odmLETctmcTRtimmestateca

1

tCuibirwu[mfb

g

Auto

A

G

1577 K

y

In the field of retina and idiagnosis of measurement Limiting MeEpithelium (RTomography (conventional techniques andmost cases of ncannot be appThis paper proRPE usingtwo-dimensionimages to impmethod, retinamorphologicaextract RPE structure of ththe proposedactual 30 normthe effectivenexperimental can extract ILand correct ex

1. Introdu

Recently, totained noncoCoherence Tousing OCT denfrared (NIR

boundaries whnterference w

reflected wavwaves, as a reure 1 is gener[1][2]. The gemacular area afore the needs been growing

Meanwhile,glaucoma, age

matic Ex

Usi

Ai YamakawHiroh

Graduate ScMie

Kurima-macJ

yamakawa@

Ab

of ophthalmoits quantitative

retinal disemethods of

embrane (ILMPE) in the re(OCT) imagesmethods empd can extract Inormal OCT imlied to some imoposes a new g morpholnal dynamic nprove extractioa layer and l operations, ain the retina

he retina can bd method. Evmal OCT imagness of thresults showe

LM and RPE xtraction rate

uction

omographic imontactly and omography (Oevices, eyegroR) laser first.here the tissue waves are genes. OCT dev

esult. The tomrated by the denerated OCT and discus neof retina diag[3][4][5]. , there are me-related macu

xtraction

ing Morp

wa, Shinji Tharu Kawanachool of Enge Universitychiya, Tsu, JAPAN @ip.elec.mi

bstract

ology, automae evaluation a

eases. Previothe thicknessM) and Rtina using Ops have been ploy basic imILM and RPE mage. Howevmages with daextraction me

logical opet model (Action accuracy. ILM are detand Active Nea. It is expebe extracted avaluation expges were conde proposed ed that the pwith damage

of 80% was o

mages of retinnoninvasive

OCT). In a ounds are irra. Incident wdensity of retnerated by thices observe

mography imagdifference amo

image can shervi optici as gnosis using O

many retinal dular degenerat

n of Lay

phologic

Tsuruoka, aka gineering, y Mie 514-85

ie-u.ac.jp

tic measuremare important fously, automas between Innetinal Pigm

ptical Coherenreported. The

mage processappropriatelyer these metho

amages of retiethod of ILM aerations aive Net) for OIn the propos

termined by et is employedected that layappropriately periments us

ducted to validmethod. T

roposed methes appropriateobtained.

na have been oely by Optiretina diagnodiated with n

waves reflect tina changes, ahe incident athe interferen

ge shown in Fong these wavhow the statusan image, the

OCT images ha

diseases such tion, retinopat

yer Struc

cal Oper

507,

ent for

aticner ent nce eseing y in ods na. andand

OCT sedthe

d to yer by

ingdate The hodely,

ob-ical osis near

on and and nce

Fig-ves s of ere-ave

as thia

diabdiseare disetina

CbetwPigmeva(FigILM

cture for

ration an

FumGraduate S

Unive1001-1 Ki

fok

betic and so eases cause vinot well unde

ease conditional diseases. Currently the ween Inner Lment Epithel

aluate the congure 2), the deM and RPE ha

Figure1. Exam

Figure

Figure3.

r Retina

nd Active

mio OkuyamSchool of Heersity of Meishioka, Suz

JAPAkuyama@su

on [6]. Althoision loss, theerstood and qun is so import

method thatLimiting Memium (RPE) [7

ndition of retievelopment os been require

mple of tomogOCT

e2. Inner limit retinal pigm

Extraction re

in OCT

e Net

ma, Toshiki Yealth Sciencedical Scienzuka, Mie 5

AN uzuka-u.ac.j

ough most ofeir pathogenicuantitative evatant in the tre

t calculates mbrane (ILM)7] is mainly inal diseases f the extractio

ed as a practica

graphic imageT

ting membranment epithelium

sult by Yagi’s

Image

Yagice, Suzuka nce 510-0293,

jp

f these retinalc mechanismsaluation of theeatment of re-

the thickness) and Retinalemployed to

quantitativelyon method foral application.

e of retina by

ne and m

method

l s e -

s l o y r .

