Zikomo pochezera Nature.com.Mukugwiritsa ntchito msakatuli wokhala ndi chithandizo chochepa cha CSS.Kuti mudziwe zambiri, tikupangira kuti mugwiritse ntchito msakatuli wosinthidwa (kapena kuletsa Compatibility Mode mu Internet Explorer).Kuphatikiza apo, kuti tithandizire kupitilizabe, tikuwonetsa tsambalo popanda masitayilo ndi JavaScript.
Masilayidi owonetsa zolemba zitatu pa slide iliyonse.Gwiritsani ntchito mabatani akumbuyo ndi ena kuti mudutse zithunzi, kapena mabatani owongolera masilayidi kumapeto kuti mudutse silayidi iliyonse.
Optical coherence tomographic angiography (OCTA) ndi njira yatsopano yowonera ziwiya za retina.Ngakhale OCTA ili ndi mapulogalamu ambiri azachipatala omwe amalonjeza, kudziwa mtundu wazithunzi kumakhalabe kovuta.Tinapanga njira yophunzirira mwakuya pogwiritsa ntchito ResNet152 neural network classifier yophunzitsidwa bwino ndi ImageNet kuyika m'magulu azithunzi zapamwamba za capillary plexus kuchokera ku ma scan 347 a odwala 134.Zithunzizo zidawunikidwanso pamanja ngati chowonadi ndi anthu awiri odziyimira pawokha panjira yoyang'aniridwa yophunzirira.Chifukwa chakuti zofunikira zamtundu wa zithunzi zimatha kusiyana malingana ndi zochitika zachipatala kapena kafukufuku, zitsanzo ziwiri zinaphunzitsidwa, imodzi yozindikiritsa zithunzi zamtundu wapamwamba ndipo ina yodziwika bwino ya zithunzi.Mtundu wathu wa neural network ukuwonetsa malo abwino kwambiri pansi pa poto (AUC), 95% CI 0.96-0.99, \(\kappa\) = 0.81), yomwe ili bwino kwambiri kuposa kuchuluka kwa siginecha komwe makinawo amanenedwa (AUC = 0.82, 95) % CI).0.77–0.86, \(\kappa\) = 0.52 ndi AUC = 0.78, 95% CI 0.73–0.83, \(\kappa\) = 0.27, motsatana).Kafukufuku wathu akuwonetsa kuti njira zophunzirira makina zitha kugwiritsidwa ntchito kupanga njira zosinthika komanso zolimba zowongolera zazithunzi za OCTA.
Optical coherence tomographic angiography (OCTA) ndi njira yatsopano yozikidwa pa optical coherence tomography (OCT) yomwe ingagwiritsidwe ntchito poyang'ana mosavutikira kwa retinal microvasculature.OCTA imayesa kusiyana kwa mawonekedwe owunikira kuchokera kumagetsi obwerezabwereza m'dera lomwelo la retina, ndipo kukonzanso kumatha kuwerengedwa kuti kuwulula mitsempha yamagazi popanda kugwiritsa ntchito utoto movutikira kapena zinthu zina zosiyanitsa.OCTA imathandizanso kulingalira kwakuya kwa mitsempha, kulola madokotala kuti ayang'ane padera ndi zigawo zakuya zotengera, zomwe zimathandiza kusiyanitsa pakati pa matenda a chorioretinal.
Ngakhale njira iyi ikulonjeza, kusiyanasiyana kwazithunzi kumakhalabe vuto lalikulu pakusanthula kodalirika kwazithunzi, kupangitsa kutanthauzira kwazithunzi kukhala kovuta ndikuletsa kufalikira kwachipatala.Chifukwa OCTA imagwiritsa ntchito masikanidwe angapo otsatizana a OCT, imakhudzidwa kwambiri ndi zinthu zakale kuposa OCT wamba.Mapulatifomu ambiri a OCTA amalonda amapereka mawonekedwe awo azithunzi omwe amatchedwa Signal Strength (SS) kapena nthawi zina Signal Strength Index (SSI).Komabe, zithunzi zokhala ndi mtengo wapamwamba wa SS kapena SSI sizikutsimikizira kusakhalapo kwa zithunzi zakale, zomwe zingakhudze kusanthula kulikonse kotsatira ndikupangitsa zisankho zolakwika zachipatala.Zithunzi zodziwika bwino zomwe zitha kuchitika muzojambula za OCTA zimaphatikizanso zinthu zakale zoyenda, zida zamagawo, zida zapa media opacity, ndi zinthu zakale zowonetsera1,2,3.
