Ndatenda nekushanyira Nature.com.Uri kushandisa shanduro yebrowser ine tsigiro shoma yeCSS.Kuti uwane ruzivo rwakanyanya, tinokurudzira kuti ushandise browser yakagadziridzwa (kana kudzima Compatibility Mode muInternet Explorer).Pamusoro pezvo, kuti tive nechokwadi chekuenderera mberi nerutsigiro, tinoratidza saiti isina masitaera uye JavaScript.
Masiraidhi anoratidza zvinyorwa zvitatu pane siraidhi.Shandisa mabhatani ekumashure neanotevera kufamba nemumasiraidhi, kana mabhatani ekudzora masiraidhi ari kumagumo kuti ufambe nemumasiraidhi ega ega.
Optical coherence tomographic angiography (OCTA) inzira itsva yekuona kusingapindike kwemidziyo yeretinal.Kunyangwe OCTA iine akawanda anovimbisa makiriniki ekushandisa, kuona mhando yemufananidzo inoramba iri dambudziko.Isu takagadzira yakadzama yekudzidza yakavakirwa sisitimu tichishandisa iyo ResNet152 neural network classifier yakafanodzidziswa neImageNet kurongedza epamusoro capillary plexus mifananidzo kubva 347 scans e134 varwere.Iwo mapikicha akaongororwawo nemaoko sechokwadi chechokwadi nevaviri vakazvimiririra reti yemhando inotariswa yekudzidza.Nekuti zvinodiwa zvemhando yemifananidzo zvinogona kusiyana zvichienderana nekiriniki kana tsvakiridzo marongero, mamodheru maviri akadzidziswa, imwe yemhando yepamusoro yekuzivikanwa kwemifananidzo uye imwe yemhando yakaderera yekuzivikanwa kwemifananidzo.Yedu neural network modhi inoratidza yakanakisa nzvimbo pasi pe curve (AUC), 95% CI 0.96-0.99, \(\ kappa\) = 0.81), iri nani zvakanyanya kupfuura chiyero chechiratidzo chakataurwa nemuchina (AUC = 0.82, 95 % CI).0.77–0.86, \(\kappa\) = 0.52 uye AUC = 0.78, 95% CI 0.73–0.83, \(\kappa\) = 0.27, zvichiteerana).Chidzidzo chedu chinoratidza kuti nzira dzekudzidza dzemuchina dzinogona kushandiswa kugadzira nzira dzinochinjika uye dzakasimba dzekudzora mhando yemifananidzo yeOCTA.
Optical coherence tomographic angiography (OCTA) inzira ichangoburwa yakavakirwa pa optical coherence tomography (OCT) iyo inogona kushandiswa kune isiri-invasive visualization ye retinal microvasculature.OCTA inoyera mutsauko wekuratidzira mapatani kubva kune akadzokororwa mwenje pulses munzvimbo imwechete ye retina, uye kuvakazve kunogona kuverengerwa kuratidza tsinga dzeropa pasina kushandiswa kwedhayi kana zvimwe zvinosiyanisa.OCTA inogonesawo kudzika-resolution vascular imaging, ichibvumira varapi kuti vaongorore zvakasiyana uye zvakadzika midziyo yemidziyo, zvichibatsira kusiyanisa pakati pechirwere chechorioretinal.
Kunyange nzira iyi iri kuvimbisa, kusiyanisa kwemhando yemifananidzo kunoramba kuri dambudziko guru rekuongorora mifananidzo yakavimbika, zvichiita kuti kududzirwa kwemifananidzo kuve kwakaoma uye kudzivirira kupararira kwekiriniki yekugamuchirwa.Nekuti OCTA inoshandisa akawanda akateedzana OCT scans, inonyanya kunzwisiswa kumifananidzo yakagadzirwa pane yakajairwa OCT.Mazhinji ekutengesa OCTA mapuratifomu anopa yavo yemhando yemhando metric inonzi Signal Strength (SS) kana dzimwe nguva Signal Strength Index (SSI).Nekudaro, mifananidzo ine yakakwira SS kana yeSSI kukosha haivimbisi kusavapo kwechigadzirwa chemufananidzo, izvo zvinogona kukanganisa chero inotevera yemifananidzo kuongororwa uye kutungamirira kune zvisiri izvo zvekiriniki sarudzo.Zvigadzirwa zvemifananidzo zvakajairika zvinogona kuitika muOCTA imaging zvinosanganisira mafambiro ekugadzira, segmentation artifacts, media opacity artifacts, uye projection artifacts1,2,3.
