Automatic CDR Estimation for Early Glaucoma Diagnosis

Resea ch A icle Au oma ic CDR Es ima ion o Ea ly Glaucoma Diagnosis M. A. Fe nandez-G ane o, 1 A. Sa mien o, 2 D. Sanchez-Mo illo, 1 S. Jiménez, 3 P. Alemany, 3 and I. Fondón 2 1 Biomedical Enginee ing and Telemedicine Resea ch G oup, Uni e si y o Cádiz, Pue o Real, Cádiz, Spain 2 Signal Theo y and Communica ion Depa men , Uni e si y o Se ille, Se ille, Spain 3 Oph halmology Uni , Pue a del Ma Hospi al, Cádiz, Spain Co espondence should be add essed o M. A. Fe nandez-G ane o; [email p o ec ed] and I. Fondón; [email p o ec ed] Recei ed 1 June 2017; Re ised 9 Sep embe 2017; Accep ed 24 Sep embe 2017; Published 27 No embe 2017 Academic Edi o : And eas Maie Copy igh © 2017 M. A. Fe nandez-G ane o e al. This is an open access a icle dis ibu ed unde he C ea i e Commons A ibu ion License, which pe mi s un es ic ed use, dis ibu ion, and ep oduc ion in any medium, p o ided he o iginal wo k is p ope ly ci ed. Glaucoma is a degene a i e disease ha cons i u es he second cause o blindness in de eloped coun ies. Al hough i canno be cu ed, i s p og ession can be p e en ed h ough ea ly diagnosis. In his pape , we p opose a new algo i hm o au oma ic glaucoma diagnosis based on e inal colou images. We ocus on cap u ing he inhe en colou changes o op ic disc (OD) and cup bo de s by compu ing se e al colou de i a i es in CIE L ∗ a ∗ b ∗ colou space wi h CIE94 colou dis ance. In addi ion, we conside spa ial in o ma ion e aining hese colou de i a i es and he o iginal CIE L ∗ a ∗ b ∗ alues o he pixel and adding o he cha ac e is ics such as i s dis ance o he OD cen e. The p oposed s a egy is obus due o a simple s uc u e ha does no need nei he ini ial segmen a ion no emo al o he ascula ee o de ec ion o essel bends. The me hod has been ex ensi ely alida ed wi h wo da ase s (one public and one p i a e), each one comp ising 60 images o high a iabili y o appea ances. Achie ed class-wise-a e aged accu acy o 95.02% and 81.19% demons a es ha his au oma ed app oach could suppo physicians in he diagnosis o glaucoma in i s ea ly s age, and he e o e, i could be seen as an oppo uni y o de eloping low-cos solu ions o mass sc eening p og ams. 1. In oduc ion The Wo ld Heal h O ganiza ion (WHO) has epo ed an inc ease o he numbe o pa ien s suffe ing om eye dis- eases due o he aging o wo ld popula ion [1]. Among all o hem, glaucoma is he second leading cause o blindness in de eloped coun ies. This disease is conside ed as a majo public heal h conce n, and i s p e alence will p obably con inue o inc ease as li e expec ancy con inues o ise [2]. Glaucoma desc ibes a g oup o ocula diso de s wi h a common cha ac e is ic: he p og essi e loss o ne e fibe s in he e ina. Al hough i canno be cu ed, i s associa ed blindness may be p e en ed h ough ea ly diagnosis. How- e e , glaucoma is known as he “silen he o sigh ”in he sense ha i p esen s no symp oms un il ision is al eady los . Glaucoma should be diagnosed ea ly in he disease cou se in o de o iden i y pa ien s ha equi e ea men o main ain quali y o li e [2]. The loss o op ical fibe s due o glaucoma p og ession is associa ed wi h a co esponding change in he op ic disc (OD). The e o e, he emp y space wi hin he OD and he so-called cup is subsequen ly enla ged. Tha is he eason why he cup o disc a io (CDR), defined as he ela ion be ween he OD and cup a ea, inc eases wi h he p og ession o he disease. OD appea ance is, he e o e, c i ical in glau- coma diagnosis, and images o he e ina a e manda o y o a co ec disease assessmen . Se e al eye imaging echnologies ha e been de eloped du ing he las 160 yea s [3]. Heidelbe g e ina omog aph (HRT) and op ical cohe ence omog aphy (OCT) along wi h angiog aphy a e widely used in he diagnosis