MVA2009 IAPR Conference on Machine Vision Applications, May 20-22, 2009, Yokohama, JAPAN3-4

42

tubRw

mtmmerwdee

2

2

agg

Gnrsit

tcpoA

2

it

Figure4. Pr

To comply wthe extraction using morpholbasic examinaRPE line is intwith disappear

In this papemethod using tour model. minimization method of a twextract the conregion enclosewe expect to edamage part. extract properlexamined the

2. ExtractMorph

2.1. PreproIn this pape

as the experimgitized to a pixgray scale and

Figure 4 shGenerally OCnoises due to mretina layer. smoothing promage as prepr

the average filThe propos

tions and the can reduce thprove the extoperations canActive Net can

2.2. SegmeIn this pape

s formed in a titive operation

roposed extrac

with such requmethod of IL

logical operatations. But thistermittently exrment parts (Fer, we proposActive Net, wActive Net theory and i

wo dimensionntour using imed with net moextract RPE uIn this paperly for retinal lextraction res

tion of Neighological Op

ocessing forer, OCT imagemental materiaxel size of 0.0

d resolution ofhows the outlCT images ofmultipath reflIn the propoocess to redurocessing. Thlter of 3×3 pixsed method usActive Net. T

he processing traction accurn extract the n extract RPE

entation of nr, the initial poquadrangle. A

n until extract

ction method o

uest, Yagi at. LM and RPE ftion and so ons method has xtracted in ret

Figure 3). se a new autowhich is one o

is based ois widely use

nal region [9]. mage texture inodel. By applyusing neighbor, we proposelayer image wult of ILM an

ghborhoodperation

r OCT imagees of 30 normaal. These OCT006 mm × 0.0f 512 × 480 piline of our prften have spelection and intosed method,uce these noihe smoothing pxels. ses the morphThe morpholo

time of Actiracy, and theneighborhood

E precisely.

neighborhooosition of nodActive Net prting object. Th

of ILM and RP

el. [8] proposfrom OCT iman, and conducthe problem ttinal layer ima

omatic extractiof dynamic coon the enered as extractiThe method c

nformation in ying Active N

or informatione the method

with damage, and RPE.

d of Retina b

eal retina are usT images are 006 mm, an 8-xels. roposed metheckle and spterference in, we employises in the Oprocess emplo

hological opeogical operatioive Net and ie morphologid of retina. T

od of retinades in Active Nocesses by repherefore, it tak

PE

sed age ted

that age

ion on-rgy ion can the

Net, n of d to and

by

sed di-

-bit

od. ike the

y a CT oys

era-ons im-ical The

Net pe-kes

so lshoprodesfromActhoometshothe bec

2.3S

Theorigpromorthistimneccomclossidecomdilaretinremupswhimovremminopeprofiguremto wregaILMextr

3.

3.1W

tionretinis lomar

Wtionposeveheiginit

long to set theorten processicessing to neicribe the metm OCT imagetive Net. As tod of retina, thod is applie

ows the result intensity of m

comes white.

. RemovalSome black coese componenginally, howevcedures. Therrphological ops step, a circues to the bina

cted componemponents in rsing operatione ILM wermponents are ation of closinal layer effec

move minimumside ILM. White connectedved by labelin

moval minimunimum compoeration. Closincessing imageure shows that

moved and moswhite pixels. Aarded as retin

M is distinguisraction proces

ExtractionActive Net

. Setting tWe ease the un of the whitena) with a maocated at the brgin (15 pixelsWe set the noned area (Figsition of everyery node. The ght 20 × widthial position of

initial net foring time by ighborhood othod to extrae as preprocesthe first step obinarization b

ed to the smoof binarizatio

most pixels in

l of extra coonnected compnts are the elever, not regardrefore closing perations, is aular filter is earized image

ents with smaretinal layer cn. But white cre existed.

linked whiteng operation,ctually. So, lam white connhite connectedd componentsng operation. Fum white connonents could ng operation we, and appliet small black cst pixels in theAs the result,

nal layer is exshed by figuressing of neighb

n of Retinal t

the initial poupper area thate area in Figargin (20 pixebottom positios). ode of Activegure 7(a)). Ay node to minumber of no

h 30). The metf Active Net.

