Ua tsaug rau koj tuaj xyuas Nature.com.Koj siv lub browser version nrog kev txhawb nqa CSS tsawg.Rau qhov kev paub zoo tshaj plaws, peb xav kom koj siv qhov browser tshiab (lossis lov tes taw Compatibility Hom hauv Internet Explorer).Tsis tas li ntawd, txhawm rau ua kom muaj kev txhawb nqa txuas ntxiv, peb qhia lub vev xaib tsis muaj qauv thiab JavaScript.
Sliders qhia peb kab lus rau ib tus swb.Siv cov nyees khawm rov qab thiab tom ntej kom txav mus los ntawm cov slides, lossis cov khawm tswj swb thaum kawg kom txav mus los ntawm txhua tus swb.
Optical coherence tomographic angiography (OCTA) yog ib txoj hauv kev tshiab rau kev tsis pom kev pom ntawm cov hlab ntsha retinal.Txawm hais tias OCTA muaj ntau daim ntawv thov kev kho mob zoo, kev txiav txim siab cov duab zoo tseem yog qhov nyuaj.Peb tsim ib qho kev kawm sib sib zog nqus raws li kev siv ResNet152 neural network classifier pretrained nrog ImageNet los faib cov duab capillary plexus ntawm 347 scans ntawm 134 tus neeg mob.Cov duab no kuj raug ntsuas los ntawm tus kheej raws li qhov tseeb los ntawm ob tus neeg ua haujlwm ywj pheej rau tus qauv saib xyuas kev kawm.Vim tias cov duab zoo yuav tsum sib txawv nyob ntawm qhov chaw kho mob lossis kev tshawb fawb, ob qho qauv tau raug cob qhia, ib qho rau kev lees paub cov duab zoo thiab lwm yam rau cov duab tsis zoo.Peb cov qauv neural network qhia tau hais tias thaj chaw zoo hauv qab qhov nkhaus (AUC), 95% CI 0.96-0.99, \(\kappa\) = 0.81), uas yog qhov zoo dua li cov teeb liab qhia los ntawm lub tshuab (AUC = 0.82, 95). % CI).0.77–0.86, \(\kappa\) = 0.52 thiab AUC = 0.78, 95% CI 0.73–0.83, \(\kappa\) = 0.27, ntsig txog).Peb txoj kev tshawb fawb qhia tau hais tias cov kev kawm tshuab tuaj yeem siv los txhim kho cov txheej txheem tswj tau zoo thiab muaj zog rau OCTA cov duab.
Optical coherence tomographic angiography (OCTA) yog cov txheej txheem tshiab raws li kev kho qhov muag coherence tomography (OCT) uas tuaj yeem siv rau qhov tsis pom kev pom ntawm lub qhov muag ntawm lub qhov muag.OCTA ntsuas qhov sib txawv ntawm cov qauv kev xav los ntawm kev rov ua lub teeb pulses nyob rau hauv tib cheeb tsam ntawm lub retina, thiab rov tsim kho dua tuaj yeem raug xam los qhia txog cov hlab ntsha yam tsis muaj kev siv cov dyes lossis lwm yam sib txawv.OCTA kuj tseem tso cai rau kev daws teeb meem vascular qhov tob, tso cai rau cov kws kho mob cais cov txheej txheem ntawm cov hlab ntsha thiab cov hlab ntsha sib sib zog nqus, pab kom sib txawv ntawm cov kab mob chorioretinal.
Thaum cov txheej txheem no tau cog lus tseg, kev hloov pauv cov duab zoo tseem yog ib qho kev sib tw loj rau kev txheeb xyuas cov duab txhim khu kev qha, ua rau kev txhais cov duab nyuaj thiab tiv thaiv kev siv tshuaj kho mob thoob plaws.Vim tias OCTA siv ntau qhov sib txuas OCT scans, nws yog qhov nkag siab rau cov duab kos duab ntau dua li OCT tus qauv.Feem ntau cov lag luam OCTA platforms muab lawv tus kheej cov duab zoo metric hu ua Signal Strength (SS) lossis qee zaum Signal Strength Index (SSI).Txawm li cas los xij, cov duab uas muaj tus nqi SS lossis SSI siab tsis tuaj yeem lav qhov tsis muaj cov duab kos duab, uas tuaj yeem cuam tshuam rau cov duab tom ntej thiab ua rau kev txiav txim siab tsis raug.Cov duab artifacts uas tuaj yeem tshwm sim hauv OCTA cov duab suav nrog cov duab kos duab, cov duab kos duab, cov duab kos duab opacity, thiab cov duab kos duab 1,2,3.
Raws li OCTA-derived ntsuas xws li vascular density tau nce siv hauv kev tshawb fawb kev txhais lus, kev sim tshuaj thiab kev siv tshuaj kho mob, nws yog qhov yuav tsum tau ua kom muaj zog thiab txhim khu kev qha cov duab zoo tswj cov txheej txheem kom tshem tawm cov duab artefacts4.Hla kev sib txuas, tseem hu ua residual kev sib txuas, yog qhov projections hauv neural network architecture uas tso cai rau cov ntaub ntawv los hla cov txheej txheem convolutional thaum khaws cov ntaub ntawv ntawm cov ntsuas sib txawv lossis kev daws teeb meem5.Vim tias cov duab kos duab tuaj yeem cuam tshuam rau cov duab me me thiab cov duab loj loj, hla-kev sib txuas neural tes hauj lwm zoo haum rau automate qhov kev tswj kom zoo no 5.Tsis ntev los no luam tawm ua hauj lwm tau qhia ib co lus cog tseg rau sib sib zog nqus convolutional neural networks kawm siv cov ntaub ntawv zoo los ntawm tib neeg kwv yees6.
