Kea leboha ha u etetse Nature.com.U sebelisa mofuta oa sebali o nang le tšehetso e fokolang ea CSS.Bakeng sa boiphihlelo bo botle ka ho fetesisa, re khothaletsa hore o sebelise sebatli se ntlafalitsoeng (kapa o tima Compatibility Mode ho Internet Explorer).Ho feta moo, ho netefatsa ts'ehetso e tsoelang pele, re bonts'a sebaka sa marang-rang ntle le mekhoa le JavaScript.
Li-slide tse bonts'ang lingoloa tse tharo selaeteng se seng le se seng.Sebelisa likonopo tse ka morao le tse latelang ho tsamaea ka har'a li-slide, kapa likonopo tsa selaoli se qetellong ho tsamaea selaeteng ka seng.
Optical coherence tomographic angiography (OCTA) ke mokhoa o mocha oa pono e sa hlaseleng ea likepe tsa retina.Leha OCTA e na le lits'ebetso tse ngata tse ts'episang tsa kliniki, ho bona boleng ba setšoantšo e ntse e le phephetso.Re thehile mokhoa o tebileng oa ho ithuta re sebelisa ResNet152 neural network classifier e koetlisitsoeng esale pele ka ImageNet ho hlophisa litšoantšo tse holimo tsa capillary plexus ho tsoa ho likheo tse 347 tsa bakuli ba 134.Litšoantšo li ile tsa boela tsa hlahlojoa ka letsoho e le 'nete ea' nete ke batho ba babeli ba ikemetseng bakeng sa mohlala oa ho ithuta o hlokometsoeng.Hobane litlhoko tsa boleng ba setšoantšo li ka fapana ho latela maemo a kliniki kapa lipatlisiso, mefuta e 'meli e ile ea koetlisetsoa, ​​e' ngoe e le ea ho lemoha litšoantšo tsa boleng bo holimo 'me e 'ngoe e le ea ho lemoha litšoantšo tsa boleng bo tlaase.Mohlala oa rona oa marang-rang oa neural o bontša sebaka se setle ka tlas'a lekhalo (AUC), 95% CI 0.96-0.99, \(\kappa\) = 0.81), e leng molemo haholo ho feta tekanyo ea pontšo e tlalehiloeng ke mochine (AUC = 0.82, 95). % CI).0.77–0.86, \(\kappa\) = 0.52 le AUC = 0.78, 95% CI 0.73–0.83, \(\kappa\) = 0,27, ka ho latellana).Boithuto ba rona bo bonts'a hore mekhoa ea ho ithuta ka mochini e ka sebelisoa ho nts'etsapele mekhoa e feto-fetohang le e matla ea ho laola boleng ba litšoantšo tsa OCTA.
Optical coherence tomographic angiography (OCTA) ke mokhoa o mocha o thehiloeng ho optical coherence tomography (OCT) o ka sebelisoang bakeng sa pono e sa hlaseleng ea retinal microvasculature.OCTA e lekanya phapang ea lipaterone tsa ponaletso ho tsoa ho maqhubu a khanya khafetsa sebakeng se le seng sa retina, 'me li-retina li ka baloa ho senola methapo ea mali ntle le ts'ebeliso e mpe ea lidae kapa lisebelisoa tse ling tse fapaneng.OCTA e boetse e thusa ho nahana ka methapo ea methapo e tebileng, e lumellang lingaka hore li hlahlobe ka thoko le likarolo tse tebileng tsa lijana, ho thusa ho khetholla lipakeng tsa lefu la chorioretinal.
Le hoja mokhoa ona o ntse o tšepisa, phapang ea boleng ba setšoantšo e ntse e le phephetso e kholo bakeng sa tlhahlobo e tšepahalang ea litšoantšo, ho etsa hore tlhaloso ea litšoantšo e be thata le ho thibela ho ata ha kliniki ho amoheloa.Hobane OCTA e sebelisa lisebelisoa tse ngata tse latellanang tsa OCT, e na le maikutlo a matle ho lintho tsa khale tsa khale ho feta OCT e tloaelehileng.Boholo ba liforomo tsa OCTA tsa khoebo li fana ka metric ea bona ea boleng ba setšoantšo e bitsoang Signal Strength (SS) kapa ka linako tse ling Signal Strength Index (SSI).Leha ho le joalo, litšoantšo tse nang le boleng bo phahameng ba SS kapa SSI ha li tiise ho ba sieo ha litšoantšo tsa maiketsetso, tse ka amang tlhahlobo efe kapa efe e latelang ea litšoantšo le ho lebisa liqetong tse fosahetseng tsa kliniki.Lisebelisoa tse tloaelehileng tsa litšoantšo tse ka hlahang litšoantšong tsa OCTA li kenyelletsa lintho tsa khale tsa motsamao, li-artifacts tsa likarolo, li-media opacity artifacts, le li-projection artifacts1,2,3.