Sezvo OCTA-yakatorwa zviyero zvakadai sevascular density zviri kuwedzera kushandiswa mukushandura tsvakurudzo, miedzo yekliniki uye maitiro ekurapa, pane chido chekukurumidzira kukudziridza maitiro akasimba uye akavimbika ekugadzirisa kwehutano hwemifananidzo kubvisa mifananidzo yemifananidzo4.Skip connections, inozivikanwawo seyasara yekubatanidza, ifungidziro mune neural network architecture inobvumira ruzivo kudarika convolutional layer uchichengeta ruzivo pazvikero zvakasiyana kana zvisungo5.Nekuti mifananidzo yakagadzirwa inogona kukanganisa diki-mwero uye yakakura-yakakura-chikero kuita kwemufananidzo, skip-yekubatanidza neural network yakanyatsokodzera kuita otomatiki iyi yemhando yekutonga basa5.Basa richangoburwa rakaratidza vimbiso yezvakadzama convolutional neural network yakadzidziswa nemhando yepamusoro data kubva kune vanhu vanofungidzira6.
Muchidzidzo ichi, tinodzidzisa yekubatanidza-kusvetuka convolutional neural network kuti itarise otomatiki kunaka kwemifananidzo yeOCTA.Isu tinovaka pabasa rekare nekugadzira mamodheru akasiyana ekuona mifananidzo yemhando yepamusoro uye mifananidzo yakaderera, sezvo zvinodikanwa zvemhando yemifananidzo zvingasiyana kune chaiyo yekiriniki kana yetsvakiridzo mamiriro.Isu tinoenzanisa mibairo yemanetiweki aya neconvolutional neural network pasina kushayikwa kubatanidza kuti tiongorore kukosha kwekubatanidza maficha pamatanho akawanda e granularity mukati mekudzidza kwakadzama.Takazoenzanisa zvatakawana nesimba rechiratidzo, chiyero chinogamuchirwa chemhando yemifananidzo inopihwa nevagadziri.
Chidzidzo chedu chaisanganisira varwere vane chirwere cheshuga vakapinda kuYale Eye Center pakati paAugust 11, 2017 naApril 11, 2019.Pakanga pasina maitiro ekubatanidza kana kusabvisa zvichienderana nezera, murume kana mukadzi, rudzi, kunaka kwemufananidzo, kana chimwe chinhu.
Mifananidzo yeOCTA yakatorwa pachishandiswa AngioPlex papuratifomu paCirrus HD-OCT 5000 (Carl Zeiss Meditec Inc, Dublin, CA) pasi pe8\(\nguva\)8 mm uye 6\(\nguva\)6 mm maprotocol ekufungidzira.Mvumo ine ruzivo rwekutora chikamu muchidzidzo ichi yakawanikwa kubva kune mumwe nemumwe anotora chidzidzo, uye Yale University Institutional Review Board (IRB) yakabvumidza kushandiswa kwemvumo ine ruzivo nekutorwa kwepasi rose kwevarwere ava.Kutevera misimboti yeDeclaration yeHelsinki.Chidzidzo ichi chakabvumidzwa neYale University IRB.
Mifananidzo yeplate yepamusoro yakaongororwa zvichibva pane yakambotsanangurwa Motion Artifact Score (MAS), iyo yakambotsanangurwa Segmentation Artifact Score (SAS), foveal centre, kuvapo kwemedia opacity, uye kuona kwakanaka kwemacapillaries madiki sekutemerwa nemuongorori wemufananidzo.Iyo mifananidzo yakaongororwa nevaviri vakazvimiririra vaongorori (RD neJW).Mufananidzo une zvibodzwa zve2 (zvakakodzera) kana zvese zvinotevera zvazadzikiswa: mufananidzo uri pakati pefovea (pasi pe100 pixels kubva pakati pemufananidzo), MAS 1 kana 2, SAS 1, uye media opacity ishoma pane 1. Ipo pamifananidzo yehukuru / 16, uye madiki capillaries anoonekwa mumifananidzo yakakura kupfuura 15/16.Mufananidzo wakapimwa 0 (hapana muyero) kana chimwe chezvinotevera chichiitwa: mufananidzo wacho uri kure-pakati, kana MAS iri 4, kana SAS iri 2, kana kupenya kwepakati kunopfuura 1/4 yemufananidzo, uye madiki capillaries haagone kugadziriswa kupfuura 1 mufananidzo / 4 kusiyanisa.Mimwe mifananidzo yese isingaite zvibodzwa 0 kana 2 inopihwa se 1 (clipping).
Pamusoro pemuonde.1 inoratidza mifananidzo yemuenzaniso kune yega yega fungidziro yakayerwa uye mifananidzo yakagadzirwa.Inter-rater kuvimbika kwezvibodzwa zvega zvakaongororwa naCohen's kappa weighting8.Zvibodzwa zvega ega zvemurengi wega wega zvinopfupikiswa kuti uwane zvibodzwa zvese zvemufananidzo wega wega, kubva pa0 kusvika 4. Mifananidzo ine zvibodzwa zvese zve4 inoonekwa seyakanaka.Mifananidzo ine zvibodzwa zvese zve0 kana 1 inoonekwa seyakaderera mhando.
