Mahalo no kou kipa ʻana iā Nature.com.Ke hoʻohana nei ʻoe i kahi polokalamu kele pūnaewele me ke kākoʻo CSS palena ʻole.No ka ʻike maikaʻi loa, manaʻo mākou e hoʻohana i kahi polokalamu kele pūnaewele hou (a i ʻole e hoʻopau i ke ʻano Compatibility Mode ma Internet Explorer).Eia kekahi, e hōʻoia i ke kākoʻo mau, hōʻike mākou i ka pūnaewele me ka ʻole o nā ʻano a me JavaScript.
ʻO nā slider e hōʻike ana i ʻekolu ʻatikala ma kēlā me kēia paheʻe.E hoʻohana i nā pihi hope a i hope no ka neʻe ʻana i nā paheʻe, a i ʻole nā ​​pihi hoʻokele paheʻe ma ka hopena e neʻe i kēlā me kēia paheʻe.
ʻO ka Optical coherence tomographic angiography (OCTA) kahi ala hou no ka nānā ʻole ʻana i nā moku retinal.ʻOiai he nui nā noi lapaʻau maikaʻi o OCTA, ʻo ka hoʻoholo ʻana i ka maikaʻi o ke kiʻi he paʻakikī.Ua hoʻomohala mākou i kahi ʻōnaehana hoʻonaʻauao hohonu e hoʻohana ana i ka ResNet152 neural network classifier i hoʻomaʻamaʻa mua ʻia me ImageNet e hoʻokaʻawale i nā kiʻi plexus capillary superficial mai 347 scans o 134 mau maʻi.Ua loiloi lima ʻia nā kiʻi ma ke ʻano he ʻoiaʻiʻo maoli e ʻelua mau loiloi kūʻokoʻa no kahi kumu hoʻonaʻauao mālama ʻia.No ka mea, ʻokoʻa paha nā koi o ka maikaʻi o ke kiʻi ma muli o nā hoʻonohonoho lapaʻau a noiʻi paha, ua aʻo ʻia ʻelua mau hiʻohiʻona, hoʻokahi no ka ʻike kiʻi kiʻekiʻe a ʻo kekahi no ka ʻike kiʻi haʻahaʻa.Hōʻike kā mākou ʻōnaehana neural i kahi wahi maikaʻi loa ma lalo o ka pihi (AUC), 95% CI 0.96-0.99, \(\kappa\) = 0.81), ʻoi aku ka maikaʻi ma mua o ka pae hōʻailona i hōʻike ʻia e ka mīkini (AUC = 0.82, 95). % CI).0.77–0.86, \(\kappa\) = 0.52 a me AUC = 0.78, 95% CI 0.73–0.83, \(\kappa\) = 0.27, pakahi).Hōʻike kā mākou haʻawina e hiki ke hoʻohana ʻia nā ʻano aʻo mīkini no ka hoʻomohala ʻana i nā ala hoʻokele maikaʻi a paʻa i nā kiʻi OCTA.
ʻO ka Optical coherence tomographic angiography (OCTA) kahi hana hou e pili ana i ka optical coherence tomography (OCT) hiki ke hoʻohana ʻia no ka ʻike ʻole ʻana o ka microvasculature retinal.ʻIke ʻo OCTA i ka ʻokoʻa o nā hiʻohiʻona noʻonoʻo mai nā puʻupuʻu māmā i ka ʻāpana like o ka retina, a laila hiki ke helu ʻia ke kūkulu hou ʻana e hōʻike i nā kīʻaha koko me ka ʻole o ka hoʻohana ʻana i nā mea pena a i ʻole nā ​​​​mea hoʻohālikelike ʻē aʻe.Hiki i ka OCTA ke hoʻonā i ka hohonu o nā kiʻi vascular, e ʻae ana i nā kauka e nānā kaʻawale i nā ʻāpana moku hohonu a me ka hohonu, e kōkua ana e hoʻokaʻawale i waena o ka maʻi chorioretinal.
ʻOiai ke hoʻohiki nei kēia ʻano hana, ʻo ka hoʻololi ʻana i ka maikaʻi o ke kiʻi he paʻakikī nui no ka nānā ʻana i nā kiʻi hilinaʻi, e paʻakikī ana i ka unuhi ʻana o nā kiʻi a me ka pale ʻana i ka lawe ʻia ʻana o ka lāʻau lapaʻau.No ka mea, hoʻohana ʻo OCTA i nā hiʻohiʻona OCT he nui, ʻoi aku ka maʻalahi o nā kiʻi kiʻi ma mua o ka OCT maʻamau.Hāʻawi ka hapa nui o nā kahua pāʻoihana OCTA i kā lākou ʻano kiʻi kiʻi ponoʻī i kapa ʻia ʻo Signal Strength (SS) a i kekahi manawa Signal Strength Index (SSI).Eia naʻe, ʻaʻole hōʻoiaʻiʻo nā kiʻi me ka waiwai SS kiʻekiʻe a i ʻole SSI i ka loaʻa ʻole o nā kiʻi kiʻi kiʻi, hiki ke hoʻopili i nā kiʻi kiʻi ma hope a alakaʻi i nā hoʻoholo lapaʻau hewa.ʻO nā kiʻi kiʻi maʻamau i hiki ke loaʻa i ka OCTA kiʻi kiʻi ʻoniʻoni, nā mea hana hoʻokaʻawale, nā mea hana opacity media, a me nā mea hana projection1,2,3.