Ha mehato e nkiloeng ho OCTA e kang vascular density e ntse e sebelisoa ka ho eketsehileng lipatlisisong tsa phetolelo, liteko tsa meriana le mekhoa ea meriana, ho na le tlhokahalo e potlakileng ea ho hlahisa mekhoa e matla le e ka tšeptjoang ea ho laola boleng ba setšoantšo ho felisa litšoantšo tsa litšoantšo4.Likhokahano tsa skip, tse tsejoang hape e le li-residual connections, ke likhakanyo tsa meralo ea neural network e lumellang tlhahisoleseling ho feta likarolo tsa convolutional ha e ntse e boloka tlhahisoleseling ka likala kapa liqeto tse fapaneng5.Hobane lintho tsa khale tsa litšoantšo li ka ama ts'ebetso ea litšoantšo tse nyane le tse kholo ka kakaretso, marang-rang a neural a skip-connection a loketse ho iketsetsa mosebetsi ona oa taolo ea boleng5.Mosebetsi o sa tsoa hatisoa o bonts'itse ts'episo e itseng bakeng sa marang-rang a tebileng a convolutional neural a koetliselitsoeng ho sebelisa data ea boleng bo holimo ho tsoa ho batho ba hakanyang6.
Thutong ena, re koetlisa marang-rang a marang-rang a convolutional neural ho fumana ka bo eona boleng ba litšoantšo tsa OCTA.Re haha ​​​​mosebetsing oa pele ka ho etsa mefuta e fapaneng ea ho khetholla litšoantšo tsa boleng bo holimo le litšoantšo tsa boleng bo tlaase, kaha litlhoko tsa boleng ba setšoantšo li ka fapana bakeng sa maemo a itseng a bongaka kapa a lipatlisiso.Re bapisa liphetho tsa marang-rang ana le marang-rang a convolutional neural ntle le likhokahano tse sieo ho lekola boleng ba ho kenyelletsa likarolo maemong a mangata a granularity ka har'a thuto e tebileng.Ka mor'a moo re ile ra bapisa liphetho tsa rona le matla a pontšo, e leng tekanyo e amohelehang ea boleng ba setšoantšo e fanoang ke bahlahisi.
Boithuto ba rona bo kenyelelitse bakuli ba nang le lefu la tsoekere ba ileng ba ea Yale Eye Center pakeng tsa la 11 Phato, 2017 le la 11 Mmesa, 2019. Bakuli ba nang le lefu la chorioretinal leo e seng la lefu la tsoekere ba ne ba qheletsoe ka thoko.Ho ne ho se na mokhoa oa ho kenyelletsa kapa oa ho qheleloa ka thoko ho latela lilemo, bong, morabe, boleng ba setšoantšo, kapa ntho efe kapa efe.
Litšoantšo tsa OCTA li ile tsa fumanoa ho sebelisoa sethala sa AngioPlex ho Cirrus HD-OCT 5000 (Carl Zeiss Meditec Inc, Dublin, CA) tlas'a li-protocol tsa 8\(\times\)8 mm le 6\(\times\)6 mm tsa litšoantšo.Tumello e nang le tsebo ea ho kenya letsoho phuputsong e ile ea fumanoa ho monkakarolo e mong le e mong oa thuto, 'me Boto ea Tlhahlobo ea Setsi sa Yale University (IRB) e amohetse tšebeliso ea tumello e nang le tsebo le ho nka lifoto lefatšeng ka bophara bakeng sa bakuli bana kaofela.Ho latela melao-motheo ea Phatlalatso ea Helsinki.Boithuto bona bo amohetsoe ke Yale University IRB.
Lits'oants'o tsa poleiti ea sefahleho li ile tsa hlahlobjoa ho ipapisitsoe le Motion Artifact Score (MAS) e hlalositsoeng pejana, Segmentation Artifact Score (SAS), setsi sa foveal, boteng ba media opacity, le pono e ntle ea li-capillaries tse nyane joalo ka ha ho khethiloe ke mohlahlobi oa setšoantšo.Litšoantšo li ile tsa hlahlojoa ke bahlahlobi ba babeli ba ikemetseng (RD le JW).Setšoantšo se na le lintlha tse hlophiloeng tsa 2 (tse loketseng) haeba lintlha tsohle tse latelang li finyelloa: setšoantšo se le bohareng ba fovea (ka tlase ho 100 pixels ho tloha bohareng ba setšoantšo), MAS ke 1 kapa 2, SAS ke 1, 'me opacity ea mecha ea litaba e ka tlase ho 1. E hlahisa litšoantšo tsa boholo / 16, 'me li-capillaries tse nyenyane li bonoa litšoantšong tse kholo ho feta 15/16.Setšoantšo se filoe lintlha tse 0 (ha ho na lintlha) haeba leha e le efe ea lintlha tse latelang e fihletsoe: setšoantšo ha se bohareng, haeba MAS e le 4, haeba SAS e le 2, kapa ponahalo ea kakaretso e kholo ho feta 1/4 ea setšoantšo, le li-capillaries tse nyane ha li khone ho fetoloa ho feta setšoantšo sa 1 / 4 ho khetholla.Lits'oants'o tse ling kaofela tse sa fihlelleng litekanyetso tsa lintlha 0 kapa 2 li fuoe lintlha e le 1 (ho penya).
Ka feiga.1 e bonts'a mehlala ea litšoantšo bakeng sa likhakanyo tse lekantsoeng le litšoantšo tsa maiketsetso.Ho ts'epahala ha lintlha tse fapaneng ho ile ha hlahlojoa ke Cohen's kappa weighting8.Lintlha tsa motho ka mong tsa rater ka 'ngoe li akaretsoa ho fumana lintlha tse akaretsang bakeng sa setšoantšo ka seng, ho tloha ho 0 ho isa ho 4. Litšoantšo tse nang le kakaretso ea 4 li nkoa li le ntle.Litšoantšo tse nang le kakaretso ea 0 kapa 1 li nkoa e le tsa boleng bo tlaase.
