Neuromuscular disease classification system

Neu omuscula disease classi ica ion sys em Au o a Sáez,aBegoña Acha,aAdo ación Mon e o-Sánchez,bEloy Ri as,cLuis M. Escude o,band Ca men Se anoa aUni e si y o Se ille, Depa men o Signal Theo y and Communica ions, ETSI, 41092, Se ille, Spain bRocío/CSIC/Uni e sidad de Se illa, Hospi al Uni e si a io Vi gen del, Ins i u o de Biomedicina de Se illa, 41013, Se ille, Spain cHospi al Uni e si a io Vi gen del Rocío, Depa men o Pa hology, 41013, Se ille, Spain Abs ac . Diagnosis o neu omuscula diseases is based on subjec i e isual assessmen o biopsies om pa ien s by he pa hologis specialis . A sys em o objec i e analysis and classi ica ion o muscula dys ophies and neu ogenic a ophies h ough muscle biopsy images o luo escence mic oscopy is p esen ed. The p ocedu e s a s wi h an accu a e segmen a ion o he muscle ibe s using ma hema ical mo phology and a wa e shed ans o m. A ea u e ex ac ion s ep is ca ied ou in wo pa s: 24 ea u es ha pa hologis s ake in o accoun o diagnose he diseases and 58 s uc u al ea u es ha he human eye canno see, based on he assump ion ha he biopsy is conside ed as a g aph, whe e he nodes a e ep esen ed by each ibe , and wo nodes a e connec ed i wo ibe s a e adjacen . A ea u e selec ion using sequen ial o wa d selec ion and sequen ial backwa d selec ion me hods, a classi ica ion using a Fuzzy ARTMAP neu al ne wo k, and a s udy o g ading he se e i y a e pe o med on hese wo se s o ea u es. A da abase consis ing o 91 images was used: 71 images o he aining s ep and 20 as he es . A clas- si ica ion e o o 0% was ob ained. I is concluded ha he addi ion o ea u es unde ec able by he human isual inspec ion imp o es he ca ego iza ion o a ophic pa e ns.©2013 Socie y o Pho o-Op ical Ins umen a ion Enginee s (SPIE) [DOI: 10 .1117/1.JBO.18.6.066017] Keywo ds: segmen a ion; wa e shed; uzzy classi ica ion; ea u e ex ac ion; neu omuscula disease; g aph heo y. Pape 130052R ecei ed Jan. 30, 2013; e ised manusc ip ecei ed May 13, 2013; accep ed o publica ion May 20, 2013; published online Jun. 26, 2013. 1 In oduc ion Neu omuscula diseases co e a la ge g oup o pa hologies wi h a huge he e ogeneous e iology and cou se. Al hough all neu o- muscula diseases a e p og essi e in na u e, hey mani es in a wide age ange wi h di e en deg ees o se e i y. The mos common classi ica ion elies on he componen o he neu o- muscula sys em ha is a ec ed. Hence, we can ind myopa- hies, in which he muscle is p ima ily a ec ed, o neu ogenic a ophies (NA), in which he mo o neu on is a ec ed. To s udy he pa ien a ec ions, he pa hologis exam- ines muscula biopsies unde mic oscope ocusing in o a mo - phological muscula ibe analysis. The e alua ion o he changes in he mo phological cha ac e is ics o a gi en biopsy wi h espec o he no mal muscle is one o he main ea u es o he diagnos ics o a neu omuscula disease. Howe e , he manual mo phome ic app oach and in e p e a ion o muscle biopsy ma e ial is a subjec i e, edious, and ime-consum- ing ask.1 Cu en ly, a issue his opa hology slide can be digi ized and s o ed in digi al-image o m. This p og ess has allowed he de elopmen o compu e -assis ed diagnosis o disease de ec- ion, diagnosis, and p ognosis p edic ion o complemen he opinion o he pa hologis .2In his sense, in he ecen li e a u e wo ks ela ed o g ading o p os a e cance ,3de ec ion o ce i- cal cance ,4classi ica ion o hepa ocellula ca cicoma,5de ec- ion o ce ical cell nuclei,6o simple de ec ion o di e en ypes o cells7–9can be ound. Focusing on he s udies o muscula ibe s, we can ind some wo ks ha add ess he seg- men a ion o ibe s in muscula biopsies,10–12 he classi ica ion o muscle- ibe ype,13–16 and he ex ac ion o mo phome ic ea u es.17 Howe e , s udies o he cha ac e iza ion o neu o- muscula disease based on image p ocessing ha e no been ound in he cu en li e a u e. This pape p esen s an analysis o luo escence mic oscopy images o muscula biopsy o ob ain use ul in o ma ion o he diagnosis and he se e i y g ading o di e en neu omuscula diseases. To achie e his, ou da abase consis s o 91 images belonging o con ol biopsies (no disease), biopsies a ec ed by muscula dys ophies (MD), and biopsies a ec ed by NA. The s udy is based no only on he ex ac ion o ea u es ela ed o he cha ac e is ics ha he pa hologis akes in o accoun o diagnosis, bu also on he sea ch o ea u es wi h inhe en p op- e ies ha escape he e alua ion o he pa hologis and which could be mo e e icien o he classi ica ion o he di e en muscula images. In his sense, he pape p oposes an ex ac ion o mo phome ic and s uc u al in o ma ion based on he assump ion ha he biopsy is conside ed