Pamene miyeso yochokera ku OCTA monga kuchulukitsidwa kwa mitsempha ikugwiritsidwa ntchito mowonjezereka mu kafukufuku womasulira, mayesero a zachipatala ndi zochitika zachipatala, pakufunika kofunika kwambiri kuti pakhale njira zoyendetsera khalidwe lachithunzithunzi zolimba komanso zodalirika kuti athetse zojambula za zithunzi4.Malumikizidwe odumphira, omwe amadziwikanso kuti maulalo otsalira, ndizomwe zimapangidwira mu neural network zomanga zomwe zimalola kuti chidziwitso chidutse zigawo zosinthika ndikusunga zambiri pamiyeso kapena malingaliro osiyanasiyana5.Chifukwa zojambula zazithunzi zimatha kukhudza magwiridwe antchito ang'onoang'ono komanso akulu, ma skip-connection neural network ali oyenerera kuti azitha kuyendetsa bwino ntchitoyi5.Ntchito yomwe yasindikizidwa posachedwa yawonetsa lonjezo la maukonde ozama a neural ophunzitsidwa kugwiritsa ntchito chidziwitso chapamwamba kwambiri kuchokera kwa oyerekeza anthu6.
M'kafukufukuyu, timaphunzitsa ma netiweki a convolutional neural network kuti adziŵe okha mtundu wa zithunzi za OCTA.Timamanga pa ntchito ya m'mbuyomu popanga zitsanzo zosiyana zodziwira zithunzi zapamwamba kwambiri ndi zithunzi zotsika, chifukwa zofunikira zazithunzi zingakhale zosiyana pazochitika zachipatala kapena kafukufuku.Timayerekeza zotsatira zama netiwekiwa ndi ma convolutional neural network popanda kulumikizana kuti tiwone kufunikira kophatikiza zinthu pamilingo yambiri ya granularity mkati mwa kuphunzira mozama.Kenako tidafanizira zotsatira zathu ndi mphamvu ya siginecha, muyeso wovomerezeka wodziwika bwino wazithunzi zoperekedwa ndi opanga.
Kafukufuku wathu adaphatikizapo odwala matenda a shuga omwe adapita ku Yale Eye Center pakati pa Ogasiti 11, 2017 ndi Epulo 11, 2019. Odwala omwe ali ndi matenda aliwonse omwe alibe matenda a shuga a chorioretinal sanaphatikizidwe.Panalibe njira zophatikizirapo kapena zopatula kutengera zaka, jenda, mtundu, mtundu wazithunzi, kapena china chilichonse.
Zithunzi za OCTA zidapezedwa pogwiritsa ntchito nsanja ya AngioPlex pa Cirrus HD-OCT 5000 (Carl Zeiss Meditec Inc, Dublin, CA) pansi pa 8\(\times\)8 mm ndi 6\(\times\)6mm maprotocol.Chilolezo chodziwitsidwa chotenga nawo gawo mu kafukufukuyu chinapezedwa kwa aliyense wochita nawo kafukufukuyu, ndipo Yale University Institutional Review Board (IRB) idavomereza kugwiritsa ntchito chilolezo chodziwitsidwa ndi kujambula padziko lonse lapansi kwa odwala onsewa.Kutsatira mfundo za Declaration of Helsinki.Kafukufukuyu adavomerezedwa ndi Yale University IRB.
Zithunzi za mbale zapamwamba zidawunikidwa kutengera zomwe zafotokozedwa kale za Motion Artifact Score (MAS), zomwe zafotokozedwa kale za Segmentation Artifact Score (SAS), foveal center, kupezeka kwa media opacity, ndikuwona bwino kwa ma capillary ang'onoang'ono monga momwe wowunika wazithunzi adatsimikiza.Zithunzizi zidawunikidwa ndi owunika awiri odziyimira pawokha (RD ndi JW).Chithunzi chili ndi zigoli 2 (zoyenera) ngati zonse zotsatirazi zakwaniritsidwa: chithunzi chili pakati pa fovea (mapikseli ochepera 100 kuchokera pakati pa chithunzi), MAS ndi 1 kapena 2, SAS ndi 1, ndipo Media opacity ndi yocheperapo 1. Ikupezeka pazithunzi za kukula / 16, ndipo ma capillaries ang'onoang'ono amawoneka pazithunzi zazikulu kuposa 15/16.Chithunzichi chidavoteredwa 0 (palibe mavoti) ngati chimodzi mwazotsatirazi chikukwaniritsidwa: chithunzicho sichili pakati, ngati MAS ndi 4, ngati SAS ndi 2, kapena kuwala kwapakati ndi kwakukulu kuposa 1/4 ya chithunzicho, ndipo ma capillaries ang'onoang'ono sangasinthidwe kuposa 1 chithunzi / 4 kuti asiyanitse.Zithunzi zina zonse zomwe sizikukwaniritsa zigoli 0 kapena 2 zigoledwa ngati 1 (kudula).