Ke hoʻohana nui ʻia nei nā ana i loaʻa mai i ka OCTA e like me ka vascular density i ka noiʻi unuhi, nā hoʻokolohua lapaʻau a me nā hana lapaʻau, pono e hoʻomohala i nā kaʻina hana hoʻomalu maikaʻi a hilinaʻi no ka hoʻopau ʻana i nā kiʻi kiʻi4.ʻO nā hoʻokuʻi ʻana, ʻike ʻia hoʻi ʻo nā koena pili, he mau manaʻo i loko o ka hoʻolālā pūnaewele neural e hiki ai i ka ʻike ke kāpae i nā papa convolutional oiai e mālama ana i ka ʻike ma nā unahi like ʻole a i ʻole nā ​​hoʻonā5.No ka mea hiki i nā mea kiʻi kiʻi ke hoʻopilikia i ka hana kiʻi liʻiliʻi a me ka laulā nui, ua kūpono nā ʻupena neural hoʻokuʻi skip-connection e hoʻokaʻawale i kēia hana hoʻomalu maikaʻi5.Ua hōʻike ʻia nā hana i paʻi ʻia i kēia manawa i kekahi ʻōlelo hoʻohiki no nā ʻupena neural convolutional hohonu i hoʻomaʻamaʻa ʻia me ka hoʻohana ʻana i ka ʻikepili kiʻekiʻe mai nā mea koho kanaka6.
Ma kēia noiʻi ʻana, hoʻomaʻamaʻa mākou i kahi pūnaewele neural convolutional convolutional connection-skipping no ka hoʻoholo ʻana i ka maikaʻi o nā kiʻi OCTA.Kūkulu mākou ma luna o ka hana mua ma o ka hoʻomohala ʻana i nā hiʻohiʻona ʻokoʻa no ka ʻike ʻana i nā kiʻi kiʻekiʻe a me nā kiʻi haʻahaʻa haʻahaʻa, no ka mea, ʻokoʻa paha nā koi o ke kiʻi no nā hiʻohiʻona kikoʻī a noiʻi paha.Hoʻohālikelike mākou i nā hopena o kēia mau ʻupena me nā convolutional neural network me ka nalowale ʻole o nā pilina e loiloi i ka waiwai o ka hoʻopili ʻana i nā hiʻohiʻona ma nā pae he nui o ka granularity i loko o ke aʻo hohonu.A laila hoʻohālikelike mākou i kā mākou hopena i ka ikaika hōʻailona, ​​kahi ana i ʻae ʻia o ka maikaʻi kiʻi i hāʻawi ʻia e nā mea hana.
Hoʻokomo ʻia kā mākou haʻawina i nā poʻe maʻi me ka maʻi diabetes i hele i ka Yale Eye Center ma waena o ʻAukake 11, 2017 a me ʻApelila 11, 2019. Ua kāpae ʻia nā maʻi me nā maʻi chorioretinal non-diabetic.ʻAʻohe mea hoʻokomo a hoʻokuʻu ʻia paha e pili ana i ka makahiki, ke kāne, ka lāhui, ka maikaʻi o ke kiʻi, a i ʻole kekahi kumu ʻē aʻe.
Ua loaʻa nā kiʻi OCTA me ka hoʻohana ʻana i ka platform AngioPlex ma kahi Cirrus HD-OCT 5000 (Carl Zeiss Meditec Inc, Dublin, CA) ma lalo o 8\(\times\)8 mm a me 6\(\times\)6 mm imaging protocols.Ua loaʻa ka ʻae ʻike no ke komo ʻana i ke aʻo ʻana mai kēlā me kēia mea noiʻi, a ua ʻae ka Yale University Institutional Review Board (IRB) i ka hoʻohana ʻana i ka ʻae ʻike me ke kiʻi honua no kēia mau maʻi āpau.Ma muli o nā loina o ka Hōʻike o Helsinki.Ua ʻae ʻia ke aʻo ʻana e ka Yale University IRB.
Ua loiloi ʻia nā kiʻi papa honua ma muli o ka Motion Artifact Score (MAS) i hōʻike mua ʻia, ka Segmentation Artifact Score (SAS), ke kikowaena foveal, ke alo o ka media opacity, a me ka ʻike maikaʻi o nā capillaries liʻiliʻi e like me ka mea i hoʻoholo ʻia e ka mea loiloi kiʻi.Ua kālailai ʻia nā kiʻi e ʻelua mau loiloi kūʻokoʻa (RD a me JW).Loaʻa i kahi kiʻi ka helu helu o 2 (kūpono) inā e hoʻokō ʻia nā pae hoʻohālike a pau: ʻo ke kiʻi ke kikowaena ma ka fovea (emi ma lalo o 100 pixels mai ke kikowaena o ke kiʻi), ʻo MAS ka 1 a i ʻole 2, ʻo SAS ka 1, a ʻOi aku ka opacity media ma lalo o 1. Hōʻike ʻia ma nā kiʻi o ka nui / 16, a ʻike ʻia nā capillaries liʻiliʻi ma nā kiʻi ʻoi aku ka nui ma mua o 15/16.Ua helu ʻia ke kiʻi i ka 0 (ʻaʻohe helu) inā e hoʻokō ʻia kekahi o kēia mau pae hoʻohālike: ʻaʻole i waena ke kiʻi, inā he 4 ka MAS, inā he 2 ʻo SAS, a ʻoi aku ka opacity ma mua o 1/4 o ke kiʻi, a ʻAʻole hiki ke hoʻololi i nā capillaries liʻiliʻi ma mua o 1 kiʻi / 4 e ʻike ai.ʻO nā kiʻi ʻē aʻe ʻē aʻe i kūpono ʻole i nā koina helu 0 a i ʻole 2 ua helu ʻia ma ke ʻano he 1 (ʻokiʻoki).