ResNet152 architecture convolutional neural network (Fig. 3A.i) e koetlisitsoeng pele ho litšoantšo tse tsoang ho database ea ImageNet e ile ea hlahisoa ho sebelisoa fast.ai le PyTorch framework5, 9, 10, 11. Convolutional neural network ke marang-rang a sebelisang ba ithutileng. li-filters bakeng sa ho lekola likhechana tsa litšoantšo ho ithuta ka libaka le likarolo tsa lehae.ResNet ea rona e koetlisitsoeng ke marang-rang a 152-layer neural a khetholloang ke likheo kapa "likhokahano tse setseng" tse fetisang ka nako e le 'ngoe tlhahisoleseling ka liqeto tse ngata.Ka ho hlahisa tlhahisoleseling ka liqeto tse fapaneng holim'a marang-rang, sethala se ka ithuta likarolo tsa litšoantšo tsa boleng bo tlase maemong a mangata a lintlha.Ho phaella ho mohlala oa rona oa ResNet, re ile ra boela ra koetlisa AlexNet, mohaho oa marang-rang o ithutoang hantle, ntle le likhokahano tse sieo bakeng sa ho bapisa (Setšoantšo sa 3A.ii)12.Ntle le likhokahano tse sieo, marang-rang ana a ke ke a khona ho nka likarolo ka granularity e phahameng.
Sets'oants'o sa mantlha sa 8\(\ times\)8mm OCTA13 se ntlafalitsoe ho sebelisoa mekhoa ea ho bonahatsa e otlolohileng le e otlolohileng.Dataset e feletseng e ile ea aroloa ka mokhoa o sa reroang boemong ba setšoantšo ho ea koetliso (51.2%), tlhahlobo (12.8%), hyperparameter tuning (16%), le lisebelisoa tsa netefatso (20%) tse sebelisang lebokose la lithulusi la scikit-learn python14.Ho ile ha shejoa linyeoe tse peli, e 'ngoe e ipapisitse le ho fumana litšoantšo tsa boleng bo holimo feela (kakaretso ea lintlha tse 4) mme e' ngoe e ipapisitse le ho fumana litšoantšo tsa boleng bo tlase (kakaretso ea lintlha 0 kapa 1).Bakeng sa ts'ebeliso e 'ngoe le e 'ngoe ea boemo bo holimo le bo tlase, marang-rang a neural a koetlisoa hape hang ho data ea rona ea litšoantšo.Ketsahalong e 'ngoe le e' ngoe ea ts'ebeliso, marang-rang a neural a ne a koetlisetsoa nako ea 10, kaofela ntle le boima bo phahameng ka ho fetisisa ba lera bo ne bo hoamisitsoe, 'me litekanyo tsa likarolo tsohle tsa ka hare li ile tsa ithutoa bakeng sa linako tse 40 ho sebelisoa mokhoa o khethollang oa ho ithuta o nang le ts'ebetso ea tahlehelo ea cross-entropy 15, 16..Mosebetsi oa tahlehelo ea cross entropy ke tekanyo ea tekanyo ea logarithmic ea ho se lumellane pakeng tsa li-label tsa marang-rang tse boletsoeng esale pele le data ea sebele.Nakong ea koetliso, ho theoha ha gradient ho etsoa ka mekhahlelo ea ka hare ea neural network ho fokotsa tahlehelo.Sekhahla sa ho ithuta, sekhahla sa ho tlohela sekolo, le li-hyperparameter tsa phokotso ea boima ba 'mele li ile tsa hlophisoa ho sebelisoa ts'ebeliso ea Bayesian e nang le lintlha tse 2 tsa ho qala tse sa reroang le tse 10, mme AUC ho dataset e ile ea hlophisoa ho sebelisoa li-hyperparameter joalo ka sepheo sa 17.
Mehlala e emelang litšoantšo tsa 8 × 8 mm OCTA tsa li-capillary plexuses tse ka holimo li fumane 2 (A, B), 1 (C, D), le 0 (E, F).Lintho tse entsoeng ka litšoantšo tse bonts'itsoeng li kenyelletsa mela e panyang (metsu), likaroloana tsa maiketsetso (asterisk), le opacity ea media (metsu).Setšoantšo (E) le sona ha se bohareng.
Litšobotsi tsa ts'ebetso ea li-receiver (ROC) li hlahisoa bakeng sa mefuta eohle ea marang-rang ea neural, 'me litlaleho tsa matla a pontšo ea enjene li hlahisoa bakeng sa nyeoe e' ngoe le e 'ngoe ea boleng bo tlaase le ea boleng bo phahameng.Sebaka se tlas'a curve (AUC) se ne se baloa ho sebelisoa sephutheloana sa pROC R, 'me linako tsa boitšepo tsa 95% le litekanyetso tsa p li ile tsa baloa ho sebelisoa mokhoa oa DeLong18,19.Lipalo tse akaretsang tsa li-reter tsa batho li sebelisoa e le motheo oa lipalo tsohle tsa ROC.Bakeng sa matla a lets'oao a tlalehiloeng ke mochini, AUC e baloa habeli: hanngoe bakeng sa sekhahla sa boleng bo holimo sa Scalability Score le hanngoe bakeng sa sekhahla sa boleng bo tlase ba Scalability Score.Neural network e bapisoa le matla a lets'oao la AUC a bonts'ang maemo a eona a koetliso le tlhahlobo.