Pa mkuyu.1 ikuwonetsa zithunzi zachitsanzo pazoyerekeza zowerengeka ndi zithunzi zakale.Kudalirika kwapakati pamagulu amunthu payekha kudawunikidwa ndi Cohen's kappa weighting8.Ziwerengero zapayekha za ovotera aliyense zimafupikitsidwa kuti mupeze chiwongolero chonse pa chithunzi chilichonse, kuyambira 0 mpaka 4. Zithunzi zokhala ndi 4 ziwerengero zimawonedwa ngati zabwino.Zithunzi zomwe zili ndi chiwerengero chonse cha 0 kapena 1 zimatengedwa ngati zabwino kwambiri.
ResNet152 architecture convolutional neural network (Fig. 3A.i) yophunzitsidwa kale zithunzi kuchokera ku database ya ImageNet idapangidwa pogwiritsa ntchito fast.ai ndi PyTorch framework5, 9, 10, 11. Convolutional neural network ndi netiweki yomwe imagwiritsa ntchito ophunzira. zosefera zosefera zidutswa za zithunzi kuti muphunzire za malo ndi zapafupi.ResNet yathu yophunzitsidwa bwino ndi 152-layer neural network yodziwika ndi mipata kapena "malumikizidwe otsalira" omwe amatumiza nthawi imodzi chidziwitso chokhala ndi malingaliro angapo.Popanga zidziwitso pazosankha zosiyanasiyana pamaneti, nsanja imatha kuphunzira mawonekedwe azithunzi zotsika pamagawo angapo atsatanetsatane.Kuwonjezera pa chitsanzo chathu cha ResNet, tinaphunzitsanso AlexNet, yophunzira bwino neural network yomangamanga, popanda kusowa kugwirizana poyerekeza (Chithunzi 3A.ii)12.Popanda malumikizidwe osowa, netiweki iyi siitha kujambula mawonekedwe awo mwachiwombankhanga kwambiri.
Chithunzi choyambirira cha 8\(\times\)8mm OCTA13 chakonzedwanso pogwiritsa ntchito njira zowonetsera zopingasa komanso zoyima.Deta yonseyo idagawika mwachisawawa pamlingo wazithunzi kukhala maphunziro (51.2%), kuyesa (12.8%), hyperparameter tuning (16%), ndi kutsimikizira (20%) ma dataset pogwiritsa ntchito scikit-learn toolbox python14.Milandu iwiri idaganiziridwa, imodzi yozikidwa pakupeza zithunzi zapamwamba kwambiri (zonse 4) ndipo ina potengera zithunzi zotsika kwambiri (zonse 0 kapena 1).Pazochitika zilizonse zapamwamba komanso zotsika, neural network imaphunzitsidwanso kamodzi pazithunzi zathu.Munthawi iliyonse yogwiritsiridwa ntchito, neural network idaphunzitsidwa ma epoch 10, zonse kupatula zolemera kwambiri zosanjikiza zidazizira, ndipo zolemera zamitundu yonse yamkati zidaphunziridwa kwa nthawi 40 pogwiritsa ntchito njira yosankha yophunzirira yokhala ndi ntchito yotayika yodutsa 15, 16..Cross entropy loss function ndi muyeso wa sikelo ya logarithmic ya kusiyana pakati pa zilembo zonenedweratu za netiweki ndi deta yeniyeni.Pakuphunzitsidwa, kutsika kwa gradient kumachitika pazigawo zamkati za neural network kuti muchepetse kutayika.Kuchuluka kwa maphunziro, kutsika kwa maphunziro, ndi ma hyperparameters ochepetsera kulemera kunakonzedwa pogwiritsa ntchito kukhathamiritsa kwa Bayesian ndi 2 poyambira mwachisawawa ndi kubwereza 10, ndipo AUC pa dataset idakonzedwa pogwiritsa ntchito ma hyperparameters monga chandamale cha 17.