Ma ka fig.Hōʻike ʻo 1 i nā laʻana kiʻi no kēlā me kēia o nā manaʻo hoʻohālikelike a me nā mea kiʻi kiʻi.Ua loiloi ʻia ka hilinaʻi waena o nā helu pākahi e kā Cohen kappa weighting8.Hoʻopili ʻia nā helu pākahi o kēlā me kēia helu helu no ka loaʻa ʻana o ka helu holoʻokoʻa no kēlā me kēia kiʻi, mai ka 0 a hiki i ka 4. Manaʻo ʻia nā kiʻi me ka huina o 4 he maikaʻi.ʻO nā kiʻi me ka helu piha o 0 a i ʻole 1 ke manaʻo ʻia he haʻahaʻa haʻahaʻa.
ʻO kahi ResNet152 architecture convolutional neural network (Fig. 3A.i) i hoʻomaʻamaʻa mua ʻia ma nā kiʻi mai ka waihona ImageNet i hana ʻia me ka hoʻohana ʻana i ka fast.ai a me ka PyTorch framework5, 9, 10, 11. ʻO kahi pūnaewele neural convolutional kahi pūnaewele e hoʻohana ana i ka mea aʻo. nā kānana no ka nānā ʻana i nā ʻāpana kiʻi e aʻo i nā hiʻohiʻona spatial a me nā hiʻohiʻona kūloko.ʻO kā mākou ResNet i hoʻomaʻamaʻa ʻia he 152-layer neural network i hōʻike ʻia e nā āpau a i ʻole "koena pili" e hoʻouna i ka ʻike me nā hoʻoholo he nui.Ma ka hoʻolālā ʻana i ka ʻike ma nā ʻōlelo hoʻoholo like ʻole ma luna o ka pūnaewele, hiki i ke kahua ke aʻo i nā hiʻohiʻona o nā kiʻi haʻahaʻa haʻahaʻa ma nā pae kikoʻī he nui.Ma waho aʻe o kā mākou kumu hoʻohālike ResNet, ua aʻo pū mākou iā AlexNet, kahi hoʻolālā pūnaewele neural i aʻo maikaʻi ʻia, me ka ʻole o nā pilina no ka hoʻohālikelike ʻana (Figure 3A.ii)12.Me ka nele i nā pilina, ʻaʻole hiki i kēia pūnaewele ke hopu i nā hiʻohiʻona ma kahi granularity kiʻekiʻe.
Ua hoʻonui ʻia ka hoʻonohonoho kiʻi kumu 8\(\times\)8mm OCTA13 me ka hoʻohana ʻana i nā ʻenehana noʻonoʻo ākea a kū pololei.Ua hoʻokaʻawale ʻia ka ʻikepili piha i ka pae kiʻi i ka hoʻomaʻamaʻa ʻana (51.2%), ka hoʻāʻo ʻana (12.8%), ka hoʻoponopono ʻana i ka hyperparameter (16%), a me ka hōʻoia ʻana (20%) me ka hoʻohana ʻana i ka pahu hana scikit-Learn python14.ʻElua mau hihia i noʻonoʻo ʻia, hoʻokahi ma muli o ka ʻike ʻana i nā kiʻi kiʻekiʻe loa (helu 4 holoʻokoʻa) a ʻo kekahi ma muli o ka ʻike ʻana i nā kiʻi haʻahaʻa haʻahaʻa wale nō (helu 0 a i ʻole 1 holoʻokoʻa).No kēlā me kēia hihia hoʻohana kiʻekiʻe a haʻahaʻa, hoʻomaʻamaʻa hou ʻia ka pūnaewele neural i hoʻokahi manawa ma kā mākou ʻikepili kiʻi.I kēlā me kēia hihia hoʻohana, ua aʻo ʻia ka ʻupena neural no 10 mau manawa, ua maloʻo nā mea a pau akā ʻo nā paona papa kiʻekiʻe loa, a ua aʻo ʻia nā paona o nā ʻāpana kūloko āpau no 40 mau manawa me ka hoʻohana ʻana i ke ʻano helu aʻo discriminative me kahi hana poho cross-entropy 15, 16..ʻO ka hana poho cross entropy he ana ia o ka unahi logarithmic o ka ʻokoʻa ma waena o nā lepili pūnaewele wānana a me ka ʻikepili maoli.I ka wā hoʻomaʻamaʻa, hana ʻia ka iho gradient ma nā ʻāpana kūloko o ka pūnaewele neural e hōʻemi i nā poho.Ua hoʻopaʻa ʻia ka nui o ke aʻo ʻana, ka haʻalele ʻana, a me nā hyperparameters hoʻēmi kaumaha me ka hoʻohana ʻana i ka Bayesian optimization me 2 mau helu hoʻomaka a me 10 iterations, a ua hoʻopaʻa ʻia ka AUC ma ka waihona me ka hoʻohana ʻana i nā hyperparameters ma ke ʻano he 17.
ʻO nā hiʻohiʻona o 8 × 8 mm OCTA kiʻi o nā plexus capillary superficial i helu 2 (A, B), 1 (C, D), a me 0 (E, F).ʻO nā mea kiʻi kiʻi i hōʻike ʻia e pili ana i nā laina ʻōlinolino (nā pua), nā mea hana hoʻokaʻawale (asterisks), a me ka opacity media (nā pua).Aia ke kiʻi (E) ma waho waena.
Hoʻokumu ʻia nā ʻano hana o ka mea hoʻokipa (ROC) no nā hiʻohiʻona neural network āpau, a hana ʻia nā hōʻike ikaika hōʻailona mīkini no kēlā me kēia hihia hoʻohana haʻahaʻa a kiʻekiʻe.Ua helu ʻia ka ʻāpana ma lalo o ka pihi (AUC) me ka hoʻohana ʻana i ka pūʻolo pROC R, a ua helu ʻia nā manawa hilinaʻi 95% a me nā waiwai p me ka hoʻohana ʻana i ke ala DeLong18,19.Hoʻohana ʻia nā helu kumulative o nā helu kanaka ma ke ʻano he kumu no nā helu ROC āpau.No ka ikaika hōʻailona i hōʻike ʻia e ka mīkini, ua helu ʻia ka AUC i ʻelua mau manawa: hoʻokahi no ka ʻoki ʻoki Scalability Score kiʻekiʻe a hoʻokahi hoʻokahi no ka ʻoki ʻana i ka Scalability Score haʻahaʻa.Hoʻohālikelike ʻia ka neural network me ka ikaika hōʻailona AUC e hōʻike ana i kāna mau kūlana hoʻomaʻamaʻa a loiloi.
