Na gode da ziyartar Nature.com.Kuna amfani da sigar burauza tare da iyakancewar tallafin CSS.Don ƙwarewa mafi kyau, muna ba da shawarar ku yi amfani da sabuntar burauza (ko kuma musaki Yanayin dacewa a cikin Internet Explorer).Bugu da ƙari, don tabbatar da ci gaba da goyon baya, muna nuna shafin ba tare da salo da JavaScript ba.
Sliders suna nuna labarai uku a kowane faifai.Yi amfani da maɓallan baya da na gaba don motsawa ta cikin nunin faifai, ko maɓallan masu sarrafa nunin faifai a ƙarshen don matsawa ta kowane faifan.
Haɗin kai na gani na gani angiography (OCTA) wata sabuwar hanya ce don hangen nesa mara ƙarfi na tasoshin ido.Kodayake OCTA tana da aikace-aikacen asibiti masu ban sha'awa da yawa, ƙayyade ingancin hoto ya kasance ƙalubale.Mun haɓaka tsarin tushen koyo mai zurfi ta amfani da mai rarraba cibiyar sadarwa na ResNet152 wanda aka riga aka horar da shi tare da ImageNet don rarraba hotunan plexus na sama na sama daga binciken 347 na marasa lafiya 134.Hotunan kuma an tantance su da hannu a matsayin gaskiya ta gaskiya ta masu kima masu zaman kansu guda biyu don tsarin ilmantarwa da ake kulawa.Saboda buƙatun ingancin hoto na iya bambanta dangane da na asibiti ko saitunan bincike, an horar da samfura guda biyu, ɗaya don ingancin hoto mai inganci da ɗayan don ƙarancin ingancin hoton hoto.Tsarin hanyar sadarwar mu na jijiyoyi yana nuna kyakkyawan yanki a ƙarƙashin lanƙwasa (AUC), 95% CI 0.96-0.99, \ (\kappa \) = 0.81), wanda ya fi mahimmanci fiye da matakin siginar da injin ya ruwaito (AUC = 0.82, 95). % CI).0.77-0.86, \(\kappa \) = 0.52 da AUC = 0.78, 95% CI 0.73-0.83, \ (\kappa \) = 0.27, bi da bi).Bincikenmu ya nuna cewa ana iya amfani da hanyoyin koyo na inji don haɓaka sassauƙa da hanyoyin sarrafa inganci masu ƙarfi don hotunan OCTA.
Haɗin kai na gani na gani na gani (OCTA) sabuwar dabara ce wacce ta dogara da tsarin haɗin kai na gani (OCT) wanda za a iya amfani da shi don hangen nesa mara ƙarfi na microvasculature na ido.OCTA tana auna bambance-bambancen tsarin tunani daga maimaita bugun haske a cikin yanki ɗaya na retina, sannan kuma ana iya ƙididdige gyare-gyare don bayyana tasoshin jini ba tare da yin amfani da rini ko wasu abubuwan ban sha'awa ba.OCTA kuma yana ba da damar zurfin-ƙudirin hoto na jijiyoyin jini, kyale likitocin su bincika daban-daban na sama da zurfin jirgin ruwa, suna taimakawa bambance tsakanin cututtukan chorioretinal.
Duk da yake wannan fasaha tana da alƙawarin, bambancin ingancin hoto ya kasance babban ƙalubale don ingantaccen bincike na hoto, yin fassarar hoto mai wahala da hana yaduwar asibiti.Saboda OCTA yana amfani da binciken OCT da yawa a jere, ya fi kula da kayan tarihi fiye da daidaitattun OCT.Yawancin dandamali na OCTA na kasuwanci suna samar da nasu awo ingancin hoton da ake kira Ƙarfin Sigina (SS) ko wani lokacin Ƙarfin Sigina (SSI).Duk da haka, hotuna tare da babban darajar SS ko SSI ba su bada garantin rashin kayan aikin hoto ba, wanda zai iya rinjayar duk wani binciken hoto na gaba kuma ya haifar da yanke shawara na asibiti ba daidai ba.Ayyukan hoto na gama-gari waɗanda zasu iya faruwa a cikin hoton OCTA sun haɗa da kayan aikin motsi, kayan tarihi na yanki, kayan aikin watsa labarai, da kayan tsinkaya1,2,3.
