Enkosi ngokundwendwela i-Nature.com.Usebenzisa uguqulelo lwebrawuza enenkxaso enyiniweyo yeCSS.Ngowona mava angcono, sicebisa ukuba usebenzise isikhangeli esihlaziyiweyo (okanye uvale iModi yokuThelela kwi-Internet Explorer).Ukongeza, ukuqinisekisa inkxaso eqhubekayo, sibonisa indawo ngaphandle kwezitayela kunye neJavaScript.
Izilayidi ezibonisa amanqaku amathathu kwisilayidi ngasinye.Sebenzisa amaqhosha angasemva nalandelayo ukuhamba kwizilayidi, okanye amaqhosha okulawula isilayidi ekupheleni ukuya kwisilayidi ngasinye.
I-Optical coherence tomographic angiography (OCTA) yindlela entsha ye-non-invasive visualization yemikhumbi ye-retinal.Nangona i-OCTA inezicelo ezininzi zeklinikhi ezithembisayo, ukumisela umgangatho wesithombe kuhlala kungumngeni.Siphuhlise inkqubo esekelwe kufundo olunzulu sisebenzisa i-ResNet152 neural network classifier eqeqeshwe kwangaphambili nge-ImageNet ukuhlela imifanekiso ye-capillary plexus engaphezulu ukusuka kwi-347 scans ye-134 yezigulane.Imifanekiso nayo yavavanywa ngesandla njengenyaniso eyinyani ngabareyitha ababini abazimeleyo bemodeli yokufunda egadiweyo.Ngenxa yokuba iimfuno zomgangatho wemifanekiso zinokwahluka ngokuxhomekeke kwizicwangciso zeklinikhi okanye zophando, iimodeli ezimbini zaqeqeshwa, enye ikumgangatho ophezulu wokuqatshelwa komfanekiso kunye nenye yokuqatshelwa komgangatho ophantsi.Imodeli yethu yenethiwekhi ye-neural ibonisa indawo egqwesileyo phantsi kwe-curve (AUC), 95% CI 0.96-0.99, \ (\ kappa \) = 0.81), eyona nto ingcono kakhulu kunomgangatho wesignali ochazwe ngumatshini (AUC = 0.82, 95 % CI).0.77–0.86, \(\kappa\) = 0.52 kunye ne-AUC = 0.78, 95% CI 0.73–0.83, \(\kappa\) = 0.27, ngokulandelanayo).Uphononongo lwethu lubonisa ukuba iindlela zokufunda zoomatshini zingasetyenziselwa ukuphuhlisa iindlela eziguquguqukayo kunye nezomeleleyo zokulawula umgangatho wemifanekiso ye-OCTA.
I-Optical coherence tomographic angiography (OCTA) bubuchule obutsha ngokwentelekiso obusekwe kwi-optical coherence tomography (OCT) enokuthi isetyenziswe kwi-non-invasive visualization ye-retinal microvasculature.I-OCTA ilinganisa umahluko kwiipateni zokubonisa ukusuka kwi-pulses yokukhanya ephindaphindiweyo kwindawo enye ye-retina, kwaye ukwakhiwa kwakhona kunokubalwa ukuveza imithambo yegazi ngaphandle kokusetyenziswa kwedayi okanye ezinye ii-agent ezichaseneyo.I-OCTA ikwavumela ukuba kubonwe ubunzulu be-vascular imaging, ivumela oogqirha ukuba bahlole ngokwahlukeneyo iileya zenqanawa ezingaphezulu kunye nezinzulu, zinceda ukwahlula phakathi kwesifo se-chorioretinal.
Ngelixa obu buchule buthembisa, ukuhluka komgangatho wemifanekiso kuhlala kungumngeni omkhulu wokuhlalutya umfanekiso othembekileyo, okwenza ukutolika komfanekiso kube nzima kwaye kuthintele ukwamkelwa kwekliniki ngokubanzi.Kuba i-OCTA isebenzisa izikena ze-OCT ezininzi ezilandelelanayo, inovakalelo ngakumbi kwimifanekiso ye-artifacts kune-OCT eqhelekileyo.Uninzi lweeplatifti ze-OCTA zorhwebo zibonelela nge-metric yazo yomgangatho womfanekiso obizwa ngokuba yi-Signal Strength (SS) okanye ngamanye amaxesha i-Signal Strength Index (SSI).Nangona kunjalo, imifanekiso enexabiso eliphezulu le-SS okanye le-SSI ayiqinisekisi ukungabikho kobugcisa bemifanekiso, enokuchaphazela naluphi na uhlalutyo lomfanekiso olulandelayo kwaye lukhokelela kwizigqibo zeklinikhi ezingalunganga.Imifanekiso eqhelekileyo yemifanekiso enokuthi yenzeke kwi-imaging ye-OCTA ibandakanya izinto ezishukumayo, ii-artifacts ze-segmentation, i-media opacity artifacts, kunye ne-projection 1,2,3.
