateo Aboy,
Jesus Poza, Daniel Abasolo (2008), ‘A novel automatic
image processing algorithm for detection of hard exudates
based on retinal image analysis’, Medical Engineering and
Physics, Vol. 30, pp. 350-357.
[31]. Conor Heneghan, John Flynn, Michael O Keefe, Mark
Cahill (2002), ‘Characterization of changes in blood vessel
215
width and tortuosity in retinopathy of prematuriy using
image analysis’, Medical Image Analysis, Vol. 6, pp. 407429.
[32]. Conor Heneghan, John Flynn, Michael O Keefe, Mark
Cahill (2002), ‘Characterization of changes in blood vessel
width and tortuosity in retinopathy of prematuriy using
image analysis’, Medical Image Analysis, Vol. 6, pp. 407429.
[33]. Cornforth D.J., Jelinek H.J., Leandro J.J.G., Soares
J.V.B., Cesar,Jr R.M., Cree M.J., Mitchell P., Bossomaier T.
(2005),‘Development of retinal blood vessel segmentation
methodology using wavelet transforms for assessment of
diabetic retinopathy’, Complexity International, pp. 50-60.
[34]. Cree M.J., Cornforth D., Jelinek H.F.(2005), ‘Vessel
segmentation and tracking using a two-dimensional model’,
Proceedings of the Image and Vision Computing
Conference, University of Otago, New Zealand, pp. 345350.
[35]. Cristian Perra, Maria Petrou and Giusto D.D. (2000),
‘Retinal Image Segmentation by Watersheds’, IAPR
Workshop on Machine Vision Applications, pp. 315-318.
[36]. Daniel Welfer, Jacob Schacanski, Cleyson M.K.,
Melissa M.D.P., Laura W.B.L., Diane Ruschel Marinho
(2010), ‘Segmentation of the optic disc in color eye fundus
images using an adaptive morphological approach’,Journal
on Computers in Biology and Medicine”, Vol. 40, pp. 124137.
[37]. Debrup Chakraborty and Nikhil R.P. (2004), ‘A neuro
fuzzy scheme for simultaneous feature selection and fuzzy
rule based classification’ IEEE Transactions on Neural
Networks, Vol. 15, No. 1, pp.110-123.
[38]. Dietrich Paulus and Serge Chastel and Tobias
Feldmann (2005), ‘Vessel segmentation in retinal images’,
Proceedings of SPIE, Vol. 5746, No.696.
[39]. Elena Martinez-Perez M., Alun D.H.,Simon A.T., Anil
A.B., Kim H.P. (2007), ‘Segmentation of blood vessels from
redfree and fluorescein retinal images’, Medical Image
Analysis, Vol. 11, pp. 47-61.
[40]. Elisa Ricci and Renzo Perfetti (2007), ‘Retinal Blood
Vessels Segmentation Using Line Operators and Support
Vector Classification’, IEEE Transaction on Medical
Imaging, Vol. 26, No.10, pp. 1357-1365.
[41]. Elizabeth M.Massey and James A.Lowell (2009),
‘Lesion boundary segmentation using level set methods’, In
Advances in Computer Graphics and Computer vision.
[42]. Enrique J. Carmona, Mariano Rincon, Julian GarciaFeijoo, Jose M.Martinez-de-la-Casa(2008), ‘Identification of
the optic nerve head with genetic algorithms’, Journal on
Artificial intelligence in medicine, Vol. 43, pp. 243-259.
[43]. Espona L., Carreira M.J., Penedo M.G.,Ortega M.
(2008), ‘Retinal vessel tree segmentation using a deformable
contour model’, IbPRIA07, pp. 178-185.
[44]. Farnell D.J.J., Hatfield F.N., Knox P.,Reakes M.,
Spencer S., Parry D., Harding S.P. (2008), ‘Enhancement of
blood vessels in digital fundus photographs via the
application of multiscale line operators’,Journal of the
Franklin Institute, pp. 1-18.
[45]Automated Image Detection of retinal pathology by
herbert f. jelinek michael j. cree .crc press
[46]J.anitha, c.kezi selva vijila and d.jude hemanth “an
overview of computational intelligence techniques for
retinal disease identification applications” International
Journal of Reviews in Computing © 2009
IJOART
Copyright © 2013 SciResPub.
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue5, May-2013
ISSN 2278-7763
[47] Crick, R.P. and Khaw, P.T., A textbook of clinical
ophthalmology, World Scientific, New Jersey, 3rd ed., 2003.
[48] Kohner, E.M., Sleightholm, M., and KROC
Collaborative Study Group, Does microaneurysm count
reflect
severity of
early diabetic
retinopathy?
Ophthalmology, 93, 586, 1986.
[49] Kohner, E.M., Stratton, I.M., Aldington, S.J., et al.,
Microaneurysms in the development of diabetic retinopathy
(UKPDS 42), Diabetologia, 42(9), 1107,1999.