as a g aph, whe e he nodes a e ep esen ed by each ibe , and wo nodes a e con- nec ed i wo ibe s a e adjacen . This was mo i a ed by he wo k o Escude o e al.,18 in which he in oduc ion o a ne wo k allowed one o desc ibe he epi helial o ganiza ion objec i ely. To ge he ea u e ex ac ion, an accu a e segmen a ion was equi ed. In his pape , ma hema ical mo phology and a wa e - shed ans o m a e used o add ess his ask. A uzzy classi ica- ion based on a neu al ne wo k a chi ec u e is p esen ed. Finally, a s udy o he se e i y g ading is ca ied ou o analyze he esul s. In Sáez e al.,19 some o hese esul s, analyzed om he biological poin o iew using di e en aining da a se s, a e p esen ed. Add ess all co espondence o: Au o a Sáez, Uni e si y o Se ille, Depa men o Signal Theo y and Communica ions, ETSI, 41092, Se ille, Spain. Tel: +34 954 486091; Fax: +34 95 448 7341; E-mail: [email p o ec ed] o Luis M. Escude o, Rocío/CSIC/Uni e sidad de Se illa, Hospi al Uni e si a io Vi gen del, Ins i u o de Biomedicina de Se illa, 41013, Se ille, Spain. Tel: +34 955923048; Fax: +34 95 461 7301; E-mail: [email p o ec ed] 0091-3286/2013/$25.00 © 2013 SPIE Jou nal o Biomedical Op ics 066017-1 June 2013 •Vol. 18(6) Jou nal o Biomedical Op ics 18(6), 066017 (June 2013) Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/25400/ on 04/06/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx The es o he pape is o ganized as ollows: he ype o images and he neu omuscula diseases a e p esen ed in Sec. 2; in Sec. 3 he p ocedu e ollowed is explained, which includes a segmen a ion me hod, a ea u e ex ac ion s ep, a ea u e selec- ion s ep, a classi ica ion, and he se e i y g ading. In u n, each sec ion con ains a subsec ion o esul s. Finally, a discussion o he esul s is p esen ed. 2 Muscle Biopsy Images and Neu omuscula Diseases The muscula ibe s a e o ganized in ascicles, which a e su - ounded by a laye o connec i e issue named he pe imysium. Be ween he ibe s wi hin a ascicle appea s he endomysium, a mesh o loose connec i e issue composed o ine collagen and e icula ibe s. Skele al muscles ibe s can be classi ied in o wo main ypes: ype I o slow ibe s ( hey ha e a slow con ac ion eloci y) and ype II o as ibe s ( as con ac ion eloci y). They a e dis ibu ed in a diso de ed mosaic pa e n, and appea wi h a simila size along he ascicles. The ans e sal sec ion o a no mal muscle ep esen s he ibe s wi h a polygonal shape su ounded by a hin mesh o collagen.20 Muscula biopsies we e p ocessed by he s anda d me hods o eezing and cu ing wi h c yos a . The muscula ibe and he collagen con en we e de ec ed by luo escence mic oscopy. The an ibodies mouse an i-myosin hea y chain (slow), mouse an i- myosin hea y chain ( as ), and abbi an i-collagen ype VI we e used using a s anda d p o ocol o immunos aining. The sec ions we e incuba ed wi h Alexa luo 488 and Alexa luo 568 sec- onda y an ibodies. All he slides we e analyzed unde a luo es- cence mic oscope (BX-61 Olympus wi h a DP70 came a) using a me cu y lamp h ough a 470 o 490 nm o 560 o 579 nm band-pass il e o exci e Alexa luo 488 o Alexa luo 568, espec i ely. The s ained cells we e pho og aphed, and wo high- esolu ion images (size o 4080 ×3072 pixels) we e ob ained. A o al o 91 images om 70 muscle biopsies, s o ed in he Tissue Bank o he Hospi al Uni e si a io Vi gen del Rocío, Se ille, Spain, we e p ocessed. An RGB image was c ea ed om he wo images ob ained. The ed componen is he image ob ained when he biopsy is exci ed a 560 o 579 nm [Fig. 1(b)], and he g een componen is he image ob ained when he biopsy is exci ed a 470 o 490 nm [Fig. 1(c)]. Figu e 1(a) shows a sample o he esul ing RGB image. Slow ibe s in a da k colo , as ibe s in a eddish colo , collagen in a g eenish colo , and capilla ies as small da k s uc u es among he collagen can be obse ed. All o iginal images ha e he same esolu ion (4080× 3072 pixels). Howe e , he igu es shown in his pape ep esen only a pa o he comple e image (1150 ×1150 pixels) o co - ec ly isualize he de ails. The yellow ba a he op o Fig. 1(a) ep esen s he image scale co esponding o 200 μm. In Fig. 1(d)–1( ), examples o he a iabili y in he images a e shown. They belong o he pa ien s wi h di e en ages and di e en diseases. In his pape , h ee mo phological pa e ns p esen ed in muscle biopsies a e s udied. The i s one is he no mal pa e n (no disease). The second is he dys ophic pa e n. MD a e a ype o myopa hy, which a e cha ac e ized by a wide a ia ion in ibe size, a ounded shape o a ophic ibe s, and ib osis (an inc ease o endomisial collagen). This con as s wi h he hi d pa e n, NA ha p esen angula ed a ophic ibe s, la