No ka hoʻāʻo hou ʻana i ke ʻano hoʻonaʻauao hohonu i hoʻomaʻamaʻa ʻia ma kahi ʻikepili ʻokoʻa, ua hoʻopili pololei ʻia nā hiʻohiʻona kiʻekiʻe a me nā hiʻohiʻona haʻahaʻa i ka loiloi hana ʻana o 32 mau kiʻi piha 6\(\times\) 6mm i hōʻiliʻili ʻia mai ke Kulanui ʻo Yale.Hoʻokumu ʻia ka nui o ka maka i ka manawa like me ke kiʻi 8 \(\times \) 8 mm.Ua helu lima ʻia nā kiʻi 6\(\×\) 6 mm e nā mea helu like (RD a me JW) ma ke ʻano like me nā kiʻi 8\(\×\) 8 mm, ua helu ʻia ʻo AUC a me ka pololei a me ka kappa a Cohen. .like like .
ʻO ka lakio kaulike ʻole o ka papa he 158:189 (\(\rho = 1.19\)) no ke kumu hoʻohālike haʻahaʻa a me 80:267 (\(\rho = 3.3\)) no ke kumu hoʻohālike kiʻekiʻe.Ma muli o ka liʻiliʻi o ka ratio like ʻole o ka papa ma mua o 1: 4, ʻaʻole i hoʻololi ʻia ka hoʻololi ʻana i ke kūkulu hale no ka hoʻoponopono ʻana i ke ʻano like ʻole o ka papa20,21.
No ka ʻike maikaʻi ʻana i ke kaʻina aʻo, ua hana ʻia nā palapala ʻāina hoʻāla papa no nā kumu hoʻonaʻauao hohonu ʻehā i aʻo ʻia: kumu hoʻohālike ResNet152 kiʻekiʻe, kumu hoʻohālike haʻahaʻa ResNet152, kumu hoʻohālike kiʻekiʻe AlexNet, a me AlexNet kumu hoʻohālike haʻahaʻa.Hoʻokumu ʻia nā palapala ʻāina hoʻāla papa mai nā papa convolutional hoʻokomo o kēia mau hiʻohiʻona ʻehā, a ua hana ʻia nā palapala ʻāina wela ma ka uhi ʻana i nā palapala hoʻāla me nā kiʻi kumu mai ka 8 × 8 mm a me 6 × 6 mm hōʻoia hōʻoia22, 23.
Ua hoʻohana ʻia ka mana R 4.0.3 no nā helu helu helu a pau, a ua hana ʻia nā hiʻohiʻona me ka hoʻohana ʻana i ka waihona mea hana kiʻi ggplot2.
Ua hōʻiliʻili mākou i nā kiʻi mua 347 o ka plexus capillary superficial i ana 8 \(\times \)8 mm mai 134 kanaka.Ua hōʻike ka mīkini i ka ikaika hōʻailona ma kahi pālākiō o 0 a 10 no nā kiʻi āpau (mean = 6.99 ± 2.29).No nā kiʻi 347 i loaʻa, ʻo ka makahiki maʻamau i ka hoʻokolokolo ʻana he 58.7 ± 14.6 mau makahiki, a ʻo 39.2% mai nā maʻi kāne.No nā kiʻi a pau, 30.8% mai Caucasians, 32.6% mai ʻEleʻele, 30.8% mai Hispanics, 4% mai ʻAsia, a me 1.7% mai nā lāhui ʻē aʻe (Table 1).).ʻOkoʻa ka māhele makahiki o nā maʻi me OCTA ma muli o ka maikaʻi o ke kiʻi (p <0.001).ʻO ka pākēneka o nā kiʻi kiʻekiʻe kiʻekiʻe i nā poʻe maʻi ʻōpio o 18-45 mau makahiki he 33.8% i hoʻohālikelike ʻia me 12.2% o nā kiʻi haʻahaʻa haʻahaʻa (Table 1).Ua ʻano like ʻole ka puʻunaue ʻana o ke kūlana retinopathy diabetic i ka maikaʻi o ke kiʻi (p <0.017).Ma waena o nā kiʻi kiʻekiʻe kiʻekiʻe, ʻo ka pākēneka o nā maʻi me PDR he 18.8% i hoʻohālikelike ʻia me 38.8% o nā kiʻi haʻahaʻa haʻahaʻa (Table 1).
Ua hōʻike ʻia nā helu pākahi o nā kiʻi āpau i ka hilinaʻi ma waena o ka poʻe e heluhelu ana i nā kiʻi (Cohen's weighted kappa = 0.79, 95% CI: 0.76-0.82), a ʻaʻohe kiʻi kiʻi kahi i ʻokoʻa ai nā mea helu ma mua o 1 (Fig. 2A)..Hoʻopili nui ʻia ka ikaika o ka hōʻailona me ka helu manual (Pearson product moment correlation = 0.58, 95% CI 0.51-0.65, p<0.001), akā nui nā kiʻi i ʻike ʻia he kiʻekiʻe ka hōʻailona hōʻailona akā haʻahaʻa ka helu lima (Fig. 2B).
I ka wā o ka hoʻomaʻamaʻa ʻana o ka ResNet152 a me AlexNet architectures, hāʻule ka nalowale cross-entropy ma ka hōʻoia a me ke aʻo ʻana ma luna o 50 mau manawa (Figure 3B,C).ʻO ka pololei o ka hōʻoia ʻana i ka wā hoʻomaʻamaʻa hope ma luna o 90% no nā hihia hoʻohana kiʻekiʻe a haʻahaʻa.
