Siyabonga ngokuvakashela i-Nature.com.Usebenzisa inguqulo yesiphequluli enosekelo olulinganiselwe lwe-CSS.Ukuze uthole ulwazi olungcono kakhulu, sincoma ukuthi usebenzise isiphequluli esibuyekeziwe (noma ukhubaze i-Compatibility Mode ku-Internet Explorer).Ngaphezu kwalokho, ukuze siqinisekise ukwesekwa okuqhubekayo, sibonisa isayithi ngaphandle kwezitayela ne-JavaScript.
Izilayidi ezibonisa izindatshana ezintathu zesilayidi ngasinye.Sebenzisa izinkinobho ezingemuva nezilandelayo ukuhamba phakathi kwamaslayidi, noma izinkinobho zokulawula ama-slide ekugcineni ukuze uhambe kuslayidi ngasinye.
I-Optical coherence tomographic angiography (OCTA) iyindlela entsha yokubukwa kwemithambo ye-retina okungahlaseli.Nakuba i-OCTA inezinhlelo zokusebenza eziningi zomtholampilo ezithembisayo, ukunquma ikhwalithi yesithombe kuseyinselele.Sakhe uhlelo olujulile olusekelwe ekufundeni sisebenzisa i-ResNet152 neural network classifier eqeqeshwe kusengaphambili nge-ImageNet ukuze sihlukanise izithombe ze-superficial capillary plexus kusukela kumaskeni angu-347 ezigulini ezingu-134.Izithombe ziphinde zahlolwa mathupha njengeqiniso langempela ngabalingani ababili abazimele bemodeli yokufunda egadiwe.Ngenxa yokuthi izimfuneko zekhwalithi yesithombe zingase zihluke kuye ngezilungiselelo zomtholampilo noma zocwaningo, amamodeli amabili aqeqeshiwe, eyodwa ngeyokuqashelwa kwesithombe sekhwalithi ephezulu kanye neyokubonwa kwesithombe sekhwalithi ephansi.Imodeli yethu yenethiwekhi ye-neural ibonisa indawo enhle kakhulu ngaphansi kwejika (AUC), 95% CI 0.96-0.99, \(\kappa\) = 0.81), engcono kakhulu kuneleveli yesiginali ebikwe umshini (AUC = 0.82, 95 % CI).0.77–0.86, \(\kappa\) = 0.52 kanye ne-AUC = 0.78, 95% CI 0.73–0.83, \(\kappa\) = 0.27, ngokulandelanayo).Ucwaningo lwethu lubonisa ukuthi izindlela zokufunda zomshini zingasetshenziswa ukuthuthukisa izindlela zokulawula ikhwalithi eziguquguqukayo neziqinile zezithombe ze-OCTA.
I-Optical coherence tomographic angiography (OCTA) iwubuchule obusha obususelwe ku-optical coherence tomography (OCT) engasetshenziselwa ukubonwa okungahlanyisi kwe-retinal microvasculature.I-OCTA ikala umehluko wamaphethini wokubonisa kusuka kumapulse okukhanya aphindaphindiwe endaweni efanayo ye-retina, futhi ukwakhiwa kabusha kungase kubalwe ukuze kwembule imithambo yegazi ngaphandle kokusetshenziswa okuvamile kodayi noma ezinye izinto ezihlukile.I-OCTA iphinde inike amandla ukujula kwesithombe semithambo, okuvumela odokotela ukuthi bahlole ngokwehlukana izingqimba zemikhumbi ezingaphezulu nezijulile, okusiza ukuhlukanisa phakathi kwesifo se-chorioretinal.
Nakuba le nqubo ithembisa, ukuhluka kwekhwalithi yesithombe kuseyinselele enkulu yokuhlaziywa kwesithombe okuthembekile, okwenza ukuhumusha kwesithombe kube nzima futhi kuvimbele ukutholwa komtholampilo okwandile.Ngenxa yokuthi i-OCTA isebenzisa izikena ze-OCT eziningi ezilandelanayo, izwela kakhulu kuma-artifact esithombe kune-OCT evamile.Izinkundla eziningi ze-OCTA ezentengiso zihlinzeka ngemethrikhi yekhwalithi yesithombe yazo ebizwa ngokuthi Amandla Esiginali (SS) noma ngezinye izikhathi Inkomba Yamandla Esiginali (SSI).Kodwa-ke, izithombe ezinenani eliphezulu le-SS noma i-SSI aziqinisekisi ukungabikho kwama-artifact esithombe, okungase kuthinte noma yikuphi ukuhlaziywa kwesithombe okulandelayo futhi kuholele ezinqumweni zomtholampilo ezingalungile.Ama-artifact esithombe avamile angenzeka emfanekisweni we-OCTA afaka ama-artifacts anyakazayo, ama-artifact ezigaba, ama-artifacts opacity wemidiya, kanye nama-artifacts okubonakala1,2,3.