Ho tsoela pele ho lekola mohlala o tebileng oa ho ithuta ho dataset e arohaneng, mehlala ea boleng bo holimo le ea boleng bo tlase e ile ea sebelisoa ka kotloloho tlhahlobong ea ts'ebetso ea 32 sefahleho se felletseng 6 \ (\ linako\) 6mm litšoantšo tsa slab holim'a metsi tse bokelitsoeng ho tsoa Univesithing ea Yale.Boima ba Mahlo bo thehiloe ka nako e le 'ngoe le setšoantšo sa 8 \(\ linako \) 8 mm.Litšoantšo tsa 6\(\×\) 6 mm li ile tsa hlahlojoa ka letsoho ke li-raters tse tšoanang (RD le JW) ka mokhoa o ts'oanang le litšoantšo tsa 8\(\×\) 8 mm, AUC e ne e baloa hammoho le ho nepahala le kappa ea Cohen. .ka ho lekana .
Karolelano ea ho se lekane ha sehlopha ke 158: 189 (\ (\ rho = 1.19 \)) bakeng sa mohlala oa boleng bo tlaase le 80: 267 (\ (\ rho = 3.3 \)) bakeng sa mohlala oa boleng bo phahameng.Hobane karo-karolelano ea ho se leka-lekane ha sehlopha e ka tlase ho 1:4, ha ho na liphetoho tse khethehileng tsa meralo tse entsoeng ho lokisa ho se leka-lekane ha sehlopha20,21.
Ho bona hantle mokhoa oa ho ithuta, limmapa tsa ts'ebetso ea sehlopha li ile tsa hlahisoa bakeng sa mefuta eohle e mene ea thuto e tebileng e koetlisitsoeng: mohlala oa boleng bo holimo oa ResNet152, oa boleng bo tlase oa ResNet152, oa boemo bo holimo oa AlexNet, le mohlala oa AlexNet oa boleng bo tlase.Limmapa tsa ts'ebetso ea sehlopha li hlahisoa ho tsoa ho likarolo tsa tlhahiso ea mefuta ena e mene, 'me limmapa tsa mocheso li hlahisoa ka ho koahela limmapa tsa ts'ebetso tse nang le litšoantšo tsa mohloli ho tsoa ho li-sets tsa netefatso tsa 8 × 8 mm le 6 × 6 mm22, 23.
Mofuta oa 4.0.3 oa R o ile oa sebelisoa bakeng sa lipalo-palo tsohle, 'me lipono li entsoe ho sebelisoa laeborari ea lisebelisoa tsa litšoantšo tsa ggplot2.
Re ile ra bokella litšoantšo tse ka pele tse 347 tsa "capillary plexus" e ka holimo-limo e ka bang 8 \ (\ linako \) 8 mm ho tloha ho batho ba 134.Mochine o tlalehile matla a pontšo ka tekanyo ea 0 ho ea ho 10 bakeng sa litšoantšo tsohle (ho bolela = 6.99 ± 2.29).Har'a litšoantšo tsa 347 tse fumanoeng, lilemo tse tloaelehileng tsa tlhahlobo e ne e le lilemo tse 58.7 ± 14.6, 'me 39.2% e ne e tsoa ho bakuli ba banna.Har'a litšoantšo tsohle, 30.8% e tsoa ho Caucasus, 32.6% e tsoa ho Batho ba Batšo, 30.8% e tsoa ho Hispanics, 4% e tsoa ho Maasia, le 1.7% ho tsoa merabeng e meng (Letlapa la 1).).Kabo ea lilemo tsa bakuli ba nang le OCTA e fapane haholo ho latela boleng ba setšoantšo (p <0.001).Palo ea litšoantšo tsa boleng bo phahameng ho bakuli ba banyenyane ba lilemo li 18-45 e ne e le 33.8% ha e bapisoa le 12.2% ea litšoantšo tse tlaase (Letlapa la 1).Kabo ea boemo ba lefu la tsoekere ea retinopathy le eona e fapane haholo ka boleng ba setšoantšo (p <0.017).Har'a litšoantšo tsohle tsa boleng bo phahameng, karolo ea bakuli ba nang le PDR e ne e le 18.8% ha e bapisoa le 38.8% ea litšoantšo tsohle tsa boleng bo tlaase (Letlapa la 1).
Litekanyetso tsa motho ka mong tsa litšoantšo tsohle li bonts'itse ts'epahalo e leka-lekaneng ho isa ho e matla lipakeng tsa batho ba balang litšoantšo (Cohen's weighted kappa = 0.79, 95% CI: 0.76-0.82), 'me ho ne ho se na lintlha tsa litšoantšo moo li-rate li fapaneng ho feta 1 (Fig. 2A)..Matla a pontšo a hokahane haholo le lintlha tsa letsoho (Pearson product moment correlation = 0.58, 95% CI 0.51-0.65, p <0.001), empa litšoantšo tse ngata li ile tsa khetholloa e le tse nang le matla a phahameng a pontšo empa lintlha tse tlaase tsa letsoho (Fig. .2B).
Nakong ea koetliso ea meralo ea ResNet152 le AlexNet, tahlehelo ea cross-entropy mabapi le netefatso le koetliso e oela ka nako ea 50 epochs (Figure 3B, C).Ho nepahala ha netefatso nakong ea koetliso ea ho qetela ho feta 90% bakeng sa linyeoe tsa ts'ebeliso ea boleng bo holimo le bo tlase.