Hōʻike nā ʻōkuhi hana hoʻokipa e ʻoi aku ka maikaʻi o ka hiʻohiʻona ResNet152 ma mua o ka mana hōʻailona i hōʻike ʻia e ka mīkini i nā hihia hoʻohana haʻahaʻa a kiʻekiʻe (p <0.001).ʻO ke kumu hoʻohālike ResNet152 ka mea nui ma mua o ka AlexNet architecture (p = 0.005 a me p = 0.014 no nā kūlana haʻahaʻa a me nā kūlana kiʻekiʻe, kēlā me kēia).ʻO nā hiʻohiʻona hopena no kēlā me kēia mau hana i hiki ke hoʻokō i nā waiwai AUC o 0.99 a me 0.97, ʻoi aku ka maikaʻi ma mua o nā koina AUC e pili ana i 0.82 a me 0.78 no ka mīkini hōʻailona ikaika index a i ʻole 0.97 a me 0.94 no AlexNet ..(Fig. 3).ʻOi aku ka ʻokoʻa ma waena o ResNet a me AUC i ka ikaika hōʻailona i ka wā e ʻike ai i nā kiʻi kiʻekiʻe, e hōʻike ana i nā pono hou o ka hoʻohana ʻana iā ResNet no kēia hana.
Hōʻike nā pakuhi i ka hiki i kēlā me kēia mea helu kūʻokoʻa ke helu a hoʻohālikelike me ka ikaika hōʻailona i hōʻike ʻia e ka mīkini.(A) Hoʻohana ʻia ka huina o nā helu e loiloi ʻia e hana i ka huina o nā helu e loiloi ʻia.Hāʻawi ʻia nā kiʻi me ka helu scalability holoʻokoʻa o 4 i ka maikaʻi kiʻekiʻe, aʻo nā kiʻi me ka helu scalability holoʻokoʻa o 1 a i ʻole ka liʻiliʻi e hāʻawi ʻia i ka maikaʻi haʻahaʻa.(B) Hoʻopili ʻia ka ikaika o ka hōʻailona me nā koho lima, akā ʻoi aku ka maikaʻi o nā kiʻi me ka ikaika o ka hōʻailona.Hōʻike ka laina kiko ʻulaʻula i ka paepae maikaʻi i ʻōlelo ʻia e ka mea hana ma muli o ka ikaika hōʻailona (ikaika hōʻailona \(\ge\)6).
Hāʻawi ke aʻo ʻana i ka ResNet transfer i kahi hoʻomaikaʻi koʻikoʻi i ka ʻike ʻana i ka maikaʻi o ke kiʻi no nā hihia hoʻohana haʻahaʻa a me nā kūlana kiʻekiʻe e hoʻohālikelike ʻia i nā pae hōʻailona hōʻike ʻia e ka mīkini.(A) Nā kiʻi hoʻolālā maʻalahi o ka hoʻomaʻamaʻa mua ʻia (i) ResNet152 a me (ii) AlexNet architecture.(B) Ka mōʻaukala hoʻomaʻamaʻa a me nā pihi hana hoʻokipa no ResNet152 i hoʻohālikelike ʻia me ka mīkini i hōʻike ʻia i ka ikaika hōʻailona a me nā pae hoʻohālike haʻahaʻa AlexNet.(C) ResNet152 ka mōʻaukala hoʻomaʻamaʻa a me nā pihi hana i hoʻohālikelike ʻia me ka mīkini i hōʻike ʻia i ka ikaika hōʻailona a me nā pae hoʻohālike kiʻekiʻe AlexNet.
Ma hope o ka hoʻoponopono ʻana i ka paepae palena hoʻoholo, ʻo 95.3% ka wānana ʻoi loa o ka wānana ResNet152 no ka hihia haʻahaʻa haʻahaʻa a me 93.5% no ka hihia kiʻekiʻe (Table 2).ʻO ka 91.0% ka wānana kiʻekiʻe loa o ka hiʻohiʻona AlexNet no ka hihia haʻahaʻa haʻahaʻa a me 90.1% no ka hihia kiʻekiʻe (Table 2).ʻO ka 76.1% ka hōʻailona hōʻailona wānana pololei no ka hihia hoʻohana haʻahaʻa a me 77.8% no ka hihia hoʻohana kiʻekiʻe.Wahi a Cohen's kappa (\(\kappa\)), o ka aelike ma waena o ka ResNet152 model a me na mea koho, he 0.90 no ka hihia ha'aha'a a me 0.81 no ka hihia ki'eki'e.ʻO ka AlexNet kappa ʻo Cohen he 0.82 a me 0.71 no nā hihia hoʻohana haʻahaʻa a me nā kūlana kiʻekiʻe.ʻO 0.52 a me 0.27 ka ikaika hōʻailona kappa o Cohen no nā hihia hoʻohana haʻahaʻa a kiʻekiʻe.
ʻO ka hōʻoia ʻana o nā hiʻohiʻona ʻike kiʻekiʻe a haʻahaʻa ma nā kiʻi 6\(\x\) o kahi pā pālahalaha 6 mm e hōʻike ana i ka hiki o ke kumu hoʻohālike i aʻo ʻia e hoʻoholo i ka maikaʻi o ke kiʻi ma nā ʻāpana kiʻi like ʻole.I ka hoʻohana ʻana i 6\(\x\) 6 mm mau papa pāpaʻu no ka maikaʻi o ke kiʻi, loaʻa ka AUC o 0.83 (95% CI: 0.69–0.98) o ke kumu hoʻohālike haʻahaʻa a he AUC o 0.85 ke kumu hoʻohālike kiʻekiʻe.(95% CI: 0.55–1.00) (Papa 2).