Njengoba izinyathelo ezithathwe ku-OCTA ezifana nokuminyana kwemithambo ziya ngokuya zisetshenziswa ocwaningweni lokuhumusha, izivivinyo zomtholampilo kanye nokusebenza komtholampilo, kunesidingo esiphuthumayo sokuthuthukisa izinqubo eziqinile nezithembekile zokulawula ikhwalithi yesithombe ukuze kuqedwe ubuciko besithombe4.Uxhumano Lweqa, okwaziwa nangokuthi ukuxhumana okuyinsalela, kungukuqagela ekwakhiweni kwenethiwekhi ye-neural evumela ulwazi ukuthi ludlule izendlalelo ze-convolutional kuyilapho lugcina ulwazi ngezikali ezihlukene noma izinqumo5.Ngenxa yokuthi ama-artifact esithombe angathinta ukusebenza kwesithombe sezinga elincane kanye nesilinganiso esikhulu esivamile, amanethiwekhi e-neural okweqa afaneleka kahle ukwenza ngokuzenzakalelayo lo msebenzi wokulawula ikhwalithi5.Umsebenzi osanda kushicilelwa ubonise isithembiso samanethiwekhi ajulile e-convolutional neural aqeqeshwe kusetshenziswa idatha yekhwalithi ephezulu evela kubalinganisi babantu6.
Kulolu cwaningo, siqeqesha inethiwekhi ye-neural ye-convolutional yokweqa ukuxhumana ukuze sinqume ngokuzenzakalela ikhwalithi yezithombe ze-OCTA.Sakhela phezu komsebenzi wangaphambili ngokwakha amamodeli ahlukene okuhlonza izithombe zekhwalithi ephezulu nezithombe zekhwalithi ephansi, njengoba izimfuneko zekhwalithi yesithombe zingase zehluke kuzimo ezithile zomtholampilo noma zocwaningo.Siqhathanisa imiphumela yalawa manethiwekhi namanethiwekhi e-convolutional neural ngaphandle kokuxhumeka okungekho ukuze sihlole inani lokufaka izici kumaleveli amaningi embudumbudu ngaphakathi kokufunda okujulile.Sibe sesiqhathanisa imiphumela yethu namandla esignali, isilinganiso esivame ukwamukelwa sekhwalithi yesithombe esinikezwa abakhiqizi.
Ucwaningo lwethu belubandakanya iziguli ezinesifo sikashukela eziye e-Yale Eye Center phakathi kuka-Agasti 11, 2017 no-Ephreli 11, 2019. Iziguli ezinanoma yisiphi isifo se-chorioretinal esingenaso isifo sikashukela azifakwanga.Bezingekho imibandela yokufakwa noma yokukhishwa esekelwe eminyakeni yobudala, ubulili, uhlanga, ikhwalithi yesithombe, nanoma iyiphi enye into.
Izithombe ze-OCTA zitholwe kusetshenziswa inkundla ye-AngioPlex ku-Cirrus HD-OCT 5000 (Carl Zeiss Meditec Inc, Dublin, CA) ngaphansi kwephrothokholi yokucabanga engu-8\(\izikhathi\)8 mm no-6\(\izikhathi\)6 mm.Imvume enolwazi yokubamba iqhaza ocwaningweni yatholwa kubahlanganyeli bocwaningo ngamunye, futhi Ibhodi Lokubuyekeza Isikhungo Senyuvesi Yale (IRB) ligunyaze ukusetshenziswa kwemvume enolwazi ngokuthwebula izithombe emhlabeni wonke kuzo zonke lezi ziguli.Ukulandela imigomo yeSimemezelo sase-Helsinki.Ucwaningo luvunywe yi-IRB University yaseYale.
Izithombe zepuleti elingaphezulu zihlolwe ngokusekelwe ku-Motion Artifact Score (MAS) echazwe ngaphambilini, i-Segmentation Artifact Score echazwe ngaphambilini (i-SAS), isikhungo se-foveal, ukuba khona kokufiphala kwemidiya, nokubonakala kahle kwama-capillary amancane njengoba kunqunywa umhloli wesithombe.Izithombe zihlaziywe ngabahloli ababili abazimele (i-RD ne-JW).Isithombe sinamaphuzu afakwe ileveli angu-2 (okufanelekekayo) uma zonke lezi zimo ezilandelayo zifinyelelwa: isithombe simaphakathi ne-fovea (ngaphansi kwamaphikseli angu-100 ukusuka maphakathi nesithombe), i-MAS ingu-1 noma 2, i-SAS ingu-1, futhi ukukhanya kwemidiya kungaphansi kuka-1. Yethula ezithombeni zosayizi / 16, futhi ama-capillary amancane abonakala ezithombeni ezinkulu kuno-15/16.Isithombe silinganiselwe ngo-0 (asikho isilinganiso) uma noma iyiphi kulezi zindlela ezilandelayo ifinyelelwa: isithombe asikho phakathi nendawo, uma i-MAS iyi-4, uma i-SAS ingu-2, noma ukukhanya okumaphakathi kungaphezu kuka-1/4 wesithombe, futhi ama-capillary amancane awakwazi ukulungiswa ngaphezu kwesithombe esingu-1/4 ukuze ahlukanise.Zonke ezinye izithombe ezingahlangabezani nemibandela yokufaka amagoli 0 noma 2 zitholwa njengo-1 (isiqeshana).
Emkhiwaneni.1 ibonisa izithombe zesampula zezilinganiso ezikaliwe ngazinye nama-artifact esithombe.Ukuthembeka kwamazinga aphakathi kwamanani angawodwana kwahlolwa i-kappa weighting8 ka-Cohen.Amaphuzu angawodwana okulinganisa ngakunye afingqwa ukuze kutholwe amaphuzu aphelele esithombe ngasinye, ukusuka ku-0 kuye ku-4. Izithombe ezinamanani aphelele angu-4 zibhekwa njengezinhle.Izithombe ezinamanani aphelele angu-0 noma angu-1 zibhekwa njengekhwalithi ephansi.