Kamar yadda matakan da aka samo daga OCTA irin su ƙananan ƙwayoyin cuta suna ƙara yin amfani da su a cikin bincike na fassarar, gwaje-gwaje na asibiti da kuma aikin asibiti, akwai buƙatar gaggawa don haɓaka ƙaƙƙarfan tsarin kula da ingancin hoto don kawar da artefacts4.Haɗin tsallake-tsallake, wanda kuma aka sani da ragowar haɗin kai, hasashe ne a cikin gine-ginen cibiyar sadarwar jijiyoyi waɗanda ke ba da damar bayanai su ketare yadudduka na juyi yayin adana bayanai a ma'auni ko kudurori daban-daban5.Saboda kayan tarihi na hoto na iya shafar ƙarami da babban aikin hoto na gabaɗaya, cibiyoyin sadarwar jijiyoyi na tsallake-tsallake sun dace sosai don sarrafa wannan aikin sarrafa inganci5.Ayyukan da aka buga kwanan nan sun nuna wasu alkawura don zurfin hanyoyin sadarwa na jujjuyawar jijiyoyi waɗanda aka horar da su ta amfani da bayanai masu inganci daga masu kimanta ɗan adam6.
A cikin wannan binciken, muna horar da hanyar sadarwa ta tsallake-tsallake na jijiyoyi don tantance ingancin hotunan OCTA ta atomatik.Mun gina kan aikin da ya gabata ta hanyar haɓaka samfura daban-daban don gano hotuna masu inganci da ƙananan hotuna, kamar yadda buƙatun ingancin hoto na iya bambanta ga takamaiman yanayin asibiti ko bincike.Muna kwatanta sakamakon waɗannan cibiyoyin sadarwa tare da hanyoyin sadarwa na jujjuyawar jijiyoyi ba tare da ɓacewar haɗin kai don kimanta ƙimar haɗa fasali a matakan girma da yawa a cikin zurfin koyo ba.Sai muka kwatanta sakamakonmu zuwa ƙarfin sigina, ma'aunin ingancin hoto da aka yarda da shi wanda masana'antun ke bayarwa.
Nazarinmu ya haɗa da marasa lafiya da ciwon sukari waɗanda suka halarci Cibiyar Yale Eye tsakanin Agusta 11, 2017 da Afrilu 11, 2019. An cire marasa lafiya da duk wani cututtukan chorioretinal marasa ciwon sukari.Babu ƙa'idodin haɗawa ko keɓance dangane da shekaru, jinsi, launin fata, ingancin hoto, ko kowane abu.
An samo hotunan OCTA ta amfani da dandalin AngioPlex akan Cirrus HD-OCT 5000 (Carl Zeiss Meditec Inc, Dublin, CA) a ƙarƙashin 8" (\ times") 8 mm da 6 \ (\ times \) 6 mm ladabi ladabi.An sami sanarwar yarda don shiga cikin binciken daga kowane ɗan takarar binciken, kuma Hukumar Kula da Cibiyoyin Kula da Makarantun Jami’ar Yale (IRB) ta amince da yin amfani da sanarwar yarda tare da daukar hoto na duniya ga duk waɗannan marasa lafiya.Bi ka'idodin sanarwar Helsinki.Jami'ar Yale IRB ta amince da binciken.
An kimanta Hotunan faranti na saman saman da aka kwatanta a baya da aka kwatanta da Motion Artifact Score (MAS), wanda aka siffanta a baya Segmentation Artifact Score (SAS), cibiyar foveal, kasancewar rashin bayyanawar kafofin watsa labarai, da kyakkyawan hangen nesa na ƙananan capillaries kamar yadda mai kimanta hoto ya ƙaddara.Masu kimantawa biyu masu zaman kansu (RD da JW) ne suka tantance hotunan.Hoton yana da maki 2 (cancanci) idan duk waɗannan sharuɗɗan sun cika: Hoton yana a tsakiya a fovea (kasa da 100 pixels daga tsakiyar hoton), MAS shine 1 ko 2, SAS shine 1, kuma Media opacity kasa da 1. Gaba a kan hotuna na size / 16, da kuma kananan capillaries ana gani a cikin hotuna girma fiye da 15/16.An ƙididdige hoto 0 (babu kima) idan an cika kowane ɗayan waɗannan sharuɗɗan masu zuwa: hoton ba ya tsakiya, idan MAS 4 ne, idan SAS shine 2, ko matsakaicin rashin ƙarfi ya fi 1/4 na hoton, kuma Ba za a iya daidaita ƙananan capillaries fiye da 1 hoto / 4 don bambanta ba.Duk sauran hotunan da basu cika ka'idojin maki 0 ​​ko 2 ana saka su azaman 1 (clipping).