Li-curve tsa ts'ebetso ea li-receiver li bontša hore mohlala oa ResNet152 o feta haholo matla a pontšo a tlalehiloeng ke mochine maemong a mabeli a tlaase le a boleng bo phahameng ba tšebeliso (p <0.001).Mohlala oa ResNet152 o boetse o feta haholo mohaho oa AlexNet (p = 0.005 le p = 0.014 bakeng sa linyeoe tse tlaase le tsa boleng bo phahameng, ka ho latellana).Mefuta e hlahisitsoeng bakeng sa e 'ngoe le e' ngoe ea mesebetsi ena e khonne ho fihlela boleng ba AUC ba 0.99 le 0.97, ka ho latellana, e leng betere haholo ho feta boleng ba AUC ba 0.82 le 0.78 bakeng sa index ea matla ea mochini kapa 0.97 le 0.94 bakeng sa AlexNet. ..(Setšoantšo sa 3).Phapang pakeng tsa ResNet le AUC ka matla a pontšo e phahame ha u lemoha litšoantšo tsa boleng bo phahameng, tse bontšang melemo e eketsehileng ea ho sebelisa ResNet bakeng sa mosebetsi ona.
Li-graph li bontša bokhoni ba motho e mong le e mong ea ikemetseng oa ho fumana lintlha le ho bapisa le matla a lets'oao a tlalehiloeng ke mochini.(A) Kakaretso ea lintlha tse tla hlahlojoa e sebelisoa ho bopa kakaretso ea lintlha tse lokelang ho hlahlojoa.Litšoantšo tse nang le kakaretso ea lintlha tse 4 li abeloa boleng bo holimo, athe litšoantšo tse nang le kakaretso ea kakaretso ea 1 kapa ka tlase li fuoa boleng bo tlaase.(B) Letšoao le matla le lumellana le likhakanyo tsa matsoho, empa litšoantšo tse nang le matla a holimo li ka 'na tsa e-ba tsa boleng bo tlaase.Mohala o matheba o mofubelu o bonts'a moeli oa boleng bo khothaletsoang ke moetsi o ipapisitse le matla a lets'oao (matla a lets'oao \(\ge\)6).
Boithuto ba phetisetso ea ResNet bo fana ka ntlafatso e kholo ea boitsebahatso ba boleng ba setšoantšo bakeng sa maemo a ts'ebeliso ea boleng bo tlase le ba boleng bo holimo ha ho bapisoa le maemo a mats'oao a tlalehiloeng ka mochini.(A) Litšoantšo tse nolofalitsoeng tsa meralo ea meralo e koetlisitsoeng esale pele (i) ResNet152 le (ii) AlexNet ea meralo.(B) Histori ea koetliso le li-curve tsa ts'ebetso ea moamoheli bakeng sa ResNet152 ha li bapisoa le matla a mats'oao a tlalehiloeng ke mochini le litekanyetso tsa boleng bo tlase ba AlexNet.(C) Histori ea koetliso ea moamoheli oa ResNet152 le li-curve tsa ts'ebetso ha li bapisoa le matla a mats'oao a tlalehiloeng ke mochini le litekanyetso tsa boleng bo holimo tsa AlexNet.
Ka mor'a ho lokisa moeli oa moeli oa liqeto, ho nepahala ho fetisisa ho bolela esale pele ea mohlala oa ResNet152 ke 95.3% bakeng sa nyeoe ea boleng bo tlaase le 93.5% bakeng sa nyeoe ea boleng bo phahameng (Letlapa la 2).Ho nepahala ho fetisisa ho bolela esale pele ka mohlala oa AlexNet ke 91.0% bakeng sa boemo bo tlaase ba boleng le 90.1% bakeng sa boemo bo phahameng ba boleng (Letlapa la 2).Ponahalo e kholo ea matla a ho bolela esale pele ke 76.1% bakeng sa boemo bo tlase ba tšebeliso le 77.8% bakeng sa boemo bo phahameng ba tšebeliso.Ho ea ka Cohen's kappa (\(\kappa\)), tumellano pakeng tsa mohlala oa ResNet152 le likhakanyo ke 0.90 bakeng sa nyeoe ea boleng bo tlaase le 0.81 bakeng sa nyeoe ea boleng bo phahameng.Cohen's AlexNet kappa ke 0.82 le 0.71 bakeng sa linyeoe tsa tšebeliso ea boleng bo tlaase le boleng bo phahameng, ka ho latellana.Cohen's signal power kappa ke 0.52 le 0.27 bakeng sa linyeoe tsa tšebeliso ea boleng bo tlaase le bo phahameng, ka ho latellana.
Netefatso ea limotlolo tsa temoho ea boleng bo holimo le bo tlase litšoantšong tse 6\(\x\) tsa poleiti e bataletseng ea 6 mm e bonts'a bokhoni ba moetso o koetlisitsoeng oa ho tseba boleng ba setšoantšo ho pholletsa le mekhahlelo e fapaneng ea litšoantšo.Ha u sebelisa li-slabs tse 6 \ (\x\) 6 mm tse sa tebang bakeng sa boleng ba litšoantšo, mohlala oa boleng bo tlaase o ne o e-na le AUC ea 0.83 (95% CI: 0.69-0.98) 'me mohlala oa boleng bo phahameng o ne o e-na le AUC ea 0.85.(95% CI: 0.55-1.00) (Letlapa la 2).