and ollow-up o pa ien s wi h diffe en ocula diseases such as diabe ic e inopa hy o macula degene a ion [3, 4]. Al hough OCT Hindawi Jou nal o Heal hca e Enginee ing Volume 2017, A icle ID 5953621, 14 pages h ps://doi.o g/10.1155/2017/5953621 p o ides he bes ep esen a ion o he e ina, de ices based on his echnique a e highly expensi e and hey canno be affo ded by local medical cen es [5]. As undus imaging is he mos es ablished way o e inal imaging in p ima y ca e se ings, an au oma ic glaucoma diagnosis sys em based on undus images could be deployed ha ing he po en ial o ea ly disease diagnosis [6]. Ne e heless, he use o undus imaging echniques alone could no be enough o a mass sc eening p og ams. The lack o specialis s in local heal h cen es makes he inspec ion o e e y pa ien ’s e inal image unaffo dable. Mo eo e , he amoun o in o ma ion would exceed he limi o clinicians’ abili y o ully u ilize i [3]. Unde hese ci cums ances, he use o au oma ed imaging classifica ion as a iage es may p o e o be cos effec i e [7]. Image-based glaucoma diagnosis is pe o med mainly wi h CDR measu emen , ha is, he compu a ion o he a io o OD and cup egion a eas. Cu en ly, his calculus is pe o med on he basis o manually delinea ed a eas o e he e inal undus image. The skilled human g ade mus ca e ully d aw he egion wi h an image edi o so - wa e, and a e wa ds, he a io o he a eas is calcula ed. This me hod is ime consuming and exhaus ing. A li le sa ing in ime is p o ided by some acquisi ion o de ices ha offe he possibili y o ex ac ing he OD and cup egion by adjus ing an ellipse o ou poin s ha should be in oduced by he expe . Ins ead o ca e ully ma king he whole egion, he physician should only ma k ou e e ence poin s. Howe e , assuming ha he a ea is ellip- ical and basing he adjus men on ou poin s makes he sys em a li le bi as e , ha is, abou eigh minu es pe eye unde he Klein p o ocol [8], bu less accu a e. I seems clea ha he medical communi y needs an au o- ma ic me hod o CDR compu a ion. A compu e -aided diagnosis ool (CAD) in eg a ing such an algo i hm could a oid p oblems o inaccu a e esul s while sa ing ime and cos s. Fo au oma ic CDR es ima ion, he OD and cup egions ha e o be segmen ed based on hei cha ac e is ic appea - ance (Figu e 1). Howe e , i mus be no iced ha he shape, size, and colou a ia ions on e inal images ac oss a popula- ion a e expec ed o be high [3], making OD and cup segmen- a ion a challenging ask (Figu e 2). Gene ally speaking, OD is an ex emely in ense egion inside he undus image and can be iden ified om ea u es such as he ollowing [9]: (i) Shape: he OD is oughly ci cula . (ii) Colou : he OD usually p esen s hues anging om o ange o yellow. (iii) B igh ness: he OD p esen s a b igh ness alue ha is usually highe han he es o he e inal image. (i ) Size: he OD a ea is usually less han 1/7 o he o al eye. The op ic cup is imme sed wi hin he OD egion. I usu- ally p esen s a oughly ci cula shape and a b igh yellowish colou as can be app ecia ed in Figu e 1. Howe e , i is well known ha i s segmen a ion om e inal undus images is a duous, due o he lack o dep h in o ma ion, which is no a ailable in he 2D images. Fu he mo e, he p esence o ill-defined and inhomogeneous op ic cup bounda ies (see Figu e 2) makes he p oblem e en mo e difficul [10]. F om he abo emen ioned cha ac e is ics o he OD and cup, colou is he mos ele an when ying o isola e bo h a eas [4]. Consequen ly, he p ope colou space selec ion is c ucial o he e en ual success o he algo i hm. Con e sely, i is a gene al end o only conside he illumina ion in o - ma ion o pixels [11–21]. Two ac s a e p esen ed by he majo i y o pape s o suppo hei selec ion: (1) The use o colou images in ol es highe complexi y due o hei h