Inethiwekhi ye-architecture ye-ResNet152 convolutional neural network (Fig. 3A.i) eqeqeshwe ngaphambilini ezithombeni ezisuka kusizindalwazi se-ImageNet yakhiwe kusetshenziswa i-fast.ai kanye ne-PyTorch framework5, 9, 10, 11. Inethiwekhi ye-convolutional neural iyinethiwekhi esebenzisa okufundiwe. izihlungi zokuskena izingcezu zesithombe ukuze kufundwe izici zendawo nezendawo.I-ResNet yethu eqeqeshiwe iyinethiwekhi ye-neural enezingqimba ezingu-152 ebonakala ngezikhala noma "ukuxhumana okusele" okudlulisa ngesikhathi esisodwa ulwazi olunezinqumo eziningi.Ngokuveza ulwazi ngezinqumo ezihlukene kunethiwekhi, inkundla ingafunda izici zezithombe zekhwalithi ephansi kumaleveli amaningi emininingwane.Ngokungeziwe kumodeli yethu ye-ResNet, siphinde saqeqesha i-AlexNet, ukwakhiwa kwenethiwekhi ye-neural efundwe kahle, ngaphandle kokuxhumeka okungekho lapho kuqhathaniswa (Umfanekiso 3A.ii)12.Ngaphandle kokuxhumeka okungekho, le nethiwekhi ngeke ikwazi ukuthwebula izici ngembudumbudu ephezulu.
Isethi yesithombe sokuqala esingu-8\(\izikhathi\)8mm OCTA13 sithuthukisiwe kusetshenziswa izindlela zokubonisa ezivundlile neziqondile.Idathasethi egcwele yabe isihlukaniswa ngokungahleliwe ezingeni lesithombe yaba ukuqeqeshwa (51.2%), ukuhlola (12.8%), i-hyperparameter tuning (16%), kanye nokuqinisekisa (20%) amasethi edatha edatha kusetshenziswa ibhokisi lamathuluzi le-scikit-learn python14.Kucatshangelwe izimo ezimbili, eyodwa isekelwe ekutholeni kuphela izithombe zekhwalithi ephezulu kakhulu (ingqikithi yamaphuzu angu-4) kanti enye isekelwe ekutholeni kuphela izithombe zekhwalithi ephansi (ingqikithi yamaphuzu angu-0 noma angu-1).Kusimo ngasinye sokusetshenziswa sekhwalithi ephezulu nekhwalithi ephansi, inethiwekhi ye-neural iqeqeshwa kabusha kanye kudatha yesithombe sethu.Esimeni ngasinye sokusetshenziswa, inethiwekhi ye-neural yaqeqeshelwa ama-epoch angu-10, zonke izisindo zesendlalelo eziphakeme kakhulu zafrizwa, futhi izisindo zawo wonke amapharamitha angaphakathi zafundwa ama-epoch angu-40 kusetshenziswa indlela yezinga lokufunda ebandlululayo ngomsebenzi wokulahlekelwa kwe-cross-entropy 15, 16..Umsebenzi wokulahlekelwa kwe-entropy uyisikali se-logarithmic yokuhluka phakathi kwamalebula enethiwekhi abikezelwe kanye nedatha yangempela.Ngesikhathi sokuqeqeshwa, ukwehla kwe-gradient kwenziwa kumapharamitha angaphakathi wenethiwekhi ye-neural ukuze kuncishiswe ukulahlekelwa.Izinga lokufunda, izinga lokuyeka, kanye nama-hyperparameter okunciphisa isisindo ashunwa kusetshenziswa ukuthuthukiswa kwe-Bayesian ngamaphoyinti okuqalwa okungahleliwe angu-2 nokuphindaphinda okungu-10, futhi i-AUC kudathasethi yashunwa kusetshenziswa ama-hyperparameter njengokuhlosiwe okungu-17.
Izibonelo ezimelelayo zezithombe ezingu-8 × 8 mm OCTA zama-plexuses we-capillary angaphezulu amaphuzu angu-2 (A, B), 1 (C, D), no-0 (E, F).Ama-artifact esithombe abonisiwe afaka phakathi imigqa elokozayo (imicibisholo), ama-artifact okuhlukanisa (izinkanyezi), kanye nokufiphala kwemidiya (imicibisholo).Isithombe (E) naso asikho maphakathi.
I-Receiver operating features (ROC) amajika abe esekhiqizwa kuwo wonke amamodeli enethiwekhi ye-neural, futhi imibiko yamandla esignali yenjini ikhiqizwa esimweni ngasinye sokusetshenziswa sekhwalithi ephansi nekhwalithi ephezulu.Indawo engaphansi kwejika (AUC) yabalwa kusetshenziswa iphakheji ye-pROC R, futhi izikhawu zokuzethemba ezingama-95% namavelu we-p abalwe kusetshenziswa indlela ye-DeLong18,19.Izikolo eziqoqiwe zabalinganisi babantu zisetshenziswa njengesisekelo sazo zonke izibalo ze-ROC.Ngamandla esignali abikwe umshini, i-AUC ibalwe kabili: kanye ngokunqanyulwa kwekhwalithi ephezulu ye-Scalability Score kanye nokunqamuka kwekhwalithi ephansi ye-Scalability Score.Inethiwekhi ye-neural iqhathaniswa namandla esignali ye-AUC ebonisa izimo zayo zokuqeqesha nokuhlola.