A kan fig.1 yana nuna samfurin hotuna don kowane ƙididdiga masu ƙima da kayan tarihi na hoto.An tantance amincin tsaka-tsakin makin daidaiku ta Cohen's kappa weighting8.Ana tara makin kowane mutum ɗaya don samun cikakken makin kowane hoto, jere daga 0 zuwa 4. Hotunan da ke da jimillar maki 4 ana ɗaukan suna da kyau.Hotuna masu jimlar maki 0 ​​ko 1 ana ɗaukar ƙarancin inganci.
ResNet152 architecture convolutional neural network (Fig. 3A.i) wanda aka riga aka horar akan hotuna daga bayanan bayanan ImageNet an samar dashi ta amfani da fast.ai da tsarin PyTorch5, 9, 10, 11. Cibiyar sadarwa ta juzu'i wata hanyar sadarwa ce wacce ke amfani da masu koyo. tacewa don duba gutsuttsura hoto don nazarin sararin samaniya da fasalin gida.ResNet ɗinmu da aka horar da shi cibiyar sadarwa ce mai Layer Layer 152 wacce ke da giɓi ko “haɗin da ya rage” waɗanda ke watsa bayanai tare da ƙuduri da yawa a lokaci guda.Ta hanyar ƙaddamar da bayanai a matakai daban-daban akan hanyar sadarwa, dandamali na iya koyon fasalulluka na ƙananan hotuna a matakan daki-daki.Baya ga tsarin mu na ResNet, mun kuma horar da AlexNet, ingantaccen tsarin gine-ginen cibiyar sadarwa na jijiyoyi, ba tare da rasa haɗin kai don kwatantawa ba (Hoto 3A.ii)12.Ba tare da ɓacewar haɗin kai ba, wannan hanyar sadarwar ba za ta iya ɗaukar fasalulluka a mafi girman girma ba.
An haɓaka saitin hoto na asali na 8mm OCTA13 ta amfani da dabarun tunani a kwance da tsaye.Sannan an raba cikakken bayanan da bazuwar a matakin hoton zuwa horo (51.2%), gwaji (12.8%), daidaitawar hyperparameter (16%), da ingantattun bayanai (20%) ta amfani da akwatin kayan aikin scikit-Learn python14.An yi la'akari da shari'o'i biyu, ɗaya ya dogara ne akan gano mafi kyawun hotuna kawai (cikakken maki 4) ɗayan kuma bisa ga gano mafi ƙarancin hotuna kawai (maki 0 ​​ko 1).Ga kowane yanayin amfani mai inganci da ƙarancin inganci, cibiyar sadarwar jijiyoyi ana sake horar da su sau ɗaya akan bayanan hoton mu.A cikin kowane yanayin amfani, an horar da cibiyar sadarwa na jijiyoyi don lokutan 10, duk sai dai mafi girman nauyin nauyin daskarewa, kuma an koyi nauyin duk sigogi na ciki don lokutan 40 ta amfani da hanyar ƙimar koyo na nuna bambanci tare da aikin asarar giciye-entropy 15, 16..Aikin asarar giciye entropy shine ma'auni na ma'aunin logarithmic na rashin daidaituwa tsakanin alamun cibiyar sadarwa da aka annabta da ainihin bayanai.A lokacin horo, ana yin saukowar gradient akan sigogin ciki na cibiyar sadarwar jijiya don rage asara.An daidaita ƙimar koyo, ƙimar ficewa, da ma'aunin rage nauyi ta amfani da ingantawar Bayesian tare da 2 bazuwar farawa da 10, kuma AUC akan bayanan bayanan an kunna ta ta amfani da hyperparameters a matsayin manufa na 17.
Misalai na wakilci na 8 × 8 mm Hotunan OCTA na ɗimbin ɗimbin ɗabi'a sun sami maki 2 (A, B), 1 (C, D), da 0 (E, F).Hotunan kayan tarihi da aka nuna sun haɗa da layukan firgita (kibiyoyi), kayan tarihi na rarrabuwa (asterisks), da baƙar magana (kibiyoyi).Hoton (E) kuma baya cikin tsakiya.
Ana haifar da halayen aiki na mai karɓa (ROC) don duk ƙirar hanyar sadarwa na jijiyoyi, kuma ana samar da rahotannin ƙarfin siginar injin don kowane ƙaramin inganci da ingancin amfani.An ƙididdige yanki a ƙarƙashin lanƙwasa (AUC) ta amfani da kunshin pROC R, kuma an ƙididdige tazarar amincewa da 95% ta amfani da hanyar DeLong18,19.Ana amfani da tarin makin na ɗan adam azaman jigon duk lissafin ROC.Don ƙarfin siginar da na'urar ta ruwaito, an ƙididdige AUC sau biyu: sau ɗaya don babban ingancin Scalability Score yanke kuma sau ɗaya don yanke madaidaicin ƙimar Scalability Score.Ana kwatanta hanyar sadarwar jijiya da ƙarfin siginar AUC wanda ke nuna nasa horo da yanayin kimantawa.