ge g oups o a ophic ibe s, ascicula a ophy, and loss o he andom Fig. 1 (a) Muscle biopsy image. The yellow ba a he op ep esen s he image scale. I co esponds o 200 μm, and i is he same o he es o he images. (b) R-componen . (c) G-componen . (d) Example o muscle biopsy image a ec ed by neu ogenic a ophy (NA). (e) Example o muscle biopsy (no disease) image belonging o a child. ( ) Example o muscle biopsy image a ec ed by muscula dys ophy (MD). Jou nal o Biomedical Op ics 066017-2 June 2013 •Vol. 18(6) Sáez e al.: Neu omuscula disease classi ica ion sys em Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/25400/ on 04/06/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx checke boa d dis ibu ion o ibe ypes o ibe - ype g ouping.21 In addi ion, i is possible o ind a muscle ha can show a com- bina ion o neu ogenic and myopa hic ea u es. Fo he s udy, we ha e 91 images om 70 subjec s; 41 con ol images, 27 dys- ophy images, and 23 a ophy images. 3 Me hodology The mo phological analysis is a undamen al ool o he diag- nosis o neu omuscula diso de s.22 The aim o his pape is o analyze muscle biopsies in a mo phological and s uc u al way, which allows one o de elop a diagnosis help ool o he neu o- muscula diseases explained in he p e ious sec ion. The p o- cedu e ollowed is desc ibed in Fig. 2. Each block o he low diag am in Fig. 2is explained in he ollowing sec ions. 3.1 Segmen a ion A p e equisi e o classi y any disease is he abili y o au oma i- cally iden i y he s uc u es p esen in he image.2Mo eo e , o achie e a obus mo phological analysis, an accu a e segmen a- ion o he ibe s in he biopsy image is equi ed. Shape desc ip- ion and accu a e segmen a ion is possible i he ini ial localiza ion o he muscle ibe s is known.23 The e o e, in his pape he segmen a ion p ocess is di ided in o wo s eps: iden i ica ion o he muscle ibe localiza ion by applying mo - phological ope a o s, and accu a e de ec ion o he ibe con- ou s by a wa e shed ans o ma ion. I is impo an o no e ha he segmen a ion me hod mus be au oma ic and alid o all ypes o images [see Fig. 1(a) and 1(d)–1( )]. 3.1.1 Fibe localiza ion The biopsy images we wo k wi h p esen di e en s uc u es such as collagen, muscle ibe s, capilla ies, and a e ac s. The aim o his s ep is he co ec iden i ica ion o he muscle ibe s. The p ocessing is pe o med on he G-componen o he image due o he high con as be ween he muscle ibe s and he collagen [see Fig. 1(c)]. Conside ing ha he ibe s a e da ke han he su ounding collagen, in ensi y alleys in he image a e sea ched. Fo his eason, he H-minima ans o m24 is applied o he G-componen image in o de o ge homogeneous minima alleys. This ans o m has been success ully used in di e en medical applica ions.25,26 The H-minima o H-maxima ans- o m is a powe ul ma hema ical ool o supp ess undesi ed min- ima o maxima. I allows one o ex ac egional minima whose dep h is lowe han o equal o he gi en h- alue. Regional min- ima a e connec ed componen s o pixels wi h a cons an in en- si y alue, and whose ex e nal bounda y pixels all ha e a highe alue. The H-minima ans o m24 is pe o med by HhðGÞ¼Rε GðGþhÞ;(1) whe e h ep esen s he gi en dep h, and Rand ε ep esen he econs uc ion and e osion ope a o s, espec i ely. The h- alue has a di ec in luence on he numbe o segmen ed egions. The la ge he h- alue, he ewe segmen ed egions. As he da kes egions o he G-componen image ep esen he muscles ibe s, we can in ui i ely assume ha he equi ed h- alue should be lowe han he a e age in ensi y o he image. Th ee di e en alues ha ook in o accoun he a e age in ensi y we e es ed: wo hi ds o he a e age in ensi y o G(h1), hal o he a e age in ensi y o G(h2), and one hi d o he a e age in ensi y o G(h3). They a e calcula ed as h¼K1 N·MX N i¼1 X M j¼1 Gði; jÞ;K¼2 3;1 2;1 3;(2) whe e Gði; jÞis he in ensi y alue o G-componen a he pixel ði; jÞand Nand Ma e he image pixel dimensions. The esul ing image is a bina y image, in which egions wi h a pixel alue o 1.0, displayed as whi e, ep esen candida e muscle ibe s. Figu e 3shows he in luence o he h ee h- alues on he numbe o segmen ed egions. 