A ResNet152 architecture convolutional neural network (Fig. 3A.i) pre-trained pamifananidzo kubva ku ImageNet database yakagadzirwa pachishandiswa fast.ai uye PyTorch framework5, 9, 10, 11. A convolutional neural network network inoshandisa vakadzidza. mafirita ekuongorora zvimedu zvemufananidzo kuti udzidze nzvimbo uye zvemuno maficha.Yedu yakadzidziswa ResNet ndeye 152-layer neural network inoratidzwa nemagapu kana "zvakasara zvinongedzo" izvo panguva imwe chete zvinofambisa ruzivo nezvakawanda zvigadziriso.Nekuronga ruzivo pazvisarudzo zvakasiyana pamusoro petiweki, chikuva chinogona kudzidza maficha emifananidzo yakaderera pamatanho akawanda eruzivo.Pamusoro pemuenzaniso wedu weResNet, takadzidzisawo AlexNet, yakanyatsodzidza neural network architecture, pasina kushayikwa kwekubatanidza kwekuenzanisa (Mufananidzo 3A.ii)12.Pasina makonekisheni asipo, network iyi haigone kutora maficha pamwero wepamusoro.
Iyo yekutanga 8\(\nguva\)8mm OCTA13 seti yemufananidzo yakakwidziridzwa uchishandisa yakachinjika uye yakatwasuka nzira yekuratidzira.Iyo dataset yakazara yakazopatsanurwa zvisina tsarukano padanho remufananidzo mukudzidziswa (51.2%), kuyedzwa (12.8%), hyperparameter tuning (16%), uye kusimbiswa (20%) datasets uchishandisa scikit-dzidza bhokisi rekushandisa python14.Mhosva mbiri dzakatariswa, imwe yakavakirwa pakuona chete yemhando yepamusoro mifananidzo (yakazara mamakisi 4) uye imwe yacho zvichibva pakuona chete yakaderera mhando mifananidzo (yakazara mamakisi 0 kana 1).Kune yega yega-yemhando yepamusoro uye yakaderera-mhando yekushandisa kesi, iyo neural network inodzidziswazve kamwe chete pamifananidzo yedu data.Muchiitiko chega chega chekushandisa, neural network yakadzidziswa kwenguva gumi, zvese kunze kwehuremu hwepamusoro hwakaomeswa nechando, uye huremu hwezvinhu zvese zvemukati zvakadzidzwa kwe40 epochs vachishandisa nzira yerusarura yekudzidza ine cross-entropy kurasikirwa basa 15, 16..Muchinjikwa entropy kurasikirwa basa chiyero cheiyo logarithmic chiyero chekusiyana pakati pezvakafanotaurwa network mavara uye data chaiyo.Panguva yekudzidziswa, gradient descent inoitwa pamatanho emukati eiyo neural network kuderedza kurasikirwa.Mwero wekudzidza, mwero wekudonhedza, uye kudzikisa huremu hyperparameters zvakagadziridzwa uchishandisa Bayesian optimization ine 2 isina kurongeka yekutanga mapoinzi uye gumi iterations, uye iyo AUC padhataset yakagadziridzwa uchishandisa hyperparameter sechinangwa chegumi nenomwe.
Mienzaniso inomiririra ye8 × 8 mm OCTA mifananidzo yepamusoro capillary plexuses yakawana 2 (A, B), 1 (C, D), uye 0 (E, F).Zvigadzirwa zvemifananidzo zvinoratidzwa zvinosanganisira mitsara inopenya (miseve), zvikamu zvekugadzira (asterisk), uye kupenya kwemedia (miseve).Mufananidzo (E) zvakare uri kure-pakati.
Receiver operating features (ROC) curves inobva yagadzirwa kune ese neural network modhi, uye injini mishumo yesimba remagetsi inogadzirwa kune yega yega-yemhando yepamusoro uye yepamusoro-mhando yekushandisa kesi.Nzvimbo iri pasi pe curve (AUC) yakaverengerwa ichishandisa pROC R pasuru, uye 95% nguva dzekuvimba uye p-maitiro akaverengerwa uchishandisa DeLong nzira18,19.Iwo akawedzera zvibodzwa zvemareti evanhu anoshandiswa sehwaro hwezvese ROC kuverenga.Nekuda kwesimba rechiratidzo rakashumwa nemuchina, iyo AUC yakaverengerwa kaviri: kamwe yemhando yepamusoro Scalability Score cutoff uye kamwe kune yakaderera mhando Scalability Score cutoff.Iyo neural network inofananidzwa neiyo AUC chiratidzo chesimba inoratidza yayo yekudzidziswa uye mamiriro ekuongorora.