Don ci gaba da gwada samfurin ilmantarwa mai zurfi da aka horar akan keɓantaccen bayanan bayanai, an yi amfani da ƙima mai inganci da ƙarancin ƙima kai tsaye zuwa kimanta aikin 32 cikakkiyar fuska 6 \ (\ sau \) 6mm saman bene hotuna da aka tattara daga Jami'ar Yale.Ido Mass yana tsakiya a lokaci guda da hoton 8 \(\ times \) 8 mm.Hotunan 6 \ (\ × \) 6 mm an tantance su da hannu ta masu ƙima iri ɗaya (RD da JW) daidai da hotuna 8 \ (\ × \) 8 mm, an ƙididdige AUC da daidaito da kappa na Cohen. .daidai .
Matsakaicin rashin daidaituwa na aji shine 158:189 (\ (\rho = 1.19 \)) don ƙima mai ƙarancin inganci da 80:267 (\ (\rho = 3.3 \)) don ƙira mai inganci.Saboda rashin daidaituwar ajin bai wuce 1:4 ba, ba a yi takamaiman canje-canjen gine-gine don gyara rashin daidaituwar aji20,21.
Don ganin tsarin ilmantarwa da kyau, an ƙirƙiri taswirorin kunna aji don duk nau'ikan horarwa mai zurfi guda huɗu: ƙirar ResNet152 mai inganci, ƙirar ResNet152 mara ƙarancin inganci, ƙirar AlexNet mai inganci, da ƙarancin ƙirar AlexNet.Ana ƙirƙira taswirorin kunna aji daga shigarwar juzu'i na waɗannan samfura huɗu, kuma ana samar da taswirorin zafi ta hanyar rufe taswirorin kunnawa tare da hotunan tushe daga 8 × 8 mm da 6 × 6 mm ingantattun saiti22, 23.
An yi amfani da sigar R 4.0.3 don duk ƙididdiga na ƙididdiga, kuma an ƙirƙiri abubuwan gani ta amfani da ɗakin karatu na kayan aikin hoto na ggplot2.
Mun tattara hotuna 347 na gaba na plexus capillary na sama mai auna 8 \ (\ lokuta \) 8 mm daga mutane 134.Injin ya ba da rahoton ƙarfin sigina akan sikelin 0 zuwa 10 don duk hotuna (ma'ana = 6.99 ± 2.29).Daga cikin hotuna 347 da aka samu, matsakaicin shekarun da aka gwada shine 58.7 ± 14.6 shekaru, kuma 39.2% sun fito ne daga marasa lafiya maza.Daga cikin dukkan hotuna, 30.8% sun fito daga Caucasians, 32.6% daga Blacks, 30.8% daga Hispanic, 4% daga Asiya, da 1.7% daga sauran jinsi (Table 1).).Rarraba shekarun marasa lafiya tare da OCTA ya bambanta sosai dangane da ingancin hoton (p <0.001).Yawan hotuna masu inganci a cikin ƙananan marasa lafiya masu shekaru 18-45 sun kasance 33.8% idan aka kwatanta da 12.2% na ƙananan hotuna (Table 1).Rarraba matsayin retinopathy na ciwon sukari shima ya bambanta sosai a ingancin hoto (p <0.017).Daga cikin dukkanin hotuna masu inganci, yawan marasa lafiya tare da PDR shine 18.8% idan aka kwatanta da 38.8% na duk ƙananan hotuna (Table 1).
Ƙididdigar ɗaiɗaikun duk hotuna sun nuna matsakaici zuwa ƙarfi mai ƙarfi tsakanin mutanen da ke karanta hotunan (Kappa mai nauyi na Cohen = 0.79, 95% CI: 0.76-0.82), kuma babu alamun hoto inda masu ƙima suka bambanta da fiye da 1 (Fig. 2 A)..Ƙarfin siginar yana da alaƙa da mahimmanci tare da ƙima na hannu (daidaitaccen lokacin samfurin Pearson = 0.58, 95% CI 0.51-0.65, p<0.001), amma an gano hotuna da yawa a matsayin suna da girman sigina amma ƙananan ƙira na hannu (Fig. .2B).
A lokacin horar da gine-ginen ResNet152 da AlexNet, asarar giciye-entropy akan ingantawa da horarwa ya faɗi akan 50 zamanin (Hoto 3B,C).Tabbatar da daidaito a lokacin horo na ƙarshe ya wuce 90% don duka inganci da ƙarancin amfani.