3.1.2 Resul s o he ibe localiza ion Resul s we e es ed in only 10 1150 ×1150 pixel images. The eason o his is ha he manual segmen a ion o each one is necessa y in o de o check he quali y o he me hod, and each image has hund eds o cells. So, he manual delinea ion o hem is a e y ime-consuming and edious ask. I should be no ed ha he numbe o cells segmen ed in hese es images ep esen s only a po ion o he numbe o cells o he en i e image. Table 1shows he numbe o cells segmen ed by he spe- cialis , and he numbe o he egions de ec ed by he H-minima ans o m wi h he h ee di e en h- alues. The numbe o de ec ed egions should be close o he numbe o manually seg- men ed cells, aking in o accoun ha he numbe o he de ec ed egions mus be highe han o equal o he numbe o segmen ed cells. I he alue is lowe , i will in ol e he loss o de ec ed cells. Al hough h1p o ides a numbe o de ec ed egions mo e simila o he numbe o cells es ima ed by he specialis (see Table 1), in some cases he numbe o de ec ed egions is lowe . This means ha some cells a e no de ec ed. Since he numbe shown he e ep esen s only a po ion o he numbe o cells in he en i e image, his numbe o los cells could inc ease. Fu he mo e, he o e segmen a ion is due o he exis - ence o a e ac s and capilla ies in he image. These egions can be emo ed by using mo phological ope a o s. Fo hese Fig. 2 Flow diag am o he sys em. Jou nal o Biomedical Op ics 066017-3 June 2013 •Vol. 18(6) Sáez e al.: Neu omuscula disease classi ica ion sys em Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/25400/ on 04/06/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx easons, he chosen h- alue is h2 ollowed by a p ocessing o he esul ing image using ma hema ical mo phology. The mo phological ope a o s used in his s ep a e summa- ized he e. Regions wi h an a ea smalle han 0.25% o he bigges dimension in pixels o he o iginal image ( aking in o accoun he esolu ion indica ed in Sec. 2) we e emo ed by applying mo phological opening.24 P esumably, small egions co espond o capilla ies. I egula i ies (holes) wi hin he egions de ec ed we e e illed by mo phological econs uc- ion.24 Finally, an e osion ope a o wi h a 3×3 pixel s uc u al elemen was applied o p e en ha adjacen cells we e joined. The esul s o he numbe o egions de ec ed a e his s ep a e shown in he ou h column o Table 2. Al hough he numbe o egions dec eased, i could be desi - able o be e adjus he numbe o de ec ed cells and he numbe o segmen ed cells. To his aim, inally, wo independen colo condi ions we e imposed on he de ec ed egions in o de o emo e hose ha did no uly co espond wi h cells. Fi s condi ion: hose egions whose a e age in ensi y o G-componen had a alue highe han 25% o he maximum in ensi y alue o Gin he da abase we e ejec ed as candida e ibe s, as he muscle ibe s p esen a low g een alue. Second condi ion: a his og am equaliza ion o he G-compo- nen was pe o med, he a e age in ensi y o his new image was calcula ed o all egions, and he maximum o hese alues was compu ed. Finally, hose egions whose a e age in ensi y was highe han he 90% o he maximum and an a e age alue o R-componen less han 20% o he maximum in ensi y alue o Rin he da abase we e emo ed. The h esholds men ioned we e expe imen ally ixed, and p o ide a co ec segmen a ion o he 91 images analysed. Wi h hese wo condi ions, capilla ies and a e ac s p esen among he collagen a e emo ed. The i h column o Table 2 p esen s he inal esul s. 3.1.3 De ec ion o ibe con ou s Once he localiza ions o he muscle ibe s a e iden i ied, an accu a e de ec ion o hei con ou s is equi ed o a la e obus mo phological analysis. Fo his objec i e, wo well-known echniques we e applied: le el se and a wa e shed ans o m. Bo h me hods a e b ie ly explained below. Le el se me hods27 ha e been widely used as a global app oach owa d he op imiza ion o ac i e con ou s o he Table 1 Numbe o manually segmen ed cells and numbe o 5 de ec ed egions by H-minima ans o m wi h di e en h- alues. Image Ncells manual h1h2h3 17776 79 83 2 103 108 128 144 3 40 250 295 366 4 45 256 530 1000 5 56 371 467 574 6 40 61 113 245 7 92 90 105 121 84948 56 60 9 310 377 419 483 10 63 136 155 184 No e: The bold alues indica e ha numbe o he egions de ec ed is lowe han numbe o manually segmen ed cells, his will in ol e loss o de ec ed cells. Table 2 Numbe o manually segmen ed cells and numbe o de ec ed egions by H-minima ans o m wi h h- alue ¼hal o he a e age in ensi y o G, a e he applica ion o mo phological ope a o s and wo colo condi ions. Image Ncells manual h2 Mo phological ope a o s Colo condi ions 1777977 77 2 103 128 103 103 3 40 295 59 40 4 45 530 104 48 5 56 467 59 56 6 40 113 47 41 7 92 105 91 91 8495652 49 9 310 419 333 310 10 63 155 74 65 Fig. 3 (a) Muscle biopsy image. (b) H-minima ans o m wi h h- alue ¼h1(h1¼60.85). (c) H-minima ans o m wi h h- alue ¼h2(h2¼45.63). (d) H- minima ans o m wi h h- alue ¼h3(h3¼30.42). Jou nal o Biomedical Op ics 066017-4 June 2013 •Vol. 18(6) Sáez e al.: Neu omuscula disease classi ica ion sys em Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/25400/ on 04/06/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx segmen a ion o objec s o in e es om he backg ound. In