21682194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/JBHI.2015.2396198, IEEE Journal of Biomedical and Health Informatics
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. XXX, NO. XXX, XXXXXXX XXXX 1
A Novel Curvature Based Algorithm for AutomaticGrading of Retinal Blood Vessel Tortuosity
Masoud AghamohamadianSharbaf, Hamid Reza Pourreza,
Senior Member
and Touka Banaee,
Abstract
—Tortuosity of retinal blood vessels is an improtantsymptom of diabetic retinopathy or retinopathy of prematurity.In this paper, we propose an automatic image based methodfor measuring single vessel and vessel network tortuosity of these vessels. Simplicity of the algorithm, low computationalburden and an excellent matching to the clinically perceivedtortuosity are the important features of the proposed algorithm.To measure tortuosity, we use curvature which is an indicatorof local inﬂection of a curve. For curvature calculation, templatedisk method is a common choice and has been utilized in mostof the state of the art. However, we show that this method doesnot possess linearity against curvature and by proposing twomodiﬁcations, we improve the method. We use the basic and themodiﬁed methods to measure tortuosity on a publicly availabledata bank and two data banks of our own. While interpreting theresults, we pursue three goals. First, to show that our algorithm ismore efﬁcient to implement than the state of the art. Second, toshow that our method possesses an excellent correlation withsubjective results (0.94 correlation for vessel tortuosity, 0.95correlation for vessel network tortuosity in diabetic retinopathyand 0.7 correlation for vessel network tortuosity in ROP). Third,to show that the tortuosity perceived by an expert and curvaturepossess a nonlinear relation.
Index Terms
—Curvature, Diabetic retinopathy, Retinal image,Tortuosity measure,.
I. I
NTRODUCTION
D
IABETES causes damage in blood vessels. Vessel damages in heart muscles are related to ischemic heart diseases and heart attacks; and vessel damages in the retina causereduction of sight. The latter called diabetic retinopathy (DR),is one of the common causes of reduction of sight. In fact,many studies show that DR dramatically increases chancesof blindness [1] and that it is the leading cause of blindnessof the working age especially in developed countries [2].Presence of numerous microaneurysms is the earliest sign of DR [3]. As the disorder develops, retinal blood vessels becomethicker, more twisted and turned [4]. In more advanced levels,neovascularization through inability to provide the requiredamount of nutrition and oxygen for the retina occurs [5].These newly generated vessels are very fragile. Therefore, theycause internal bleeding in the retina which endangers the visualsystem and might ultimately result in blindness.Early diagnosis of DR helps controlling its side effects.This is possible with the aid of technical examination of
Masoud AghamohamdianSharbaf and Hamid Reza Pourreza are withthe Department of Electrical Engineering, Ferdowsi University of Mashhad, Mashhad, Iran (emails: masoudas@hotmail.com; hpourreza@ieee.org).Hamid Reza Pourreza is the corresponding authorT. Banaee is with the Retina Research Centre, Mashhad University of Medical Sciences, Mashhad, Iran (email: banaeet@mums.ac.ir)Manuscript received ——, —; revised —–, —.
(a) Normal tortuosity
(b) High tortuosityFig. 1. Two images of the retina with different tortuosity level
the ocular fundus. Experimentally, it has been shown thatwhen there is no critical symptom of retinal damage, anescalation in retinal blood vessel tortuosity is an early sign of DR [4]. Moreover, presence of tortuous retinal blood vesselsis an indicator of retinopathy of prematurity (ROP) in preterminfants [6]. It is wellknown that in serious cases, ROP causesretinal detachment and blindness [7]. For a detailed reviewon applications of image processing to diagnosing ROP andcomparison of different methods see [8], [9].Qualitatively speaking, tortuosity is an indication of howwinding a blood vessel is [10]. In Fig. 1, two ocular fundusimages with normal and tortuous vessels are shown. As isclear, the vessel network in Fig. 1(b) is more twisted than anormal network which alerts probable presence of DR. Notethat vessel tortuosity can happen locally in small portions of blood vessels, or throughout the blood vessel network.For quantitative measurement of tortuosity, the vessel ismodeled as a smooth connected curve. Based on this model,different tortuosity measurement algorithms have been proposed in the literature that could be divided into four generalgroups:
A. Arc Length over Chord Length Ratio Methods
Methods of this group have simple mathematical expressions. Lotmar
et al.