Maɓallin aikin mai karɓa ya nuna cewa samfurin ResNet152 ya fi ƙarfin ƙarfin siginar da injin ya ruwaito a cikin ƙananan ƙarancin amfani da inganci (p <0.001).Samfurin ResNet152 shima ya fi karfin tsarin gine-ginen AlexNet (p = 0.005 da p = 0.014 don karancin inganci da lokuta masu inganci, bi da bi).Samfuran da aka samo don kowane ɗayan waɗannan ayyukan sun sami damar cimma ƙimar AUC na 0.99 da 0.97, bi da bi, wanda ya fi daidai da ƙimar AUC na 0.82 da 0.78 don alamar ƙarfin siginar injin ko 0.97 da 0.94 don AlexNet ..(Hoto na 3).Bambanci tsakanin ResNet da AUC a cikin ƙarfin sigina ya fi girma yayin gane hotuna masu inganci, yana nuna ƙarin fa'idodin amfani da ResNet don wannan aikin.
Hotunan suna nuna ikon kowane mai ƙididdigewa mai zaman kansa don yin ƙima da kwatanta shi da ƙarfin siginar da injin ya ruwaito.(A) Ana amfani da jimillar abubuwan da za a tantance don ƙirƙirar adadin adadin da za a tantance.Hotunan da ke da maƙiyan ma'auni na 4 gabaɗaya an sanya su masu inganci, yayin da hotunan da ke da ƙima mai ƙima na 1 ko ƙasa da haka ana sanya ƙarancin inganci.(B) Ƙarfin siginar ya yi daidai da ƙididdiga na hannu, amma hotuna masu girman sigina na iya zama mafi ƙarancin inganci.Layin jajayen dige-dige yana nuna madaidaicin shawarar masana'anta dangane da ƙarfin sigina (ƙarfin sigina \(\ge\)6).
Koyon canja wurin ResNet yana ba da gagarumin ci gaba a cikin tantance ingancin hoto don duka ƙarancin inganci da yanayin amfani mai inganci idan aka kwatanta da matakan siginar da injin ya ruwaito.(A) Sauƙaƙe zane-zane na gine-ginen da aka riga aka horar (i) ResNet152 da (ii) gine-ginen AlexNet.(B) Tarihin horarwa da ƙwanƙwasa aikin mai karɓa don ResNet152 idan aka kwatanta da na'ura da aka ruwaito ƙarfin sigina da ƙananan ma'auni na AlexNet.(C) Tarihin horar da mai karɓar ResNet152 da madaidaicin aiki idan aka kwatanta da ƙarfin siginar na'ura da aka bayar da rahoton AlexNet.
Bayan daidaita madaidaicin iyakar yanke shawara, matsakaicin daidaiton tsinkayar samfurin ResNet152 shine 95.3% don ƙaramin inganci da 93.5% don ƙarar inganci (Table 2).Matsakaicin daidaiton tsinkaya na ƙirar AlexNet shine 91.0% don ƙaramin inganci da 90.1% don ƙarar inganci (Table 2).Madaidaicin ƙimar ƙarfin sigina shine 76.1% don ƙaramin ingancin amfani da 77.8% don yanayin amfani mai inganci.A cewar Cohen's kappa (\ (\kappa \)), yarjejeniya tsakanin samfurin ResNet152 da masu ƙididdigewa shine 0.90 don ƙaramin inganci da 0.81 don ƙarar inganci.Cohen's AlexNet kappa shine 0.82 da 0.71 don ƙarancin inganci da ingancin amfani, bi da bi.Ƙarfin siginar Cohen kappa shine 0.52 da 0.27 don ƙarancin amfani da inganci mai inganci, bi da bi.
Tabbatar da samfura masu inganci da ƙarancin inganci akan hotuna 6 (\ x) na faranti mai faɗin mm 6 yana nuna ƙarfin ƙirar ƙirar don tantance ingancin hoto a cikin sigogin hoto daban-daban.Lokacin amfani da 6 \ (\ x \) 6 mm slabs maras tushe don ingancin hoto, ƙaramin ƙirar ƙirar yana da AUC na 0.83 (95% CI: 0.69-0.98) kuma babban ingancin ƙirar yana da AUC na 0.85.(95% CI: 0.55-1.00) (Table 2).
Duban gani na taswirar kunna aji ajin shigarwa ya nuna cewa duk hanyoyin sadarwar jijiyoyi da aka horar sun yi amfani da fasalin hoto yayin rarraba hoto (Fig. 4A, B).Domin 8 (\ times \) 8 mm da 6 \ (\ times \) 6 mm hotuna, Hotunan kunna ResNet suna bin vasculature na ido.Taswirorin kunnawa AlexNet suma suna bin tasoshin ido, amma tare da matsananciyar ƙuduri.