Tlhahlobo ea pono ea limmapa tsa ts'ebetso ea sehlopha sa ho kenya li bonts'itse hore marang-rang ohle a koetlisitsoeng a neural a sebelisa likarolo tsa setšoantšo nakong ea ho arola litšoantšo (setšoantšo sa 4A, B).Bakeng sa litšoantšo tse 8 \(\ linako \) 8 mm le 6 \(\ linako \) 6 mm, litšoantšo tsa ts'ebetso ea ResNet li latela haufi-ufi le methapo ea methapo ea retina.Limmapa tsa ts'ebetso ea AlexNet le tsona li latela likepe tsa retinal, empa ka tharollo e matla.
Limmapa tsa ts'ebetso ea sehlopha bakeng sa mefuta ea ResNet152 le AlexNet li totobatsa likarolo tse amanang le boleng ba setšoantšo.aLQ ea mohlala e koetlisitsoeng ka litekanyetso tsa boleng bo tlaase, mohlala oa HQ o koetlisitsoeng ka litekanyetso tsa boleng bo phahameng.
Ho 'nile ha bontšoa hore boleng ba setšoantšo bo ka ama haholo palo leha e le efe ea litšoantšo tsa OCTA.Ho phaella moo, ho ba teng ha retinopathy ho eketsa liketsahalo tsa litšoantšo tsa litšoantšo7,26.Ha e le hantle, boitsebisong ba rona, bo lumellanang le lithuto tse fetileng, re fumane kamano e kholo pakeng tsa lilemo tse ntseng li eketseha le ho teba ha lefu la retinal le ho senyeha ha boleng ba setšoantšo (p <0.001, p = 0.017 bakeng sa lilemo le boemo ba DR, ka ho latellana; Letlapa la 1) 27 Ka hona, ho bohlokoa ho hlahloba boleng ba setšoantšo pele u etsa tlhahlobo ea palo ea litšoantšo tsa OCTA.Lithuto tse ngata tse hlahlobang litšoantšo tsa OCTA li sebelisa liteishene tse tlalehiloeng ke mochini ho thibela litšoantšo tsa boleng bo tlase.Le hoja matla a matšoao a 'nile a bontšoa hore a ama palo ea li-parameter tsa OCTA, matla a holimo a le mong a ka' na a se ke a lekana ho tlosa litšoantšo tse nang le litšoantšo tse entsoeng ka litšoantšo2,3,28,29.Ka hona, hoa hlokahala ho hlahisa mokhoa o ka tšeptjoang haholoanyane oa ho laola boleng ba setšoantšo.Ho fihlela sena, re hlahloba ts'ebetso ea mekhoa e tebileng ea ho ithuta e hlokometsoeng khahlanong le matla a pontšo a tlalehiloeng ke mochine.
Re thehile mefuta e mengata ea ho lekola boleng ba setšoantšo hobane linyeoe tse fapaneng tsa tšebeliso ea OCTA li ka ba le litlhoko tse fapaneng tsa boleng ba setšoantšo.Ka mohlala, litšoantšo li lokela ho ba tsa boleng bo holimo.Ho phaella moo, litekanyo tse khethehileng tsa palo ea thahasello le tsona li bohlokoa.Ka mohlala, sebaka sa foveal avascular zone ha se itšetlehe ka moferefere oa sebaka se seng sa bohareng, empa se ama boima ba likepe.Leha lipatlisiso tsa rona li ntse li tsoela pele ho lebisa tlhokomelo ho mokhoa o akaretsang oa boleng ba setšoantšo, o sa amahanngoang le litlhoko tsa tlhahlobo e itseng, empa o reretsoe ho nka sebaka ka ho toba matla a lets'oao a tlalehiloeng ke mochini, re ts'epa ho fa basebelisi matla a maholo a taolo e le hore e ka khetha metric e ikhethang eo o e ratang ho mosebelisi.khetha mohlala o lumellanang le tekanyo e phahameng ea litšoantšo tsa maiketsetso tse nkoang li amoheleha.
Bakeng sa lits'oants'o tsa boleng bo tlase le tsa boleng bo holimo, re bonts'a ts'ebetso e ntle haholo ea marang-rang a tebileng a convolutional neural a sieo, ka li-AUC tsa 0.97 le 0.99 le mefuta ea boleng bo tlase, ka ho latellana.Re boetse re bonts'a ts'ebetso e phahameng ea mokhoa oa rona oa ho ithuta o tebileng ha o bapisoa le maemo a mats'oao a tlalehiloeng feela ke mechini.Tlola likhokahano li lumella neural network ho ithuta likarolo maemong a mangata a lintlha, ho hapa likarolo tse ntle tsa litšoantšo (mohlala, phapang) hammoho le likarolo tse akaretsang (mohlala, centering30,31).Kaha lintho tse entsoeng ka matsoho tse amang boleng ba setšoantšo li ka 'na tsa tsejoa hamolemo ho feta mefuta e mengata, meralo ea neural network e nang le likhokahano tse sieo e ka bonts'a ts'ebetso e ntle ho feta e se nang mesebetsi ea ho tseba boleng ba setšoantšo.
Ha re hlahloba mohlala oa rona litšoantšong tsa 6 \ (\ × 6mm) OCTA, re hlokometse ho fokotseha ha ts'ebetso ea lihlopha bakeng sa mefuta e 'meli ea boleng bo phahameng le e tlaase (setšoantšo sa 2), ho fapana le boholo ba mohlala o koetliselitsoeng ho arola.Ha ho bapisoa le mohlala oa ResNet, mohlala oa AlexNet o na le ho oela ho hoholo.Ts'ebetso e batlang e le betere ea ResNet e kanna ea ba ka lebaka la bokhoni ba likhokahano tse setseng ho fetisa tlhahisoleseling ka litekanyo tse ngata, e leng se etsang hore mohlala o be matla haholoanyane bakeng sa ho hlophisa litšoantšo tse hapiloeng ka litekanyo tse fapaneng le / kapa ho holisa.