ee-dimensional na u e. (2) G ey le el images allow using well-known algo- i hms [22]. Some o he me hods using g ey scale images selec only one colou plane o he h ee a ailable in any colou image ep esen a ion (RGB, HIS, e c.). Mos o hese a icles claim ha he OD can be easily disc imina ed om he G channel when analysing he RGB componen s o he image [9, 23–33]. F equen ly, blood essels need o be p e iously inpain ed o p e en an in e e ence wi h he OD segmen a ion algo i hm [23, 27, 33]. Likewise, he e is a common end o using basic image p ocessing echniques such as his o- g am h esholding alone [30] o combined wi h o he me hods [23–25, 31–34]. Equally impo an is he use o he R colou plane [35–38], he V channel om HSV, [5, 39, 40], o he M colou channel o CMY [41]. Only a mino i y o me hods elies on he luminance coo dina e (L ∗ ) o CIE L ∗ a ∗ b ∗ colou space [42]. Se e al au ho s p e e he use o mo e han wo colou planes usually om RGB colou space. The p ocessing is pe o med sepa a ely, as i diffe en g ey le el images we e a ailable [8, 10, 43, 44]. Op ic disc Cup Figu e 1: The OD and cup as seen on a ypical e inal undus image. The OD is p esen ed as an almos ci cula egion wi h a colou anging om o ange o yellow. The cup is he b igh es egion wi hin i , wi h a diffuse bo de only dis inguishable by essel bends. 2 Jou nal o Heal hca e Enginee ing The abo emen ioned echniques a e based on g ey le el image p ocessing. F om a compu a ional poin o iew, he use o scala alues may educe p ocessing ime. Ne e heless, he co ela ion among diffe en colou planes is neglec ed by hese app oaches, and hence, some use ul ea u es may be los . Only ac oss he in eg a ion o he in o ma ion in all channels, a colou image can be effec i ely segmen ed. The use o he ull colou ep esen a ion on au oma ic diagnosis assessmen is impo an . Howe e , equally sig- nifican is he ole o colou pe cep ion in objec ecog- ni ion and scene unde s anding bo h o humans and in elligen ision sys ems [45, 46]. Among all he possibil- i ies o colou image ep esen a ion, he use o a uni o m colou space, ha is, a ep esen a ion o he image whe e colou dis ances a e co ela ed o pe cep i e diffe ences could benefi he quali y o he esul s. Some au ho s ha e used CIE L ∗ a ∗ b ∗ colou space due o i s uni o mi y and he possibili y o using ad ance colou me ics [47–49]. Au ho s in [50] used JCh colou space om he CIE- CAM02 colou appea ance model o OD ex ac ion. Al hough hese me hods p e end o ake ad an age o he comple e colou in o ma ion a ailable while being as close o human pe cep ion as possible, OD segmen a ion is a p oblem ye o be sol ed. Fo ins ance, au ho s in [50] p ocess only he g ey le el plane J. Me hods illus a ed in [34] and [48] would no wo k in images whe e colou diffe ences be ween OD and backg ound a e no signifi- can . The compu a ion o colou de i a i es, only in ce - ain pixels loca ed in a adius cen ed on he OD, was p esen ed in [47]. In such app oach, he final ob ained bo de is dependen on he sepa a ion o he adial lines. Rega ding cup segmen a ion, he same limi a ions abou colou spaces and human pe cep ion could be applied. I is impo an o no e ha he cup a ea is mo e difficul o seg- men han OD due o essels, bo de asymme y, and colou a iabili y. I usually happens ha images p esen no b igh yellowish a ea a all bu he cup is s ill he e. In hese cases, he cup edge is dic a ed by essel bends. Fo his eason, he majo i y o app oaches p esen ed in he pas may no gi e accu a e esul s when dealing wi h complex image da abases [5, 8, 10, 26–28, 30–34, 40, 41, 44, 49, 51–54]. Al hough he abo emen ioned echniques p esen ele- an esul s, he e a e s ill some weak poin s ha should be add essed: (1) The use o colou in o ma ion is usually limi ed o sepa