Ua hōʻike ʻia ka nānā ʻana i nā palapala hoʻāla papa papa hoʻokomo i ka hoʻohana ʻana o nā ʻupena neural i hoʻomaʻamaʻa ʻia i nā hiʻohiʻona kiʻi i ka wā hoʻohālikelike kiʻi (Fig. 4A, B).No nā kiʻi 8 \(\times \) 8 mm a me 6 \(\times \) 6 mm, pili pono nā kiʻi hoʻonā ResNet i ka retinal vasculature.Hoʻopili pū nā palapala hoʻonā AlexNet i nā moku retinal, akā me ka hoʻonā ʻoi loa.
Hōʻike nā palapala hoʻāla papa no nā hiʻohiʻona ResNet152 a me AlexNet i nā hiʻohiʻona e pili ana i ka maikaʻi o ke kiʻi.(A) Papa ho'āla palapala e hōʻike ana i ka coherent ho'ā 'ana ma hope superficial retinal vasculature ma 8 \(\times \) 8 mm hōʻoia kiʻi a me (B) laulā ma ka liʻiliʻi 6 \(\ manawa \) 6 mm hōʻoia kiʻi.Hoʻomaʻamaʻa ʻia ke kumu hoʻohālike LQ ma nā koina haʻahaʻa haʻahaʻa, ʻo ka hiʻohiʻona HQ i aʻo ʻia ma nā pae kiʻekiʻe.
Ua hōʻike mua ʻia e hiki i ka maikaʻi o ke kiʻi ke hoʻopilikia nui i nā helu o nā kiʻi OCTA.Eia kekahi,ʻo ka heleʻana o ka retinopathy e hoʻonui i ka nui o nā kiʻi kiʻi kiʻi7,26.ʻO kaʻoiaʻiʻo, i kā mākouʻikepili, e like me nā haʻawina mua, uaʻike mākou i kahi pilina nui ma waena o ka hoʻonuiʻana i ka makahiki a me ka paʻakikī o ka maʻi retinal a me ka emiʻana o ka maikaʻi o ke kiʻi (p <0.001, p = 0.017 no ka makahiki a me ke kūlana DR, kēlā me kēia; Papa 1) 27 .Hoʻohana ka hapa nui o nā haʻawina e kālailai ana i nā kiʻi OCTA i nā paepae hōʻailona hōʻike i hōʻike ʻia e ka mīkini e kāpae i nā kiʻi haʻahaʻa.ʻOiai ua hōʻike ʻia ka ikaika o ka hōʻailona e pili ana i ka helu ʻana o nā ʻāpana OCTA, ʻaʻole lawa ka ikaika o ka hōʻailona kiʻekiʻe e hoʻopau i nā kiʻi me nā kiʻi kiʻi2,3,28,29.No laila, pono e hoʻomohala i kahi ʻano hilinaʻi ʻoi aku ka maikaʻi o ke kiʻi.I kēia hopena, loiloi mākou i ka hana o nā ʻano aʻo hohonu i mālama ʻia e kūʻē i ka ikaika hōʻailona i hōʻike ʻia e ka mīkini.
Ua hoʻomohala mākou i kekahi mau hiʻohiʻona no ka loiloi ʻana i ka maikaʻi o ke kiʻi no ka mea he ʻokoʻa nā koi o ka maikaʻi o ke kiʻi o ka OCTA.No ka laʻana, pono nā kiʻi i ʻoi aku ka maikaʻi.Eia kekahi, koʻikoʻi nō hoʻi nā ʻāpana quantitative o ka hoihoi.No ka laʻana, ʻaʻole hilinaʻi ka ʻāpana o ka foveal avascular zone i ka turbidity o ke kikowaena non-central, akā pili i ka nui o nā moku.ʻOiai ke hoʻomau nei kā mākou noiʻi i ke ʻano maʻamau i ka maikaʻi o ke kiʻi, ʻaʻole i pili i nā koi o kekahi hoʻokolohua kūikawā, akā i manaʻo ʻia e hoʻololi pololei i ka ikaika hōʻailona i hōʻike ʻia e ka mīkini, ke manaʻo nei mākou e hāʻawi i nā mea hoʻohana i kahi kiʻekiʻe o ka mana i hiki ai iā lākou. hiki ke koho i ka metric kiko'ī o ka hoihoi i ka mea hoʻohana.e koho i ke kŘkohu e pili ana i ka pae ki'eki'e loa o ke ki'i ki'i i 'ae 'ia.
No nā hiʻohiʻona haʻahaʻa a kiʻekiʻe, hōʻike mākou i ka hana maikaʻi loa o nā ʻupena neural convolutional hohonu pili i ka pilina, me nā AUC o 0.97 a me 0.99 a me nā hiʻohiʻona haʻahaʻa.Hōʻike pū mākou i ka hana ʻoi loa o kā mākou ala hoʻonaʻauao hohonu ke hoʻohālikelike ʻia i nā pae hōʻailona i hōʻike ʻia e nā mīkini wale nō.ʻO ka hoʻokuʻi ʻana e hiki i nā pūnaewele neural ke aʻo i nā hiʻohiʻona ma nā pae kikoʻī he nui, e hopu ana i nā hiʻohiʻona ʻoi aku ka maikaʻi o nā kiʻi (e laʻa me ka ʻokoʻa) a me nā hiʻohiʻona maʻamau (e laʻa me ke kikowaena kiʻi30,31).No ka mea, ʻoi aku ka maikaʻi o nā kiʻi kiʻi e pili ana i ka maikaʻi o ke kiʻi ma kahi ākea, ʻoi aku ka maikaʻi o ka hana ʻana o nā neural network architecture me nā pilina nalo ma mua o ka poʻe me ka ʻole o nā hana hoʻoholo kiʻi.
