A cellular automata based semi-automatic algorithm for segmentation of choroidal blood vessels from ultrahigh resolution optical coherence images of rat retina

Abstract

Abnormal changes in choroidal blood flow have been linked to various retinal diseases, such as Diabetic Retinopathy (DR) and Age related Macular Degeneration (AMD), which at later stages can lead to blindness. Therefore non-invasive and precise evaluation of choroidal blood flow can aid the diagnosis, treatment and monitoring of retinal disease progression. Doppler Optical Coherence Tomography is an imaging technique capable of measuring blood flow velocity and visualization of retinal and choroidal blood vessles. However accurate assessment of retinal and choroidal blood flow requires precise measurement of the blood vessel thickness. The presence of speckle noise and low image contrast of OCT tomograms makes this task very challenging. This paper proposes a cellular automata based semi-automatic algorithm for the segmentation of choroidal blood vessels. The proposed approach propagates user-defined points in order to identify the vessel boundaries, allowing a thickness profile to be extracted. The performance of the algorithm was tested on a series of retinal images acquired from living rats with a high speed, ultrahigh resolution OCT system (UHROCT). Experimental results show that the proposed approach provides precise thickness profiles even in the suboptimal conditions of low image contrast in the UHROCT images.

Full text

Page 1

A CELLULAR AUTOMATA BASED SEMI-AUTOMATIC ALGORITHM FOR
SEGMENTATION OF CHOROIDAL BLOOD VESSELS FROM ULTRAHIGH RESOLUTION
OPTICAL COHERENCE IMAGES OF RAT RETINA
Akshaya Mishra, Sepideh Hariri, Alireza A. Moayed, Kostadinka Bizheva, Paul Fieguth and David Clausi
University Of Waterloo
Department of Systems Design Engineering, and Physics and Astronomy
Waterloo, ON, Canada
ABSTRACT
Abnormal changes in choroidal blood flow have been linked
to various retinal diseases, such as Diabetic Retinopathy (DR)
and Age related Macular Degeneration (AMD), which at later
stages can lead to blindness. Therefore non-invasive and pre-
cise evaluation of choroidal blood flow can aid the diagno-
sis, treatment and monitoring of retinal disease progression.
Doppler Optical Coherence Tomography is an imaging tech-
nique capable of measuring blood flow velocity and visualiza-
tion of retinal and choroidal blood vessles. However accurate
assesment of retinal and choroidal blood flow requires pre-
cise measurement of the blood vessel thickness. The presence
of speckle noise and low image contrast of OCT tomograms
makes this task very challenging.
Thispaperproposesacellularautomatabasedsemi-automatic
algorithm for the segmentation of choroidal blood vessels.
The proposed approach propagates user-defined points in or-
der to identify the vessel boundaries, allowing a thickness
profile to be extracted. The performance of the algorithm
was tested on a series of retinal images acquired from liv-
ing rats with a high speed, ultrahigh resolution OCT system
(UHROCT). Experimental results show that the proposed ap-
proach provides precise thickness profiles even in the sub-
optimal conditions of low image contrast in the UHROCT
images.
Index Terms— Cellular automata, semi automatic seg-
mentation, ultra high resolution optical coherence tomogra-
phy (UHROCT), choroidal blood vessels.
1. INTRODUCTION
Opticalcoherencetomography(OCT)isanimagingtechnique
that has the ability to acquire non-invasive, high resolution
and 3-dimensional images of the structural composition of
biological tissue [3,4]. One of the most advanced biomed-
ical applications of this imaging technique is use for non-
contact, 3D imaging of human and animal retina. Morpho-
akmishra, dclausi, pfieguth @uwaterloo.ca
logical features that can be viewed and measured from retinal
OCT tomograms — drusen, cysts, macular holes, and retinal
/ choroidal blood vessels — can be used as markers in retinal
disease diagnostics and progression. For example, the reti-
nal nerve fiber layer thickness is an important marker for the
clinical diagnosis of glaucoma [5,6].
Since retinal diseases can affect blood flow in the retinal
vessels, precise assessment of flow changes can prove a use-
ful clinical diagnostic. Doppler OCT has been developed in
the past 15 years as an expansion of traditional OCT technol-
ogy, allowing for imaging and measurement of blood flow ve-
locity in retinal vessels. Although Doppler OCT can provide
very precise measurement of flow velocity, calculating the to-
tal blood flow in a retinal vessel requires knowing the blood
vessel thickness or cross-section. Because retinal OCT to-
mograms are affected by speckle noise, the task of automatic
segmentation and sizing of retinal blood vessles is difficult.
Speckle noise in OCT images causes difficulty in the precise
identification of the boundaries of layers or other structural
features in the image either through direct observation or use
of segmentation algorithms. Furthermore, the low contrast
in some OCT retinal images, related to sub-optimal imaging
conditions or from the high haemoglobin absorption of light,
can also cause malfunction in the segmentation algorithms or
reduce their precision. Even more challenging is the segmen-
tation of choroidal blood vessels due to further reduced image
contrast at greater imaging depth.
The field of image segmentation is very mature and many
methods [7,8] are available for biomedical image segmenta-
tion. Realizing the fact that unsupervised segmentation meth-
ods are not suitable for biomedical image segmentation, sev-
eral model-based problem-specific and semi-automatic seg-
mentation methods [9, 11] have been proposed for biomed-
ical image segmentation. However, the low signal noise to
ratio of OCT imaging along with a small difference in refrac-
tive index between blood vessels and water content make the
choroidbloodvesselboundaryindistinctfromtheforeground.
As a result, none of the existing segmentation methods can be
directly applied to identify the boundaries of choroid blood
978-1-4244-5377-1/10/$26.00 ©2010 IEEE