Njengoko imilinganiselo ephuma kwi-OCTA efana ne-vascular density isetyenziswa ngakumbi kuphando lokuguqulela, izilingo zeklinikhi kunye nokusebenza kweklinikhi, kukho imfuneko engxamisekileyo yokuphuhlisa iinkqubo zokulawula umgangatho womfanekiso oqinileyo kunye nothembekileyo ukuze kupheliswe imifanekiso ye-artefacts4.Uqhagamshelo lokutsiba, olukwabizwa ngokuba luqhagamshelo olushiyekileyo, luqikelelo kulwakhiwo lwenethiwekhi ye-neural evumela ulwazi ukuba ludlule kumaleko we-convolutional ngelixa ugcina ulwazi kwizikali ezahlukeneyo okanye izisombululo5.Ngenxa yokuba ii-artifacts zemifanekiso zinokuchaphazela umgangatho omncinci kunye nokusebenza komfanekiso omkhulu ngokubanzi, i-skip-connection neural networks ifaneleke ngokufanelekileyo ukwenza lo msebenzi wokulawula umgangatho5.Umsebenzi osandula ukupapashwa ubonise isithembiso sothungelwano olunzulu lwe-neural convolutional oluqeqeshwe kusetyenziswa idatha ekumgangatho ophezulu evela kubaqikeleli babantu6.
Kolu phononongo, siqeqesha unxibelelwano-lokutsiba inethiwekhi ye-neural ye-convolutional ukumisela ngokuzenzekelayo umgangatho wemifanekiso ye-OCTA.Sakha kumsebenzi wangaphambili ngokuphuhlisa iimodeli ezahlukeneyo zokuchonga imifanekiso esemgangathweni ophezulu kunye nemifanekiso ephantsi, njengoko iimfuno zomgangatho wemifanekiso zinokwahluka kwiimeko ezithile zeklinikhi okanye zophando.Sithelekisa iziphumo zolu thungelwano kunye nothungelwano lwe-neural lwe-convolutional ngaphandle koqhagamshelwano olulahlekileyo ukuvavanya ixabiso lokubandakanya iimpawu kumanqanaba amaninzi e-granularity ngaphakathi kokufunda okunzulu.Siye sathelekisa iziphumo zethu kunye namandla omqondiso, umlinganiselo owamkelwa ngokuqhelekileyo womgangatho womfanekiso onikwa ngabavelisi.
Uphononongo lwethu lubandakanya izigulane ezinesifo seswekile eziye kwi-Yale Eye Centre phakathi kwe-11 ka-Agasti 2017 kunye no-Aprili 11, i-2019.Kwakungekho milinganiselo yokubandakanywa okanye yokukhutshwa ngokusekelwe kwiminyaka yobudala, isini, uhlanga, umgangatho womfanekiso, okanye nayiphi na enye into.
Imifanekiso ye-OCTA ifunyenwe kusetyenziswa iqonga le-AngioPlex kwi-Cirrus HD-OCT 5000 (Carl Zeiss Meditec Inc, Dublin, CA) phantsi kwe-8 \ (\ amaxesha \) 8 mm kunye ne-6 \ (\ amaxesha \) 6 mm iiprothokholi zokucinga.Imvume enolwazi yokuthatha inxaxheba kuphononongo ifunyenwe kumthathi-nxaxheba ngamnye wophando, kwaye iBhodi yokuHlola yeZiko leYunivesithi yaseYale (IRB) yavuma ukusetyenziswa kwemvume enolwazi ngokufota kwihlabathi kuzo zonke ezi zigulana.Ukulandela imigaqo yeSibhengezo saseHelsinki.Uphononongo lwamkelwe yi-IRB yeYunivesithi yaseYale.
Imifanekiso yepleyiti yomphezulu yavavanywa ngokusekelwe kwiNqaku elichazwe ngaphambili le-Motion Artifact Score (MAS), i-Segmentation Artifact Score echazwe ngaphambili (i-SAS), iziko le-foveal, ubukho be-media opacity, kunye nokubonwa kakuhle kwee-capillaries ezincinci njengoko kunqunywe ngumvavanyi womfanekiso.Imifanekiso yahlalutywa ngabavavanyi ababini abazimeleyo (RD kunye neJW).Umfanekiso unamanqaku akwinqanaba lesi-2 (ufanelekile) ukuba zonke ezi nqobo zilandelayo zihlangatyeziwe: umfanekiso uphakathi kwifovea (ngaphantsi kwe-100 ipixels ukusuka embindini womfanekiso), iMAS yi-1 okanye 2, SAS yi-1, kwaye i-media opacity ingaphantsi kwe-1. Ibonisa kwimifanekiso yobukhulu / i-16, kunye nee-capillaries ezincinci zibonwa kwimifanekiso emikhulu kune-15/16.Umfanekiso ukalwe 0 (akukho rating) ukuba naziphi na kwezi zilandelayo zifezekisiweyo: umfanekiso awukho embindini, ukuba iMAS yi-4, ukuba i-SAS yi-2, okanye i-avareji yokukhanya ingaphezulu kwe-1/4 yomfanekiso, kwaye i-capillaries encinci ayikwazi ukulungiswa ngaphezu komfanekiso we-1 / 4 ukuhlula.Yonke eminye imifanekiso engahambelani nemilinganiselo yokukora i-0 okanye i-2 ifakwe amanqaku njenge-1 (ikliphu).
Kwikhiwane.I-1 ibonisa imifanekiso yesampulu kuqikelelo olulinganisiweyo kunye nobugcisa bemifanekiso.Ukuthembeka kwamanqaku aphakathi kwamanqaku ngamnye kwavavanywa nguCohen's kappa weighting8.Amanqaku omntu ngamnye womlinganiso ngamnye ashwankathelwa ukufumana amanqaku apheleleyo kumfanekiso ngamnye, ukusuka kwi-0 ukuya kwisi-4. Imifanekiso enamanqaku ewonke angama-4 ithathwa njengelungileyo.Imifanekiso enenqaku elipheleleyo le-0 okanye i-1 ithathwa njengomgangatho ophantsi.