I ka hoʻāʻo ʻana i kā mākou kŘkohu ma nā kiʻi OCTA 6\(\×6mm), ua ʻike mākou i ka emi ʻana o ka hana hoʻokaʻawale ʻana no nā hiʻohiʻona kiʻekiʻe a me nā hiʻohiʻona haʻahaʻa (Fig.Ke hoʻohālikelike ʻia i ka model ResNet, ʻoi aku ka nui o ka hāʻule ʻana o ka model AlexNet.ʻO ka hana ʻoi aku ka maikaʻi o ResNet ma muli paha o ka hiki ʻana o nā pili koena e hoʻouna i ka ʻike ma nā unahi he nui, ʻo ia ka mea e ʻoi aku ka ikaika o ke kumu hoʻohālike no ka hoʻokaʻawale ʻana i nā kiʻi i hopu ʻia ma nā unahi like ʻole a/a i ʻole nā ​​mea hoʻonui.
ʻO kekahi mau ʻokoʻa ma waena o nā kiʻi 8 \(\×\) 8 mm a me nā kiʻi 6 \(\×\) 6 mm hiki ke alakaʻi i ka hoʻokaʻawale maikaʻi ʻole ʻana, me ke kiʻekiʻe kiʻekiʻe o nā kiʻi i loaʻa nā wahi foveal avascular, nā loli i ka ʻike ʻia, nā arcade vascular, a ʻaʻohe aʻalolo optic ma ke kiʻi 6×6 mm.ʻOiai kēia, ua hiki i kā mākou kumu hoʻohālike kiʻekiʻe ResNet ke hoʻokō i kahi AUC o 85% no nā kiʻi 6 \(\x\) 6 mm, kahi hoʻonohonoho i hoʻomaʻamaʻa ʻole ʻia ke kumu hoʻohālike, e hōʻike ana e hoʻopili ʻia ka ʻike kiʻi kiʻi i loko o ka neural network. kūpono.no ka nui kiʻi hoʻokahi a i ʻole ka hoʻonohonoho mīkini ma waho o kāna aʻo ʻana (Papa 2).ʻO ka hōʻoluʻolu, ResNet- a me AlexNet-like activation palapala o 8 \(\times \) 8 mm a me 6 \(\times \) 6 mm kiʻi i hiki ke hoʻokolo i nā moku retinal i nā hihiaʻelua, e hōʻike ana heʻike koʻikoʻi ko ke kŘkohu.kūpono no ka hoʻokaʻawale ʻana i nā ʻano kiʻi ʻelua o OCTA (Fig. 4).
Lauerman et al.Ua like ka loiloi ʻana i ka maikaʻi o nā kiʻi ma nā kiʻi OCTA me ka hoʻohana ʻana i ka Inception architecture, kahi skip-connection convolutional neural network6,32 me ka hoʻohana ʻana i nā ʻenehana aʻo hohonu.Ua kaupalena lākou i ke aʻo ʻana i nā kiʻi o ka plexus capillary superficial, akā hoʻohana wale i nā kiʻi liʻiliʻi 3 × 3 mm mai Optovue AngioVue, ʻoiai ua hoʻokomo pū ʻia nā maʻi me nā maʻi chorioretinal.Kūkulu ʻia kā mākou hana ma luna o kā lākou mau kumu, me nā hiʻohiʻona he nui e hoʻoponopono i nā paepae ʻano kiʻi like ʻole a hōʻoia i nā hopena no nā kiʻi o nā ʻano nui like ʻole.Hōʻike pū mākou i ka AUC metric o nā mīkini aʻo ʻana a hoʻonui i ko lākou pololei kupaianaha (90%)6 no nā hiʻohiʻona haʻahaʻa (96%) a me nā hiʻohiʻona kiʻekiʻe (95.7%)6.
He mau palena kēia aʻo.ʻO ka mea mua, ua loaʻa nā kiʻi me hoʻokahi mīkini OCTA, me nā kiʻi wale nō o ka plexus capillary superficial ma 8\(\times\)8 mm a me 6\(\times\)6 mm.ʻO ke kumu o ka haʻalele ʻana i nā kiʻi mai nā papa hohonu, ʻo ia ka mea hiki i nā kiʻi kiʻi projection ke hana i ka loiloi manual o nā kiʻi i ʻoi aku ka paʻakikī a i ʻole ke kūlike.Eia kekahi, ua kiʻi wale ʻia nā kiʻi i nā maʻi maʻi diabetic, nona ka OCTA e puka mai nei ma ke ʻano he mea diagnostic nui a me ka prognostic tool33,34.ʻOiai ua hiki iā mākou ke hoʻāʻo i kā mākou hiʻohiʻona ma nā kiʻi o nā ʻano nui like ʻole e hōʻoia i ka paʻa o nā hopena, ʻaʻole hiki iā mākou ke ʻike i nā ʻikepili kūpono mai nā kikowaena like ʻole, i kaupalena ʻia kā mākou loiloi i ka laulā o ke kumu hoʻohālike.ʻOiai ua loaʻa nā kiʻi mai kahi kikowaena hoʻokahi, ua loaʻa iā lākou mai nā poʻe maʻi o nā ʻano lāhui like ʻole a me nā ʻano lāhui, kahi ikaika kūʻokoʻa o kā mākou noiʻi.Ma ka hoʻokomo ʻana i ka ʻokoʻa i loko o kā mākou kaʻina aʻo ʻana, manaʻolana mākou e hoʻonui ʻia kā mākou hopena ma ke ʻano ākea, a e pale aku mākou i ka hoʻopili ʻana i ka manaʻo hoʻohālikelike i nā kumu hoʻohālike a mākou e aʻo ai.