he li e a u e, nume ous wo ks ela ed o his echnique ha e been p oposed. In his pape , he me hod de eloped by Li e al.28 is used. In he epo ed wo k, an ene gy unc ion ies o main ain he le el se unc ion nea he signed dis ance unc ion, hus a oiding he need o e-ini ializa ion o he le el se unc ion.28 The ini ial cu e equi ed is he con ou o he bina y image esul ing om he muscle ibe localiza ion de ec ion, explained in he p e ious sec ion. In he wa e shed p ocedu e,29 an image is iewed as a opo- g aphic su ace: he highe he alue o a pixel, he highe he al i ude a he co esponding poin on he opog aphic su ace o elie . The wa e shed ans o m is usually applied o he g adien image. The minima in he g adien image will co espond o he si es wi hin homogeneous egions in he o iginal image. Howe e , he wa e shed algo i hm yields esul s wi h subs an ial o e segmen a ion; ha is, he numbe o segmen ed egions could be much la ge han desi ed, wi h egions being b oken in o mul iple smalle egions. This undesi able esul is due o he ac ha he g adien image used in he p ocess is sensi i e o noise. The p oblem o o e segmen a ion can be o e come wi h he use o ma ke s ha iden i y he objec s. The objec con- ou s in he g adien image can be seen as he highes c es -lines a ound he objec ma ke s. In ou case, he image g adien is calcula ed in he G-componen , and he in e nal and ex e nal ma ke s a e de i ed om he p e ious sec ion, whe e he aim was o iden i y he muscles ibe s. The bina y image esul ing om he p e ious sec ion cons i u es he in e nal make s. Meanwhile, he ex e nal ma ke s we e ex ac ed by applying ano he wa e shed ans o m o he bina y image used as in e nal ma ke s. 3.1.4 Resul s o he de ec ion o he ibe con ou s To e alua e he pe o mance o bo h me hods, 10 images man- ually segmen ed by he specialis we e used. The esul s we e e alua ed wi h he Jacca d coe icien and he Dice coe icien . The Jacca d index, also known as he Jacca d simila i y coe i- cien , is a s a is ical pa ame e used o compa ing he simila i y and di e si y o sample se s. I is de ined as he size o he in e - sec ion di ided by he size o he union o he segmen ed cells. Assuming wo se s co esponding o he segmen ed pixels ob ained om he manual segmen a ion (Pm) and he segmen ed pixels ob ained om he au oma ic me hod (Ps), Jacca d’s coe - icien is de ined as J¼jPm∩Psj jPm∪Psj:(3) The Dice coe icien , D, is also a simila i y measu e, which is de ined as D¼2jPm∩Psj jPmjþjPsj:(4) Resul s a e p esen ed in Table 3. As can be seen, he wa e - shed ans o m ou pe o ms le el se echnique in all images. Some image examples o bo h segmen a ions a e shown in Fig. 4. I is impo an o no e ha besides he accu acy in de ec ing he con ou s o he ibe s is highe in wa e shed [Fig. 4(b)] han in le el se s [Fig. 4(c)], he wa e shed ans o m is able o sepa a e wo independen cells al hough hey seem o be linked (see yellow ec angles). Table 3 Segmen a ion esul s o he wa e shed ans o m and le el se segmen a ion e alua ed by he Dice coe icien and he Jacca d coe icien . Image Dice coe icien Jacca d index Wa e shed Le el se s Wa e shed Le el se s 1 0.963 0.94 0.927 0.8969 2 0.966 0.93 0.933 0.868 3 0.97 0.95 0.942 0.912 4 0.969 0.952 0.939 0.908 5 0.969 0.949 0.94 0.903 6 0.956 0.938 0.915 0.881 7 0.96 0.917 0.922 0.843 8 0.975 0.947 0.951 0.89 9 0.971 0.917 0.94 0.84 10 0.973 0.922 0.948 0.854 A e age 0.967 0.936 0.934 0.879 No e: The bold alues indica e he bes esul . Fig. 4 (a) Muscle biopsy image. The yellow ec angle indica es linked cells. (b) Wa e shed segmen a ion esul . The yellow ec angle indica es a good esul . (c) Le el se segmen a ion esul . The yellow ec angle indica es a bad esul , because wo linked cells ha e been segmen ed as only one. Jou nal o Biomedical Op ics 066017-5 June 2013 •Vol. 18(6) Sáez e al.: Neu omuscula disease classi ica ion sys em Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/25400/ on 04/06/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx Fu he mo e, he compu a ional cos is also lowe o wa e - shed. Figu e 5shows ha he a io o compu a ional ime be ween bo h segmen a ions (consumed ime by le el se s/con- sumed ime by wa e shed) inc eases loga i hmically wi h he numbe o cells o segmen in he image. Fu he mo e, i should be no ed ha he ime spen on segmen ing he images by he p oposed me hod is signi ican ly less han ha needed by he specialis (see Table 4). The s eps ollowed in inal segmen a ion p ocess a e shown in Fig. 6. 