[11] were the ﬁrst to introduce methodsof this category and their method was widely utilized thereafter(e.g. [12]–[15]). However, it is apparent that the arc over chordlength ratio, on its own, is insufﬁcient for determiniation of vessels with smooth curvature and vessels with variation incurvature direction. For compensation, Bullit
et al.
[16] andGrisan
et al.
[17] proposed modiﬁcations on the approach.In [17], vessels are partitioned into segments with the sameconvexity and a weighted sum of arc over chord length of allsegments is proposed as a tortuosity measure.
21682194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/JBHI.2015.2396198, IEEE Journal of Biomedical and Health Informatics
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. XXX, NO. XXX, XXXXXXX XXXX 2
B. Methods Based on Curvature
Curvature is a mathematical measure for how inﬂected acurve is at a certain coordinate. Hart
et al.
[13] use curvatureto propose two tortuosity measures which are the integral of curvature and the integral of curvature squared. For future reference, as in [17], we call these measures
τ
c
and
τ
sc
. Moreover,the ratio of these integrals over arch or chord length have alsobeen proposed as tortuosity measures in [13]. In [18], the integral of squared curvature derivative is suggested as a measureof tortuosity. These or other curvaturebased algorithms havebeen used in most of the recent works including [19], [20] aswell. Curvaturebased tortuosity measures are more reliable,but they impose a heavy computational burden compared tothe methods of the ﬁrst group.
C. Methods Based on Angle Variation
These methods compute the direction variations of thevessel to measure tortuosity. In [21], the average of the anglesbetween sample center points that describe the vessel (calledlocal direction variation) were used to measure tortuosity.In [22], the same method is used to measure local angles andthe number of times a local angle surpasses
π/
6
is consideredas tortuosity index.
D. Methods Based on Other Domain
These methods are in fact a subgroup of the curvaturebased methods. The difference is that unlike the ﬁrst group,they calculate curvature in domains beside the space domain.Kaupp
et al.
[23] use Fourier analysis and Ghadiri
et al.
[24]use circular Hough transform to calculate curvature. Moreover,in [25] Non Subsampled Contourlet Transform (NSCT) is usedfor curvature calculation. The key feature of these methods isevaluation of tortuosity without vessel extraction. However,they suffer from heavy computational burden as well.There have also been some special cases of tortuositymeasurement algorithms. For example, Wallace
et al.
[26] usecubicspline interpolation for measuring tortuosity. Other work include [10], [27], [28]. For a more detailed review on different tortuosity measurement algorithms and their applicationssee [29].
E. Our Method
Our method is a curvaturebased tortuosity measurement;therefore it falls into the second group. To illustrate themethod, in section II we deﬁne curvature as a mathematicaltool for measuring local inﬂection. To calculate curvature, anovel approach called the template disk method is commonlyused. In section III, after showing the pitfalls of the templatedisk method, we propose two modiﬁcations for amendment of the method. Later in this section, we examine the accuracyof the modiﬁed methods using synthetic functions. SectionIV deals with how to use the template disk method andits modiﬁcations for tortuosity measurement and ﬁnally insections V and VI the test results, the concluding remarksand possible future works are given.
Fig. 2. Curvature calculation with the template disk method
II. C
URVATURE
C
ALCULATION
Curvature is an indication of local twistedness of acurve [30]. In the continuous case, for a planar curve
y
=
f
(
x
)
, curvature is given by:
κ
=
y
′′
(1 +
y
′
2
)
3
/
2
.
(1)Moreover, for the parametric representation
x
=
x
(
t
)
and
y
=
y
(
t
)
, we have:
κ
=
y
′′
(
t
)
x
′
(
t
)
−
x
′′
(
t
)
y
′
(
t
)(
x
′
(
t
)
2
+
y
′
(
t
)
2
)
3
/
2
.