Ukuze kuqhutshekwe kuhlolwe imodeli yokufunda ejulile eqeqeshiwe kudathasethi ehlukile, amamodeli ekhwalithi ephezulu nawekhwalithi ephansi asetshenziswe ngokuqondile ekuhlolweni kokusebenza okungu-32 okugcwele ubuso 6\(\izikhathi\) izithombe ze-surface slab ezingu-6mm eziqoqwe eNyuvesi yase-Yale.I-Eye Mass igxile ngesikhathi esifanayo nesithombe 8 \ (\ izikhathi \) 8 mm.Izithombe ezingu-6\(\×\) 6 mm zihlolwe ngesandla abalingani abafanayo (i-RD ne-JW) ngendlela efanayo nezithombe ezingu-8\(\×\) 8 mm, i-AUC yabalwa kanye nokunemba kanye ne-kappa ka-Cohen. .ngokulinganayo .
Isilinganiso sokungalingani kwekilasi singu-158:189 (\(\rho = 1.19\)) kumodeli yekhwalithi ephansi kanye no-80:267 (\(\rho = 3.3\)) kumodeli yekhwalithi ephezulu.Ngenxa yokuthi isilinganiso sokungalingani kwekilasi singaphansi kuka-1:4, azikho izinguquko ezithile zezakhiwo ezenziwe ukuze kulungiswe ukungalingani kwekilasi20,21.
Ukuze ubone ngeso lengqondo inqubo yokufunda kangcono, amamephu wokuvula ikilasi akhiqizwe kuwo wonke amamodeli okufunda ajulile amane aqeqeshiwe: imodeli yekhwalithi ephezulu ye-ResNet152, imodeli yekhwalithi ephansi ye-ResNet152, imodeli yekhwalithi ephezulu ye-AlexNet, nemodeli ye-AlexNet yekhwalithi ephansi.Amamephu okusebenza ekilasi akhiqizwa kusukela kuzendlalelo zokufakwayo ze-convolution zalawa mamodeli amane, futhi amamephu okushisa akhiqizwa ngokumboza amamephu okwenza asebenze anezithombe zomthombo kusukela kumasethi okuqinisekisa angu-8 × 8 mm no-6 × 6 mm22, 23.
Inguqulo engu-R engu-4.0.3 yasetshenziselwa zonke izibalo zezibalo, futhi okubonwayo kwadalwa kusetshenziswa umtapo wezincwadi wamathuluzi wezithombe we-ggplot2.
Siqoqe izithombe ezingaphambili ezingu-347 ze-superficial capillary plexus enesilinganiso esingu-8 \(\izikhathi \)8 mm kubantu abangu-134.Umshini ubike amandla esignali esikalini sika-0 kuye ku-10 kuzo zonke izithombe (okusho = 6.99 ± 2.29).Ezithombeni ezingu-347 ezitholiwe, iminyaka yobudala ekuhlolweni yayiyiminyaka engu-58.7 ± 14.6, kanti ama-39.2% ayevela ezigulini zabesilisa.Kuzo zonke izithombe, u-30.8% wawuvela eCaucasus, u-32.6% wawuvela kwabaNsundu, u-30.8% wawuvela e-Hispanics, u-4% wawuvela e-Asia, u-1.7% wawuvela kwezinye izinhlanga (Ithebula 1).).Ukusatshalaliswa kweminyaka yeziguli ezine-OCTA kuhluke kakhulu kuye ngekhwalithi yesithombe (p <0.001).Iphesenti lezithombe zekhwalithi ephezulu ezigulini ezincane ezineminyaka engu-18-45 lalingu-33.8% uma kuqhathaniswa no-12.2% wezithombe zekhwalithi ephansi (Ithebula 1).Ukusatshalaliswa kwesimo se-retinopathy yesifo sikashukela nakho kuhluke kakhulu ngekhwalithi yesithombe (p <0.017).Phakathi kwazo zonke izithombe zekhwalithi ephezulu, iphesenti leziguli ezine-PDR laliyi-18.8% uma kuqhathaniswa ne-38.8% yazo zonke izithombe ezisezingeni eliphansi (Ithebula 1).
Izilinganiso ngazinye zazo zonke izithombe zibonise ukwethembeka okumaphakathi kuya kokuqinile phakathi kwabantu abafunda izithombe (i-kappa enesisindo sika-Cohen = 0.79, 95% CI: 0.76-0.82), futhi awekho amaphoyinti esithombe lapho abalingani behluke ngokungaphezu kuka-1 (Fig. 2A)..Ukuqina kwesignali kuhlotshaniswa ngokuphawulekayo nokushaya amagoli okwenziwa ngesandla (ukulinganisa komzuzu womkhiqizo we-Pearson = 0.58, 95% CI 0.51–0.65, p<0.001), kodwa izithombe eziningi zihlonzwe njengezinomfutho wesignali ophezulu kodwa amaphuzu aphansi okwenziwa ngesandla (Fig. .2B).
Ngesikhathi sokuqeqeshwa kwezakhiwo ze-ResNet152 kanye ne-AlexNet, ukulahlekelwa kwe-cross-entropy ekuqinisekisweni nasekuqeqeshweni kwehlela ngaphezu kwama-epoch angu-50 (Umfanekiso 3B, C).Ukunemba kokuqinisekisa esikhathini sokuqeqesha sokugcina kungaphezu kuka-90% kuzo zombili izimo zokusebenzisa ikhwalithi ephezulu nekhwalithi ephansi.