a ely p ocessing each colou plane. Howe e , e inal undus images a e ec o - alued colou images and he e o e, hei analysis in a scala ash- ion could add some e o s o he p ocess. (2) Medical image pe cep ion is no add essed by he majo i y o he app oaches. The use o uni o m col- ou spaces o ad anced colou dis ances is limi ed. (3) The complexi y o he p oposed echniques is high making he ools unconnec ed and he me hod inelegan . (4) The p oposed me hods ely on essel de ec ion and inpain ing equen ly. In many app oaches, essel bends mus be compu ed as well. The e o s in his ini ial s age will p opaga e o he es o he algo i hm. (5) The me hods a e designed and es ed in he same image da abases, wi h a limi ed numbe o images. These da abases a e p i a e in mos cases. The gold s anda d is usually no a ailable. The e o e, he eal quali y o he ool canno be accessed. To add ess hese issues, he p esen echnique has he ollowing key poin s: (1) The me hod is simple. I has h ee s ages only. (2) I does no ely on he segmen a ion, inpain ing, o de ec ion o essel bends o o he e inal image s uc u es. (3) The me hod makes use o a uni o m colou model along wi h a colou pe cep ion-adap ed dis ance image. (4) The echnique has been ex ensi ely alida ed. I has been designed on public image da abases. The esul (a) (b) (c) (d) Figu e 2: The OD p esen s a gene al appea ance making i sui able o i s au oma ic de ec ion. Howe e , he e is a high a iabili y among popula ion: (a) clea colou change and ed hue, well-defined bo de ; (b) sub le colou change and ed hue, diffuse bo de ; (c) sub le colou change and pale yellow uzzy bo de ; and (d) b igh yellow diffuse bo de p esence o pe ipapilla y a ophy. 3Jou nal o Heal hca e Enginee ing o he es on hese da abases is p esen ed. Once he ool has been ained, a second expe imen is pe - o med using a comple ely diffe en da abase. (5) As glaucoma diagnosis on e inal undus images is cu en ly pe o med mainly by manual inspec ion, eehand, o ellipse fi ed, we do no p esen only he segmen a ions o hese a eas bu also he CDR measu emen s ha a e au oma ically calcula ed. We compa e he esul s o he echnique wi h he gold s anda d p o ided by expe s wi h bo h o he app oxima ion me hods gene ally used. 2. Ma e ials and Me hods 2.1. Image Da abase. We cons uc ed wo image da abases, namely, Da ase 1 and Da ase 2, each con aining 60 e inal undus images. These 120 images spanned a g ea di e si y o e inal con en . The key poin on selec ing he images was ha hey needed o be ep esen a i e o he con en ha he algo i hm will encoun e on i s p ac ical use. The e o e, we explo ed se en publicly a ailable da abases [55–61] o c ea e Da ase 1. Six y images ha offe a wide ange o appea ances, illumina ion, and colou s we e selec ed as shown in Figu e 3 (a de ailed lis o images can be accessed in he supplemen a y ma e ial a ailable online a h ps://doi. o g/10.1155/2017/5953621). The image da abase comp ised heal hy and unheal hy images o pa ien s suffe ing om glaucoma in some cases and also diabe ic e inopa hy. Two expe s pe o med manual anno a ion o all o he e inog a- phies since public gold s anda d was no a ailable. Da ase 2 included 60 images om he Su ge y Depa - men and Glaucoma Uni o he Uni e si y Hospi al Pue a del Ma o Cadiz (Spain). Images we e anno a ed by wo expe s and we e used as an independen es se . The complexi y o he images o Da ase 2 was high, as can be app ecia ed in Figu e 4, including challenging cases wi h no isible cup, p esence o abno mali ies, o diffuse bo de s. In Da ase 2, wo gold s anda ds we e used: (a) The fi s gold s anda d consis ed o he eehand d awing on he e inal image. I was a edious and ime-consuming ask due o he difficul y o selec ing he p ecise bo de o he OD and cup