Nyob rau hauv txoj kev tshawb no, peb cob qhia kev sib txuas-kev sib txuas ntawm cov neural network kom txiav txim siab qhov zoo ntawm OCTA cov duab.Peb tsim kev ua haujlwm dhau los los ntawm kev tsim cov qauv sib cais los txheeb xyuas cov duab zoo thiab cov duab tsis zoo, vim tias cov duab zoo yuav txawv ntawm qhov chaw kho mob lossis kev tshawb fawb.Peb sib piv cov txiaj ntsig ntawm cov tes hauj lwm no nrog cov kev sib txuas ntawm cov neural uas tsis muaj kev sib txuas uas ploj lawm los ntsuas cov txiaj ntsig ntawm kev suav nrog cov yam ntxwv ntawm ntau theem ntawm granularity hauv kev kawm tob.Peb mam li muab piv peb cov txiaj ntsig rau lub teeb liab lub zog, ib qho kev lees txais kev ntsuas ntawm cov duab zoo los ntawm cov tuam ntxhab.
Peb txoj kev tshawb fawb suav nrog cov neeg mob ntshav qab zib uas tuaj koom Yale Eye Center thaum lub Yim Hli 11, 2017 thiab Lub Plaub Hlis 11, 2019. Cov neeg mob uas tsis yog mob ntshav qab zib chorioretinal tsis suav nrog.Tsis muaj kev suav lossis cais tawm raws li hnub nyoog, poj niam txiv neej, haiv neeg, duab zoo, lossis lwm yam.
OCTA cov duab tau txais los ntawm AngioPlex platform ntawm Cirrus HD-OCT 5000 (Carl Zeiss Meditec Inc, Dublin, CA) nyob rau hauv 8\(\times\)8 hli thiab 6\(\times\)6 hli imaging raws tu qauv.Cov ntaub ntawv tso cai rau kev koom nrog hauv txoj kev tshawb fawb tau txais los ntawm txhua tus neeg koom nrog txoj kev tshawb fawb, thiab Yale University Institutional Review Board (IRB) tau pom zoo siv cov ntawv pom zoo nrog kev yees duab thoob ntiaj teb rau tag nrho cov neeg mob no.Ua raws li cov hauv paus ntsiab lus ntawm Kev Tshaj Tawm ntawm Helsinki.Txoj kev tshawb no tau pom zoo los ntawm Yale University IRB.
Cov duab phaj saum npoo tau raug soj ntsuam raws li cov lus piav qhia yav dhau los Motion Artifact Score (MAS), yav dhau los tau piav qhia Segmentation Artifact Score (SAS), qhov chaw foveal, muaj qhov tsis pom kev tawm, thiab kev pom zoo ntawm cov hlab ntsha me me raws li kev txiav txim siab los ntawm cov duab ntsuas.Cov duab no tau txheeb xyuas los ntawm ob tus neeg soj ntsuam ywj pheej (RD thiab JW).Ib daim duab muaj qhov qhab nia ntawm 2 (tsim nyog) yog tias tag nrho cov qauv hauv qab no tau ua tiav: daim duab yog nyob nruab nrab ntawm fovea (tsawg dua 100 pixels ntawm qhov nruab nrab ntawm daim duab), MAS yog 1 lossis 2, SAS yog 1, thiab media opacity tsawg dua 1. Tam sim no ntawm cov duab ntawm qhov loj / 16, thiab cov capillaries me tau pom hauv cov duab loj dua 15/16.Ib daim duab raug ntsuas 0 (tsis muaj kev ntsuam xyuas) yog tias ib qho ntawm cov qauv hauv qab no tau ua tiav: daim duab tsis nyob hauv nruab nrab, yog MAS yog 4, yog SAS yog 2, lossis qhov nruab nrab opacity ntau dua 1/4 ntawm daim duab, thiab cov capillaries me me tsis tuaj yeem hloov kho ntau dua 1 daim duab / 4 kom paub qhov txawv.Tag nrho lwm cov duab uas tsis ua raws li cov qhab nia 0 lossis 2 tau qhab nia raws li 1 (clipping).
Ntawm daim duab.1 qhia cov duab piv txwv rau txhua qhov ntsuas ntsuas thiab cov duab kos duab.Inter-rater kev ntseeg siab ntawm tus kheej cov qhab nia raug soj ntsuam los ntawm Cohen's kappa hnyav 8.Cov qhab nia ntawm txhua tus neeg ntsuas tau suav nrog kom tau txais tag nrho cov qhab nia rau txhua daim duab, xws li 0 txog 4. Cov duab nrog tag nrho cov qhab nia ntawm 4 yog suav tias yog qhov zoo.Cov duab nrog tag nrho cov qhab nia ntawm 0 lossis 1 yog suav tias tsis zoo.