I-ResNet152 i-architecture convolutional network ye-neural (Fig. 3A.i) eqeqeshwe kwangaphambili kwimifanekiso esuka kwi-database ye-ImageNet yaveliswa kusetyenziswa i-fast.ai ne-PyTorch framework5, 9, 10, 11. Inethiwekhi ye-convolutional neural yinethiwekhi esebenzisa okufundiweyo. izihluzi zokuskena amaqhekeza emifanekiso ukufunda isithuba kunye neempawu zasekhaya.I-ResNet yethu eqeqeshiweyo yi-152-layer neural network ebonakaliswe zizikhewu okanye "uqhagamshelo olushiyekileyo" oluthumela ngaxeshanye ulwazi ngezigqibo ezininzi.Ngokubonisa ulwazi kwizisombululo ezahlukeneyo kwinethiwekhi, iqonga linokufunda iimpawu zemifanekiso ekumgangatho ophantsi kumanqanaba amaninzi eenkcukacha.Ukongeza kwimodeli yethu ye-ResNet, siye saqeqesha i-AlexNet, i-neural network architecture efundwe kakuhle, ngaphandle koqhagamshelwano olulahlekileyo lokuthelekisa (Umfanekiso 3A.ii)12.Ngaphandle koqhagamshelo olungekhoyo, le nethiwekhi ayizukwazi ukuthatha iimpawu kwigranularity ephezulu.
Uqobo lwe-8\(\ amaxesha\)8mm OCTA13 iseti yomfanekiso yandisiwe kusetyenziswa iindlela zokubonisa ezithe tye nezithe nkqo.I-dataset epheleleyo yahlulwe ngokungaqhelekanga kwinqanaba lomfanekiso ekuqeqesheni (51.2%), uvavanyo (12.8%), i-hyperparameter tuning (16%), kunye nokuqinisekiswa (20%) iiseti zedatha usebenzisa i-scikit-learn toolbox python14.Amatyala amabini acatshangelwa, enye isekelwe ekufumaneni kuphela imifanekiso ephezulu kakhulu (amanqaku apheleleyo 4) kunye nomnye ngokusekelwe ekuboneni kuphela imifanekiso ephantsi kakhulu (inqaku elipheleleyo 0 okanye 1).Kwimeko nganye yomgangatho ophezulu kunye nomgangatho ophantsi wokusetyenziswa, inethiwekhi ye-neural iphinda iqeqeshwe kube kanye kwidatha yomfanekiso wethu.Kwimeko nganye yokusetyenziswa, inethiwekhi ye-neural yaqeqeshelwa ii-epochs ze-10, zonke kodwa izisindo eziphakamileyo zomaleko zazikhenkcezisiwe, kwaye imilinganiselo yazo zonke iiparitha zangaphakathi zafundwa kwii-epoch ze-40 kusetyenziswa indlela yokufunda ecalucalulo kunye ne-cross-entropy loss loss function 15, 16..Umsebenzi wokulahleka kwe-entropy ngumlinganiselo wesikali se-logarithmic yokungafani phakathi kweelebhile zenethiwekhi eziqikelelweyo kunye nedatha yangempela.Ngexesha loqeqesho, ukuhla kwe-gradient kwenziwa kwiiparameters zangaphakathi ze-neural network ukunciphisa ilahleko.Izinga lokufunda, izinga lokuyeka, kunye ne-hyperparameters yokunciphisa ubunzima kusetyenziswa ukusetyenziswa kwe-Bayesian optimization kunye ne-2 i-random starting points and 10 iterations, kunye ne-AUC kwi-dataset yalungiswa ngokusebenzisa i-hyperparameters njengethagethi ye-17.
Imizekelo emele i-8 × 8 mm ye-OCTA yemifanekiso ye-capillary plexuses ephezulu ifumene amanqaku e-2 (A, B), 1 (C, D), kunye ne-0 (E, F).Ii-artifacts zomfanekiso obonisiweyo ziquka iilayini ezidayizayo (iintolo), ii-artifacts zecandelo (iinkwenkwezi), kunye nokungafihli kwemidiya (iintolo).Umfanekiso (E) ukwangekho embindini.
Iimpawu zokusebenza ze-receiver (ROC) ii-curves zenziwa kuzo zonke iimodeli zenethiwekhi ye-neural, kunye neengxelo zamandla omqondiso we-injini zenziwe kwimeko nganye yomgangatho ophantsi kunye nomgangatho ophezulu wokusetyenziswa.Indawo ephantsi kwegophe (AUC) yabalwa kusetyenziswa iphakheji ye-pROC R, kwaye i-95% yamathuba okuzithemba kunye namaxabiso e-p abalwe kusetyenziswa indlela ye-DeLong18,19.Amanqaku aqokelelweyo ereyitha zabantu asetyenziswa njengesiseko sazo zonke izibalo zeROC.Ngamandla omqondiso oxelwe ngumatshini, i-AUC ibalwa kabini: kanye kumgangatho ophezulu we-Scalability Score cutoff kwaye kube kanye kumgangatho ophantsi we-Scalability Score cutoff.Inethiwekhi ye-neural ithelekiswa namandla omqondiso we-AUC ebonisa uqeqesho lwayo kunye neemeko zokuvavanya.