egions. (b) As a second gold s anda d, he OCT so wa e pe - o med ellipse fi ing. Expe s ma ked up ou poin s o bo h egions. I mus no be con used wi h image p ocessing-based ellipse fi ing. The OCT so wa e only compu es he equa ion o an ellipse based on he ou manually ma ked poin s. No image in o ma- ion is aking in o accoun . Once he da abases we e buil , a egion o in e es (ROI) was au oma ically selec ed in o de o educe compu a ional (a) (b) (c) (d) Figu e 3: Da ase 1 comp ises a wide ange o OD and cup appea ances due o hei diffe en na u e, popula ion, and acquisi ion de ices. (a) (b) (c) (d) Figu e 4: Da ase 2 is a p i a e da abase compounded by e inal undus images acqui ed by he same de ice. The o e all complexi y is high due o he p esence o many diffe en appea ances: uzzy edges, sub le colou changes, a ophies, and so o h. 4 Jou nal o Heal hca e Enginee ing ime [11, 17, 18, 21, 27]. The ROI a ea co esponded o a egion wi h he ollowing cha ac e is ics: (i) Squa e shape (ii) Cen ed on he OD (iii) Wi h an a ea equal o 1/7 o he e ina size I mus be no iced ha any o he me hods p esen ed in li e a u e abou OD loca ion could be used in his s ep [9–47]. Howe e , he con ibu ion o he p oposed ech- nique ela es on OD and cup de ec ion and no OD localiza ion. In o de o effec i ely e alua e he pe o - mance o he echnique no dis u bing i wi h possible e o p opaga ion, we ha e manually inpu OD cen es o all o he images. 2.2. Vec o -Based Colou De i a i es. Image de i a i es we e used in o de o iden i y OD and cup bounda ies, due o hei capabili y o cap u ing changes on a ce ain pixel neighbou hood. De i a i es can be compu ed in se - e al di ec ions by o a ing he ke nel be o e pe o ming he con olu ion. Re inog aphies a e colou images. Consequen ly, he edges should be ound by looking o colou changes. Edge de ec ion in colou images is usually pe o med by applying he de i a i e ke nels o he h ee colou channels indepen- den ly and hen by combining he esul s. These kinds o me hods do no ake in o accoun he co ela ion among colou channels, and, he e o e, hey end o miss edges ha ha e he same s eng h bu in opposi e di ec ions in wo o hei colou componen s [62]. In an a emp o a oid his issue, we ha e adop ed he echnique p oposed in [62], whe e colou images a e ea ed as wo dimensional (pixel loca ion), h ee-channel (colou planes) ec o fields. Then, hey can be cha ac e ized by a disc e e in ege unc ion I(x,y) ha can be w i en as ollows: Ix,y=CP 1x,y,CP 2x,y,CP 3x,y, 1 whe e CP1x,y,CP 2x,y, and CP3x,yco espond o colou channels and x,y o pixels’loca ions. Fo ins ance, in RGB colou space CP1x,y=Rx,y,2 CP2x,y=Gx,y, 3 CP3x,y=Bx,y4 The magni ude o maximum a ia ion a pixel x,ywi h an o ien a ion o 0 ° is defined as ollows [62]: Bx,y=ΔVx,y, 5 whe e ΔV,i Euclidean dis ance (ΔE) is used, is defined as ollows: ΔVx,y=ΔEV +x,y,V−x,y6 The quan i ies V+,V−,H+,and H−a e he con olu ion ke nels whose ou pu s a e ec o s co esponding o he local a e age colou s. Le he edge masks (k)be − 1 − 2 − 3 000 + 1 + 2 + 3 , 7 and he image neighbou hood (W) w1w2w3 w4w5w6 w7w8w9 8 Each wi,i=1,…,9, is a ec o wi h h ee componen s co esponding o each colou plane. Then, V+= + 1w7+ + 2w8+ + 3w9, 9 V−= − 1w1+ − 2w2+ − 3w310 To imp o e adap a ion o human pe cep ion o he me hod, ins ead o using Euclidean dis ance o mula as in (3), he echnique p esen ed in [63] was ollowed. The colou space CIE L ∗ a ∗ b ∗ was selec ed due o i s uni o mi y. Fo he colou dis ance o mula, CIE94 was adop ed ins ead o Euclidean, due o i s bes pe o mance and lowe compu a- ional ime when compa ed o o he pe cep ion-adap ed colou diffe ences such as CIEDE2000. Then, wi=wL∗ i,wa∗ i,wb∗ i, i=1,…,9 11 Equa ion (9) can be ew i en as V+= + 1wL∗ 7+ + 2wL∗ 8+ + 3wL∗ 9, + 1wa∗ 