3.2 Fea u e Ex ac ion To iden i y i a biopsy is a ec ed by a pa hology, an objec i e analysis o he biopsy is needed. The mo phological and s uc u al cha ac e is ics o he whole biopsy cons i u e a ec o , also called biosigna u e.23 The objec i e o his sec ion is o ex ac he ea- u es ha co espond o he isual a ibu es de ined by clinicians as pa icula ly impo an o he men ioned pa hology g ading and diagnosis as well as a s udy o new ea u es wi h inhe en p ope - ies ha escape om he e alua ion o he pa hologis and ha could be use ul o he cha ac e iza ion o hese diseases. In his sense, bo h mo phological and s uc u al ea u es a e p oposed. 3.2.1 Mo phological ea u es The mo phological cha ac e is ics can be desc ibed by shape o geome y.23 Fo mula ion o mo phological ea u es is an easy Fig. 5 Ra io be ween compu a ional cos o bo h segmen a ion me hods and numbe o cells segmen ed. Table 4 Time spen on segmen ing he images by he p oposed me hod and by he specialis . Image Manual segmen a ion (min) P oposed me hod (s) 120 33 2 33 35.5 314 33 4 22 33.7 5 11 30.14 68 30 729 41 815 34 940 45 10 20 33 A e age 21.2 31.5 Table 5 Fou een mo phological ea u es o he cells. 1 A e age a ea 2 S d. de . a ea 3 A e age a ea o slow cells 4 S d. de . a ea o slow cells 5 A e age a ea o as cells 6 S d. de . a ea o as cells 7 A e age majo axis 8 A e age mino axis 9 A e age a io axis 10 S d. De . a io axis 11 A e age con ex hull 12 S d. De . con ex hull 13 A e age angles 14 S d. de . angles Fig. 6 S eps ollowed in he ibe segmen a ion. Jou nal o Biomedical Op ics 066017-6 June 2013 •Vol. 18(6) Sáez e al.: Neu omuscula disease classi ica ion sys em Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/25400/ on 04/06/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx and as way o au oma ize he manual mo phological quan i i- ca ion, which is e y labo ious and subjec i e. Fea u es such as he a ea o each ibe , he a e age o he a eas o ype I (slow) and ype II ( as ) ibe s, iden i ied by he in ensi y a e age o R-componen , o he majo and he mino axis leng hs o he cells a e calcula ed. In Table 5, he 14 cha ac e is ics compu ed a e shown. 3.2.2 S uc u al ea u es To add ess his app oach, each biopsy image was in e p e ed as a g aph. G aphs a e e icien da a s uc u es o ep esen spa ial da a, and an e ec i e way o ep esen s uc u al in o ma ion by de ining a la ge se o opological ea u es.2Fo mally, a sim- ple g aph G¼ðV;EÞis an undi ec ed and unweighed g aph wi hou sel -loops, wi h Vand Ebeing he node and edge se o g aph G, espec i ely. Applying his concep o ou p ob- lem, we gene a ed a cellula ne wo k in which he muscle ibe s a e ep esen ed by nodes, and wo nodes a e connec ed i wo ibe s a e adjacen . To iden i y he neighbo hood o each ibe , we gene a ed a mosaic such ha each ibe con ou is expanded o each he expanded con ou o he adjacen ibe . This concep was add essed by applying a wa e shed ans o m o a bina y image esul ing om he ibe de ec ion s ep (Sec. 3.1.3). An example is shown in Fig. 7(c). I should be no ed ha an impo an ea u e was ex ac ed om his mosaic. The a io be ween he a ea o a ibe (A2) and he a ea when i s con ou is expanded (A1) ( ea u es 15 and 16 in Table 6) is indica i e o he amoun o he collagen, and he e o e indica i e o he exis ence o ib osis, one o he main a ibu es o he MD. In his poin , a ec o o 24 ea u es (included he 14 geo- me ical ones) was ixed. Cha ac e is ics such as he numbe o neighbo s o he numbe o neighbo s o a de e mined ype o ibe we e added. Table 6shows he added ea u es. These 24 ea u es y o emula e he cha ac e is ics ha he pa hologis akes in o accoun o diagnosis, howe e , in his pape ano he se o ea u es ha escape om he e alua ion o human ision is ex ac ed. Fo his pu pose, a weigh ed g aph de i ed om he muscula biopsy was gene a ed, whe e each node was ep esen ed by he mass cen e o he each ibe , and he neighbo hood ela ions a e mapped in o weigh ed edges, whe e each weigh co esponds o he Euclidean dis ance be ween he nodes [see Fig. 7(d)]. Fi y-eigh new cha ac e is ics we e inco po a ed. The i s 14 cha ac e is ics ou o 58 ones we e ob ained om he pa ame e s ha we e al eady compu ed (a ea, axis, con ex hull, angles, and a io A1/A2), bu aking in o accoun he neigh- bo hood o each ibe , i.e., he a io be ween he alue o he pa ame e s o each ibe and he a e age o he alues o he co esponding adjacen ibe s cons i u ing hese new 14 ea u es ( ea u es 25 o 38 in Table 7). The 44 emaining ea u es a e compu ed om g aphs heo y ( ea u es 39 o 82 in Table 7) when i is applied o an undi ec ed and weigh ed g aph. Table 7shows hese 58 new ea u es. To a oid e o s due o he lack o neighbo s o he ibe s a he image edge, he cha ac e is ics a e calcula ed on a egion o in e es (ROI) chosen by he use s, such ha a leas one ow o ibe s a ound he ROI exis s [see Fig. 7(d)]. 