Ukuvavanya ngakumbi imodeli yokufunda enzulu eqeqeshiweyo kwiseti yedatha eyahlukileyo, umgangatho ophezulu kunye neemodeli zomgangatho ophantsi ziye zasetyenziswa ngokuthe ngqo kuvandlakanyo lokusebenza 32 ubuso obugcweleyo 6\(\ amaxesha\) 6mm imifanekiso yesilabhu yomphezulu eqokelelwe kwiYunivesithi yaseYale.I-Eye Mass igxile ngexesha elifanayo nomfanekiso 8 \ (\ amaxesha \) 8 mm.I-6\(\×\) imifanekiso ye-6 mm ihlolwe ngesandla ngamaxabiso afanayo (RD kunye ne-JW) ngendlela efanayo ne-8\(\×\) ye-8 mm imifanekiso, i-AUC yabalwa kunye nokuchaneka kunye ne-kappa kaCohen. .ngokulinganayo .
Umlinganiselo wokungalingani kweklasi ngu-158: 189 (\ (\ rho = 1.19 \)) kwimodeli yekhwalithi ephantsi kunye ne-80: 267 (\ (\ rho = 3.3 \)) kwimodeli ephezulu.Ngenxa yokuba umlinganiselo wokungalingani kweklasi ungaphantsi kwe-1: 4, akukho lutshintsho oluthile lwezakhiwo lwenziwe ukulungisa ukungalingani kweklasi20,21.
Ukuze ube nombono ongcono wenkqubo yokufunda, iimephu zokuvula iklasi zenziwa kuzo zone iimodeli zokufunda ezinzulu eziqeqeshiweyo: imodeli yeResNet152 ekumgangatho ophezulu, imodeli ephantsi yeResNet152, imodeli ye-AlexNet ekumgangatho ophezulu, kunye nemodeli ye-AlexNet ephantsi.Iimephu zokuvula zeklasi ziveliswa kulwaleko lwe-convolutional yegalelo lale mifuziselo mine, kwaye iimephu zobushushu ziveliswa ngokugquma iimephu zokuvula ezinemifanekiso ephuma kwi-8 × 8 mm kunye ne-6 × 6 mm yokuqinisekisa iiseti22, 23.
Uguqulelo lwe-R 4.0.3 lusetyenziselwe lonke ubalo lwamanani, kunye nokubonwayo kwenziwa kusetyenziswa ithala leencwadi leggplot2 yesixhobo segrafiki.
Siqokelele imifanekiso engaphambili ye-347 ye-capillary plexus engaphezulu enomlinganiselo we-8 \ (\ amaxesha \) 8 mm ukusuka kubantu abayi-134.Umatshini wabika amandla omqondiso kwisikali se-0 ukuya kwi-10 kuyo yonke imifanekiso (ithetha = 6.99 ± 2.29).Kwimifanekiso ye-347 efunyenweyo, iminyaka yobudala ekuhlolweni yayiyi-58.7 ± 14.6 iminyaka, kwaye i-39.2% yayisuka kwizigulane zamadoda.Kuyo yonke imifanekiso, i-30.8% isuka eCaucasus, 32.6% isuka kwabaNtsundu, 30.8% isuka kwiHispanics, 4% ivela kumaAsia, kunye ne-1.7% yezinye iintlanga (iTheyibhile 1).).Ukuhanjiswa kweminyaka yezigulane nge-OCTA kwahluke kakhulu kuxhomekeke kumgangatho womfanekiso (p <0.001).Ipesenti yemifanekiso esemgangathweni ophezulu kwizigulane ezincinci ezineminyaka eyi-18-45 iminyaka yayingu-33.8% xa kuthelekiswa ne-12.2% yemifanekiso ephantsi (iThebhile 1).Ukuhanjiswa kwesimo se-diabetes retinopathy kwakhona kwahluka kakhulu kumgangatho wesithombe (p <0.017).Phakathi kwayo yonke imifanekiso esemgangathweni ophezulu, ipesenti yezigulane ezine-PDR yi-18.8% xa kuthelekiswa ne-38.8% yazo zonke imifanekiso ephantsi (iThebhile 1).
Ukulinganisa komntu ngamnye kuyo yonke imifanekiso kubonise ukuthembeka okuphakathi ukuya kokuqina phakathi kwabantu abafunda imifanekiso (i-Cohen's weighted kappa = 0.79, 95% CI: 0.76-0.82), kwaye akukho manqaku emifanekiso apho abalingani bahluke ngaphezu kwe-1 (Fig. 2A)..Ubungakanani besignali buhambelana ngokuphawulekayo kunye nokufaka amanqaku okwenziwa ngesandla (i-Pearson product moment correlation = 0.58, 95% CI 0.51-0.65, p <0.001), kodwa imifanekiso emininzi ichongiwe njengokuba inomlinganiselo ophezulu wesignali kodwa i-low manual scoring (Fig. .2B).
Ngethuba loqeqesho lwe-ResNet152 kunye ne-AlexNet ye-architectures, ilahleko ye-cross-entropy ekuqinisekiseni kunye noqeqesho iwela kwii-epochs ze-50 (Umfanekiso 3B, C).Ukuchaneka kokuqinisekiswa kwixesha lokugqibela loqeqesho kungaphezulu kwe-90% kuzo zombini iimeko ezikumgangatho ophezulu kunye nomgangatho ophantsi wokusetyenziswa.