A ResNet152 architecture convolutional neural network (Fig. 3A.i) ua ntej kev cob qhia ntawm cov duab los ntawm ImageNet database tau tsim los siv fast.ai thiab PyTorch lub moj khaum5, 9, 10, 11. Lub convolutional neural network yog lub network uas siv cov kev kawm. cov ntxaij lim dej rau scanning duab fragments los kawm spatial thiab hauv zos nta.Peb qhov kev cob qhia ResNet yog 152-txheej neural network cim los ntawm qhov khoob lossis "kev sib txuas seem" uas ib txhij xa cov ntaub ntawv nrog ntau qhov kev daws teeb meem.Los ntawm kev npaj cov ntaub ntawv ntawm cov kev daws teeb meem sib txawv hauv lub network, lub platform tuaj yeem kawm cov yam ntxwv ntawm cov duab tsis zoo ntawm ntau theem ntawm kev nthuav dav.Ntxiv rau peb tus qauv ResNet, peb kuj tau cob qhia AlexNet, ib qho kev kawm zoo neural network architecture, tsis muaj kev sib txuas rau kev sib piv (Daim duab 3A.ii)12.Yog tsis muaj kev sib txuas uas ploj lawm, lub network no yuav tsis tuaj yeem ntes cov yam ntxwv ntawm qhov ntau dua.
Thawj 8 \ (\times\) 8mm OCTA13 duab teeb tau raug kho kom zoo siv cov txheej txheem rov qab thiab ntsug.Cov ntaub ntawv tag nrho tau muab faib ua ntu zus ntawm cov duab theem rau hauv kev cob qhia (51.2%), kev sim (12.8%), hyperparameter tuning (16%), thiab validation (20%) datasets siv lub scikit-kawm toolbox python14.Ob qhov xwm txheej tau txiav txim siab, ib qho raws li kev kuaj pom tsuas yog cov duab zoo tshaj plaws (tag nrho cov qhab nia 4) thiab lwm qhov raws li kev kuaj pom cov duab qis tshaj plaws (tag nrho cov qhab nia 0 lossis 1).Rau txhua qhov kev siv tau zoo thiab tsis zoo, lub neural network tau rov ua dua ib zaug ntawm peb cov ntaub ntawv duab.Nyob rau hauv txhua rooj plaub, lub neural network tau raug cob qhia rau 10 lub sijhawm, tag nrho tab sis cov txheej txheem siab tshaj plaws tau khov, thiab qhov hnyav ntawm tag nrho cov kev ntsuas sab hauv tau kawm rau 40 epochs siv cov txheej txheem kev ntxub ntxaug nrog kev sib txawv ntawm entropy poob ua haujlwm 15, 16..Tus ntoo khaub lig entropy poob muaj nuj nqi yog ib qho kev ntsuas ntawm lub logarithmic nplai ntawm qhov tsis sib xws ntawm kev kwv yees network ntawv thiab cov ntaub ntawv tiag.Thaum lub sij hawm kev cob qhia, gradient qhovntsej thiaj tsis mob nyob rau hauv lub internal parameters ntawm lub neural network kom txo tau losses.Qhov kev kawm, tus nqi tso tawm, thiab qhov hnyav txo qhov hyperparameters tau kho siv Bayesian optimization nrog 2 random pib cov ntsiab lus thiab 10 iterations, thiab AUC ntawm cov ntaub ntawv tau hloov kho siv cov hyperparameters raws li lub hom phiaj ntawm 17.
Tus neeg sawv cev piv txwv ntawm 8 × 8 hli OCTA cov duab ntawm cov duab capillary plexuses tau qhab nia 2 (A, B), 1 (C, D), thiab 0 (E, F).Cov duab kos duab pom muaj xws li cov kab flickering (xub xub), segmentation artifacts (asterisks), thiab media opacity ( xub).Duab (E) kuj tsis yog qhov chaw.
Tus neeg txais kev khiav hauj lwm yam ntxwv (ROC) curves yog tom qab ntawd tsim rau txhua tus qauv neural network, thiab cov ntawv ceeb toom lub zog ntawm lub cav yog tsim rau txhua qhov kev siv qis thiab zoo.Thaj chaw hauv qab qhov nkhaus (AUC) tau suav nrog siv pROC R pob, thiab 95% kev ntseeg siab lub sijhawm thiab p-tus nqi raug suav nrog siv txoj kev DeLong18,19.Cov qhab-nees sib npaug ntawm tib neeg tus neeg ntsuas yog siv los ua lub hauv paus rau txhua qhov kev suav ROC.Rau lub teeb liab lub zog qhia los ntawm lub tshuab, AUC tau xam ob zaug: ib zaug rau qhov zoo Scalability Score txiav thiab ib zaug rau qhov tsis zoo Scalability Score txiav.Lub neural network piv rau AUC lub teeb liab lub zog qhia txog nws tus kheej kev cob qhia thiab kev soj ntsuam.