r full of OCT iconfining th

f retina. In thct neighborho

ssing to set theof the extractby a discrimioothed imageon. The Figurthe neighborh

omponents ponents exist iements of the

ded as them in processing, w

applied to theemployed and

to remove thall size. Blacan be removconnected comThese whit

e components, therefore it abeling was apnected compond componentss contained ILFig.5(b) shownected compo

be removedwas applied tod result was

connected come retinal layer only the proc

xtracted (figur 5(c), we extraborhood of re

Pigment Ep

osition of Act is located at

gure 5(c) (neigels), and the lon of the whi

e Net for theActive Net tr

nimize the toode of Active thod extracts R

image. So, wehe object of

his section, weood of retinae initial net ofting neighbor-inant analysise. Figure 5(a)re shows thathood of retina

in Figure 5(a).e retinal layerthe following

which is one ofe next step. Ind applied fivehe black con-ck connectedved by usingmponents up-te connecteds of ILM bycan’t extract

pplied so as tonents existings upside largeLM were re-

ws the result ofonents. Thesed by labelingo the labelingFig.5(c). The

mponents werewere changed

cessing regionre 5(c)). Also,act ILM using

etina.

pithelium by

ctive Nett the top posi-ghborhood oflower are thatte area with a

e above men-ransforms theotal energy of

Net is 600 (=RPE using the

e f e a f -s ) t a

. r g f n e -d g -d y t o g e -f e g g e e d n ,

g

y

-f t a

-e f = e

43

3

mmscfcrrmc

hbt

icve

3.2. Confoage

Active Net model based omethod that estability statecharacteristicsfunction of Accorresponds wregion in an imregion by defmarkably charconformity ch

In OCT imahigher than thbeams stronglythe strength of

In this papemage is defin

can be extractevariance of inenergy of imag

Figure5. R

Fi

ormity char

represents theon mechanics extracts the oe with energys energy of nective Net. The

with the powermage. It is pofining energyracteristics ofaracteristics eages, intensity

he surroundingy reflect at thef interference er, the conformned as the folloed properly byntensity in thge.

(a) Bina

(b) L

(c) ClosiResult of extra

gure 6. Node

acteristics e

e behavior of dequation as e

object from imy minimizatiet and image e power lesser drawing net ossible to extr

y function thaf region hopinenergy of imagy in retinal layg tissue becaue boundary of waves becommity characteowing equatioy the regionalhe conformity

arized image

Labeling

ing operation acted neighbor

points of Acti

energy of im

dynamic contonergy, and is

mage by findiion. Conformis one of ener

ening this enerto characteris

ract the selecat represents ng to extract ge. yer are genera

use the NIR laf retina tissue a

me high. eristics energyon, because Rl average and y characterist

rhood of retin

ive Net

m-

our the ing

mity rgy rgy stic ted re-for

ally aser and

y of RPE

the tics

na

EI

V

�

Fneigthe borhstepsitychalayethe neginsilayegavweidefi(1),Alsdiretrac

),( qpEimageInf(p,q) :Avera

at a noVnf(p,q):Varian

at a n�1,�2 :Weigh

For each nodeghborhood reaverage and

hood region, aps. In this papy are the smaracteristics ener in OCT imvariance is

gative weight ide node pointer. Also the o

ve weight (�1

ight) so as to fining conform, Active Net tso Active Netection, becausction to vertica

(a)

(b)

(c) ResulF

,(1 pInf��ge of intensityode point(p,q)nce of intensitode point(p,q)ht (2.0, 2.0)

e points (Figurgion (9 pixelthe variance

and regularizeper, if the avermaller, then nergy of imagage, the averasmall. Therefand �2 is po

ts of Active Noutside node 1 is positive w

be let to boumity characteritransforms to t is not madese Active Net al direction.