Iingqungquthela zokusebenza kwe-receiver zibonisa ukuba imodeli ye-ResNet152 idlula kakhulu amandla omqondiso ochazwe ngumatshini kwiimeko zombini eziphantsi kunye eziphezulu zokusetyenziswa (p <0.001).Imodeli ye-ResNet152 iphinda iphumelele kakhulu kwi-architecture ye-AlexNet (p = 0.005 kunye ne-p = 0.014 kumgangatho ophantsi kunye neemeko eziphezulu, ngokulandelanayo).Iimodeli eziye zaphumela kulo msebenzi ngamnye zakwazi ukufikelela kumaxabiso e-AUC ye-0.99 kunye ne-0.97, ngokulandelelana, nto leyo ingcono kakhulu kunexabiso le-AUC elihambelana ne-0.82 kunye ne-0.78 yesalathiso samandla omatshini okanye i-0.97 kunye ne-0.94 ye-AlexNet ..(Umfanekiso wesi-3).Umahluko phakathi kwe-ResNet kunye ne-AUC kumandla omqondiso uphezulu xa uqaphela imifanekiso esemgangathweni ophezulu, ebonisa inzuzo eyongezelelweyo yokusebenzisa i-ResNet kulo msebenzi.
Iigrafu zibonisa amandla omlinganisi ngamnye ozimeleyo wokufumana amanqaku kwaye athelekise amandla omqondiso oxelwe ngumatshini.(A) Isimbuku samanqaku aza kuvavanywa sisetyenziselwa ukwenza inani lilonke lamanqaku aza kuvavanywa.Imifanekiso ene-scalability score ye-4 inikwe umgangatho ophezulu, ngelixa imifanekiso ene-scalability score ye-1 okanye ngaphantsi inikwe umgangatho ophantsi.(B) Ukuqina komqondiso kuhambelana noqikelelo lwezandla, kodwa imifanekiso enobunzulu bophawu oluphezulu inokuba kumgangatho ophantsi.Umgca onamachaphaza obomvu ubonisa umgangatho womgangatho ocetyiswayo womenzi ngokusekelwe kumandla omqondiso (amandla omqondiso \(\ge\)6).
Ukufundwa kokudluliselwa kwe-ResNet kubonelela ngophuculo olubalulekileyo ekuchongeni umgangatho womfanekiso kuzo zombini umgangatho ophantsi kunye neemeko zokusetyenziswa ezikumgangatho ophezulu xa kuthelekiswa namanqanaba omqondiso axelwe ngumatshini.(A) Idayagramu yoyilo olululiweyo yoqeqesho lwangaphambili (i) ResNet152 kunye (ii) noyilo lwe-AlexNet.(B) Imbali yoqeqesho kunye neengqungquthela zokusebenza kommkeli we-ResNet152 xa kuthelekiswa nomatshini oxeliweyo wamandla omqondiso kunye ne-AlexNet yemigangatho ephantsi yomgangatho.(C) Imbali yoqeqesho lwabamkeli be-ResNet152 kunye neengqungquthela zokusebenza xa kuthelekiswa nomatshini oxeliweyo wamandla omqondiso kunye ne-AlexNet yomgangatho ophezulu.
Emva kokulungelelanisa umda wesigqibo somda, ukuchaneka okuphezulu kokuchaneka kwemodeli ye-ResNet152 yi-95.3% kwimeko ephantsi kunye ne-93.5% kwimeko ephezulu (iThebhile 2).Ubuninzi bokuchaneka kokuchaneka kwemodeli ye-AlexNet yi-91.0% yecala eliphantsi kunye ne-90.1% kwimeko ephezulu (iThebhile 2).Ubuninzi besibonakaliso sokuchaneka kokuchaneka kokuchaneka kwe-76.1% kwimeko yokusetyenziswa komgangatho ophantsi kunye ne-77.8% kwimeko yokusetyenziswa komgangatho ophezulu.Ngokutsho kweCohen's kappa (\(\kappa\)), isivumelwano phakathi kwemodeli ye-ResNet152 kunye nabaqikelelo ngu-0.90 kwimeko ephantsi kunye ne-0.81 kwimeko ephezulu.I-Cohen's AlexNet kappa yi-0.82 kunye ne-0.71 yomgangatho ophantsi kunye neemeko zokusetyenziswa komgangatho ophezulu, ngokulandelanayo.I-kappa yamandla omqondiso we-Cohen yi-0.52 kunye ne-0.27 kwiimeko eziphantsi kunye neziphezulu zokusetyenziswa, ngokulandelanayo.
Ukuqinisekiswa kwemifuziselo yokuqaphela umgangatho ophezulu nophantsi kwimifanekiso ye-6\(\x\) ye-flat plate eyi-6 mm ibonisa isakhono semodeli eqeqeshiweyo ukumisela umgangatho womfanekiso kwiiparamitha ezahlukeneyo zokucinga.Xa usebenzisa i-6 \ (\x\) i-6 mm slabs enzulu yomgangatho we-imaging, imodeli ephantsi yomgangatho ophantsi yayine-AUC ye-0.83 (95% CI: 0.69-0.98) kunye nemodeli ephezulu ene-AUC ye-0.85.(95% CI: 0.55-1.00) (Itheyibhile 2).
Ukuhlolwa okubonwayo kweemephu zokwenziwa kusebenze kweklasi yomaleko kubonise ukuba zonke iinethiwekhi ze-neural eziqeqeshiweyo zisebenzisa iimpawu zemifanekiso ngexesha lokuhlelwa komfanekiso (Umfanekiso 4A, B).Kwi-8 \ (\ amaxesha \) 8 mm kunye ne-6 \ (\ amaxesha \) 6 mm imifanekiso, imifanekiso yokuvuselela i-ResNet ilandele ngokusondeleyo i-vasculature ye-retinal.Iimephu zokuvula i-AlexNet zikwalandela iinqanawa zeretina, kodwa zinesisombululo esirhabaxa.