3.3 Fea u e Selec ion Fea u e selec ion has wo bene i s: i educes he cos o da a collec ion and compu a ional cos o ecogni ion, and i usually imp o es he gene aliza ion pe o mance o he classi- ie . Ac ually, a la ge se o ea u es may possibly be de imen al o he classi ica ion pe o mance, a phenomenon known as “ he cu se o dimensionali y.”Fea u e selec ion is a means o selec he ele an and impo an ea u es om a la ge se o ea u es. An op imal ea u e selec ion me hod would equi e an exhaus- i e sea ch, which is no p ac ical o a la ge se o ea u es gen- e a ed om a la ge da ase . The e o e, se e al heu is ic Fig. 7 (a) Muscle biopsy image. (b) Mask o he wa e shed segmen a ion esul . (c) Mosaic, whe e each ibe con ou is expanded o each he expanded con ou o he adjacen ibe . (d) G aph, whe e each node is ep esen ed by he mass cen e o each ibe , and he neighbo hood ela ions a e mapped in o he edges. Table 6 S uc u al ea u es. 15 A e age a io A1/A2 16 S d. de . a io A1/A2 17 A e age neighbo s 18 S d. De . neighbo s 19 S d. de . neighbo s o slow ibe s 20 S d. de . neighbo s o as ibe s 21 Slow neighbo s o slow ibe s 22 Fas neighbo s o slow ibe s 23 Slow neighbo s o as ibe s 24 Fas neighbo s o as ibe s Jou nal o Biomedical Op ics 066017-7 June 2013 •Vol. 18(6) Sáez e al.: Neu omuscula disease classi ica ion sys em Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/25400/ on 04/06/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx algo i hms ha e been de eloped which use classi ica ion accu- acy as he op imali y c i e ion.2 In his pape , he well-known ea u e selec ion me hods, namely sequen ial o wa d selec ion (SFS) and sequen ial back- wa d selec ion (SBS),30 a e used. SFS wo ks by sequen ially adding he ea u e ha mos imp o es he classi ica ion pe o - mance; simila ly, SBS begins wi h he en i e ea u e se and sequen ially emo es he ea u e ha mos imp o es he classi- ica ion pe o mance. While hese me hods s ill canno gua an- ee op imali y o he selec ed ea u e subse , hey ha e been shown o pe o m e y well compa ed wi h o he ea u e selec- ion me hods31 and a e, u he mo e, much mo e compu a ion- ally e icien .32 As i has been men ioned, hese me hods use classi ica ion accu acy as he op imali y c i e ion. In his case, he classi ica- ion was pe o med by a Fuzzy-ARTMAP neu al ne wo k. I is a neu al ne wo k a chi ec u e de eloped by Ca pen e e al.,33 and i is based on adap i e esonance heo y (ART). Fuzzy- ARTMAP is a supe ised lea ning classi ica ion a chi ec u e o analogue- alue inpu pai s o pa e ns, whe e each indi idual inpu is mapped o a class label. To e alua e he classi ica ion accu acy, a se o 71 images was used. Thi y- ou con ol biopsies om quad iceps and Table 7 Fi y-eigh new s uc u al ea u es. 25 A e age ela ion neighbo s a ea 26 S d. de . ela ion neighbo s a ea 27 A e age ela ion neighbo s majo axis 28 S d. de . ela ion neighbo s majo axis 29 A e age ela ion neighbo s mino axis 30 S d. de . ela ion neighbo s mino axis 31 A e age ela ion neighbo s ela ion axis 32 S d. de . ela ion neighbo s ela ion axis 33 A e age ela ion neighbo s con ex hull 34 S d. de . ela ion neighbo s con ex hull 35 A e age ela ion neighbo s angles 36 S d. de . ela ion neighbo s angles 37 A e age ela ion neighbo s a io A1/A2 38 S d. de . ela ion neighbo s a io A1/A2 39 A e age s eng hs 40 S d. de . s eng hs 41 A e age s eng hs o as cells 42 S d. de . s eng hs o as cells 43 A e age s eng hs o slow cells 44 S d. de . s eng hs o slow cells 45 A e age clus e ing coe icien 46 S d. de . clus e ing coe icien 47 A e age clus e ing coe icien o as cells 48 S d. de . clus e ing coe icien o as cells 49 A e age clus e ing coe icien o slow cells 50 S d. de . clus e ing coe icien o slow cells 51 A e age eccen ici y 52 S d. de . eccen ici y 53 A e age eccen ici y o as cells 54 S d. de . eccen ici y o as cells 55 A e age eccen ici y o slow cells 56 S d. de . eccen ici y o slow cells 57 A e age be weenness cen ali y 58 S d. de . be weenness cen ali y 59 A e age be weenness cen ali y o as cells 60 S d. de . be weenness cen ali y o as cells 61 A e age be weenness cen ali y o slow cells 62 S d. de . be weenness cen ali y o slow cells 63 A e age sho es pa hs leng hs 64 S d. de . sho es pa hs leng hs 65 A e age sho es pa hs leng hs om as cells o as cells 66 S d. de . sho es pa hs leng hs om as cells o as cells 67 A e age sho es pa hs leng hs om as cells o slow cells 68 S d. de . sho es pa hs leng hs om as cells o slow cells 69 A e age sho es pa hs leng hs om slow cells o slow cells 70 S d. de . sho es pa hs leng hs om slow cells o slow cells 71 A e age sho es pa hs leng hs om slow cells o as