Kuti uwedzere kuyedza iyo yakadzidziswa yakadzika modhi yekudzidza pane yakaparadzana dataset, yemhando yepamusoro uye yakaderera mhando mhando dzakashandiswa zvakananga kuongororo yekushanda kwe32 yakazara kumeso 6 \ (\ nguva\) 6mm pamusoro slab mifananidzo yakaunganidzwa kubva kuYale University.Misa yeziso yakatarisana panguva imwechete nemufananidzo 8 \ (\ nguva \) 8 mm.Iyo 6\(\×\) 6 mm mifananidzo yakaongororwa nemaoko neakaenzana mareta (RD neJW) nenzira yakafanana ne8\(\×\) 8 mm mifananidzo, AUC yakaverengerwa pamwe nekurongeka uye Cohen's kappa. .zvakaenzana .
Chiyero chekusaenzana kwekirasi ndeye 158: 189 (\(\ rho = 1.19\)) yemhando yakaderera yemhando uye 80: 267 (\(\ rho = 3.3\)) yemhando yepamusoro yemhando.Nekuti kusaenzana kwekirasi kwakaderera pane 1: 4, hapana shanduko yekuvaka yakaitwa kugadzirisa kusaenzana kwekirasi20,21.
Kuti uone zviri nani maitirwo ekudzidza, mepu dzekuita zvekirasi dzakagadzirwa kune ese mana akadzidziswa mamodhi ekudzidza zvakadzama: yemhando yepamusoro ResNet152 modhi, yakaderera mhando ResNet152 modhi, yemhando yepamusoro yeAlexNet modhi, uye yakaderera mhando yeAlexNet.Mamepu ekushandisa ekirasi anogadzirwa kubva kune yekuisa convolutional layers yeaya mamodheru mana, uye kupisa mepu anogadzirwa nekufukidza mamepu e activation ane masosi emifananidzo kubva ku8 × 8 mm uye 6 × 6 mm yekusimbisa sets22, 23.
R vhezheni 4.0.3 yakashandiswa pakuverenga kwese kwenhamba, uye maonesheni akagadzirwa pachishandiswa raibhurari yeggplot2 graphics tool.
Takaunganidza 347 mifananidzo yepamberi yepamusoro capillary plexus inoyera 8 \ (\ nguva \) 8 mm kubva kuvanhu 134.Muchina wakataura simba rechiratidzo pachiyero che 0 kusvika 10 yemifananidzo yese (zvinoreva = 6.99 ± 2.29).Pamifananidzo ye347 yakawanikwa, zera rekurevesa pakuongororwa raive 58.7 ± 14.6 makore, uye 39.2% vaibva kuvarwere vechirume.Pamifananidzo yose, 30.8% yakabva kuCaucasus, 32.6% kubva kuvatema, 30.8% kubva kuHispanics, 4% kubva kuAsia, uye 1.7% kubva kune mamwe marudzi (Tafura 1).)Kugoverwa kwezera kwevarwere vane OCTA kwakasiyana zvakanyanya zvichienderana nehutano hwemufananidzo (p <0.001).Nhamba yemifananidzo yepamusoro-soro muvarwere vaduku vane makore 18-45 vaiva 33.8% kana ichienzaniswa ne12.2% yemifananidzo yakaderera (Tafura 1).Kugoverwa kwechirwere cheshuga retinopathy chimiro chakasiyanawo zvakanyanya mumhando yemufananidzo (p <0.017).Pakati pemifananidzo yose yepamusoro, chikamu chevarwere vane PDR chaiva 18.8% chichienzaniswa ne38.8% yemifananidzo yose yakaderera (Tafura 1).
Maonero ega ega emifananidzo yese airatidza kuvimbika kusvika kwakasimba pakati pevanhu vanoverenga mifananidzo (Cohen's weighted kappa = 0.79, 95% CI: 0.76-0.82), uye pakanga pasina mapoinzi emifananidzo apo mareti akasiyana neanopfuura 1 (Fig. 2A)..Signal intensity yakabatana zvakanyanya nemanual scoring (Pearson product moment correlation = 0.58, 95% CI 0.51-0.65, p <0.001), asi mifananidzo yakawanda yakaonekwa seine simba repamusoro rechiratidzo asi yakaderera manual scoring (Fig. .2B).
Munguva yekudzidziswa kweResNet152 uye AlexNet architectures, iyo muchinjika-entropy kurasikirwa pakusimbisa uye kudzidziswa kunowira pamusoro pe50 epochs (Mufananidzo 3B, C).Kutendeseka kwechokwadi munguva yekupedzisira yekudzidziswa kwapfuura 90% kune ese emhando yepamusoro uye yakaderera mhando yemakesi ekushandisa.