[50] Klein, R., Meuer, S.M., Moss, S.E., et al., The
relationship of retinal microaneurysm counts to the 4-year
progresssion of diabetic retinopathy, Archives of
Ophthalmology, 107, 1780, 1989.
[51] Klein, R., Meuer, S.M., Moss, S.E., et al., Retinal
microaneurysm counts and 10-year progression of diabetic
retinopathy, Archives of Ophthalmology,113(11), 1386,
1995.
[52] Goatman, K.A., Cree, M.J., Olson, J.A., et al.,
Automated measurement of microaneurysm turnover,
Investigative Ophthalmology & Visual Science,44(12),
5335, 2003.
[53] Kohner, E.M. and Dollery, C.T., The rate of formation
and disapearance of microaneurysms in diabetic retinopathy,
Eur J Clin Invest, 1(3), 167, 1970.
[54] La¨y, B., Baudoin, C., and Klein, J.C., Automatic
detection of microaneurysms in retinopathy fluoroangiogram, Proceedings of the SPIE, 432, 165, 1983
[55] Baudoin, C.E., Lay, B.J., and Klein, J.C., Automatic
detection of microaneurysms in diabetic fluorescein
angiography, Rev. Epid´em. et. Sant´e Publ.32, 254, 1984.
[56] Spencer, T., Phillips, R.P., Sharp, P.F., et al.,
Automated detection and quantification of microaneurysms
in fluorescein angiograms, Graefes Arch Clin Exp
Ophthalmol, 230(1), 36, 1992.
[57] Spencer, T., Olson, J.A., McHardy, K.C., et al., An
image-processing strategy for the segmentation and
quantification of microaneurysms in fluorescein angiograms
of the ocular fundus, Computers and Biomedical Research,
29, 284,1996.
[58] Cree, M.J., Olson, J.A., McHardy, K.C., et al.,
Automated microaneurysm detection,in International
Conference
on
Image
Processing,
Lausanne,
Switzerland,1996, vol. 3, 699–702.
[59] Cree, M.J., Olson, J.A., McHardy, K.C., et al., A fully
automated comparative microaneurysm digital detection
system, Eye, 11, 622, 1997.
[60] Cree, M.J., Olson, J.A., McHardy, K.C., et al., The
preprocessing of retinal images for the detection of
fluorescein leakage, Physics in Medicine and Biology,44(1),
293, 1999.
[61] Frame, A.J., Undrill, P.E., Cree, M.J., et al., A
comparison of computer based classification methods
applied to the detection of microaneurysms in ophthalmic
fluorescein angiograms, Computers in Biology and
Medicine, 28(3),
225, 1998.
[62] Hipwell, J.H., Strachan, F., Olson, J.A., et al.,
Automated detection of microaneurysms in digital red-free
photographs: A diabetic retinopathy screening tool, Diabetic
Medicine, 17(8), 588, 2000
[63] MacMahon, S., Blood pressure and the risk of
cardiovascular disease, N Engl J Med, 342, 50, 2000.
216
[64] Stamler, J., Dyer, A.R., Shekelle, R.B., et al.,
Relationship of baseline major risk factors to coronary and
all-cause mortality, and to longevity: findings from longterm follow-up of Chicago cohorts, Cardiology, 82, 191,
1993.
[65] Wolf, P.A., D’Agostino, R.B., Kannel, W.B., et al.,
Cigarette smoking as a risk factor for stroke. the
Framingham study, JAMA, 259, 1025, 1988.
[66] Langer, R.D., Ganiats, T.G., and Barrett-Connor, E.,
Paradoxical survival of elderly men with high blood
pressure, BMJ, 298, 1356, 1989.
[67] Pekkanen, J., Tervahauta, M., Nissinen, A., et al., Does
the predictive value of baseline coronary risk factors change
over a 30-year follow-up? Cardiology,82, 181, 1993
[68] Wong, T.Y., Klein, R., Klein, B.E., et al., Retinal vessel
diameters and their associations with age and blood
pressure, Invest Ophthalmol Vis Sci, 44, 4644, 2003.
[69] Wong, T.Y., Hubbard, L.D., Klein, R., et al., Retinal
microvascular abnormalities
and blood pressure in older people: the Cardiovascular
Health Study, BrJ Ophthalmol, 86, 1007, 2002.
[70]Michael D. Abràmoff, Senior Member, IEEE, Mona K.
Garvin, Member, IEEE, and Milan Sonka, Retinal Imaging
and Image Analysis IEEE Trans Med Imaging. 2010
January 1; 3: 169–208.[76]Niall Patton,Tariq M. Aslam,
Thomas
MacGillivray
Retinal
image
analysis:
Concepts,applications and potential Progress in Retinal and
Eye Research 25 (2006) 99–127
[71]http://www.ncbi.nlm.nih.gov/books/NBK11530/
[72](http://www.eri.harvard.edu/faculty/peli/
[73]http://www.biomed-imaging.uiowa.edu/research.html
[74]http://ulincoln.academia.edu/BashirAlDir
IJOART
Copyright © 2013 SciResPub.
IJOART