Liphapang tse ling pakeng tsa litšoantšo tsa 8 \(\×\) 8 mm le 6 \(\×\) litšoantšo tsa 6 \(\×\) 6 mm li ka lebisa ho lihlopha tse fokolang, ho kenyelletsa le karolo e batlang e phahame ea litšoantšo tse nang le libaka tsa foveal avascular, liphetoho tsa ponahalo, li-vascular arcades, le ha ho na methapo ea optic setšoantšong sa 6 × 6 mm.Ho sa tsotellehe sena, mohlala oa rona oa boleng bo phahameng oa ResNet o ile oa khona ho finyella AUC ea 85% bakeng sa litšoantšo tsa 6 \ (\x\) 6 mm, sebopeho seo mohlala o sa kang oa koetlisoa, o fana ka maikutlo a hore boitsebiso ba boleng ba setšoantšo bo kenyelelitsoe marang-rang a neural. e loketse.bakeng sa boholo ba setšoantšo kapa mochine o hlophisitsoeng ka ntle ho koetliso ea oona (Letlapa la 2).Ka mokhoa o ts'oanelang, ResNet- le AlexNet-like activation limmapa tsa 8 \ (\ linako \) 8 mm le 6 \ (\ linako \) 6 mm litšoantšo li khonne ho totobatsa likepe tsa retinal maemong a mabeli, ho fana ka maikutlo a hore mohlala o na le boitsebiso ba bohlokoa.li sebetsa bakeng sa ho arola mefuta e 'meli ea litšoantšo tsa OCTA (setšoantšo sa 4).
Lauerman le ba bang.Tlhahlobo ea boleng ba setšoantšo litšoantšong tsa OCTA e ile ea etsoa ka mokhoa o tšoanang ho sebelisoa mohaho oa Inception, e 'ngoe ea skip-connection convolutional neural network6,32 e sebelisang mekhoa e tebileng ea ho ithuta.Ba boetse ba lekanyetsa thuto ho litšoantšo tsa "capillary plexus" e holimo, empa ba sebelisa litšoantšo tse nyane tsa 3 × 3 mm ho tsoa ho Optovue AngioVue, leha bakuli ba nang le mafu a fapaneng a chorioretinal le bona ba ne ba kenyelelitsoe.Mosebetsi oa rona o thehiloe holim'a metheo ea bona, ho kenyeletsoa mefuta e mengata ea ho sebetsana le maemo a fapaneng a boleng ba setšoantšo le ho netefatsa liphetho tsa litšoantšo tsa boholo bo fapaneng.Re boetse re tlaleha metric ea AUC ea mefuta ea ho ithuta ka mochini mme re eketsa ho nepahala ha eona ho seng ho ntse ho khahla (90%)6 bakeng sa mefuta ea bobeli ea boleng bo tlase (96%) le boleng bo holimo (95.7%)6.
Koetliso ena e na le mefokolo e mengata.Ntlha ea pele, litšoantšo li ile tsa fumanoa ka mochine o le mong feela oa OCTA, ho kenyeletsa feela litšoantšo tsa plexus ea capillary e ka holimo ho 8 \ (\ linako \) 8 mm le 6 \ (\ linako \) 6 mm.Lebaka la ho se kenyeletse litšoantšo ho tsoa ho likarolo tse tebileng ke hore li-artifacts tsa projeke li ka etsa hore tlhahlobo ea matsoho ea litšoantšo e be thata le ho feta ho se tsitsane.Ho feta moo, litšoantšo li fumanoe feela ho bakuli ba nang le lefu la tsoekere, bao OCTA e hlahang e le sesebelisoa sa bohlokoa sa ho hlahloba le ho hlahloba pele33,34.Le hoja re ile ra khona ho hlahloba mohlala oa rona litšoantšong tsa boholo bo fapaneng ho etsa bonnete ba hore liphello li matla, ha rea ​​ka ra khona ho khetholla li-datasets tse loketseng tse tsoang litsing tse fapaneng, tse neng li fokotsa tlhahlobo ea rona ea kakaretso ea mohlala.Le hoja litšoantšo li ile tsa fumanoa setsing se le seng feela, li ile tsa fumanoa ho bakuli ba merabe e fapaneng le merabe e fapaneng, e leng matla a ikhethang a thuto ea rona.Ka ho kenyelletsa mefuta e fapaneng ea lithupelo tsa rona, re tšepa hore liphetho tsa rona li tla atolosoa ka mokhoa o pharaletseng, le hore re tla qoba ho kenyelletsa khethollo ea morabe mefuteng eo re e koetlisang.
Boithuto ba rona bo bonts'a hore marang-rang a marang-rang a marang-rang a ka koetlisetsoa ho fihlela ts'ebetso e phahameng ho khetholla boleng ba setšoantšo sa OCTA.Re fana ka mehlala ena e le lisebelisoa bakeng sa lipatlisiso tse eketsehileng.Hobane metrics e fapaneng e ka ba le litlhoko tse fapaneng tsa boleng ba setšoantšo, ho ka etsoa mohlala oa motho ka mong oa taolo ea boleng bakeng sa metric ka 'ngoe ho sebelisoa sebopeho se thehiloeng mona.