Txhawm rau kuaj ntxiv cov qauv kev kawm sib sib zog nqus ntawm cov ntaub ntawv sib cais, cov qauv zoo thiab cov qauv tsis zoo tau siv ncaj qha rau kev ntsuas kev ua tau zoo ntawm 32 lub ntsej muag 6\(\times\) 6mm nto slab duab sau los ntawm Yale University.Qhov muag loj yog nyob nruab nrab ntawm tib lub sijhawm raws li daim duab 8 \(\times \) 8 mm.Cov duab 6\(\×\) 6 hli tau raug ntsuas los ntawm tib tus neeg ntsuas (RD thiab JW) tib yam li 8\(\×\) 8 hli cov duab, AUC tau suav nrog qhov tseeb thiab Cohen's kappa .sib npaug.
Cov chav kawm tsis sib npaug yog 158:189 (\(\rho = 1.19\)) rau cov qauv tsis zoo thiab 80:267 (\(\rho = 3.3\)) rau cov qauv zoo.Vim hais tias cov chav kawm tsis sib npaug tsawg dua 1: 4, tsis muaj kev hloov kho vaj tse tshwj xeeb tau ua los kho cov chav kawm tsis sib npaug 20,21.
Txhawm rau kom pom cov txheej txheem kev kawm zoo dua, daim ntawv qhia kev ua haujlwm hauv chav kawm tau tsim rau tag nrho plaub tus qauv kawm sib sib zog nqus: tus qauv zoo ResNet152, tus qauv tsis zoo ResNet152, tus qauv AlexNet zoo, thiab tus qauv AlexNet tsis zoo.Daim ntawv qhia kev ua haujlwm hauv chav kawm yog tsim los ntawm cov txheej txheem input convolutional ntawm plaub qauv no, thiab cov ntawv qhia tshav kub yog tsim los ntawm kev sib tshooj ua kom cov duab qhia chaw nrog cov duab los ntawm 8 × 8 mm thiab 6 × 6 mm validation sets22, 23.
R version 4.0.3 tau siv rau txhua qhov kev suav suav, thiab kev pom pom tau tsim siv lub tsev qiv ntawv ggplot2 cov duab duab.
Peb sau 347 frontal dluab ntawm superficial capillary plexus ntsuas 8 \ (\times \) 8 mm los ntawm 134 tus neeg.Lub tshuab tau tshaj tawm cov teeb liab lub zog ntawm qhov ntsuas ntawm 0 txog 10 rau tag nrho cov duab (txhais tau tias = 6.99 ± 2.29).Ntawm 347 cov duab tau txais, qhov nruab nrab hnub nyoog ntawm kev ntsuam xyuas yog 58.7 ± 14.6 xyoo, thiab 39.2% yog los ntawm cov txiv neej cov neeg mob.Ntawm tag nrho cov duab, 30.8% yog los ntawm Caucasians, 32.6% los ntawm Dub, 30.8% ntawm Hispanics, 4% los ntawm Neeg Esxias, thiab 1.7% los ntawm lwm haiv neeg (Table 1).).Lub hnub nyoog faib ntawm cov neeg mob nrog OCTA sib txawv heev nyob ntawm qhov zoo ntawm daim duab (p <0.001).Qhov feem pua ​​​​ntawm cov duab zoo hauv cov neeg mob hluas hnub nyoog 18-45 xyoo yog 33.8% piv rau 12.2% ntawm cov duab tsis zoo (Table 1).Kev faib tawm ntawm tus mob ntshav qab zib retinopathy tseem muaj qhov sib txawv ntawm cov duab zoo (p <0.017).Ntawm tag nrho cov duab zoo, feem pua ​​​​ntawm cov neeg mob PDR yog 18.8% piv rau 38.8% ntawm tag nrho cov duab tsis zoo (Table 1).
Kev ntsuas tus kheej ntawm tag nrho cov duab tau pom muaj qhov nruab nrab rau qhov muaj zog ntawm qhov kev ntseeg siab ntawm cov neeg nyeem cov duab (Cohen's weighted kappa = 0.79, 95% CI: 0.76-0.82), thiab tsis muaj cov ntsiab lus duab uas cov neeg ntsuas sib txawv ntau dua 1 (Fig. 2 A)..Cov teeb liab siv sib cuam tshuam nrog kev sib tw ntawm phau ntawv (Pearson khoom lub sij hawm sib raug zoo = 0.58, 95% CI 0.51–0.65, p<0.001), tab sis ntau cov duab tau txheeb xyuas tias muaj teeb liab siab siv tab sis phau ntawv qhab nia qis (Fig. .2B).
Thaum lub sij hawm kev cob qhia ntawm ResNet152 thiab AlexNet architectures, qhov hla-enttropy poob ntawm kev siv tau thiab kev cob qhia poob ntau dua 50 lub sijhawm (Daim duab 3B, C).Kev lees paub qhov tseeb nyob rau hauv qhov kev cob qhia zaum kawg yog tshaj 90% rau ob qho tib si zoo thiab siv tsis zoo.