cells 72 S d. de . sho es pa hs leng hs om slow cells o as cells 73 Radius 74 Diame e 75 E iciency 76 Pea son co ela ion 77 Algeb aic connec i i y 78 S me ic 79 Asso a i i y 80 Densi y 81 T ansi i i y 82 Modula i y Jou nal o Biomedical Op ics 066017-8 June 2013 •Vol. 18(6) Sáez e al.: Neu omuscula disease classi ica ion sys em Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/25400/ on 04/06/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx biceps, 20 biopsies images a ec ed by dys ophy, and 17 biop- sies images a ec ed by NA. Th ee s udies we e ca ied ou : •Compa ison be ween he h ee g oups o images [con ols (C)−MD−NA] •Compa ison be ween con ol and dys ophies (biopsies o muscle a ec ed by dys ophy belong o quad iceps) •Compa ison be ween con ol and NA (biopsies o muscle a ec ed by a ophy belong o biceps) Fo each compa ison, he selec ion pe o mance was e alu- a ed by ou old c oss- alida ion (XVAL).30 In his sense, he disad an age o sensi i eness o he o de o p esen a ion o he aining se ha he SBS and SFS me hods p esen was diminished. To pe o m he XVAL me hod, ou disjoin subse s o each class (con ol, dys ophy, NA) we e used. Th ee o hese subse s se ed as aining se s o he neu al ne wo k, while he o he one was used as a alida ion se . Then, he p ocedu e was epea ed in e changing he alida ion subse wi h one o he aining subse s, and so on, ill all ou subse s we e used as alida ion se s. The inal classi ica ion e o was calcula ed as he mean o he e o s o each XVAL un. 3.3.1 Resul s o ea u e selec ion The ea u e selec ion p ocedu e was pe o med wice o he h ee compa isons men ioned abo e. The i s selec ion was ca - ied ou on he 24 i s ea u es desc ibed in he p e ious sec ion (see Tables 5and 6), and he second selec ion was pe o med on he 82 ea u es (see Table 7). The esul s a e shown in Table 8,in which he selec ed ea u es and he classi ica ion e o ob ained wi h his selec ion a e shown. The selec ed ea u es a e p e- sen ed in ascending o de by disc imina ion powe . As can be seen, when we compa e be ween bo h quad iceps (dys ophies) and biceps (a ophies), he classi ica ion success is 100%, he e o e, adding new ea u es does no imp o e he clas- si ica ion e o in his s age o aining. In he ollowing sec ion, we will s udy wha classi ica ion e o is ob ained when we clas- si y new biopsies, ha we e no included in he s age o ea u e selec ion. Howe e , in he case o he dis inc ion be ween he h ee ca ego ies, he classi ica ion e o dec eases when we add s uc u al ea u es. E en so, in he nex sec ion we check how hese se s o ea u es a e good o classi y new biopsies. 3.4 Classi ica ion The ex ac ed ea u es ep esen he inpu s o a classi ica ion p ocedu e. The classi ie used in his pape is a Fuzzy ARTMAP neu al ne wo k. The Fuzzy ARTMAP sys em inco po a es wo uzzy ART modules, ARTaand ARTb, ha a e linked oge he ia an in e -ART module, Fab, called a map ield (see Fig. 8). In he p edic ion s age, “a”( ea u es) is he inpu ec o , and i is hoped ha he sys em esponds wi h he “b” ec o (ca ego- ies). In ou implemen a ion, Fa 2is composed by N:j’ h nodes (j¼1;:::;N,N¼numbe o aining images) and Wab j deno es he weigh ec o associa ed o he j’ h node o Fa 2. In he p edic ion s age, he jnode o Fa 2is ac i a ed h oughou he maximum o a choice unc ion Tj. The Wab jassocia ed o his node ac i a es a Knode o Fb 2. This Knode is he ca ego y chosen ha he sys em p edic s is associa ed o he “a”inpu . Howe e , in ou implemen a ion, he ca ego y choice is modi ied. As each jnode in Fa 2is associa ed o a aining image, we ha e aken in o accoun he a e age o he alues Tjby ca ego ies ins ead o he maximum. Le cl be he ca ego y (cl ¼1;:::; M,M¼numbe o he ca ego ies), in ou case, cl ¼1 ep esen s con ol, cl ¼2 ep esen s dys ophy, and cl ¼3 ep esen s a ophy, espec i ely. Le ncl be he num- be o he aining images o he cl ca ego y. The ca ego y choice p ocedu e is: o cl ¼1∶M kcl (Tj), j¼1:ncl End. The ca ego y chosen ul ills: KCL ¼max kcl∶cl ¼1∶Mg:(5) Table 8 Fea u e selec ion esul s. Compa ison 24 ea u es 82 ea u es Selec ed ea u es Classi ica ion e o (%) Selec ed ea u es Classi ica ion e o (%) Con ols-dys ophies (C/MD) 19 18 15 0 25 9 0 Con ols-a ophies (C/NA) 12 20 21 22 0 34 16 21 45 0 Con ols-dys ophies- a ophies (C/MD/NA) 20 9 19 18 16 21 14 17 13.22 25 32 14 16 58 62 15 30 26 1.48 Fig. 8 Fuzzy ARTMAP a chi ec u e. Jou nal o Biomedical Op ics 066017-9 June 2013 •Vol. 18(6) Sáez e al.: Neu omuscula disease classi ica ion sys em Downloaded F om: h p://spiedigi allib a y.o g/pd access.ashx?u l=/da a/jou nals/biomedo/25400/ on 04/06/2017 Te ms o Use: h p://spiedigi allib a y.o g/ss/ e mso use.aspx