Lipatlisiso tsa nako e tlang li lokela ho kenyelletsa litšoantšo tsa boholo bo fapaneng ho tloha botebong bo fapaneng le metjhini e fapaneng ea OCTA ho fumana ts'ebetso e tebileng ea tlhahlobo ea boleng ba setšoantšo e ka kenyelletsoang liforomong tsa OCTA le liprothokholo tsa litšoantšo.Lipatlisiso tsa morao-rao li boetse li ipapisitse le mekhoa ea ho ithuta e tebileng e hlokometsoeng e hlokang tlhahlobo ea batho le tlhahlobo ea litšoantšo, e ka bang matla le ho ja nako bakeng sa li-dataset tse kholo.Ho sa ntse ho tla bonahala hore na mekhoa ea ho ithuta e tebileng e sa laoleheng e ka khetholla ka ho lekaneng pakeng tsa litšoantšo tsa boleng bo tlaase le litšoantšo tsa boleng bo phahameng.
Ha theknoloji ea OCTA e ntse e tsoela pele ho fetoha le lebelo la ho skena le eketseha, liketsahalo tsa litšoantšo tsa maiketsetso le litšoantšo tsa boleng bo tlase li ka fokotseha.Lintlafatso ho software, joalo ka karolo e sa tsoa hlahisoa ea projeke ea ho tlosa lintho tsa khale, le eona e ka fokotsa meeli ena.Leha ho le joalo, mathata a mangata a sala e le ho nahana ka bakuli ba nang le ts'ebetso e mpe kapa bofokoli bo boholo ba mecha ea litaba ka linako tsohle bo fella ka litšoantšo tsa maiketsetso.Ha OCTA e ntse e sebelisoa haholo litekong tsa bongaka, ho hlokahala ho nahanoa ka hloko ho theha litataiso tse hlakileng bakeng sa maemo a amohelehang a maiketsetso a litšoantšo bakeng sa tlhahlobo ea litšoantšo.Tšebeliso ea mekhoa e tebileng ea ho ithuta ho litšoantšo tsa OCTA e na le tšepiso e kholo 'me lipatlisiso tse ling li hlokahala sebakeng sena ho hlahisa mokhoa o matla oa ho laola boleng ba setšoantšo.
Khoutu e sebelisitsoeng liphuputsong tsa hajoale e fumaneha sebakeng sa polokelo ea octa-qc, https://github.com/rahuldhodapkar/octa-qc.Lintlha tse hlahisitsoeng le / kapa tse hlahlobiloeng nakong ea thuto ea hajoale li fumaneha ho tsoa ho bangoli ba fapaneng ka kopo e utloahalang.
Spaide, RF, Fujimoto, JG & Waheed, NK Litšoantšo tsa maiketsetso tsa optical coherence angiography.Retina 35, 2163-2180 (2015).
Fenner, BJ le ba bang.Ho tsebahatsa likarolo tsa litšoantšo tse khethollang boleng le ho ikatisa ha litekanyo tsa methapo ea methapo ea methapo ea kutlo ho OCT angiography.BR.J. Ophthalmol.102, 509–514 (2018).
Lauerman, JL le ba bang.Tšusumetso ea theknoloji ea ho latela leihlo boleng ba setšoantšo sa OCT angiography ho degeneration ea macular e amanang le lilemo.Lebitla la lebitla.tleleniki.Exp.thuto ea mahlo.255, 1535-1542 (2017).
Babyuch AS et al.Litekanyo tsa OCTA capillary perfusion density li sebelisetsoa ho lemoha le ho hlahloba macular ischemia.ophthalmic ophthalmic.Retinal Laser Imaging 51, S30–S36 (2020).
He, K., Zhang, X., Ren, S., le Sun, J. Boithuto bo Tebileng ba Masala bakeng sa ho Tsebahatsa Litšoantšo.Ka 2016 Sebokeng sa IEEE mabapi le Pono ea Khomphutha le Tlhokomelo ea Mohlala (2016).
Lauerman, JL le ba bang.Tekolo ea boleng ba setšoantšo sa OCT angiographic e sebelisang mekhoa e tebileng ea ho ithuta.Lebitla la lebitla.tleleniki.Exp.thuto ea mahlo.257, 1641–1648 (2019).
Lauermann, J. et al.Ho ata ha liphoso tsa likarolo le lintho tse entsoeng ka letsoho ho OCT angiography ho itšetlehile ka lefu la retina.Lebitla la lebitla.tleleniki.Exp.thuto ea mahlo.256, 1807-1816 (2018).
Pask, Adam et al.Pytorch: Laeborari ea Boithuto bo Tebileng bo Bohlokoa, bo Phahameng.Ts'ebetso e tsoetseng pele ea tlhaiso-leseling ea methapo.tsamaiso.32, 8026–8037 (2019).
Deng, J. et al.ImageNet: Sebaka se Seholo sa Litšoantšo sa Hierarchical.Seboka sa 2009 sa IEEE mabapi le Pono ea Khomphutha le Kamohelo ea Paterone.248–255.(2009).
Krizhevsky A., Sutzkever I. le Hinton GE Imagenet classification sebelisa marang-rang a tebileng a convolutional neural.Ts'ebetso e tsoetseng pele ea tlhaiso-leseling ea methapo.tsamaiso.La 25, 1 (2012).