Tus txais kev ua tau zoo nkhaus qhia tau hais tias tus qauv ResNet152 ua tau zoo tshaj lub teeb liab lub zog qhia los ntawm lub tshuab hauv ob qho tib si qis thiab siv tau zoo (p <0.001).Tus qauv ResNet152 kuj tseem ua tau zoo tshaj plaws ntawm AlexNet architecture (p = 0.005 thiab p = 0.014 rau cov teeb meem tsis zoo thiab zoo, feem).Cov qauv tsim tawm rau txhua qhov haujlwm no tuaj yeem ua tiav AUC qhov tseem ceeb ntawm 0.99 thiab 0.97, feem, uas yog qhov zoo dua li qhov sib thooj AUC qhov tseem ceeb ntawm 0.82 thiab 0.78 rau lub tshuab teeb liab lub zog Performance index lossis 0.97 thiab 0.94 rau AlexNet. ..(Daim duab 3).Qhov sib txawv ntawm ResNet thiab AUC hauv cov teeb liab muaj zog dua thaum pom cov duab zoo, qhia txog cov txiaj ntsig ntxiv ntawm kev siv ResNet rau txoj haujlwm no.
Cov duab qhia txhua tus neeg ua haujlwm ywj pheej muaj peev xwm ua kom tau qhab nia thiab sib piv nrog lub teeb liab lub zog qhia los ntawm lub tshuab.(A) Cov ntsiab lus ntawm cov ntsiab lus yuav tsum tau ntsuas yog siv los tsim tag nrho cov ntsiab lus los ntsuas.Cov duab nrog rau tag nrho cov qhab nia scalability ntawm 4 yog muab siab zoo, thaum cov duab nrog ib tug tag nrho scalability qhab nia ntawm 1 los yog tsawg dua yog muab tsis zoo.(B) Kev siv lub teeb liab cuam tshuam nrog kev kwv yees ntawm phau ntawv, tab sis cov duab uas muaj lub teeb liab siv zog yuav tsis zoo.Cov kab liab dotted qhia cov chaw tsim khoom tau pom zoo qhov pib zoo raws li lub teeb liab lub zog (lub zog teeb liab \(\ge\)6).
ResNet kev hloov pauv kev kawm muab kev txhim kho tseem ceeb hauv kev txheeb xyuas cov duab zoo rau ob qho tib si tsis zoo thiab siv tau zoo piv rau cov teeb liab qhia txog tshuab.(A) Cov duab kos duab yooj yim ntawm kev kawm ua ntej (i) ResNet152 thiab (ii) AlexNet architectures.(B) Kev cob qhia keeb kwm thiab kev txais kev ua tau zoo nkhaus rau ResNet152 piv rau lub tshuab qhia lub zog thiab AlexNet cov qauv tsis zoo.(C) ResNet152 txais kev cob qhia keeb kwm thiab kev ua haujlwm nkhaus piv rau lub tshuab qhia lub zog thiab AlexNet cov qauv zoo.
Tom qab kho qhov kev txiav txim ntawm ciam teb, qhov siab tshaj plaws kev kwv yees qhov tseeb ntawm ResNet152 qauv yog 95.3% rau cov ntaub ntawv tsis zoo thiab 93.5% rau rooj plaub zoo (Table 2).Qhov siab tshaj plaws kev kwv yees qhov tseeb ntawm tus qauv AlexNet yog 91.0% rau cov ntaub ntawv tsis zoo thiab 90.1% rau cov ntaub ntawv zoo (Table 2).Qhov siab tshaj plaws lub teeb liab lub zog kwv yees qhov tseeb yog 76.1% rau cov ntaub ntawv siv tsis zoo thiab 77.8% rau cov ntaub ntawv siv tau zoo.Raws li Cohen's kappa (\(\kappa\)), qhov kev pom zoo ntawm ResNet152 qauv thiab cov kwv yees yog 0.90 rau cov ntaub ntawv tsis zoo thiab 0.81 rau rooj plaub zoo.Cohen's AlexNet kappa yog 0.82 thiab 0.71 rau qhov tsis zoo thiab siv tau zoo, feem.Cohen lub teeb liab lub zog kappa yog 0.52 thiab 0.27 rau qhov siv tsawg thiab zoo, raws li.
Kev lees paub ntawm cov qauv zoo thiab tsis tshua muaj kev lees paub ntawm 6\(\x\) cov duab ntawm 6 hli ca phaj qhia lub peev xwm ntawm cov qauv kev cob qhia los txiav txim siab cov duab zoo thoob plaws ntau yam kev ntsuas.Thaum siv 6 \ (\x\) 6 hli ntiav slabs rau cov duab zoo, cov qauv tsis zoo muaj AUC ntawm 0.83 (95% CI: 0.69–0.98) thiab cov qauv zoo muaj AUC ntawm 0.85.(95% CI: 0.55–1.00) (Table 2).
Kev tshuaj xyuas qhov pom ntawm cov txheej txheem nkag hauv chav kawm ua kom pom tau tias txhua qhov kev cob qhia neural tes hauj lwm siv cov duab nta thaum muab faib cov duab (Fig. 4A, B).Rau 8 \(\times \) 8 mm thiab 6 \(\times \) 6 mm cov duab, ResNet ua kom cov dluab ua raws li cov vasculature retinal.AlexNet ua kom daim duab qhia chaw kuj ua raws li cov hlab ntsha retinal, tab sis nrog kev daws teeb meem coarser.