Page 2

vessels. As a result, such tasks have been carried out, manu-
ally, by trained experts. Such manual segmentation by trained
experts is not only time consuming, but also tedious and error
prone.
In this paper, a novel semi-automatic algorithm is pro-
posed to precisely measure the thickness of the choroid blood
vessel. The proposed approach uses a cellular automata based
semi-automated segmentation algorithm to identify the outer
boundary of the choroid blood vessel and then the orienta-
tion of the boundary of the choroid vessel is identified using
an eigen-decomposition. The thickness profile of the choroid
vessel is computed with respect to the direction of the largest
eigenvalue.
The rest of the paper is organized as follows: A brief de-
scriptionaboutcellularautomatabasedimagesegmentationis
provided in Section. 2. The proposed approach is described in
Section. 3. Finally, the experimental results are demonstrated
in Section. 4
2. CELLULAR AUTOMATA BASED
SEGMENTATION ALGORITHM
Cellular automata (CA) first prossed by Von-Neuman. [12] is
a popular discrete model widely used in computability theory,
mathematics, theoretical biology and microstructure model-
ing. CA is a regular lattice consisting of a set of sites. Each
sites can have finite number of states, for example for a two
class case, the states of each site can be modeled as ”On”
or ”Off”. Recently, researchers have been attempting to for-
mulate image segmentation problem using finite cellular au-
tomata theory [10,11].
ThefundamentalsofCAbasedsegmentationfollows. Con-
sider P and p ∈ P be set of sites in a discreet lattice £.
Assume I = {I(p)|p ∈ P}, Lt = {Lt(p)|p ∈ P}, Ω =
{Ω(p)|p ∈ P}, Si= {Si(p)|p ∈ P} be set of random fields
on P representing the measured entity (usually an image), la-
bel field associated with the measured entity, neighbor hood
of the measured entity, and strength of the label field Lt, re-
spectively. The term t represents the iteration number. A CA
is usually defined using a triplet A = [Lt,St,Ψ] respectively,
where Ψ is a similarity measure that controls the likelihood
of state transition for t − 1 to t. The change in state of the
level field L from t − 1 to t are function of Lt−1(p), St−1(p)
and Ψ(Ω(p),p),p ∈ P. Therefore, the state or label field
of site p,p ∈ P at t Lt(p) can be expressed as a triplet
{Lt−1(p),St−1(p),Ψ(Ω(p),p)}.
This paper have proposed a novel noise robust technique
to update the strength of each sites St(p), and to define a sim-
ilarity measure Ψ of CA proposed by Rother et. al. [11].
3. PROPOSED APPROACH
Proposed approach consists of two steps: 1) identifying the
boundaryofthechoroidvesselusingacellularautomatabased
Algorithm 1 vc= Function CA(I, τ)
1: Initialize the S0and L0as
S0(p) =
⎧
⎪
⎪
⎩
⎨
1
if p is a seed pixel
or foreground seed
otherwise
0
,
L0(p) =
⎧
⎪
⎪
⎩
⎨
−1
1
0
if p is a user defined background seed
if p is a user defined foreground seed
otherwise
,
2: t = 1; Lt= {};
3: while Lt?= Lt−1do
4:
St= St−1,Lt= Lt−1,
5:
for p = P (all pixels) do
6:
Ωl= Ω(p), Ωl∗= find(Ωl?= 0),
7:
Update Ψ(Ωl∗,p) using (1
8:
9:
if abs(T − St(p)) ≤ τ then
10:
Lt(p)
St(p) = T,
11:
end if
12:
end for
13:
t = t + 1.
14: end while
T = max?St
?Ωl∗?ψ?Ωl∗,p??
= argmax?St
?Ωl∗?ψ?Ωl∗,p??,
semiautomatic segmentation algorithm and 2) measuring the
thickness profile of the choroid vessel by first determining the
orientation of the identified boundary.
3.1. Cellular Automata choroid Vessel Segmentation
Since, proposed approach is designed as a semi-automatic
segmentationtechnique, auserdefinessetpixelasbackground
and another set of pixels as foreground (choroid vessel ) as
shown in Fig.1. Then we employ a cellular automata based
propagation technique as described in Algorithm. 1 to label a
given retinal image as background and choroid vessel. Pro-
posed cellular automata based segmentation algorithm is a
different propagation techniques than the traditional cellular
automata based image segmentation techniques. Addition-
ally, proposed approach incorporate noise robust similarity
functiontopreciselyidentifyingthechoroidbloodvessel. The
propagationtechniquesusedinproposedapproachispresented
in Algorithm. 1. The similarity function used in the proposed
approch can be decribed as:
Ψ(Ωl∗,p) =
1
Zc
exp(−(c1f1+ c2f2+ c3f3))
(1)
The term c1, c2and c3are relative weight factor for f1, f2
and f3, respectively. The user defined constant Zcis used to