Amajika okusebenza komamukeli abonisa ukuthi imodeli ye-ResNet152 idlula kakhulu amandla esignali abikwe umshini kuzo zombili izimo zokusebenzisa ikhwalithi ephansi nephezulu (p <0.001).Imodeli ye-ResNet152 iphinde iphumelele kakhulu ukwakheka kwe-AlexNet (p = 0.005 kanye ne-p = 0.014 ngamacala ekhwalithi ephansi kanye nekhwalithi ephezulu, ngokulandelana).Amamodeli avela komunye nomunye wale misebenzi akwazile ukufeza amanani e-AUC ka-0.99 no-0.97, ngokulandelana, okungcono kakhulu kunamanani ahambisanayo we-AUC ka-0.82 no-0.78 wenkomba yamandla esiginali yomshini noma i-0.97 ne-0.94 ye-AlexNet ..(Umdwebo 3).Umehluko phakathi kwe-ResNet ne-AUC emandleni esignali uphezulu uma uqaphela izithombe zekhwalithi ephezulu, okubonisa izinzuzo ezengeziwe zokusebenzisa i-ResNet kulo msebenzi.
Amagrafu abonisa ikhono lomlinganisi ngamunye ozimele lokuthola amaphuzu nokuqhathanisa namandla esignali abikwe umshini.(A) Isamba samaphuzu azohlolwa sisetshenziselwa ukwakha isamba senani lamaphuzu azohlolwa.Izithombe ezinenani eliphelele lokukala okungu-4 zinikezwa ikhwalithi ephezulu, kuyilapho izithombe ezinenani eliphelele lokukala elingu-1 noma ngaphansi zinikezwa ikhwalithi ephansi.(B) Ukuqina kwesignali kuhlobana nezilinganiso ezenziwa ngesandla, kodwa izithombe ezinomfutho wesignali ephezulu zingase zibe ezekhwalithi embi kakhulu.Ulayini onamachashazi abomvu ubonisa umkhawulo wekhwalithi onconyiwe womkhiqizi ngokusekelwe kumandla esignali (amandla esiginali \(\ge\)6).
Ukufunda ngokudluliswa kwe-ResNet kunikeza ukuthuthuka okuphawulekayo ekuhlonzweni kwekhwalithi yesithombe kukho kokubili izinga eliphansi nezimo zokusetshenziswa zekhwalithi ephezulu uma kuqhathaniswa namazinga esignali abikwe ngomshini.(A) Imidwebo yezakhiwo eyenziwe lula yezakhiwo eziqeqeshwe ngaphambilini (i) ResNet152 kanye (ii) ne-AlexNet.(B) Umlando wokuqeqesha kanye namajika okusebenza komamukeli we-ResNet152 uma kuqhathaniswa namandla esignali abikiwe ngomshini kanye nenqubo yekhwalithi ephansi ye-AlexNet.(C) Umlando wokuqeqeshwa komamukeli we-ResNet152 kanye namajika okusebenza uma kuqhathaniswa namandla esignali abikiwe umshini kanye nenqubo yekhwalithi ephezulu ye-AlexNet.
Ngemva kokulungisa umkhawulo wesinqumo, ukunemba okuphezulu kokubikezela kwemodeli ye-ResNet152 kungu-95.3% wecala lekhwalithi ephansi kanye no-93.5% wecala lekhwalithi ephezulu (Ithebula 2).Ukunemba okuphezulu kokubikezela kwemodeli ye-AlexNet ngu-91.0% wecala lekhwalithi ephansi kanye no-90.1% wecala eliphezulu (Ithebula 2).Ubukhulu bokunemba kwesibikezelo samandla esignali ngu-76.1% esimweni sokusetshenziswa sekhwalithi ephansi kanye no-77.8% esimweni sokusebenzisa sekhwalithi ephezulu.Ngokuvumelana ne-kappa ka-Cohen (\(\kappa\)), isivumelwano phakathi kwemodeli ye-ResNet152 nabalinganisi singu-0.90 ngekesi lekhwalithi ephansi kanye no-0.81 ngekesi lekhwalithi ephezulu.I-AlexNet kappa ka-Cohen ingu-0.82 kanye no-0.71 ngekhwalithi ephansi kanye nezimo zokusebenzisa zekhwalithi ephezulu, ngokulandelanayo.I-kappa yamandla esignali ka-Cohen ingu-0.52 kanye no-0.27 kumakesi okusetshenziswa kwekhwalithi ephansi nephezulu, ngokulandelanayo.
Ukuqinisekiswa kwamamodeli wokuqashelwa kwekhwalithi ephezulu nephansi ezithombeni ezingu-6\(\x\) zepuleti eliyisicaba elingu-6 mm kubonisa ikhono lemodeli eqeqeshiwe lokunquma ikhwalithi yesithombe kuwo wonke amapharamitha esithombe ahlukahlukene.Uma usebenzisa ama-slabs angu-6\(\x\) 6 mm angashoni ngekhwalithi yokuthwebula, imodeli yekhwalithi ephansi yayine-AUC engu-0.83 (95% CI: 0.69–0.98) futhi imodeli yekhwalithi ephezulu yayine-AUC engu-0.85.(95% CI: 0.55–1.00) (Ithebula 2).