Daim ntawv qhia kev ua haujlwm hauv chav kawm rau ResNet152 thiab AlexNet qauv qhia txog cov yam ntxwv ntsig txog cov duab zoo.(A) Daim ntawv qhia kev ua haujlwm hauv chav kawm uas qhia txog kev ua kom muaj kev sib koom ua ke tom qab lub vasculature retinal superficial ntawm 8 \(\times \) 8 mm validation dluab thiab (B) feem ntawm me 6 \(\times \) 6 mm validation dluab.LQ qauv kawm ntawm cov qauv tsis zoo, HQ qauv tau kawm ntawm cov qauv zoo.
Nws tau dhau los ua pov thawj tias cov duab zoo tuaj yeem cuam tshuam rau txhua qhov ntau ntawm OCTA cov duab.Tsis tas li ntawd, lub xub ntiag ntawm retinopathy nce qhov tshwm sim ntawm cov duab kos duab 7,26.Qhov tseeb, hauv peb cov ntaub ntawv, raws li kev tshawb fawb yav dhau los, peb pom muaj kev koom tes tseem ceeb ntawm lub hnub nyoog nce thiab qhov hnyav ntawm cov kab mob retinal thiab deterioration ntawm cov duab zoo (p <0.001, p = 0.017 rau hnub nyoog thiab DR raws li txoj cai; Table 1) 27 Yog li ntawd, nws yog ib qho tseem ceeb rau kev ntsuam xyuas cov duab zoo ua ntej ua ib qho kev ntsuam xyuas ntau ntawm OCTA cov duab.Feem ntau cov kev tshawb fawb soj ntsuam OCTA cov duab siv lub tshuab qhia cov teeb liab siv qhov pib los txiav tawm cov duab tsis zoo.Txawm hais tias lub teeb liab siv tau pom tias muaj feem cuam tshuam rau qhov muaj nuj nqis ntawm OCTA tsis, lub teeb liab siab siv ib leeg yuav tsis txaus los txiav txim cov duab nrog cov duab kos duab 2,3,28,29.Yog li ntawd, nws yog ib qhov tsim nyog los tsim ib txoj kev txhim khu kev qha ntau dua ntawm kev tswj cov duab zoo.Txog qhov kawg no, peb ntsuas qhov kev ua tau zoo ntawm kev saib xyuas kev kawm sib sib zog nqus tawm tsam lub teeb liab lub zog qhia los ntawm lub tshuab.
Peb tau tsim ntau yam qauv rau kev soj ntsuam cov duab zoo vim hais tias cov kev siv OCTA sib txawv yuav muaj cov duab zoo sib txawv.Piv txwv li, cov duab yuav tsum zoo dua.Tsis tas li ntawd, cov ntsiab lus tshwj xeeb ntawm kev txaus siab kuj tseem ceeb.Piv txwv li, thaj tsam ntawm foveal avascular cheeb tsam tsis nyob ntawm qhov turbidity ntawm qhov nruab nrab tsis yog nruab nrab, tab sis cuam tshuam rau qhov ntom ntawm cov hlab ntsha.Thaum peb cov kev tshawb fawb txuas ntxiv tsom mus rau txoj hauv kev dav dav rau cov duab zoo, tsis khi rau qhov yuav tsum tau ua ntawm ib qho kev sim tshwj xeeb, tab sis npaj los hloov lub teeb liab lub zog qhia los ntawm lub tshuab, peb cia siab tias yuav muab cov neeg siv ntau dua kev tswj kom lawv. tuaj yeem xaiv qhov kev ntsuas tshwj xeeb ntawm kev txaus siab rau tus neeg siv.xaiv tus qauv uas sib haum rau qhov siab tshaj plaws ntawm cov duab artifacts suav tias yog txais.
Rau qhov tsis zoo thiab qhov zoo tshaj plaws, peb qhia tau zoo heev ntawm kev sib txuas-tsis muaj qhov sib sib zog nqus convolutional neural networks, nrog AUCs ntawm 0.97 thiab 0.99 thiab cov qauv tsis zoo, feem.Peb kuj ua tau zoo tshaj qhov kev ua tau zoo ntawm peb txoj kev kawm sib sib zog nqus thaum piv rau theem teeb liab qhia los ntawm cov tshuab xwb.Hla kev sib txuas tso cai rau neural tes hauj lwm kawm cov yam ntxwv ntawm ntau theem ntawm kev nthuav dav, ntes cov duab zoo dua (xws li qhov sib txawv) nrog rau cov yam ntxwv dav dav (xws li duab centering30,31).Txij li cov duab artifacts uas cuam tshuam rau cov duab zoo yog tej zaum zoo tshaj plaws txheeb xyuas nyob rau hauv ntau yam, neural network architectures uas ploj lawm kev twb kev txuas yuav ua tau zoo dua cov uas tsis muaj cov duab zoo txiav txim siab ua hauj lwm.