Hōʻike kā mākou haʻawina e hiki ke hoʻomaʻamaʻa ʻia nā ʻupena neural pili i ka hoʻokō ʻana i ka hana kiʻekiʻe i ka hoʻoholo ʻana i ka maikaʻi o ke kiʻi OCTA.Hāʻawi mākou i kēia mau hiʻohiʻona i mea hana no ka noiʻi hou aku.No ka mea he ʻokoʻa nā koi o ka maikaʻi o nā kiʻi like ʻole, hiki ke hoʻomohala ʻia kahi kumu hoʻohālike no kēlā me kēia metric me ka hoʻohana ʻana i ka hale i kūkulu ʻia ma aneʻi.
Pono nā noiʻi e hiki mai ana e hoʻokomo i nā kiʻi like ʻole mai nā hohonu like ʻole a me nā mīkini OCTA like ʻole no ka loaʻa ʻana o kahi kaʻina loiloi maikaʻi o ke kiʻi aʻo hohonu i hiki ke hoʻonui ʻia i nā paepae OCTA a me nā protocol kiʻi.Hoʻokumu pū ʻia ka noiʻi ʻana i kēia manawa ma nā ʻano hoʻonaʻauao hohonu i mālama ʻia e koi ana i ka loiloi kanaka a me ka loiloi kiʻi, hiki ke hana ikaika a hoʻopau i ka manawa no nā ʻikepili nui.E ʻike ʻia inā hiki i nā ʻano hoʻonaʻauao hohonu ʻole ke hoʻokaʻawale i waena o nā kiʻi haʻahaʻa a me nā kiʻi kiʻekiʻe.
Ke hoʻomau nei ka ʻenehana OCTA a piʻi ka wikiwiki o ka nānā ʻana, e emi ana ka nui o nā kiʻi kiʻi a me nā kiʻi maikaʻi ʻole.ʻO ka hoʻomaikaʻi ʻana i ka polokalamu, e like me ka hiʻohiʻona hoʻoneʻe projection artifact hou, hiki ke hoʻēmi i kēia mau palena.Eia nō naʻe, nui nā pilikia i koe e like me ke kiʻi ʻana i nā mea maʻi me ka hoʻoponopono maikaʻi ʻole a i ʻole ka nui o ka turbidity media e hopena mau i nā kiʻi kiʻi.Ke hoʻohana nui ʻia ka OCTA i nā hoʻokolohua lapaʻau, pono e noʻonoʻo pono e hoʻokumu i nā alakaʻi akaka no nā pae kiʻi kiʻi kūpono no ka nānā ʻana i nā kiʻi.ʻO ka hoʻohana ʻana i nā ʻano aʻo hohonu i nā kiʻi OCTA e paʻa i ka ʻōlelo hoʻohiki nui a pono ka noiʻi hou ʻana ma kēia wahi e hoʻomohala i kahi ala paʻa i ka hoʻomalu ʻana i ke kiʻi.
Loaʻa ka code i hoʻohana ʻia i ka noiʻi o kēia manawa ma ka waihona octa-qc, https://github.com/rahuldhodapkar/octa-qc.Loaʻa nā waihona ʻikepili i hana ʻia a/a i ʻike ʻia i ka wā o ke aʻo ʻana i kēia manawa mai nā mea kākau ma muli o ke noi kūpono.
Spaide, RF, Fujimoto, JG & Waheed, NK Nā mea kiʻi kiʻi i ka angiography coherence.Retina 35, 2163–2180 (2015).
Fenner, BJ et al.ʻO ka ʻike ʻana i nā hiʻohiʻona kiʻi e hoʻoholo ai i ka maikaʻi a me ka reproducibility o ke ana ʻana o ka retinal capillary plexus density ma OCT angiography.BR.J. Ophthalmol.102, 509–514 (2018).
Lauerman, JL et al.Ka hopena o ka ʻenehana nānā maka i ka maikaʻi o ke kiʻi o OCT angiography i ka macular degeneration pili makahiki.Poku lua.lāʻau lapaʻau.ʻO Exp.ophthalmology.255, 1535–1542 (2017).
Babyuch AS et al.Hoʻohana ʻia nā ana hoʻopaʻapaʻa capillary perfusion OCTA e ʻike a loiloi i ka macular ischemia.ʻoki ophthalmic.Kiʻi Laser Retinal 51, S30–S36 (2020).
ʻO ia, K., Zhang, X., Ren, S., a me Sun, J. Ke aʻo ʻana o ke koena hohonu no ka ʻike kiʻi.Ma 2016 ma ka IEEE Conference on Computer Vision and Pattern Recognition (2016).
Lauerman, JL et al.ʻO ka loiloi kiʻi kiʻi angiographic OCT me ka hoʻohana ʻana i nā algorithm aʻo hohonu.Poku lua.lāʻau lapaʻau.ʻO Exp.ophthalmology.257, 1641–1648 (2019).
Lauermann, J. et al.ʻO ka nui o nā hewa o ka māhele a me nā mea hana i ka OCT angiography e pili ana i ka maʻi o ka retina.Poku lua.lāʻau lapaʻau.ʻO Exp.ophthalmology.256, 1807–1816 (2018).
Pask, Adamu et al.Pytorch: He Hale Waihona Puke Aʻo Hohonu Pono, Kiʻekiʻe.ʻO ka hana kiʻekiʻe o ka ʻike neural.ʻōnaehana.32, 8026–8037 (2019).
Deng, J. et al.ImageNet: He waihona kiʻi hierarchical nui.2009 IEEE Conference on Computer Vision and Pattern Recognition.248–255.(2009).
Krizhevsky A., Sutzkever I. a me Hinton GE Imagenet hoʻohālikelike me ka hoʻohana ʻana i nā pūnaewele neural convolutional hohonu.ʻO ka hana kiʻekiʻe o ka ʻike neural.ʻōnaehana.25, 1 (2012).