Ukuhlolwa okubonakalayo kwamamephu okusebenza ekilasi lesendlalelo sokufakwayo kubonise ukuthi wonke amanethiwekhi e-neural aqeqeshiwe asebenzisa izici zesithombe phakathi nokuhlukaniswa kwesithombe (Fig. 4A, B).Ezithombeni ezingu-8 \(\izikhathi \) ezingu-8 mm nezingu-6 \(\izikhathi \) ezingu-6 mm, izithombe zokuvula i-ResNet zilandela eduze umthambo we-retina.Amamephu wokuvula we-AlexNet nawo alandela imikhumbi ye-retinal, kodwa enokulungiswa okuqinile.
Amamephu ekilasi okuvula amamodeli we-ResNet152 kanye ne-AlexNet agqamisa izici ezihlobene nekhwalithi yesithombe.(A) Imephu yekilasi yokwenza kusebenze ebonisa ukwenza kusebenze okuhambisanayo ngemva kwe-vasculature ye-retina engaphezulu kuzithombe zokuqinisekisa ezingu-8 \(\izikhathi \) ezingu-8 mm kanye (B) nesilinganiso sezithombe zokuqinisekisa ezingu-6 \(\izikhathi \) ezingu-6 mm ezincane.Imodeli ye-LQ eqeqeshwe ngemibandela yekhwalithi ephansi, imodeli ye-HQ eqeqeshwe ngemibandela yekhwalithi ephezulu.
Kuye kwaboniswa ngaphambilini ukuthi ikhwalithi yesithombe ingathinta kakhulu noma yikuphi ukulinganisa kwezithombe ze-OCTA.Ngaphezu kwalokho, ukuba khona kwe-retinopathy kwandisa izehlakalo zezithombe zobuciko7,26.Eqinisweni, kudatha yethu, ngokuhambisana nezifundo zangaphambilini, sithole ukuhlobana okubalulekile phakathi kokukhula kweminyaka kanye nobunzima besifo se-retinal kanye nokuwohloka kwekhwalithi yesithombe (p <0.001, p = 0.017 yobudala kanye nesimo se-DR, ngokulandelana; Ithebula 1) 27 . Ngakho-ke, kubalulekile ukuhlola ikhwalithi yesithombe ngaphambi kokwenza noma yikuphi ukuhlaziya komthamo wezithombe ze-OCTA.Ucwaningo oluningi oluhlaziya izithombe ze-OCTA lisebenzisa imikhawulo yesignali ebikwe ngomshini ukuze ikhiphe izithombe zekhwalithi ephansi.Nakuba ukushuba kwesignali kuboniswe ukuthi kuthinte ukulinganisa kwamapharamitha we-OCTA, ukuqina kwesignali ephezulu kukodwa kungase kunganele ukukhipha izithombe ezinezithombe zobuciko bezithombe2,3,28,29.Ngakho-ke, kuyadingeka ukuthuthukisa indlela enokwethenjelwa yokulawula ikhwalithi yesithombe.Kulokhu, sihlola ukusebenza kwezindlela zokufunda ezijulile ezigadiwe ngokumelene namandla esignali abikwe umshini.
Senze amamodeli ambalwa okuhlola ikhwalithi yesithombe ngoba izimo ezihlukene zokusebenzisa i-OCTA zingase zibe nezidingo ezihlukile zekhwalithi yesithombe.Isibonelo, izithombe kufanele zibe zekhwalithi ephezulu.Ngaphezu kwalokho, amapharamitha athile omthamo wenzuzo nawo abalulekile.Isibonelo, indawo ye-foveal avascular zone ayincikile ku-turbidity ye-non-central medium, kodwa ithinta ukuminyana kwemikhumbi.Nakuba ucwaningo lwethu luqhubeka nokugxila endleleni evamile yekhwalithi yesithombe, engaboshelwe ezimfuneko zanoma ikuphi ukuhlola okuthile, kodwa okuhloselwe ukufaka ngokuqondile amandla esiginali ebikwe umshini, sithemba ukunikeza abasebenzisi izinga elikhulu lokulawula ukuze ingakhetha imethrikhi ethile yentshisekelo kumsebenzisi.khetha imodeli ehambisana nezinga eliphezulu lezinto zokwenziwa zesithombe ezibhekwa njengezamukelekayo.
Ezigcawini zekhwalithi ephansi nezekhwalithi ephezulu, sibonisa ukusebenza okuhle kakhulu kwamanethiwekhi ajulile e-convolutional neural angekho uxhumano, anama-AUC angu-0.97 no-0.99 namamodeli ekhwalithi ephansi, ngokulandelana.Futhi sibonisa ukusebenza okuphezulu kwendlela yethu yokufunda ejulile uma kuqhathaniswa namaleveli esignali abikwe imishini kuphela.Ukweqa ukuxhumana kuvumela amanethiwekhi e-neural ukuthi afunde izici kumaleveli amaningi emininingwane, ukuthwebula izici ezingcono kakhulu zezithombe (isb. ukugqama) kanye nezici ezijwayelekile (isb. ukumisa phakathi kwesithombe30,31).Njengoba ama-artifact esithombe athinta ikhwalithi yesithombe cishe ekhonjwa kangcono ebangeni elibanzi, izakhiwo zenethiwekhi ye-neural ezinoxhumo olungekho zingabonisa ukusebenza okungcono kunalezo ezingenayo imisebenzi yokunquma ikhwalithi yesithombe.