Taswirorin kunna aji don ƙirar ResNet152 da AlexNet suna haskaka fasali masu alaƙa da ingancin hoto.(A) Taswirar kunna aji yana nuna kunnawa mai daidaituwa bayan vasculature na ido na sama akan 8 \ (\ times \) 8mm hotuna ingantacce da (B) kan ƙaramin 6 \ (\ sau \) 6 mm hotuna ingantacce.Samfurin LQ wanda aka horar akan ƙananan ma'auni, samfurin HQ ya horar da ma'auni masu inganci.
A baya an nuna cewa ingancin hoto na iya tasiri sosai ga kowane adadin hotunan OCTA.Bugu da ƙari, kasancewar ciwon ƙwayar cuta yana ƙara yawan abubuwan da suka faru na hotuna7,26.A gaskiya ma, a cikin bayananmu, daidai da binciken da aka yi a baya, mun sami wata ƙungiya mai mahimmanci tsakanin ƙara yawan shekaru da tsanani na cututtukan cututtuka da kuma lalacewa a cikin ingancin hoto (p <0.001, p = 0.017 don shekaru da matsayi na DR, bi da bi; Table 1) 27 Don haka, yana da mahimmanci a tantance ingancin hoto kafin yin kowane ƙididdiga na hotuna OCTA.Yawancin binciken da ke nazarin hotunan OCTA suna amfani da mashin da aka ba da rahoton na'ura don ƙaddamar da ƙananan hotuna masu inganci.Ko da yake an nuna ƙarfin sigina don rinjayar ƙididdige ma'auni na OCTA, babban ƙarfin sigina kadai bazai isa ya kawar da hotuna tare da kayan aikin hoto2,3,28,29 ba.Sabili da haka, ya zama dole don haɓaka ingantaccen ingantaccen hanyar sarrafa hoto.Don wannan, muna kimanta aikin hanyoyin ilmantarwa mai zurfi da ake kulawa akan ƙarfin siginar da injin ya ruwaito.
Mun ƙirƙira samfura da yawa don kimanta ingancin hoto saboda nau'ikan amfani da OCTA na iya samun buƙatun ingancin hoto daban-daban.Misali, ya kamata hotuna su kasance mafi inganci.Bugu da ƙari, ƙayyadaddun ƙayyadaddun ƙididdiga na sha'awa suna da mahimmanci.Alal misali, yankin na foveal avascular zone ba ya dogara ne a kan turbidity na wadanda ba tsakiya matsakaici, amma rinjayar da yawa daga cikin tasoshin.Yayin da bincikenmu ya ci gaba da mayar da hankali kan tsarin gaba ɗaya ga ingancin hoto, ba a haɗa shi da buƙatun kowane gwaji na musamman ba, amma an yi niyya don maye gurbin ƙarfin siginar da injin ya ruwaito kai tsaye, muna fatan ba masu amfani da ƙimar iko sosai don su zai iya zaɓar takamaiman awo na sha'awa ga mai amfani.zaɓi samfurin da ya dace da matsakaicin matsayi na kayan tarihi na hoto da aka ɗauka karɓuwa.
Don ƙananan wurare masu inganci da kyawawan wurare, muna nuna kyakkyawan aikin haɗin kai-bacewar hanyoyin sadarwar jijiyoyi masu zurfi, tare da AUC na 0.97 da 0.99 da ƙananan ƙira, bi da bi.Muna kuma nuna kyakkyawan aiki na tsarin ilmantarwa mai zurfi idan aka kwatanta da matakan sigina da inji kawai ya ruwaito.Haɗin tsallake-tsallake yana ba da damar cibiyoyin sadarwa na jijiyoyi don koyan fasali a matakan daki-daki da yawa, suna ɗaukar mafi kyawun abubuwan hotuna (misali bambanci) da kuma fasali na gaba ɗaya (misali hoton tsakiya30,31).Tunda kayan tarihi na hoto waɗanda ke shafar ingancin hoto mai yiwuwa an fi gano su a cikin kewayon da yawa, gine-ginen cibiyar sadarwar jijiyoyi tare da ɓacewar haɗin kai na iya nuna kyakkyawan aiki fiye da waɗanda ba su da ayyukan tantance ingancin hoto.