7+ + 2wa∗ 8+ + 3wa∗ 9, + 1wb∗ 7+ + 2wb∗ 8+ + 3wb∗ 9 =VL∗ +,Va∗ +,Vb∗ +, 12 ha is, each local a e age colou will ha e i s h ee-colou componen s. Following he same p ocedu e, (10) can be ew i en as ollows: V−=VL∗ −,Va∗ −,Vb∗ −13 Subsequen ly, ΔV=ΔE94 V+,V−, 14 whe e ΔE94 is he CIE94 colou dis ance be ween he co e- sponding ec o s: 5Jou nal o Heal hca e Enginee ing ΔE94 =ΔL∗ kLSL 2 +ΔC∗ ab kCSC 2 +ΔH∗ ab kHSH 2 15 Fo he case o ΔV, ΔL∗=VL∗ +−VL∗ −, Δa∗=Va∗ +−Va∗ −, Δb∗=Vb∗ +−Vb∗ −, C∗ +=Va∗ + 2+Vb∗ + 2, C∗ −=Va∗ − 2+Vb∗ − 2, ΔC∗ ab =C∗ +−C∗ −, ΔH∗ ab =Δa∗2+Δb∗2+ΔC∗ ab 2, SL=1, SC=1+00045C∗ +, SH=1+00015C∗ −, kL=kC=kH=1 16 The magni ude o maximum a ia ion can be compu ed using (4). In he case ha we wan o calcula e colou changes in o he di ec ions a he han e ical (0 ° ), we only need o o a e he mask on (7) o he desi ed o ien a ion. As s a ed on he in oduc ion, he OD and cup egions p esen cha ac e is ic appea ances di ec ly ela ed o colou . Howe e , he absolu e colou alue o a pixel should no be us ed. Figu es 2–5 show how colou a iabili y is oo high o de e mine a specific colou ange o e e y OD and cup a ea. On he con a y, e e y OD and cup bo de p esen a change o colou when compa ed o hei su ounding pixels. In o he wo ds, absolu e colou alues a e no disc imina i e bu hei ela i e changes on he e ina can be (Figu e 6). In he p esen app oach, we ha e aken ad an age o his ela i e change o colou o de ec pixels belonging o OD and cup edges. We compu ed Sobel ec o -based colou de i a i es in 25 o ien a ions ( om 0 ° o 360 ° wi h a sepa a- ion in e al o 15 ° ) o e e y pixel wi hin he image. To implemen he Sobel ope a o , he mask o (10) a 0 ° was −1−2−1 000 121 , 17 (a) (b) (c) (d) Figu e 5: Cup egion segmen a ion is a challenging ask due o i s wide ange o appea ances: (a) b igh yellow, well-defined bo de and small size, (b) no pe cep ible colou change, (c) pale yellow, well-defined bo de and medium size, and (d) pale yellow, diffuse bo de and la ge size. Figu e 6: Colou changes ep esen ed by g adien a ows ma ked in blue offe he necessa y in o ma ion o OD and cup segmen a ion. 6 Jou nal o Heal hca e Enginee ing while o 45 ° , he mask was −2−20 −101 012 18 Equa ion (2) was e alua ed o measu e he maximum colou a ia ion o e e y pixel and o ien a ion. Figu e 7 shows some examples whe e each pixel alue co esponds o i s B alue on ha di ec ion. 2.3. Classifica ion Based on Bagged T ees. OD de ec ion and cup de ec ion we e achie ed by classi ying each pixel on he image ega ding i s ec o -based colou de i a i es and i s dis ance o he OD cen e. Se e al classifie s we e used. The bes pe o mance was ob ained wi h a bagged ees classifie , as i will be explained in Resul s. The e o e, he bagged ees classifie s will be desc ibed b iefly in his sec ion. The idea o bagging is o ob ain he bes model by com- bining he esul s o mul iple weak classifie s in o a single and s ong one [64]. In a bagged ee, he basic classifie is a decision ee. Da a is di ided in o T aining se s o size n, and T decision ees a e ained wi h hose se s, each one ying o fi he model. The Tdecisions a e finally combined wi h a majo i y o ing ule. Bagging leads o imp o emen s o uns able p ocedu es [65]. 3. Resul s and Discussion The disc imina i e powe o colou de i a i es when he OD and cup a e de ec ed has been es ed. To ha pu pose, a ea- u e ec o mus be buil o e e y pixel on each ROI. This ea u e ec o is he inpu o he classifie ha will assign a p obabili y alue o ha pixel ega ding i s sui abili y o belonging o he OD, cup, o backg ound. The ea u e ec o con ains all o he colou a ia ion alues o he 25 o ien a- ions, o iginal CIE L ∗ a ∗ b ∗ alues o he pixel, i s dis ance, and he angle ega ding he cen e o he OD and i s posi ion. These 32 ea u es combine he a p io i colou and spa ial knowledge abou he OD and he cup. The me hod has been ex ensi ely alida ed and es ed wi h wo expe imen s ca ied ou on bo h o he da abases de ailed in Sec ion 2. 