Iimephu zeklasi zokuvula iimodeli ze-ResNet152 kunye ne-AlexNet zigxininisa iimpawu ezinxulumene nomgangatho womfanekiso.(A) Imephu yokuvula yeklasi ebonisa ukusebenza okuhambelanayo emva kwe-vasculature ye-retinal engaphezulu kwi-8 \ (\ amaxesha \) 8 mm yokuqinisekisa imifanekiso kunye (B) nobukhulu obuncinane 6 \ (\ amaxesha \) 6 mm imifanekiso yokuqinisekisa.Imodeli ye-LQ eqeqeshwe kwiikhrayitheriya ezikumgangatho ophantsi, imodeli ye-HQ eqeqeshwe kwiikhrayitheriya eziphezulu.
Kuye kwaboniswa ngaphambili ukuba umgangatho wesithombe unokuchaphazela kakhulu nayiphi na imilinganiselo yemifanekiso ye-OCTA.Ukongezelela, ubukho be-retinopathy kwandisa iziganeko zemifanekiso ye-artifacts7,26.Enyanisweni, kwidatha yethu, ngokuhambelana nezifundo zangaphambili, sifumene umbutho obalulekileyo phakathi kokunyuka kweminyaka kunye nobunzima besifo se-retinal kunye nokuwohloka komgangatho womfanekiso (p <0.001, p = 0.017 ubudala kunye nesimo se-DR, ngokulandelanayo; Itheyibhile 1) 27 Ngoko ke, kubalulekile ukuvavanya umgangatho womfanekiso ngaphambi kokwenza naluphi na uhlalutyo lobungakanani bemifanekiso ye-OCTA.Uninzi lwezifundo ezihlalutya imifanekiso ye-OCTA zisebenzisa imida yesignali exelwe ngumatshini ukuze ikhuphe imifanekiso ekumgangatho ophantsi.Nangona ubungakanani besignali bubonakaliswe ukuba buchaphazela ubungakanani beeparamitha ze-OCTA, ukunyanzeliswa kwesignali ephezulu yodwa ayinakwanela ukukhupha imifanekiso enemifanekiso ye-artifacts2,3,28,29.Ngoko ke, kuyimfuneko ukuphuhlisa indlela ethembekileyo yokulawula umgangatho womfanekiso.Ukuza kuthi ga ngoku, sivavanya ukusebenza kweendlela zokufunda ezinzulu ezigadiweyo ngokuchasene namandla omqondiso oxelwe ngumatshini.
Siphuhlise iimodeli ezininzi zokuvavanya umgangatho womfanekiso kuba iimeko ezahlukeneyo zokusetyenziswa kwe-OCTA zinokuba neemfuno ezahlukeneyo zomgangatho wemifanekiso.Umzekelo, imifanekiso kufuneka ibe yeyomgangatho ophezulu.Ukongeza, iiparameters ezithile zobungakanani bomdla nazo zibalulekile.Umzekelo, indawo ye-foveal avascular zone ayixhomekeke kwi-turbidity ye-non-central medium, kodwa ichaphazela ukuxinana kweenqanawa.Ngelixa uphando lwethu luqhubeka lujolise kwindlela eqhelekileyo yomgangatho womfanekiso, engabotshelelwanga kwiimfuno zalo naluphi na uvavanyo oluthile, kodwa ijolise ekubuyiseleni ngokuthe ngqo amandla omqondiso oxelwe ngumatshini, sinethemba lokunika abasebenzisi iqondo elikhulu lolawulo ukuze ingakhetha umlinganiselo othile womdla kumsebenzisi.khetha imodeli ehambelana neqondo eliphezulu lomfanekiso wezinto zakudala ezithathwa njengezamkelekileyo.
Kwimiboniso ekumgangatho ophantsi kunye nomgangatho ophezulu, sibonisa ukusebenza okugqwesileyo konxibelelwano-olulahlekileyo olunzulu lwe-neural networks, kunye ne-AUCs ye-0.97 kunye ne-0.99 kunye neemodeli ezikumgangatho ophantsi, ngokulandelanayo.Sikwabonisa ukusebenza okuphezulu kwendlela yethu yokufunda enzulu xa kuthelekiswa namanqanaba omqondiso axelwe ngoomatshini kuphela.Uqhagamshelo lokutsiba luvumela uthungelwano lwe-neural ukuba lufunde iimpawu kumanqanaba amaninzi eenkcukacha, ukuthabatha imiba ephucukileyo yemifanekiso (umzekelo, umahluko) kunye neempawu eziqhelekileyo (umz. umfanekiso ophakathi30,31).Kuba izinto ezenziwe ngemifanekiso ezichaphazela umgangatho womfanekiso zichongwa ngcono kuluhlu olubanzi, i-neural network architectures ezinoqhagamshelo olungekhoyo zinokubonisa ukusebenza okungcono kunezo zingenayo imisebenzi yokumisela umgangatho womfanekiso.