Receiver performance curves inoratidza kuti ResNet152 modhi inodarika zvakanyanya simba rechiratidzo rinotaurwa nemuchina mune zvese zvakaderera uye zvemhando yepamusoro zvekushandisa kesi (p <0.001).Iyo ResNet152 modhi zvakare inonyanya kukunda iyo AlexNet architecture (p = 0.005 uye p = 0.014 yemhando yakaderera uye yepamusoro mhando kesi, zvichiteerana).Mamodheru akakonzeresa kune rimwe nerimwe remabasa aya akakwanisa kuwana AUC kukosha kwe0.99 uye 0.97, zvichiteerana, iyo iri nani zvakanyanya pane inoenderana AUC kukosha kwe0.82 uye 0.78 yemuchina chiratidzo chesimba index kana 0.97 uye 0.94 yeAlexNet. ..(Mufananidzo 3).Musiyano uripo pakati peResNet neAUC musimba rechiratidzo wakakwira kana uchiziva mifananidzo yemhando yepamusoro, zvichiratidza mamwe mabhenefiti ekushandisa ResNet yebasa iri.
Iwo magirafu anoratidza yega yega yakazvimiririra rater kugona uye kuenzanisa nesimba rechiratidzo rinotaurwa nemuchina.(A) Huwandu hwemapoinzi achaongororwa hunoshandiswa kugadzira huwandu hwemapoinzi anofanirwa kuongororwa.Mifananidzo ine chibodzwa chese che 4 inopihwa zvemhando yepamusoro, ukuwo mifananidzo ine chibodzwa chese scalability che1 kana pasi ichipihwa yakaderera.(B) Signal intensity inopindirana nefungidziro yemanyorero, asi mifananidzo ine simba repamusoro rechiratidzo inogona kunge iri yehurombo.Mutsetse wakatsvuka une doti unoratidza mugadziri akakurudzirwa hunhu hwakavakirwa pasimba rechiratidzo (signal simba \(\ge\)6).
ResNet kutamisa kudzidza kunopa kuvandudzwa kwakakosha mukuzivikanwa kwemhando yemifananidzo kune ese ari maviri emhando yakaderera uye emhando yepamusoro makesi ekushandisa zvichienzaniswa nemashina-akashumwa masaini mazinga.(A) Madhizaini ekuvaka akareruka easati adzidziswa (i) ResNet152 uye (ii) AlexNet architectures.(B) Kudzidzira nhoroondo uye mutevedzeri wekuita curves yeResNet152 kana ichienzaniswa nemichina yakashuma simba rechiratidzo uye AlexNet yakaderera maitiro.(C) ResNet152 mutori wekudzidziswa nhoroondo uye maitiro ekuita zvichienzaniswa nemuchina wakashumwa simba rechiratidzo uye AlexNet yepamusoro yemhando yepamusoro.
Mushure mekugadzirisa chikumbaridzo chemuganho wesarudzo, iyo yakanyanya kufanotaura kurongeka kweiyo ResNet152 modhi ndeye 95.3% yeiyo yakaderera mhando kesi uye 93.5% yemhando yepamusoro kesi (Tafura 2).Iyo yakanyanya kufanotaura kururamisa kweAlexNet modhi ndeye 91.0% yeiyo yakaderera mhando kesi uye 90.1% yemhando yepamusoro kesi (Tafura 2).Iyo yakanyanya chiratidzo chesimba rekufanotaura chokwadi ndeye 76.1% yeiyo yakaderera mhando yekushandisa kesi uye 77.8% yemhando yepamusoro yekushandisa kesi.Zvinoenderana neCohen's kappa (\(\kappa\)), chibvumirano pakati peResNet152 modhi uye vanofungidzira ndeye 0.90 yemhando yakaderera kesi uye 0.81 yemhando yepamusoro kesi.Cohen's AlexNet kappa ndeye 0.82 uye 0.71 yemhando yakaderera uye yepamusoro mhando dzekushandisa kesi, zvichiteerana.Cohen's sign simba kappa ndeye 0.52 uye 0.27 kune yakaderera uye yemhando yepamusoro makesi ekushandisa, zvichiteerana.
Kusimbiswa kwemhando yepamusoro uye yakaderera yemhando yekuzivikanwa pamifananidzo 6\(\x\) ye6 mm flat plate inoratidza kugona kweiyo yakadzidziswa modhi yekuona kunaka kwemufananidzo pamaparamita akasiyana ekufungidzira.Paunenge uchishandisa 6\(\x\) 6 mm shallow slabs yemhando yekufungidzira, iyo yakaderera mhando modhi yaive neAUC ye0.83 (95% CI: 0.69–0.98) uye yemhando yepamusoro modhi yaive neAUC ye0.85.(95% CI: 0.55-1.00) (Tafura 2).
Kutariswa kwekuona kweiyo yekuisa layer kirasi activation mepu yakaratidza kuti ese akadzidziswa neural network akashandisa maficha emifananidzo panguva yekuisa mufananidzo (Fig. 4A, B).Kune 8 \ (\ nguva \) 8 mm uye 6 \ (\ nguva \) 6 mm mifananidzo, iyo ResNet activation mifananidzo inoteedzera retinal vasculature.AlexNet activation mepu zvakare inotevera midziyo yeretinal, asi ine coarser resolution.