Lokacin gwada samfurin mu akan 6 \ (\ × 6mm) Hotunan OCTA, mun lura da raguwar aikin rarrabawa don duka inganci da ƙananan ƙira (Fig. 2), da bambanci da girman samfurin da aka horar da su don rarrabawa.Idan aka kwatanta da ƙirar ResNet, ƙirar AlexNet tana da faɗuwa mafi girma.Ingantacciyar ingantacciyar aikin ResNet na iya kasancewa saboda iyawar sauran hanyoyin haɗin kai don watsa bayanai a ma'auni da yawa, wanda ke sa ƙirar ta fi ƙarfi don rarraba hotunan da aka ɗauka a ma'auni daban-daban da/ko girma.
Wasu bambance-bambance tsakanin 8 \ (\ × \) 8 mm hotuna da 6 \ (\ × \) 6 mm hotuna na iya haifar da rashin kyau rarrabuwa, ciki har da wani in mun gwada da babban rabo na hotuna dauke da foveal avascular yankunan, canje-canje a ganuwa, jijiyoyin bugun gini arcades, da kuma babu jijiyar gani akan hoton 6 × 6 mm.Duk da wannan, babban ingancin samfurin mu na ResNet ya sami damar cimma AUC na 85% don hotuna 6 \ (\ x) 6 mm, tsarin da ba a horar da samfurin ba, yana ba da shawarar cewa bayanan ingancin hoto da aka sanya a cikin hanyar sadarwar jijiyoyi. ya dace.don girman hoto ɗaya ko ƙayyadaddun injin a waje da horonsa (Table 2).Tabbatarwa, ResNet- da AlexNet-kamar taswirar kunnawa na 8 \ (\ times \) 8 mm da 6 \ (\ times \) 6 mm hotuna sun sami damar haskaka tasoshin retinal a cikin lokuta biyu, suna nuna cewa samfurin yana da mahimman bayanai.sun dace don rarraba nau'ikan hotunan OCTA guda biyu (Fig. 4).
Lauerman et al.Hakanan an yi kimanta ingancin hoto akan hotunan OCTA ta amfani da gine-ginen Inception, wata hanyar tsallake-tsallake mai jujjuyawar jijiya6,32 ta amfani da dabarun koyo mai zurfi.Har ila yau, sun iyakance binciken zuwa hotuna na plexus na capillary capillary, amma kawai ta amfani da ƙananan hotuna 3 × 3 mm daga Optovue AngioVue, kodayake an haɗa marasa lafiya da cututtuka daban-daban na chorioretinal.Ayyukanmu yana ginawa akan tushen su, gami da ƙira da yawa don magance ƙofofin ingancin hoto daban-daban da kuma tabbatar da sakamako don hotuna masu girma dabam.Hakanan muna ba da rahoton awo na AUC na ƙirar injuna da haɓaka daidaiton su da suka rigaya (90%) 6 don ƙarancin inganci (96%) da ƙima mai inganci (95.7%) model6.
Wannan horon yana da iyakoki da yawa.Na farko, an samo hotunan tare da injin OCTA guda ɗaya kawai, gami da hotuna kawai na plexus na capillary capillary a 8 (\ times) 8 mm da 6 \ (\ times \) 6 mm.Dalilin keɓance hotuna daga yadudduka masu zurfi shine cewa kayan aikin tsinkaya na iya sa kimanta hotuna da hannu da wuya kuma mai yuwuwa ƙasa da daidaito.Bugu da ƙari kuma, an samo hotuna ne kawai a cikin marasa lafiya masu ciwon sukari, waɗanda OCTA ke fitowa a matsayin muhimmin kayan bincike da tsinkaye33,34.Ko da yake mun iya gwada samfurin mu akan hotuna masu girma dabam don tabbatar da cewa sakamakon ya kasance mai ƙarfi, ba mu iya gano bayanan da suka dace daga cibiyoyi daban-daban ba, wanda ya iyakance ƙididdigar mu game da ƙaddamar da samfurin.Ko da yake an samo hotunan ne daga cibiya guda ɗaya kawai, an samo su ne daga marasa lafiya na kabilanci da launin fata daban-daban, wanda shine ƙarfin musamman na bincikenmu.Ta haɗa da bambance-bambance a cikin tsarin horonmu, muna fatan cewa sakamakonmu zai zama gama gari a cikin ma'ana mai fa'ida, kuma za mu guji sanya bambancin launin fata a cikin ƙirar da muke horarwa.
Bincikenmu ya nuna cewa ana iya horar da hanyoyin sadarwa na tsallake-tsallake don cimma babban aiki wajen tantance ingancin hoton OCTA.Muna ba da waɗannan samfuran azaman kayan aiki don ƙarin bincike.Saboda ma'auni daban-daban na iya samun buƙatun ingancin hoto daban-daban, ana iya haɓaka ƙirar sarrafa ingancin mutum ɗaya don kowane awo ta amfani da tsarin da aka kafa anan.