) The initial po

) Result of Act

lt of morpholoigure 7. Exper

() 2Vq nf��y in neighborh) ty in neighbor)

re 6). The mls). The methe of intensity es an obtainedrage and variathe value o

ge is the smalage of intensitfore we gave ositive weightNet were led to

points of Actweight and �undary of retiistics energy o

surround thee to contract can extract R

osition of nod

tive Net only

ogy and Activrimental resul

),( qp (1)hood region

hood region

ethod sets thehod calculates

in the neigh-d value in 256ance of inten-

of conformityler. At retinalty is high andweight (�1 ist) so that theo inner retinaltive Net were

�2 is negativeinal layer. By

of image as Eqe retinal layer.

to horizontalRPE only con-

des

ve Net lt

e s -6 -y l d s e l e e y q. . l -

44

(a) Example 1

(b) Example 2

Figure 8. Experimental result for retinal layer image that have damage

Figure 9. Failure example

3.3. The reservation of the end of Active NetActive Net transforms to make the energy minimum by

repetitive operation and extracts the object. When the energy of Active Net is the minimum value (or converges), it observes that Active Net extracts the object. So, the reservation of the end of Active Net is set to continue convergent condition of energy in this paper.

4. Experimental Results and DiscussionFigure 7 shows an experimental result by the proposed

method. Figure 7(b) shows result by Active Net only, and Figure. 7(c) shows the result that put together ILM by morphological result of extracting neighborhood of retina and RPE by result of Active Net. Figure 7(c) shows the extracted result that ILM and RPE can be extracted by the proposed method. Figure 7(b) shows that the effective use of Active Net is depend on the initial position of nodes, and we propose a new method of setting the initial nodes automatically.

Figure 8 shows the experimental result for retinal layer image with damage. Figure 8 shows that Active Net can resolve the intermitted problem by Yagi’s method[8].

However, Figure 9 shows a failure example. As the cause, Active Net in this paper has the problem that it depends on the initial net and node points. For that reason,

we should consider the conformity characteristics energy of image as it does not depend on the initial net and node points.

As the result of the experiment, also, the extraction rate of 80% (24/30) was obtained. The extraction rate is de-fined as error of retinal thickness (differential between manual answer and experimental result) is smaller than 1/10 of retinal thickness by manual answer.

5. Conclusion and Future Works In this paper, we propose an automatic extraction me-

thod using morphological operation and Active Net, and consider some basic examination. A result of the experi-ment for OCT image suggests that ILM and RPE can be extracted for retinal layer image with damage by our proposed method. The experimental result shows that our method can measure at the extracted rate of 80% for 30 normal images correctly.

In the future, conformity characteristics energy of im-age should be reconsidered, and we should experiment using many more experimental materials to improve the efficacy.

References [1] N. Tanno, T. Ichikawa, A. Saeki, “Lightwave Reflec-

tion Measurement”: Japanese Patent #2010042 (1990) (in Japanese)

[2] K. Tsubota, “Ophthalmology”: Nanzando (1995) [3] S. Kishi, “Optical Coherence Tomography in Diag-

nosis of Retinal Diseases”: Elsevier Japan (2006) [4] U. Schmidt-Erfurth, R. A.Leitgeb, S. Michels et al,

“Three-Dimensional Ultrahigh-Resolution Optical Coherence Tomography of Macular Diseases”: Inves-tigative Ophthalmology & Visual Science, Vol.46, No.9, pp.3393-3402 (Sep.2005)

[5] M. Hangai, Y. Ojima, N. Gotoh, et al, “Three-Dimensional Imaging of Macular Holes with High-speed Optical Coherence Tomography”: Oph-thalmology, Vol.114, No.4, pp.763-773 (Apr.2007)

[6] T. Tokoro, A. Kanai, “MODERN TEXT BOOK OF OPHTHALMOLOGY”: Kanehara, pp.113-138 (1983)

[7] S. Ohno, S. Kinoshita, “Standard Ophthalmology”: Igaku-Shoin, p125 (1981)

[8] T. Yagi, F. Okuyama, H. Kawanaka, S. Tsuruoka, “A Study on Extraction Method of Internal Limiting Membrane and Retinal Pigment Epithelium from OCT Images”: Proc. of Joint 4th International Con-ference on Soft Computing and Intelligent Systems and 9th International Symposium on Advanced Intel-ligent Systems (SCIS&ISIS 2008), SU-G2-5, pp.2008-2013 (2008)

[9] K. Sakaue, K. Yamamoto, “Active Net Model and Its Application to Region Extraction”: The Journal of the Institute of Television Engineers of Japan, vol.45, No.10, pp1155-1163 (1991)

45