Lapho sihlola imodeli yethu ezithombeni ezingu-6\(\×6mm) ze-OCTA, siqaphele ukwehla kokusebenza kwezigaba kokubili kwekhwalithi ephezulu nekhwalithi ephansi yamamodeli (Umfanekiso 2), ngokungafani nosayizi wemodeli eqeqeshelwe ukuhlukaniswa.Uma kuqhathaniswa nemodeli ye-ResNet, imodeli ye-AlexNet ine-falloff enkulu.Ukusebenza okungcono kakhulu kwe-ResNet kungase kube ngenxa yekhono lokuxhumana okuyinsalela ukudlulisa ulwazi ngezikali eziningi, okwenza imodeli iqine kakhulu ekuhlukaniseni izithombe ezithwetshulwe ngezikali ezihlukene kanye/noma ukukhulisa.
Omunye umehluko phakathi kwezithombe ezingu-8 \(\×\) 8 mm kanye nezithombe ezingu-6 \(\×\) ezingu-6 mm zingaholela ekuhlukaniseni okungalungile, okuhlanganisa ingxenye ephezulu kakhulu yezithombe eziqukethe izindawo ze-foveal avascular, izinguquko ekubonakaleni, i-vascular arcade, kanye akukho nerve ye-optic esithombeni 6 × 6 mm.Ngaphandle kwalokhu, imodeli yethu yekhwalithi ephezulu ye-ResNet ikwazile ukuzuza i-AUC engu-85% yezithombe ezingu-6 \(\x\) ezingu-6 mm, ukucushwa imodeli engazange iqeqeshwe ngakho, okuphakamisa ukuthi ulwazi lwekhwalithi yesithombe lufakwe ikhodi kunethiwekhi ye-neural. kufanelekile.ngosayizi wesithombe esisodwa noma ukucushwa komshini ngaphandle kokuqeqeshwa kwawo (Ithebula 2).Okuqinisekisayo, i-ResNet- kanye ne-AlexNet-like activation map of 8 \(\ times \) 8 mm kanye 6 \(\ times \) 6 mm izithombe zikwazile ukugqamisa imikhumbi ye-retinal kuzo zombili izimo, ephakamisa ukuthi imodeli inolwazi olubalulekile.ziyasebenza ekuhlukaniseni zombili izinhlobo zezithombe ze-OCTA (Fig. 4).
Lauerman et al.Ukuhlolwa kwekhwalithi yesithombe ezithombeni ze-OCTA kwenziwa ngokufanayo kusetshenziswa i-Architecture yokuQala, enye inethiwekhi ye-neural ye-skip-connection6,32 kusetshenziswa amasu okufunda ajulile.Baphinde balinganisela ucwaningo ezithombeni ze-superficial capillary plexus, kodwa basebenzisa kuphela izithombe ezincane ezingu-3×3 mm ezivela ku-Optovue AngioVue, nakuba iziguli ezinezifo ezihlukahlukene ze-chorioretinal nazo zazifakiwe.Umsebenzi wethu wakhela phezu kwezisekelo zawo, okuhlanganisa amamodeli amaningi okubhekana nemikhawulo ehlukahlukene yekhwalithi yesithombe nokuqinisekisa imiphumela yezithombe ezinosayizi abahlukahlukene.Siphinde futhi sibike imethrikhi ye-AUC yamamodeli okufunda omshini futhi sikhulise ukunemba kwawo osekuvele kuhlaba umxhwele (90%)6 kukho kokubili ikhwalithi ephansi (96%) namamodeli ekhwalithi ephezulu (95.7%)6.
Lokhu kuqeqeshwa kunemikhawulo eminingana.Okokuqala, izithombe zitholwe ngomshini owodwa kuphela we-OCTA, okuhlanganisa kuphela izithombe ze-superficial capillary plexus ku-8\(\izikhathi\)8 mm kanye no-6\(\izikhathi\)6 mm.Isizathu sokungabandakanyi izithombe ezendlalelo ezijulile ukuthi ama-artifact e-projection angenza ukuhlola okwenziwa ngesandla kwezithombe kube nzima kakhulu futhi kungenzeki okufanayo.Ngaphezu kwalokho, izithombe zitholwe kuphela ezigulini ezinesifo sikashukela, lapho i-OCTA ivela njengethuluzi elibalulekile lokuxilonga kanye ne-prognostic33,34.Nakuba sikwazile ukuhlola imodeli yethu ezithombeni zosayizi abahlukahlukene ukuze siqinisekise ukuthi imiphumela iqinile, asikwazanga ukuhlonza amasethi edatha afanelekile avela ezikhungweni ezihlukene, okukhawulele ukuhlola kwethu ukutholakala kwemodeli jikelele.Nakuba izithombe zitholwe esikhungweni esisodwa kuphela, zitholwe ezigulini ezinezizinda ezihlukene zobuhlanga nezinhlanga, okungamandla ayingqayizivele ocwaningo lwethu.Ngokufaka ukuhlukahluka kunqubo yethu yokuqeqesha, sithemba ukuthi imiphumela yethu izokwenziwa ngendlela ebanzi ngomqondo obanzi, nokuthi sizogwema ukufaka ikhodi ukuchema kobuhlanga kumamodeli esiwaqeqeshayo.
Ucwaningo lwethu lubonisa ukuthi amanethiwekhi e-neural eqa uxhumano angaqeqeshelwa ukuzuza ukusebenza okuphezulu ekunqumeni ikhwalithi yesithombe se-OCTA.Sihlinzeka ngalawa mamodeli njengamathuluzi ocwaningo olwengeziwe.Ngenxa yokuthi amamethrikhi ahlukene angase abe nezidingo ezihlukile zekhwalithi yesithombe, imodeli yokulawula ikhwalithi ingathuthukiswa kumethrikhi ngayinye kusetshenziswa isakhiwo esisungulwe lapha.