(2)Note that rigid transformations would not effect the absolutevalue of curvature. However, the sign may change by rotation.To calculate curvature using a discrete model of a curve,numerical approaches have been exploited in [31]. However,numerical methods are nontrivial and time consuming forlarge data sets [32]. An elegant method for estimation of curvature that has a relatively low computational complexitywas introduced in [33]. This approach, known as the templatedisk method, is further explored for 2D images in [34].Through its simplicity, this method has been widely utilizedin most applications requiring curvature calculation. The basicidea of the method is to relate the area between the curve and atemplate disk of a suitable radius with curvature. To illustratethe method, note that for a point with zero ﬁrst derivative wehave:
κ
=
f
′′
(
x
)
.
(3)Hence, to calculate curvature for a point
(
x,y
)
, ﬁrst a templatedisk of radius
b
is sketched around the point (Fig. 2). Next, thecenter of the Cartesian coordinate system is set on the point toensure that
f
(0) =
f
′
(0) = 0
. Under this assumption, Taylorseries for points in the vicinity of zero is:
f
(
x
) = 0 + 0 + 12
κx
2
+
o
(
x
3
)
(4)where
κ
denotes curvature at the srcin and
o
(
x
3
)
expresseshigher order terms of the series. Using polar coordinates
(
r,θ
)
,we have:
rsinθ
= 12
κr
2
cos
2
θ
+
o
(
r
3
cos
3
θ
)
.
(5)Normalizing the coordinate system by the radius of the template disk
b
, we get:
Rsinθ
= 12
KR
2
cos
2
θ
+
o
(
R
3
cos
3
θ
)
.
(6)
21682194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/JBHI.2015.2396198, IEEE Journal of Biomedical and Health Informatics
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. XXX, NO. XXX, XXXXXXX XXXX 3
2c
κ
b
0 2 4 6 8 10012345
Fig. 3. Estimated curvature versus analytical curvature for a parabola
where
R
=
r/b
and
K
=
κb
. Now, assume that
θ
is near zero,we can approximate
cosθ
by one and
sinθ
by
θ
and deﬁne
θ
as a function of
R
and
K
:
θ
(
R
)
≈
sin
−
1
12
KR
+
o
(
R
2
)
.
(7)Using this expression, we could approximate area
A
which isthe area between the disk and the curve. Deﬁne:
a
=
10
RdR
π
−
θ
(
R
)
θ
(
R
)
dθ
(8)then
A
=
ab
2
. As it turns out, analytical calculation of thisintegral does not yield a linear relation between curvature andthis area. However, if
θ
(
R
)
is further approximated by itsTaylor series around zero, we get:
θ
(
R
)
≈
12
KR
+
o
(
R
2
)
.
(9)Ignoring higher order terms, the normalized area using (8) is:
a
≈
π
2
−
K
3
,
(10)and rearranging terms yield:
κ
≈
3
π
2
b
−
3
Ab
3
= 3
A
c
b
3
−
3
π
2
b .
(11)where
A
c
is the complement of
A
. Therefore
κ
∝
A
c
. Basedon this relation and the argument given in the next section,the nonlinear estimation of curvature
κ
nl
is deﬁned as:
κ
nl
A
c
.