Thaum kuaj peb cov qauv ntawm 6 \ (\ × 6mm) OCTA cov duab, peb pom qhov txo qis hauv kev faib ua haujlwm rau ob qho tib si zoo thiab tsis zoo qauv (Fig. 2), nyob rau hauv sib piv rau qhov luaj li cas ntawm cov qauv kawm rau kev faib.Piv nrog rau tus qauv ResNet, tus qauv AlexNet muaj qhov poob loj dua.Qhov kev ua tau zoo dua ntawm ResNet tuaj yeem yog vim muaj peev xwm ntawm cov kev sib txuas ntawm qhov seem los xa cov ntaub ntawv ntawm ntau qhov teev, uas ua rau tus qauv muaj zog dua rau kev faib cov duab ntes ntawm cov nplai sib txawv thiab / lossis qhov loj.
Qee qhov sib txawv ntawm 8 \(\×\) 8 hli cov duab thiab 6 \(\×\) 6 hli cov duab tuaj yeem ua rau muaj kev faib tawm tsis zoo, suav nrog kev faib ua feem ntau ntawm cov duab uas muaj thaj chaw foveal avascular, kev hloov pauv hauv kev pom, vascular arcades, thiab tsis muaj optic paj hlwb ntawm daim duab 6 × 6 mm.Txawm li cas los xij, peb cov qauv zoo ResNet muaj peev xwm ua tiav AUC ntawm 85% rau 6 \(\x\) 6 hli cov duab, ib qho kev teeb tsa uas tus qauv tsis tau kawm, qhia tias cov duab zoo cov ntaub ntawv encoded hauv neural network. yog tsim nyog.rau ib daim duab loj lossis lub tshuab teeb tsa sab nraud ntawm nws txoj kev kawm (Table 2).Reassuringly, ResNet- thiab AlexNet-zoo li daim ntawv qhia ua haujlwm ntawm 8 \(\times \) 8 mm thiab 6 \(\times \) 6 mm cov duab tuaj yeem qhia txog cov hlab ntsha hauv ob qho tib si, qhia tias tus qauv muaj cov ntaub ntawv tseem ceeb.muaj feem xyuam rau kev faib ob hom OCTA cov duab (Fig. 4).
Lauerman et al.Kev ntsuam xyuas cov duab zoo ntawm OCTA cov duab tau zoo sib xws siv Inception architecture, lwm qhov hla-kev sib txuas convolutional neural network6,32 siv cov kev kawm tob.Lawv kuj txwv txoj kev tshawb no rau cov duab ntawm lub ntsej muag capillary plexus, tab sis tsuas yog siv cov duab me me 3 × 3 mm los ntawm Optovue AngioVue, txawm tias cov neeg mob uas muaj ntau yam kab mob chorioretinal kuj suav nrog.Peb txoj haujlwm tsim los ntawm lawv lub hauv paus, suav nrog ntau tus qauv los hais txog ntau yam duab zoo thiab ua kom pom tseeb cov txiaj ntsig rau cov duab sib txawv.Peb kuj qhia txog AUC metric ntawm cov qauv kev kawm tshuab thiab ua kom lawv qhov tseeb qhov tseeb (90%)6 rau ob qho tib si tsis zoo (96%) thiab siab zoo (95.7%) qauv6.
Qhov kev cob qhia no muaj ntau yam kev txwv.Ua ntej, cov duab tau txais nrog tsuas yog ib lub tshuab OCTA, suav nrog tsuas yog cov duab ntawm lub ntsej muag capillary plexus ntawm 8\(\times\)8 hli thiab 6\(\times\)6 hli.Yog vim li cas tsis suav nrog cov duab los ntawm cov khaubncaws sab nraud povtseg yog qhov projection artifacts tuaj yeem ua rau kev soj ntsuam ntawm cov duab nyuaj thiab tejzaum nws tsis sib xws.Tsis tas li ntawd, cov duab tsuas yog tau txais hauv cov neeg mob ntshav qab zib, rau leej twg OCTA tau tshwm sim los ua ib qho tseem ceeb rau kev kuaj mob thiab cov cuab yeej prognostic33,34.Txawm hais tias peb tuaj yeem sim peb cov qauv ntawm cov duab sib txawv kom ntseeg tau tias cov txiaj ntsig tau muaj zog, peb tsis tuaj yeem txheeb xyuas cov ntaub ntawv tsim nyog los ntawm cov chaw sib txawv, uas txwv peb qhov kev ntsuas ntawm qhov dav dav ntawm tus qauv.Txawm hais tias cov duab tau txais los ntawm ib lub chaw xwb, lawv tau txais los ntawm cov neeg mob ntawm ntau haiv neeg thiab haiv neeg keeb kwm, uas yog lub zog tshwj xeeb ntawm peb txoj kev kawm.Los ntawm kev suav nrog ntau haiv neeg hauv peb cov txheej txheem kev cob qhia, peb cia siab tias peb cov txiaj ntsig yuav raug nthuav dav hauv kev txiav txim siab dav dav, thiab peb yuav zam tsis txhob muaj kev ntxub ntxaug lwm haiv neeg hauv cov qauv peb qhia.