Bincike na gaba yakamata ya haɗa da hotuna masu girma dabam daga zurfafa daban-daban da injunan OCTA daban-daban don samun tsarin kimanta ingancin hoto mai zurfi wanda za'a iya haɗa shi zuwa dandamali na OCTA da ka'idojin hoto.Binciken na yanzu kuma ya dogara ne akan hanyoyin ilmantarwa mai zurfi da ake kulawa waɗanda ke buƙatar kimanta ɗan adam da kimanta hoto, wanda zai iya zama babban aiki da ɗaukar lokaci don manyan bayanai.Ya rage a gani ko hanyoyin ilmantarwa mai zurfi ba tare da kulawa ba za su iya bambanta tsakanin ƙananan hotuna da hotuna masu inganci.
Yayin da fasahar OCTA ke ci gaba da haɓakawa kuma saurin dubawa yana ƙaruwa, abubuwan da ke faruwa na kayan tarihi da hotuna marasa inganci na iya raguwa.Haɓakawa a cikin software, kamar fasalin cire kayan tarihi da aka gabatar kwanan nan, kuma na iya rage waɗannan iyakoki.Koyaya, matsaloli da yawa suna kasancewa azaman hoton marasa lafiya tare da gyare-gyare mara kyau ko ƙaƙƙarfan ruɗani na kafofin watsa labarai koyaushe suna haifar da kayan tarihi.Yayin da OCTA ke ƙara yin amfani da shi a cikin gwaje-gwajen asibiti, ana buƙatar yin la'akari da kyau don kafa ƙayyadaddun ƙa'idodi don karɓuwar matakan kayan aikin hoto don nazarin hoto.Aiwatar da hanyoyin ilmantarwa mai zurfi zuwa hotunan OCTA suna ɗaukar babban alkawari kuma ana buƙatar ƙarin bincike a wannan yanki don haɓaka ingantaccen tsarin kula da ingancin hoto.
Lambar da aka yi amfani da ita a cikin binciken na yanzu yana samuwa a cikin ma'ajiyar octa-qc, https://github.com/rahuldhodapkar/octa-qc.Abubuwan da aka ƙirƙira da/ko tantancewa yayin binciken na yanzu suna samuwa daga mawallafa bisa ga buƙata mai ma'ana.
Spaide, RF, Fujimoto, JG & Waheed, NK Hotunan kayan tarihi a cikin angiography hadewar gani.Retina 35, 2163-2180 (2015).
Fenner, BJ et al.Gane fasalulluka na hoto waɗanda ke ƙayyadad da inganci da sake haifuwa na ma'aunin ma'aunin ma'aunin plexus na retinal capillary a cikin OCT angiography.BR.J. Ophthalmol.102, 509-514 (2018).
Lauerman, JL et al.Tasirin fasahar sa ido kan ingancin hoto na OCT angiography a cikin macular degeneration mai alaƙa da shekaru.Kabari baka.na asibiti.Exp.ophthalmology.255, 1535-1542 (2017).
Babyuch AS et al.Ana amfani da ma'auni mai yawa na perfusion na OCTA don ganowa da kimanta ischemia macular.tiyatar ido.Hoton Laser na Ganewa 51, S30–S36 (2020).
He, K., Zhang, X., Ren, S., da Sun, J. Zurfafa Rago Koyo don Gane Hoto.A cikin 2016 a taron IEEE akan hangen nesa na Kwamfuta da Gane Tsarin (2016).
Lauerman, JL et al.Ƙimar ingancin hoton angiographic OCT ta atomatik ta amfani da algorithms mai zurfi na koyo.Kabari baka.na asibiti.Exp.ophthalmology.257, 1641-1648 (2019).
Lauermann, J. et al.Yawaitar kurakuran rabuwa da kayan motsi a cikin OCT angiography ya dogara da cutar ta ido.Kabari baka.na asibiti.Exp.ophthalmology.256, 1807-1816 (2018).
Pask, Adam et al.Pytorch: Laburaren Koyo Mai Zurfi Mai Mahimmanci.Babban aiki na bayanan jijiya.tsarin.32, 8026-8037 (2019).
Daga, J. et al.ImageNet: Babban Ma'auni na Ma'auni na Hotuna.Taron IEEE na 2009 akan hangen nesa na Kwamfuta da Gane Tsarin.248-255.(2009).
Krizhevsky A., Sutzkever I. da Hinton GE Imagenet rarrabuwa ta amfani da zurfin juzu'i na jijiyoyi.Babban aiki na bayanan jijiya.tsarin.25, 1 (2012).