Iwo ekirasi activation mepu yeResNet152 uye AlexNet modhi inosimbisa maficha ane chekuita nemhando yemufananidzo.(A) Kirasi activation mepu inoratidza inobatika activation mushure mepamusoro retinal vasculature pa 8 \(\ nguva \) 8 mm yekusimbisa mifananidzo uye (B) nhanho padiki 6 \(\ nguva \) 6 mm yekusimbisa mifananidzo.LQ modhi yakadzidziswa pane yakaderera mhando maitiro, HQ modhi yakadzidziswa pamhando yepamusoro maitiro.
Zvakamboratidzwa kuti mhando yemufananidzo inogona kukanganisa zvakanyanya chero quantification yemifananidzo yeOCTA.Mukuwedzera, kuvapo kweretinopathy kunowedzera chiitiko chekugadzira mifananidzo7,26.Muchokwadi, mune yedu data, zvinoenderana nezvidzidzo zvekare, takawana kushamwaridzana kwakakosha pakati pekuwedzera zera uye kuomarara kwechirwere cheretinal uye kuderera kwemhando yemufananidzo (p <0.001, p = 0.017 yezera uye DR mamiriro, zvichiteerana; Tafura 1) 27 . Naizvozvo, zvakakosha kuti uongorore kunaka kwemufananidzo usati waita chero huwandu hwekuongorora kwemifananidzo yeOCTA.Zvidzidzo zvizhinji zvinoongorora mifananidzo yeOCTA zvinoshandisa muchina-chakashumwa chiratidzo chekusimba zvikumbaridzo kutonga kunze kwemhando yakaderera mifananidzo.Kunyange zvazvo simba rechiratidzo rave richiratidzwa kuti rinokanganisa kuwanda kweOCTA parameters, kusimba kwechiratidzo chepamusoro chete kunogona kunge kusina kukwana kubvisa mifananidzo ine mifananidzo yakagadzirwa2,3,28,29.Naizvozvo, zvinodikanwa kukudziridza nzira yakavimbika yekudzora kwemhando yemufananidzo.Kuti izvi zviitike, isu tinoongorora mashandiro emaitiro akatariswa akadzama ekudzidza achipikisa simba rechiratidzo rinotaurwa nemuchina.
Isu takagadzira akati wandei mamodheru ekuongorora mhando yemufananidzo nekuti akasiyana OCTA makesi ekushandisa anogona kunge aine akasiyana emhando yemifananidzo.Semuenzaniso, mifananidzo inofanira kuva yemhando yepamusoro.Mukuwedzera, chaiyo quantitative parameters yekufarira inokoshawo.Semuenzaniso, nzvimbo yefoveal avascular zone haina kutsamira pane turbidity yeiyo isiri yepakati svikiro, asi inokanganisa kuwanda kwemidziyo.Nepo tsvagiridzo yedu ichiramba ichitarisa pane yakajairika nzira yemhando yemufananidzo, isina kusungirirwa kune zvinodikanwa zvechero bvunzo, asi yakanangana kutsiva zvakananga simba rechiratidzo rinotaurwa nemuchina, isu tinotarisira kupa vashandisi hukuru hwekutonga kuitira kuti inogona kusarudza iyo chaiyo metric yekufarira kumushandisi.sarudza modhi inoenderana nehupamhi hwemhando yezvigadzirwa zvemifananidzo zvinoonekwa sezvinogamuchirika.
Kune yakaderera-yemhando yepamusoro uye yemhando yepamusoro zviratidziro, isu tinoratidza kuita kwakanakisa kwekubatanidza-kusina yakadzika convolutional neural network, ine maAUCs e0.97 uye 0.99 uye yakaderera-mhando mhando, zvichiteerana.Isu tinoratidzawo kuita kwepamusoro kwemaitiro edu ekudzidza kwakadzama kana tichienzaniswa nemazinga echiratidzo anongotaurwa nemichina chete.Skip connections inobvumira neural network kuti idzidze maficha pamatanho akati wandei ehudzame, kutora akakwenenzverwa maficha emifananidzo (semusiyano) pamwe neakajairwa maficha (semuenzaniso image centering30,31).Sezvo zviumbwa zvemifananidzo zvinokanganisa kunaka kwemufananidzo zvichinyanya kuzivikanwa pamusoro pehupamhi hwakasiyana, neural network architectures ine hukama husipo inogona kuratidza kuita kuri nani pane izvo zvisina basa remhando yemifananidzo.
Patinenge tichiedza muenzaniso wedu pamifananidzo ye 6 \ (\ × 6mm) OCTA, takaona kuderera kwekuita kwemaitiro emhando dzose dzepamusoro uye dzakaderera mhando (Fig. 2), kusiyana nehukuru hwemuenzaniso wakadzidziswa kurongedza.Kuenzaniswa neiyo ResNet modhi, iyo AlexNet modhi ine yakakura falloff.Kuita kuri nani kweResNet kunogona kunge kuri nekuda kwekugona kwezvakasara zvinongedzo kufambisa ruzivo pazvikero zvakawanda, izvo zvinoita kuti modhi ive yakasimba pakuronga mifananidzo yakatorwa pamakero akasiyana uye/kana magnification.