Page 3

60 708090100110
10
15
20
25
30
35
Column
Thickness
(a) Cropped image (b) Seed pixels (c) Segmented vessel (red) (d) Estimated ellipse (yellow)(e) Thickness of the segmented vessel
as a function of column.
Fig. 1. Steps involved in the proposed approach. The seed pixels corresponding to the back-ground and fore-ground are
shown in black and white. The segmented choroid blood vessel and the estimated ellipse are shown in third and fourth rows
respectively (red and yellow). The thickness of the choroid blood vessel as function of the column is show in fifth column.
makeΨaprobabilityfunction. Thesimilarityfunctionsf1, f2
and f3are used to measure the similarity between pixel p and
q and are defined as follows: f1= ?E (I (p)) − E (I (q))?,
f2=
− E
and f3penalize the the first order intensity moment variation,
second order intensity moment variation and spatial variation,
respectively.
???E
?
I (p)2??
I (q)2???? and f3= ?p − q?. f1, f2
3.2. Choroid Vessel Thickness Profile Measurement
The pixel corresponding to the choroid vessel of a given reti-
nal image is computed from the converged label field Lt. Let
X and Y are the locations corresponding to choroid vessel
of the retinal image. Let Σ = cov(X,Y ) be the covariance
matrix representing the locations of the choroid vessel. We
apply a singular value decomposition method to extract the
two eigen vector e1and e2corresponding to the two eigen
values λ1and λ2. As shown in Fig. 1, proposed approach ap-
proximate the choroid vessel as an ellipse using these eigen
values and eigen vectors. Then the thickness profile of the
choroid vessel is computed in a direction perpendicular to the
eigen vector corresponding to the major axis of the ellipse.
4. RESULTS
The segmentation and measurement accuracy of the proposed
approach is demonstrated on a series of OCT tomograms ac-
quired from the retinas of living rats with a research grade
UHROCT system. All animal experiments were conducted in
accordance with an ethics protocol approved by the Univer-
sity of Waterloo Animal Ethics Committee. A detailed de-
scription of the UHROCT system and more information on
the animal experiments can be found in [1, 2]. Briefly, the
imaging system provides 3um axial and 5um lateral imaging
resolution in the rat retina at an image acquisition rate of 46
frames/s. The system’s sensitivity is 100dB for 1.4mW op-
tical power of the imaging beam incident on the rat’s cornea,
which is in accordance with laser safety regulations outlines
in the ANSI standard. Each retinal OCT image is comprised
of 1000 A-scans and 512 pixels and allows for clear visualiza-
tion of all intra retinal layers, as well as the choroidal vascu-
Fig. 2. Two dimensional image of the rat retina acquired in-
vivo away (left) from the optic disk. The image was acquired
at a rate of 47 A-scans /s with a state of the art UHROCT
system operating in the 1060nm wavelength region. The sys-
tem provides 3um axial and ¡5um lateral resolution in the rat
retina with SNR of 100 dB at 1.4mW optical power incident
on the rats cornea. Blue arrows point at choroidal blood ves-
sels. Redarrowspointatsurfacebloodvessels(retinalarteries
and veins).
lature (Fig. 2) As shown in Fig. 2, presence of speckle noise
blurs the boundary between the choroidal blood vessels and
the surrounding tissue, thus limiting the precision of direct
assessment of the blood vessels thickness.
Fig. 1 shows a small section of a rat retina image near
the retinal pigmented epithelium (RPE) that includes a cross-
section of a choroidal blood vessel. When the proposed al-
gorithm is applied to the image, automatic segmentation of
the blood vessel cross-section is achived. The segmented area
(Fig. 1(c)) is fitted with an ellipse (Fig. 1(d)) to allow precise
measurement of the vessel’s cross-section and size (Fig. 1(e)).
Some more segmentation results using proposed approach are
demonstrated in Fig. 3.