Xa sivavanya imodeli yethu kwi-6 \ (\ × 6mm) imifanekiso ye-OCTA, siqaphele ukuncipha kokusebenza kokuhlelwa kwiimodeli zombini ezikumgangatho ophezulu kunye nomgangatho ophantsi (umzobo 2), ngokungafaniyo nobukhulu bomzekelo oqeqeshelwe ukuhlelwa.Xa kuthelekiswa nemodeli ye-ResNet, imodeli ye-AlexNet ine-falloff enkulu.Ukusebenza okungcono kakhulu kwe-ResNet kunokuthi kube ngenxa yokukwazi ukudibanisa okushiyekileyo ukuhambisa ulwazi kwizikali ezininzi, okwenza imodeli ibe yomelele ngakumbi ekuhleleni imifanekiso ebanjwe kwizikali ezahlukeneyo kunye / okanye ukukhulisa.
Ezinye iiyantlukwano phakathi kwe-8 \ (\×\) 8 mm imifanekiso kunye ne-6 \ (\×\) 6 mm imifanekiso ingakhokelela ekuhleleni okubi, kubandakanywa umlinganiselo ophezulu wemifanekiso equkethe iindawo ze-foveal avascular, utshintsho olubonakalayo, i-vascular arcades, kunye akukho luvo optic kumfanekiso 6 × 6 mm.Ngaphandle koku, imodeli yethu ephezulu ye-ResNet yakwazi ukufikelela kwi-AUC ye-85% ye-6 \ (\x\) 6 mm imifanekiso, uqwalaselo apho imodeli ingaqeqeshwanga, iphakamisa ukuba ulwazi lomgangatho wesithombe lufakwe kwi-neural network. ifanelekile.ubungakanani bomfanekiso omnye okanye ukucwangciswa komatshini ngaphandle koqeqesho lwayo (Itheyibhile 2).Ukuqinisekisa, iimephu ze-ResNet- kunye ne-AlexNet-ezifana ne-activation ye-8 \ (\ amaxesha \) 8 mm kunye ne-6 \ (\ amaxesha \) Imifanekiso ye-6 mm yakwazi ukugqamisa iinqanawa ze-retinal kwiimeko zombini, ebonisa ukuba imodeli inolwazi olubalulekileyo.ziyasebenza ekuhleleni zombini iintlobo zemifanekiso ye-OCTA (umzobo 4).
Lauerman et al.Uvavanyo lomgangatho womfanekiso kwimifanekiso ye-OCTA lwenziwa ngokufanayo kusetyenziswa i-architecture yokuQala, enye i-skip-connection convolutional neural network6,32 isebenzisa ubuchule bokufunda obunzulu.Baphinde banciphise uphononongo kwimifanekiso ye-capillary plexus engaphezulu, kodwa kusetyenziswa kuphela imifanekiso emincinci ye-3 × 3 mm evela kwi-Optovue AngioVue, nangona izigulane ezinezifo ezahlukeneyo ze-chorioretinal nazo zibandakanyiwe.Umsebenzi wethu wakhela kwiziseko zabo, kubandakanywa iimodeli ezininzi zokujongana nemigangatho eyahlukeneyo yomgangatho wemifanekiso kunye nokuqinisekisa iziphumo zemifanekiso yobukhulu obahlukeneyo.Sikwanika ingxelo ye-AUC yeemodeli zokufunda koomatshini kunye nokwandisa ukuchaneka kwayo esele kunomtsalane (90%) 6 kuzo zombini umgangatho ophantsi (96%) kunye nomgangatho ophezulu (95.7%) imodeli6.
Olu qeqesho lunemida eliqela.Okokuqala, imifanekiso yafunyanwa ngomatshini omnye we-OCTA, kubandakanywa kuphela imifanekiso ye-capillary plexus engaphezulu kwi-8 \ (\ amaxesha \) 8 mm kunye ne-6 \ (\ amaxesha \) 6 mm.Isizathu sokungabandakanyi imifanekiso kwiileya ezinzulu kukuba ii-artifacts zentelekelelo zinokwenza uphononongo lwezandla lwemifanekiso lubenzima kwaye lungenzeki lula.Ngaphezu koko, imifanekiso ifunyenwe kuphela kwizigulane zesifo sikashukela, apho i-OCTA ivela njengesixhobo esibalulekileyo sokuxilonga kunye ne-prognostic33,34.Nangona sakwazi ukuvavanya imodeli yethu kwimifanekiso yobukhulu obuhlukeneyo ukuqinisekisa ukuba iziphumo zomelele, asikwazanga ukuchonga iiseti zedatha ezifanelekileyo ezivela kumaziko ahlukeneyo, okunciphisa ukuvavanya kwethu ngokubanzi imodeli.Nangona imifanekiso ifunyenwe kwiziko elinye kuphela, ifunyenwe kwizigulane zeentlanga ezahlukeneyo kunye nobuhlanga, obunamandla akhethekileyo kwisifundo sethu.Ngokubandakanya iyantlukwano kwinkqubo yethu yoqeqesho, sinethemba lokuba iziphumo zethu ziya kwenziwa ngokubanzi ngendlela ebanzi, kwaye siya kukuphepha ukufaka iikhowudi ucalucalulo lobuhlanga kwiimodeli esiziqeqeshayo.
Uphononongo lwethu lubonisa ukuba ukudibanisa-ukunqumla i-neural networks kunokuqeqeshwa ukuphumeza ukusebenza okuphezulu ekumiseleni umgangatho wesithombe se-OCTA.Sinikezela ngale mizekelo njengezixhobo zophando olongezelelweyo.Ngenxa yokuba iimetriki ezahlukeneyo zinokuba neemfuno ezahlukeneyo zomgangatho womfanekiso, imodeli yolawulo lomgangatho ngamnye inokuphuhliswa kwimetric nganye kusetyenziswa ubume obusekwe apha.