Peb txoj kev tshawb fawb qhia tau hais tias kev sib txuas-kev sib txuas ntawm neural network tuaj yeem raug cob qhia kom ua tiav kev ua haujlwm siab hauv kev txiav txim siab OCTA duab zoo.Peb muab cov qauv no ua cov cuab yeej rau kev tshawb fawb ntxiv.Vim tias cov metrics sib txawv yuav muaj cov duab zoo sib txawv, tus qauv tswj tus kheej tuaj yeem tsim tau rau txhua qhov kev ntsuas uas siv cov qauv tsim ntawm no.
Kev tshawb fawb yav tom ntej yuav tsum suav nrog cov duab sib txawv ntawm qhov sib txawv ntawm qhov tob thiab sib txawv OCTA tshuab kom tau txais cov txheej txheem kev ntsuam xyuas cov duab sib sib zog nqus uas tuaj yeem ua tau dav dav rau OCTA platforms thiab cov txheej txheem duab.Cov kev tshawb fawb tam sim no tseem ua raws li kev saib xyuas kev kawm tob uas xav tau kev soj ntsuam tib neeg thiab kev ntsuas duab, uas tuaj yeem ua haujlwm hnyav thiab siv sijhawm ntau rau cov ntaub ntawv loj.Nws tseem yuav pom tias cov kev kawm tob uas tsis muaj kev saib xyuas tuaj yeem paub qhov txawv ntawm cov duab tsis zoo thiab cov duab zoo.
Raws li OCTA thev naus laus zis txuas ntxiv txhim kho thiab tshawb xyuas nrawm nce, qhov xwm txheej ntawm cov duab kos duab thiab cov duab tsis zoo yuav txo qis.Kev txhim kho hauv software, xws li qhov nyuam qhuav qhia txog qhov kev tshem tawm cov khoom cuav, tuaj yeem txo cov kev txwv no.Txawm li cas los xij, ntau qhov teeb meem tseem nyob raws li kev yees duab ntawm cov neeg mob uas muaj kev kho tsis zoo los yog qhov tseem ceeb tshaj tawm xov xwm turbidity invariably ua rau cov duab artifacts.Raws li OCTA tau dhau los siv dav hauv kev sim tshuaj, yuav tsum tau ua tib zoo txiav txim siab los tsim cov lus qhia meej rau cov duab kos duab zoo rau cov duab tsom xam.Daim ntawv thov kev kawm tob rau OCTA cov duab tuav cov lus cog tseg zoo thiab xav tau kev tshawb fawb ntxiv hauv cheeb tsam no los tsim txoj hauv kev zoo rau kev tswj cov duab zoo.
Cov cai siv hauv kev tshawb fawb tam sim no muaj nyob rau hauv octa-qc repository, https://github.com/rahuldhodapkar/octa-qc.Cov ntaub ntawv tsim tawm thiab / lossis tshuaj xyuas thaum lub sijhawm kawm tam sim no muaj los ntawm cov kws sau ntawv raws li qhov kev thov tsim nyog.
Spaide, RF, Fujimoto, JG & Waheed, NK Duab artifacts hauv optical coherence angiography.Retina 35, 2163–2180 (2015).
Fenner, BJ et al.Kev txheeb xyuas cov yam ntxwv ntawm cov duab uas txiav txim siab qhov ua tau zoo thiab rov ua dua tshiab ntawm retinal capillary plexus ntom ntom ntsuas hauv OCT angiography.BR.J. Ophthalmol.102, 509–514 (2018).
Lauerman, JL et al.Kev cuam tshuam ntawm qhov muag-nrhiav thev naus laus zis ntawm cov duab zoo ntawm OCT angiography hauv cov hnub nyoog ntsig txog macular degeneration.Lub ntxa.kho mob.Exp.ophthalmology.255, 1535–1542 (2017).
Babyuch AS et al.OCTA capillary perfusion density ntsuas yog siv los kuaj thiab ntsuas macular ischemia.ophthalmic phais.Retinal Laser Imaging 51, S30–S36 (2020).
He, K., Zhang, X., Ren, S., thiab Sun, J. Deep Residual Learning for Image Recognition.Hauv 2016 ntawm IEEE Conference on Computer Vision and Pattern Recognition (2016).
Lauerman, JL et al.Automated OCT angiographic duab kev soj ntsuam zoo siv cov kev kawm sib sib zog nqus algorithms.Lub ntxa.kho mob.Exp.ophthalmology.257, 1641–1648 (2019).
Lauermann, J. et al.Feem ntau ntawm segmentation yuam kev thiab cov khoom siv dag zog hauv OCT angiography nyob ntawm tus kab mob ntawm retina.Lub ntxa.kho mob.Exp.ophthalmology.256, 1807–1816 (2018).
Pask, Adam et al.Pytoch: Ib qho tseem ceeb, Kev Ua Haujlwm Siab Tshaj Plaws Kawm Qib Siab.Advanced processing ntawm neural ntaub ntawv.qhov system.32, 8026–8037 (2019).
Deng, J. et al.ImageNet: Qhov Loj-Scale Hierarchical Image Database.2009 IEEE Conference on Computer Vision and Pattern Recognition.248–255 : kuv.(2009).
Krizhevsky A., Sutzkever I. thiab Hinton GE Imagenet kev faib tawm siv kev sib sib zog nqus ntawm cov neural networks.Advanced processing ntawm neural ntaub ntawv.qhov system.25, 1 (2012).