0°45°270° Figu e 7: Vec o -based colou de i a i es o h ee o he compu ed o ien a ions: 0 ° ,45 ° , and 270 ° . 7Jou nal o Heal hca e Enginee ing 3.1. Image Da abase Da ase 1. This da abase is composed o 60 e inal undus images om six diffe en public da abases. Da ase 1 images p esen a a ie y o appea ances, illumina- ion condi ions, e inal s uc u es, and so o h. Conse- quen ly, i is expec ed ha an algo i hm de eloped using his da abase will be highly obus . A o al o six classifie s we e ained and alida ed: (i) Simple ee (ST). (ii) Bagged ee (BT). This classifie was in oduced in Sec ion 2.3. (iii) Complex ee (CT). This is a decision ee wi h many lea es ha makes many fine dis inc ions be ween classes. (i ) Linea disc iminan (LD).Decisions a e made by es ima ing, wi h Bayes heo em, he p obabili y ha a new se o samples belongs o each class. ( ) Quad a ic disc iminan (QD).This classifie is an ex ension o LD whe e he e ogeneous a iance- co a iance ma ices a e conside ed. ( i) kNN Euclidean. kNN does no use a model o fi he aining da a and subsequen ly classi y he new samples [66]. Model selec ion was pe o med using c oss- alida ion. Fo each classifie , he accu acy o he bes pa ame e se ing was compa ed. Da ase 1 was used as a old c oss- alida ion se , which was subsequen ly and epea edly di ided in o ain and alida ion se s. Fo his in e nal alida ion, 10- old c oss- alida ion was pe o med. In each o he en olds, he classifie was ebuil om sc a ch. This en i e p ocedu e is epea ed 10 imes. The epo ed accu acy was he a e age o e he accu acies o each old. The abili y o he algo i hm o pe o m an accu a e clas- sifica ion o OD, cup, and backg ound pixels was measu ed by i s sensi i i y while i s abili y o de e mine he pixels ha do no belong o each o he h ee classes was exp essed by i s specifici y. Posi i e p edic i e alue (PPV) and nega i e p edic i e alue (NPV) we e also compu ed o gi e an idea o he p opo ions o ue posi i es and ue nega i es. Table 1 shows he pe o mance me ics ha we e used in his s udy o e alua e he use o ec o -based colou de i a i es in combina ion wi h each o he six classifie s. The analysis o Table 1 e eals ha BT classifie , which showed a class-wise-a e aged accu acy o 95.02%, p o ides he bes combina ion o classifie and ec o -based colou de i a i es. This classifie showed a specifici y o 99.23% o he OD class and 99.80% o he cup class, while p ese ing a sensi i i y o 91.75% o he OD and 90.63% o he cup. PPV alues we e 90.74% o he OD class and 94.83% o he cup class. NPV was 99.32% and 99.62% o he OD and cup classes, espec i ely. 3.2. Image Da abase Da ase 2. This image da abase com- p ised 60 images p o ided by he Uni e si y Hospi al Pue a del Ma , Cadiz, Spain. Images we e acqui ed in a ou ine sc eening p ocess wi h he same acquisi ion de ice. Ex e nal alida ion es ablishes models’ anspo abili y and gene alizabili y [67]. In his s udy, independen Da ase 2 was used o ex e nally alida e he ully ained classifie p e iously selec ed using c oss- alida ion in Da ase 1. All samples in Da ase 1 we e used o ain he BT model. Two expe s made wo anno a ions o each o he images on he da abase: (i) F eehand: Two glaucoma specialis s me iculously anno a ed he exac bo de o he OD and he cup. This p ocess was ime consuming al hough cons i u - ing he mos exac e e ence o e o calcula ion. This anno a ion was conside ed he gold s anda d o ou expe imen s. (ii) Ellipse based: The OD and cup edges we e ob ained by building an ellipse ha con ained ou poin s Table 1: Model pe o mance e alua ion unde