Mimwe misiyano pakati pe8 \(\×\) 8 mm mifananidzo uye 6 \(\×\) 6 mm mifananidzo inogona kutungamira mukusarudzika kwakashata, kusanganisira chikamu chepamusoro chemifananidzo ine foveal avascular nzvimbo, shanduko yekuonekwa, vascular arcades, uye hapana optic nerve pamufananidzo 6 × 6 mm.Pasinei neizvi, yedu yepamusoro yemhando yeResNet modhi yakakwanisa kuwana AUC ye85% ye6 \ (\x\) 6 mm mifananidzo, gadziriso iyo iyo modhi haina kudzidziswa, zvichiratidza kuti iyo yemhando yemhando yeruzivo yakavharirwa muneural network. kwakakodzera.kune imwe saizi yemufananidzo kana gadziriso yemuchina kunze kwekudzidziswa kwayo (Tafura 2).Sezvineiwo, ResNet- uye AlexNet-kufanana activation mamepu e8 \ (\ nguva \) 8 mm uye 6 \ (\ nguva \) 6 mm mifananidzo yakakwanisa kujekesa retinal midziyo mumatambudziko ose maviri, zvichiratidza kuti modhi ine ruzivo rwakakosha.zvinoshandiswa pakuronga marudzi ese eOCTA mifananidzo (Fig. 4).
Lauerman et al.Kuongororwa kwemhando yemifananidzo pamifananidzo yeOCTA yakaitwa nenzira yakafanana pachishandiswa Inception architecture, imwe skip-connection convolutional neural network6,32 uchishandisa nzira dzekudzidza dzakadzama.Ivo zvakare vakaganhurira chidzidzo kumifananidzo yepamusoro capillary plexus, asi vachingoshandisa madiki mapikicha e3 × 3 mm kubva kuOptovue AngioVue, kunyangwe varwere vane zvirwere zvakasiyana-siyana zvechorioretinal vakabatanidzwawo.Basa redu rinovaka panheyo dzadzo, kusanganisira akawanda mamodheru ekugadzirisa akasiyana emhando yemifananidzo zvikumbaridzo uye kusimbisa mhedzisiro yemifananidzo yehukuru hwakasiyana.Isu tinoshuma zvakare iyo AUC metric yemamodhi ekudzidza emuchina uye kuwedzera iwo atove anokatyamadza chokwadi (90%) 6 kune ese akaderera mhando (96%) uye yepamusoro mhando (95.7%) modhi6.
Kudzidziswa uku kune zvisingakwanisi.Kutanga, iyo mifananidzo yakawanikwa nemuchina mumwe chete weOCTA, kusanganisira chete mifananidzo yepamusoro capillary plexus pa8 \ (\ nguva \) 8 mm uye 6 \ (\ nguva \) 6 mm.Chikonzero chekusabvisa mifananidzo kubva muzvikamu zvakadzika ndechekuti mafungidziro ezvigadzirwa anogona kuita kuti kuongororwa kwemifananidzo kunyanye kuoma uye pamwe kusaenderana.Uyezve, mifananidzo yakangowanikwa muvarwere vane chirwere cheshuga, avo OCTA iri kubuda sechinhu chakakosha chekuongorora uye chekufungidzira33,34.Kunyangwe isu takakwanisa kuyedza modhi yedu pamifananidzo yehukuru hwakasiyana kuti tive nechokwadi chekuti mhedzisiro yacho yaive yakasimba, isu hatina kukwanisa kuona datasets dzakakodzera kubva kunzvimbo dzakasiyana, izvo zvakaganhurira ongororo yedu ye generalizability yemuenzaniso.Kunyange zvazvo mifananidzo yacho yakawanikwa kubva panzvimbo imwe chete, yakawanikwa kubva kuvarwere vemarudzi akasiyana-siyana uye madzinza akasiyana, iyo isimba rakasiyana rekudzidza kwedu.Nekusanganisira kusiyana-siyana mumaitirwo edu ekudzidzisa, tinovimba kuti mhedzisiro yedu ichave yakajairika munzira yakakura, uye kuti isu tinodzivirira encoding rusaruraganda mumhando dzatinodzidzisa.
Chidzidzo chedu chinoratidza kuti yekubatanidza-kusvetuka neural network inogona kudzidziswa kuti iwane yakakwirira kuita mukuona OCTA mufananidzo wemhando.Isu tinopa aya mamodheru sezvishandiso zvekuwedzera tsvakiridzo.Nekuti mametric akasiyana anogona kunge aine akasiyana emhando yemhando zvinodiwa, yemhando yekudzora modhi inogona kugadzirwa kune yega yega metric uchishandisa chimiro chakagadzwa pano.