Uphando lwexesha elizayo kufuneka lubandakanye imifanekiso yobukhulu obahlukeneyo ukusuka kubunzulu obuhlukeneyo kunye noomatshini abahlukeneyo be-OCTA ukufumana inkqubo yokuvavanya umgangatho womfanekiso wokufunda nzulu onokuthi wenziwe ngokubanzi kwiiplatifti ze-OCTA kunye neeprothokholi zokucinga.Uphando lwangoku lukwasekelwe kwiindlela zokufunda ezinzulu eziphantsi kweliso ezifuna ukuphononongwa komntu kunye novavanyo lomfanekiso, onokuthi ube ngumsebenzi onzima kwaye udle ixesha kwiidatha ezinkulu.Kuhlala kubonwa ukuba iindlela zokufunda ezinzulu ezingajongwanga ziyakwazi ukwahlula ngokwanelisayo phakathi kwemifanekiso ekumgangatho ophantsi kunye nemifanekiso ekumgangatho ophezulu.
Njengoko itekhnoloji ye-OCTA iqhubeka nokuvela kunye nezantya zokuskena zinyuka, izehlo zemifanekiso eqingqiweyo kunye nemifanekiso ekumgangatho ophantsi inokuhla.Uphuculo lwesoftware, olufana nokususwa kwe-artifact esanda kwaziswa, inokunciphisa le mida.Nangona kunjalo, iingxaki ezininzi zihlala ziyimifanekiso yezigulana ezingalungiswanga kakuhle okanye i-turbidity ebalulekileyo yemithombo yeendaba ngokungaguquguqukiyo iphumela kubugcisa bemifanekiso.Njengoko i-OCTA isetyenziswa ngokubanzi kwizilingo zeklinikhi, ukuqwalaselwa ngenyameko kuyafuneka ukuseka izikhokelo ezicacileyo kumanqanaba omfanekiso owamkelekileyo we-artifact yohlalutyo lomfanekiso.Ukusetyenziswa kweendlela zokufunda ezinzulu kwimifanekiso ye-OCTA kunesithembiso esikhulu kwaye uphando olongezelelweyo luyafuneka kule ndawo ukuphuhlisa indlela eyomeleleyo yokulawula umgangatho womfanekiso.
Ikhowudi esetyenziswe kuphando lwangoku ifumaneka kwi-octa-qc repository, https://github.com/rahuldhodapkar/octa-qc.Iisethi zedatha ezenziwe kunye / okanye ezihlalutyiweyo ngexesha lophononongo lwangoku ziyafumaneka kubabhali abachaphazelekayo ngesicelo esinengqiqo.
I-Spaide, i-RF, i-Fujimoto, i-JG kunye ne-Waheed, i-NK Iimveliso zemifanekiso kwi-optical coherence angiography.I-Retina 35, 2163-2180 (2015).
Fenner, BJ et al.Ukuchongwa kweempawu ze-imaging ezimisela umgangatho kunye nokuveliswa kwakhona kwemilinganiselo ye-retinal capillary plexus density kwi-OCT angiography.BR.J. Ophthalmol.102, 509-514 (2018).
Lauerman, JL et al.Impembelelo yetekhnoloji yokulandelela iliso kumgangatho womfanekiso we-OCT angiography kwi-macular degeneration enxulumene nobudala.Ingca yengcwaba.klinikhi.Exp.i-ophthalmology.255, 1535-1542 (2017).
Babyuch AS et al.Imilinganiselo ye-OCTA ye-capillary perfusion density isetyenziselwa ukufumanisa nokuvavanya i-macular ischemia.utyando lwamehlo.IRetinal Laser imaging 51, S30–S36 (2020).
Yena, K., Zhang, X., Ren, S., kunye neLanga, J. UkuFundisa okuNgqongileyo okuNgqongileyo kwiNgcaciso yoMfanekiso.Kwi-2016 kwiNkomfa ye-IEEE kwi-Computer Vision kunye ne-Pattern Recognition (2016).
Lauerman, JL et al.Uvavanyo oluzenzekelayo lwe-OCT angiographic yomgangatho womfanekiso usebenzisa i-algorithms yokufunda nzulu.Ingca yengcwaba.klinikhi.Exp.i-ophthalmology.257, 1641-1648 (2019).
Lauermann, J. et al.Ukuxhaphaka kweempazamo zezahlulo kunye nezixhobo zokuhamba kwi-OCT angiography kuxhomekeke kwisifo se-retina.Ingca yengcwaba.klinikhi.Exp.i-ophthalmology.256, 1807-1816 (2018).
Pask, uAdam et al.I-Pytorch: Ithala leencwadi eliFundisa ngokuNzululwazi eliBanzi, eliSebenza ngokuNzululwazi.Ukuqhutyelwa phambili kolwazi lwe-neural.inkqubo.32, 8026–8037 (2019).
Deng, J. et al.I-ImageNet: IDatha yeDatha yeMifanekiso ye-Hierarchical Enkulu.INkomfa ye-IEEE yowama-2009 kuMbono weKhompyutha kunye nokuNakwa kwePhatheni.248–255.(2009).
Krizhevsky A., Sutzkever I. kunye nokuhlelwa kwe-Hinton GE Imagenet kusetyenziswa uthungelwano olunzulu lwe-neural convolutional.Ukuqhutyelwa phambili kolwazi lwe-neural.inkqubo.25, 1 (2012).