Da ase 1 da abase. The h ee classes a e backg ound (class 1), op ic disc (class 2), and cup (class 3). (a) Simple ee, (b) bagged ees, (c) KNN, (d) complex ee, (e) linea disc iminan , and ( ) quad a ic disc iminan . Indica o a (%) b (%) c (%) d (%) e (%) (%) Accu acy 95.54 98.66 97.22 96.63 85.55 89.11 Class-wise-a e aged accu acy 82.70 95.02 89.39 86.99 53.28 61.39 Sensi i i y class 1 98.17 99.61 99.31 98.88 90.59 95.19 Sensi i i y class 2 74.35 91.75 82.23 79.16 46.84 40.70 Sensi i i y class 3 77.21 90.63 78.94 79.55 46.42 45.16 Specifici y class 1 89.43 96.20 91.24 90.90 74.78 63.71 Specifici y class 2 97.29 99.23 98.61 98.07 91.73 94.56 Specifici y class 3 99.23 99.80 99.50 99.43 95.93 98.23 PPV class 1 98.62 99.51 98.87 98.82 96.52 95.29 PPV class 2 69.30 90.74 82.91 77.12 31.77 38.07 PPV class 3 80.17 94.83 86.40 85.04 31.55 50.81 NPV class 1 86.35 96.93 94.51 91.30 50.76 63.20 NPV class 2 97.88 99.32 98.54 98.28 95.45 95.10 NPV class 3 99.08 99.62 99.15 99.18 97.79 97.79 8 Jou nal o Heal hca e Enginee ing manually ma ked by he expe s. This s a egy showed a lowe compu a ional ime. Howe e , he ob ained bo de was no as accu a e as ha in he eehand app oach. 3.2.1. Quali y o he Segmen a ion. Afi s expe imen was pe o med o effec i ely know whe he a pos p ocessing s ep would e en ually imp o e he quali y o he esul s. To his pu pose, we ook he p obabili ies o each pixel o belonging o each class and, om his in o ma ion, we buil wo p oba- bili y images (see Figu e 8) ha a e he basis o he final pos p ocessing. This las s ep was pe o med in wo ways: (i) Ac i e con ou (AC) based: The p obabili y images ha co esponded o he OD and he cup we e h e- sholded o ob ain an ini ial mask. The h eshold was au oma ically ob ained wi h O su’s echnique. The final OD and cup we e segmen ed on he base o his bina y image and e ol ing on he co esponding p obabili y alues. The Chan-Vese model [68] wi h a numbe o i e a ions expe imen ally fixed o 20 was used o AC. (ii) AC and ellipse fi ing: This pos p ocessing consis ed on au oma ically adap ing an ellipse o he bounda y pixels ob ained wi h he p e ious pos p ocessing. These wo pos p ocessing s eps we e added o emula e expe s’segmen a ions, which we e in gene al smoo he han ou algo i hm’s esul s (Figu e 9(a)). AC p o ides he so ness equi ed while p ese ing he shapes (Figu e 9(b)). The ellipse-fi ed esul was in ended o be e compa e wi h manually ma ked ellipses p o ided wi h he da abase (Figu e 9(c)). (a) (b) (c) Figu e 8: (a) O iginal ROI image, (b) OD p obabili y image, and (c) cup p obabili y image. (a) (b) (c) Figu e 9: Resul images o he o iginal ROI o Figu e 8. Whi e colou co esponds o he cup, g ey colou o he OD, and black colou o he backg ound. Labels assigned by he classifie : (a) wi hou pos p ocessing, (b) wi h AC, and (c) wi h AC and ellipse fi ing. Table 2: Bagged ee model pe o mance e alua ion unde Da ase 2 da abase. The h ee classes a e backg ound (class 1), op ic disc (class 2), and cup (class 3). (a) Wi hou pos p ocessing, (b) smoo hed wi h AC, and (c) smoo hed and ellipse fi ing. Indica o a (%) b (%) c (%) Accu acy 94.54 94.75 94.61 Class-wise-a e aged accu acy 79.66 81.95 81.19 Sensibili y class 1 97.01 96.63 96.34 Sensibili y class 2 73.07 83.02 83.18 Sensibili y class 3 81.45 76.20 78.75 Specifici y class 1 90.30 93.61 94.47 Specifici y class 2 96.36 95.73 95.56 Specifici y class 3 99.01 99.43 99.33 Posi i e p edic i e alue class 1 98.68 99.13 99.24 Posi i e p edic i e alue class 2 62.37 61.61 60.75 Posi i e p edic i e alue class 3 77.93 85.10 83.58 Nega i e p edic i e alue class 1 80.11 78.74 77.47 Nega i e p edic i e alue class 2 97.74 98.55 98.57 Nega i e p edic i e alue class 3 99.20 98.98 99.09 9Jou nal o Heal hca e Enginee ing