(12)III. M
ODIFICATIONS OF THE
T
EMPLATE
D
ISK
M
ETHOD
As we saw in the previous section, the assumption that
θ
is small simpliﬁes (6) and results in (7). However, thisassumption is valid merely if curvature is sufﬁciently small. Toshow that (11) fails to deliver for large curvature, assume that
κ
nl
is used to estimate the curvature of a parabola
y
=
cx
2
atthe srcin. Fig. 3 is a comparison of the analytical curvature
2c
(horizontal axis) versus curvature derived by the method.As is clear, though for small curvature (small
c
) estimatedcurvature and actual curvature are proportional, the relationfor large curvature is far from being linear. On the other hand,as indicated by the test results of [34], for error free calculationof curvature in digital images, the radius of the template disk should be large. This is due to the fact that for small radius,area calculation may contain signiﬁcant error. However, as
b
increases, area
A
has weaker dependency on
κ
. Therefore, weface a contradiction between the assumption that simpliﬁes (7)
Fig. 4. Approximating area
A
using the crossover point of the curve and thetemplate disk
(which assumes a small radius for the disk) and the one that isrequired for accurate implementation (which requires a largeradius for the template disk).Although we showed that the template disk method is inaccurate when estimation of high curvature is concerned, throughits simplicity, we try to perform the necessary modiﬁcationsin order to make it applicable to all curvature values. In whatfollows, two modiﬁcations are given.
A. First Modiﬁcation: Based on Crossover Point (
κ
cp
)
As Fig. 4 shows, irrespective of
θ
c
(the cross over angle of the curve and the disk), the area between the curve and thetemplate disk is the gray area plus the green area. Therefore,area
A
could be approximated by the gray area or
˜
A
. In whatfollows, we try to relate
˜
A
and
θ
c
.To derive
θ
c
, it sufﬁces to set
R
= 1
in (6). We have:
sinθ
c
= 12
Kcos
2
θ
c
+
o
(
cos
3
θ
c
)
.
(13)Now, we assess three different cases:
1)
θ
c
near zero:
In this case,
sinθ
c
is approximated by
θ
c
and
cosθ
c
by one. We have:
θ
c
= 12
K.
(14)Thus
˜
A
is:
˜
A
≈
πb
2
2
−
Kb
2
2
.
(15)Rearranging terms yields the following relation for small crossover angle:
κ
≈
πb
−
2 ˜
Ab
3
.
(16)Note that as was expected, this equation is similar to (11)neglecting a scaling factor.
2)
θ
c
near
π/
2
:
If
θ
c
is near
π/
2
, we can approximate
cosθ
c
by
π
2
−
θ
c
and
sinθ
c
by one which yields:
θ
c
≈
π
2
−
2
K .
(17)Hence, curvature for such angles has the following relationwith
˜
A
:
κ
≈
2
b
3
˜
A
2
.
(18)As (16) and (18) indicate, for small curvature (i.e. small crossover angle),
κ
has a linear relation with
A
(or
˜
A
). However, forlarge curvature (i.e. large cross over angle),
κ
is proportional
21682194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/JBHI.2015.2396198, IEEE Journal of Biomedical and Health Informatics
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. XXX, NO. XXX, XXXXXXX XXXX 4
A
c
A
θ
c
(a)(b)(c)yyyxxx
Fig. 5. Curvature estimation with the template disk method a) Drawing thetemplate disk b) Eliminating the inferior segment from inside the disk c)Determinig
A
,
A
c
and
θ
c
to the squared reciprocal of
A
(or
˜
A
). We deﬁne the squaredreciprocal of
A
as an approximation of curvature:
κ
cp
1
A
2
.
(19)Note that for small curvature,
A
is relatively large which resultsin a
κ
cp
close to zero. Therefore, (19) should have satisfactoryaccuracy for small curvature as well. For this reason, weconsider (19) it as a suitable alternative for (12). The testresults of section IIIC will manifest that this presumption isin fact valid.
3) Other values of
θ
c
:
As we will show with sufﬁcienttests, the metric
κ
cp
has sufﬁcient accuracy for values whichare neither close to
0
nor close to
π/
2
. Therefore, no furtherexpression is required to treat such other angles.
B. Second Modiﬁcation: Trigonometrical Method (
κ
tr
)
Returning to (13), if we neglect
o
(
R
3
cos
3
θ
c
)
, we have thefollowing relation between
θ
c
and
K
:
K
≈
2
sinθ
c
cos
2
θ
c
=
⇒
κ
≈
2
sinθ
c
bcos
2
θ
c
.
(20)Using (20), we can directly relate curvature to the cross overangle. Hence, deﬁne:
κ
tr
2
sinθ
c
1
−
sin
2
θ
c
.