Page 4

(a) Cropped images(c) Segmented vessels (red)
Fig. 3. The segmented vessel (red) of two retinal images ob-
tained from the 3D scan of a rat retina.
5. CONCLUSIONS
A semiautomatic segmentation algorithm based on cellular
automata theory is proposed to identify the choroid blood ves-
sel of retinal images. The thickness profile of the choroid
bloodvesseliscomputedbyapproximatingthetheretinallay-
ers using an elliptical model. The performance of proposed
method in evaluating the thickness profile of choroid blood
vessel is demonstrated using several retinal images.
The preliminary results are very encouraging and the fu-
ture work includes extending the proposed approach to mul-
tiple dimensions and making the propose algorithm fully au-
tomatic.
Acknowledgment
This research has been sponsored by the Natural Sciences and
Engineering Research Council (NSERC) of Canada through
individual Discovery Grants as well as GEOIDE (GEOmat-
ics for Informed Decisions) which is a Network of Centres
of Excellence under NSERC. A part of the funding for this
research has also been supported by the Department of Sys-
tems Design Engineering and the Graduate studies office of
University of Waterloo.
6. REFERENCES
[1] P. Puvanathasan, P. Forbes, Z. Ren, D. Malchow, S.
Boyd and K. Bizheva, “High-speed, high-resolution
Fourier-domain optical coherence tomography system
for retinal imaging in the 1060 nm wavelength region,”
Opt. Lett, Vol. 33 Issue 21, pp.2479-2481 (2008).
[2] P. Puvanathasan, P. Forbes, Z. Ren, D. Malchow, S.
Boyd and K. Bizheva, “ High speed high resolution in-
vivo imaging of healthy rat retinas with FD-OCT oper-
ating at 1060nm,”SPIE Proc. BIOS, Photonics West, v.
7163, p.716305 (2009).
[3] D. Huang, E. Swanson, C. Lin, J. Schuman, W. Stinson,
W. Chang, M. Hee, T. Flotte, K. Gregory, C. Puliafito,
and J. Fujimoto, “Optical coherence tomography,” Sci-
ence 254, 1178-1181 (1991).
[4] A.F. Fercher, “Optical coherence tomography,” J.
Biomed. Opt. 1, 157 (1996).
[5] C. Leung, C. Cheung, R. Weinreb, K. Qiu, S. Liu,
H. Li, G. Xu, N. Fan, C. Pang, R. Tse, and D.
Lam, “Evaluation of retinal nerve fiber layer progres-
sion in glaucoma with optical coherence tomography
guided progression analysis (GPA),” Invest. Ophthamal.
Visual Sci. (2009), http://www.iovs.org/cgi/
rapidpdf/iovs.09-3468v1.pdf.
[6] S. Taliantzis, D. Papaconstantinou, C. Koutsandrea, M.
Moschos, M. Apostolopoulos, and G. Georgopoulos,
“Comparative studies of RNFL thickness measured by
OCT with global index of visual fields in patients with
ocular hypertension and early open angle glaucoma,”
Clin. Ophthalmol. 3, 373-379 (2009).
[7] E. G¨ otzinger, M. Pircher, W. Geitzenauer, C. Ahlers,
B. Baumann, S. Michels, U. Schmidt-Erfurth, and
C. Hitzenberger, ”Retinal pigment epithelium seg-
mentation by polarization sensitive optical coher-
ence tomography,” Opt. Express 16(21),
16422 (2008), http://www.opticsinfobase.
org/abstract.cfm?URI=oe-16-21-16410.
16410-
[8] A. Mishra, P. Fieguth, and D. Clausi, “Robust snake
convergence based on dynamic programming,” Proc.
IEEE International Conference on Image Processing,
1092-1095 (2008).
[9] E.N. Mortensen and W.A. Barrett, ”Toboggan-based in-
telligent scissors with a four-parameter edge model”. In
CVPR, 2452?2458, 1999.
[10] A. Popovici and D. Popovici, Cellular automata in im-
age processing. In Fifteenth International Symposium
on Mathematical Theory of Networks and Systems,
2002
[11] C. Rother, V. Kolmogorov, and A. Blake, ”Grabcut -
interactive foreground extraction using iterated graph
cuts”. Proc. ACM Siggraph, 2004.
[12] J. V. Neumann, ”Theory of self-reproducing automata”.
University of Illinois Press. Ed. and Completed by A.
Burks, 1966.