Ucwaningo lwesikhathi esizayo kufanele lufake izithombe zosayizi abahlukene kusukela ekujuleni okuhlukahlukene kanye nemishini ehlukene ye-OCTA ukuze kutholwe inqubo yokuhlola ikhwalithi yesithombe yokufunda ejulile engenziwa jikelele kumaplathifomu e-OCTA namaphrothokholi ezithombe.Ucwaningo lwamanje futhi lusekelwe ezindleleni zokufunda ezijulile ezigadiwe ezidinga ukuhlolwa komuntu nokuhlolwa kwesithombe, okungase kusebenze kanzima futhi kudle isikhathi kumadathasethi amakhulu.Kusazobonakala ukuthi izindlela zokufunda ezijulile ezingagadiwe zingakwazi yini ukuhlukanisa phakathi kwezithombe zekhwalithi ephansi nezithombe zekhwalithi ephezulu.
Njengoba ubuchwepheshe be-OCTA buqhubeka nokuvela futhi nesivinini sokuskena sikhuphuka, izehlakalo zama-artifact ezithombe nezithombe zekhwalithi ephansi zingase zehle.Ukuthuthukiswa kwesofthiwe, okufana nesici esisanda kwethulwa sokususwa kwe-artifact, kungase futhi kunciphise le mikhawulo.Kodwa-ke, izinkinga eziningi zihlala zinjengokucabanga kweziguli ezingalungiseki kahle noma ukonakala kwemidiya okuphawulekayo kuholela ezithombeni zobuciko.Njengoba i-OCTA isetshenziswa kabanzi ezivivinyweni zomtholampilo, ukucatshangelwa ngokucophelela kuyadingeka ukuze kutholwe imihlahlandlela ecacile yamazinga amukelekayo we-artifact yesithombe ukuze kuhlaziywe isithombe.Ukusetshenziswa kwezindlela zokufunda ezijulile ezithombeni ze-OCTA kunesithembiso esikhulu futhi ucwaningo olwengeziwe luyadingeka kule ndawo ukuze kuthuthukiswe indlela eqinile yokulawula ikhwalithi yesithombe.
Ikhodi esetshenziswe ocwaningweni lwamanje iyatholakala ku-octa-qc repository, https://github.com/rahuldhodapkar/octa-qc.Amasethi edatha akhiqizwe kanye/noma ahlaziywa ngesikhathi socwaningo lwamanje ayatholakala kubabhali abafanele uma kunesicelo esifanele.
I-Spaide, RF, Fujimoto, JG & Waheed, NK Ama-artifacts wesithombe ku-optical coherence angiography.I-Retina 35, 2163–2180 (2015).
Fenner, BJ et al.Ukuhlonzwa kwezici zokuthwebula ezinquma ikhwalithi nokuphindaphindeka kwezilinganiso zokuminyana kwe-retinal capillary plexus ku-OCT angiography.BR.J. Ophthalmol.102, 509–514 (2018).
Lauerman, JL et al.Umthelela wobuchwepheshe bokulandelela iso kwikhwalithi yesithombe ye-OCT angiography ekuwohlokeni kwe-macular okuhlobene neminyaka.Ikhothamo lengcwaba.emtholampilo.Iphelelwa yisikhathi.i-ophthalmology.255, 1535–1542 (2017).
Babyuch AS et al.Izilinganiso ze-OCTA ze-capillary perfusion density zisetshenziselwa ukuthola nokuhlola i-macular ischemia.ukuhlinzwa kwe-ophthalmic.I-Retinal Laser Imaging 51, S30–S36 (2020).
Yena, K., Zhang, X., Ren, S., kanye neLanga, J. Ukufunda Okusalela Okujulile Kokubonwa Kwesithombe.Ngo-2016 eNgqungqutheleni ye-IEEE ye-Computer Vision kanye nokuqashelwa kwephethini (2016).
Lauerman, JL et al.Ukuhlolwa kwekhwalithi yesithombe okuzenzakalelayo kwe-OCT angiographic kusetshenziswa ama-algorithms okufunda ajulile.Ikhothamo lengcwaba.emtholampilo.Iphelelwa yisikhathi.i-ophthalmology.257, 1641–1648 (2019).
Lauermann, J. et al.Ukusabalala kwamaphutha okuhlukaniswa kanye nezinto zokwenziwa ezinyakazayo ku-OCT angiography kuncike esifweni se-retina.Ikhothamo lengcwaba.emtholampilo.Iphelelwa yisikhathi.i-ophthalmology.256, 1807–1816 (2018).
Pask, Adam et al.I-Pytorch: Umtapowolwazi Wokufunda Okujulile Obalulekile, Osebenza Kakhulu.Ukucubungula okuthuthukisiwe kolwazi lwe-neural.uhlelo.32, 8026–8037 (2019).
Deng, J. et al.I-ImageNet: Isizindalwazi Sezithombe Ezisezingeni Elikhulu.2009 IEEE Conference on Computer Vision kanye Nokuqashelwa Iphethini.248–255.(2009).
U-Krizhevsky A., u-Sutzkever I. kanye nokuhlelwa kwe-Hinton GE Imagenet kusetshenziswa amanethiwekhi ajulile e-convolutional neural.Ukucubungula okuthuthukisiwe kolwazi lwe-neural.uhlelo.25, 1 (2012).