(21)To compare the three curvature estimation methods, notethat the ﬁrst two methods rely on area calculation which is aform of integration. Therefore, we expect these two methodsto be resilient against noise and return smoother results. Onthe other hand, to get accurate results using the third method,knowledge of the exact cross over point is a fundamentalrequirement. Therefore, even quantization error may widelyaffect the results. This phenomen will be explored in the nextsection. However, this method does not need area calculationwhich is an advantage in terms of implementation. Based onthe above facts, if we should deal with noisefree signals of relatively low curvature,
κ
tr
should produce better results.However, for signals sampled in the presence of noise,
κ
cp
should function better.
C. Accuracy Examination Using Synthetic Functions
In this section, we compare the two modiﬁed template disk methods with the basic method. To do so, we use syntheticparabolas, circles and sinusoids. Note that even though foreach method the resultant range of values may be different,in our analysis of tortuosity linear proportionality to realcurvature is of mere importance. Hence, all test results to bepresented are suitably normalized.In order to estimate curvature for a point on the curve
y
=
f
(
x
)
, ﬁrst a circle with center
(
x,y
)
and radius
b
is depicted onthe curve (Fig. 5.a). Next, the segment of the curve inside thedisk that is connected to the point
(
x,y
)
is saved and the restof the curve is eliminated (Fig. 5.b). At this point, the twoareas conﬁned within the disk and the curve are calculated.The larger area is
A
c
and the smaller one is
A
. Moreover, thecrossover point of the curve and the disk is used to derive
θ
c
and this angle is used for calculation of
κ
tr
.
1) Parabola:
For a parabola
y
=
cx
2
with
c >
0
, curvaturefor
x
= 0
using (1) is:
κ
= 2
c.
(22)Fig. 6 shows the estimated curvature versus analytical curvature for
0
.
001
c
1
.
5
for four different templatedisk radii. As can be seen (and this was shown beforehand)curvature and
κ
nl
are not proportional. On the other hand,
κ
cp
and
κ
tr
elevate linearly with analytical curvature. In fact,for small curvature, all three methods show proportionalityto real curvature. However, as curvature increases, only thetwo modiﬁed methods maintain their linearity. Note that thestepwise behavior of
κ
tr
is due to the quantization of valuesof
θ
c
. For larger values of
c
, even small changes in
θ
c
highlyaffect the denominator of (21). Therefore, the quantizationerror for such values is signiﬁcant.
2) Circle :
To estimate curvature for a circle, assume thefollowing equation for a circle that is tangent to the srcin:
(
y
−
r
)
2
+
x
2
=
r
2
r
≫
b
(23)where r is the radius of the circle. Using (1), curvature at thesrcin is:
κ
= 1
/r.
(24)Fig. 7 shows the estimated curvature versus analytic curvaturefor
20
r
200
for three different values of
b
. Relatively lowvalues of curvature for a circle imply that all three methodsshould return acceptable results. Note that
κ
tr
is still sufferingfrom quantization error.
3) Sinusoid:
One of the functions that is commonly usedfor examination of the accuracy of a tortuosity measure isa sinusoid. If a sinusoid is given by
y
=
A
m
sin
(
ωx
)
, thencurvature for every point using (2) is:

κ

=
A
m
ω
2
sin
(
ωx
)(1 +
A
2
m
ω
2
cos
2
(
ωx
))
3
/
2
.
(25)With three tests, we compare the analytical curvature with
κ
nl
,
κ
cp
and
κ
tr
. In the ﬁrst test, we compare (25) with theoutcome of the three methods for a speciﬁed frequency andamplitude. In the next two test, we employ an image of asinusoid and observe the effect of amplitude and frequencyvariation on the estimated curvature. To do so, the averagevalue of curvature for each image is calculated. Note that asGrisan
et al.
[17] claim, increase in amplitude or frequencyshould increase average curvature. Detailed description of eachtest is given below.