Tsvagiridzo yeramangwana inofanira kusanganisira mifananidzo yehukuru hwakasiyana kubva pakadzika dzakasiyana uye akasiyana OCTA michina kuti uwane yakadzama yekudzidza yemhando yemhando yekuongorora maitiro ayo anogona kugadzirwa kune OCTA mapuratifomu uye ekufungidzira mapuroteni.Tsvagiridzo yemazuva ano zvakare yakavakirwa pane inotariswa yakadzama nzira dzekudzidza dzinoda kuongororwa kwevanhu uye kuongororwa kwemifananidzo, iyo inogona kuve yakanyanya kushanda uye inopedza nguva kune makuru dataset.Zvinoramba zvichionekwa kana nzira dzisina kutariswa dzakadzika dzekudzidza dzinogona kusiyanisa zvakaringana pakati pemifananidzo yakaderera uye mifananidzo yemhando yepamusoro.
Sezvo tekinoroji yeOCTA ichiramba ichishanduka uye kumhanya kwekuongorora kunowedzera, chiitiko chezvigadzirwa zvemifananidzo uye mifananidzo yakashata inogona kudzikira.Kuvandudzwa kwesoftware, senge ichangoburwa fungidziro yekubvisa artifact ficha, inogona zvakare kudzikamisa izvi zvipimo.Nekudaro, matambudziko mazhinji anoramba ari ekufungidzira kwevarwere vasina kugadzirisa zvakanaka kana yakakosha midhiya turbidity nguva dzose inoguma nemifananidzo yakagadzirwa.Sezvo OCTA ichiwedzera kushandiswa mumakiriniki ekuedzwa, kunyatsotariswa kunodiwa kuti uise nhungamiro dzakajeka dzematanho anogamuchirwa echifananidzo chekuongorora mifananidzo.Kushandiswa kwemaitiro akadzama ekudzidza kumifananidzo yeOCTA kune vimbiso huru uye kumwe kutsvagisa kunodiwa munzvimbo ino kukudziridza nzira yakasimba yekudzora kunaka kwemufananidzo.
Iyo kodhi inoshandiswa mukutsvagisa kwazvino inowanikwa mune octa-qc repository, https://github.com/rahuldhodapkar/octa-qc.Datasets akagadzirwa uye / kana kuongororwa panguva yechidzidzo chazvino anowanikwa kubva kune vakasiyana vanyori pakukumbira zvine musoro.
Spaide, RF, Fujimoto, JG & Waheed, NK Image artifacts mu optical coherence angiography.Retina 35, 2163–2180 (2015).
Fenner, BJ nevamwe.Kuzivikanwa kwezvimiro zvekufungidzira zvinotaridza kunaka uye kuberekana kweretinal capillary plexus density zviyero muOCT angiography.BR.J. Ophthalmol.102, 509–514 (2018).
Lauerman, JL nevamwe.Kupesvedzera tekinoroji yekutevera ziso pamhando yemufananidzo weOCT angiography mune zera-ane hukama macular degeneration.Guva arch.kliniki.Exp.ophthalmology.255, 1535–1542 (2017).
Babyuch AS et al.OCTA capillary perfusion density zviyero zvinoshandiswa kuona uye kuongorora macular ischemia.ophthalmic kuvhiyiwa.Retinal Laser Imaging 51, S30–S36 (2020).
Iye, K., Zhang, X., Ren, S., uye Zuva, J. Deep Residual Learning for Image Recognition.Muna 2016 paIEEE Musangano paComputer Vision uye Pattern Recognition (2016).
Lauerman, JL nevamwe.Automated OCT angiographic mufananidzo wemhando yekuongorora uchishandisa yakadzika yekudzidza algorithms.Guva arch.kliniki.Exp.ophthalmology.257, 1641–1648 (2019).
Lauermann, J. nevamwe.Kuwanda kwekukanganisa kwezvikamu uye zvigadziriso zvekufamba muOCT angiography zvinoenderana nechirwere che retina.Guva arch.kliniki.Exp.ophthalmology.256, 1807–1816 (2018).
Pask, Adam et al.Pytorch: Yakakosha, Yepamusoro-Kuita Kwakadzika Kudzidza Raibhurari.Yepamberi yekugadzirisa yeruzivo rweneural.system.32, 8026–8037 (2019).
Deng, J. nevamwe.ImageNet: Yakakura-Chikero Hierarchical Image Database.2009 IEEE Musangano paComputer Vision uye Pateni Kuzivikanwa.248–255.(2009).
Krizhevsky A., Sutzkever I. uye Hinton GE Imagenet classification uchishandisa yakadzika convolutional neural network.Yepamberi yekugadzirisa yeruzivo rweneural.system.25, 1 (2012).