21682194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/JBHI.2015.2396198, IEEE Journal of Biomedical and Health Informatics
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. XXX, NO. XXX, XXXXXXX XXXX 5
κ
n l
κ
c p
κ
n l
κ
n l
κ
n l
κ
c p
κ
c p
κ
c p
κ
t r
κ
t r
κ
t r
κ
t r
2c 2c 2c 2c2c2c2c2c2c2c
2c 2c
0 0.5 1 1.5 2 2.5 30.50.60.70.80.91Kb
0 0.5 1 1.5 2 2.5 300.20.40.60.81Kcp
0 0.5 1 1.5 2 2.5 300.20.40.60.81Ktr
0 0.5 1 1.5 2 2.5 300.20.40.60.81
0 0.5 1 1.5 2 2.5 300.20.40.60.81
0 0.5 1 1.5 2 2.5 30.50.60.70.80.91
0 0.5 1 1.5 2 2.5 300.20.40.60.81
0 0.5 1 1.5 2 2.5 300.20.40.60.81
0 0.5 1 1.5 2 2.5 30.50.60.70.80.91
0 0.5 1 1.5 2 2.5 300.20.40.60.81
0 0.5 1 1.5 2 2.5 300.20.40.60.81
0 0.5 1 1.5 2 2.5 30.60.70.80.91
b=5 b=11 b=25 b=51b=5 b=11 b=25 b=51b=5 b=11 b=25 b=51
Fig. 6. Estimated curvature with
κ
nl
,
κ
cp
and
κ
tr
versus actual curvature for a parabola
1/r1/r1/r1/r 1/r 1/r1/r1/r1/r
0.01 0.02 0.03 0.04 0.050.850.90.9510.01 0.02 0.03 0.04 0.050.20.40.60.81
0.01 0.02 0.03 0.04 0.050.860.880.90.920.940.960.981
0.01 0.02 0.03 0.04 0.050.20.40.60.81
0.01 0.02 0.03 0.04 0.0500.20.40.60.81
0.01 0.02 0.03 0.04 0.050.90.920.940.960.98
0.01 0.02 0.03 0.04 0.050.80.850.90.951c
0.01 0.02 0.03 0.04 0.050.70.750.80.850.90.951c
0.01 0.02 0.03 0.04 0.050.650.70.750.80.850.90.951c
κ
b
κ
b
κ
c p
κ
c p
κ
c p
κ
t r
κ
t r
κ
t r
b=7b=7b=11b=11 b=15b=15b=7 b=11 b=15
κ
n l
κ
n l
κ
n l
Fig. 7. Estimated curvature with
κ
nl
,
κ
cp
and
κ
tr
versus actual curvature for a circle
4) First test:
Fig. 8 shows the graph of

κ

and resultsof pointwise curvature calculation with each method for asinusoid with
A
m
= 200
and
ω
= 2
π/
100
where
b
is 11pixels. It is clear that
κ
cp
is most similar to real curvature andcurves of
κ
tr
and
κ
nl
are very much alike.
5) Second Test:
For this test, ﬁrst a
500
×
500
pixel imagecontaining a sinusoidal curve is created. Next, the period of the sinusoid is held ﬁxed at 100 pixels and its amplitude isvaried from 50 to 200 pixels. For each method, mean curvatureof each image using a template disk of radius 11 is depictedin Fig. 9. As the ﬁgure shows, elevation in average curvaturewith amplitude increment is more apparent for
κ
tr
than
κ
cp
and
κ
nl
.
6) Third Test:
For this test, the same image as the secondtest was used. Only this time the amplitude is ﬁxed at 100pixels and the period is changed from 50 to 250 pixels. Usinga template disk of radius 11, mean curvature of each image wascalculated (Fig. 10). As can be seen, proportionality betweenfrequency and curvature is better seen for
κ
tr
and
κ
cp
.IV. T
ORTUOSITY
E
VALUATION
In this section we propose an algorithm for automaticevaluation of tortuosity in retinal images. Since curvaturecalculation is the core of this algorithm, all three curvatureestimation methods given in section III are used so as to test