Zitsanzo zoyimira za 8 × 8 mm OCTA zithunzi za plexuses zapamwamba za capillary zagoletsa 2 (A, B), 1 (C, D), ndi 0 (E, F).Zithunzi zomwe zawonetsedwa zikuphatikiza mizere yokulirakulira (mivi), zotsalira za magawo (asterisk), ndi mawonekedwe a media (mivi).Chithunzi (E) chilinso chapakati.
Mawonekedwe a Receiver Opaleshoni (ROC) ndiye amapangidwa pamawonekedwe onse a neural network, ndipo malipoti a mphamvu zama injini amapangidwa pachomera chilichonse chotsika komanso chapamwamba kwambiri.Dera lomwe lili pansi pa curve (AUC) linawerengedwa pogwiritsa ntchito phukusi la pROC R, ndipo 95% nthawi zodalirika ndi ma p-values ​​adawerengedwa pogwiritsa ntchito njira ya DeLong18,19.Ziwerengero zowerengera za anthu zimagwiritsidwa ntchito ngati maziko pazowerengera zonse za ROC.Pa mphamvu ya siginecha yomwe idanenedwa ndi makinawo, AUC idawerengeredwa kawiri: kamodzi pamtundu wapamwamba kwambiri wa Scalability Score cutoff ndipo kamodzinso pakutsika kwapamwamba kwa Scalability Score.Neural network imafanizidwa ndi mphamvu ya siginecha ya AUC yowonetsa momwe amaphunzitsira ndikuwunika.
Kuti ayesenso chitsanzo chakuya chophunzitsidwa pa dataset yapadera, zitsanzo zamtundu wapamwamba komanso zotsika kwambiri zinagwiritsidwa ntchito mwachindunji pakuwunika ntchito kwa 32 nkhope yonse 6 \ (\ nthawi \) zithunzi za slab za 6mm zotengedwa kuchokera ku yunivesite ya Yale.Misa ya Diso imakhazikika nthawi imodzi ndi chithunzi 8 \ (\ nthawi \) 8 mm.Zithunzi za 6\(\×\) 6 mm zinayesedwa pamanja ndi owerengera omwewo (RD ndi JW) mofanana ndi zithunzi za 8\(\×\) 8 mm, AUC inawerengedwa komanso kulondola ndi kappa ya Cohen. .mofanana .
Chiŵerengero cha kusamvana kwa kalasi ndi 158: 189 (\(\ rho = 1.19\)) kwa chitsanzo chochepa kwambiri ndi 80: 267 (\(\ rho = 3.3\)) pa chitsanzo chapamwamba.Chifukwa chiŵerengero cha kusamvana kwa kalasi ndi chochepera 1: 4, palibe zosintha zinazake zomwe zapangidwa kuti ziwongolere kusalinganika kwamagulu20,21.
Kuti muwone bwino njira yophunzirira, mamapu otsegulira kalasi adapangidwa pamitundu yonse inayi yophunzitsidwa mozama: mtundu wapamwamba kwambiri wa ResNet152, mtundu wa ResNet152 wotsika kwambiri, mtundu wapamwamba kwambiri wa AlexNet, ndi mtundu wa AlexNet wotsika.Mamapu oyambitsa kalasi amapangidwa kuchokera pazowonjezera zosinthira zamitundu inayi, ndipo mamapu otentha amapangidwa ndikukuta mamapu otsegulira okhala ndi zithunzi zochokera ku 8 × 8 mm ndi 6 × 6 mm zotsimikizira sets22, 23.
R mtundu 4.0.3 unagwiritsidwa ntchito powerengera zonse, ndipo zowonera zidapangidwa pogwiritsa ntchito laibulale ya zida za ggplot2.
Tidasonkhanitsa zithunzi za 347 zakutsogolo za plexus ya capillary yapamwamba kwambiri 8 \ (\ nthawi \) 8 mm kuchokera kwa anthu 134.Makinawa adafotokoza mphamvu yazizindikiro pamlingo wa 0 mpaka 10 pazithunzi zonse (kutanthauza = 6.99 ± 2.29).Pazithunzi za 347 zomwe zidapezedwa, zaka zowerengeka pakuwunika zinali zaka 58.7 ± 14.6, ndipo 39.2% zidachokera kwa odwala amuna.Pazithunzi zonse, 30.8% inachokera ku Caucasus, 32.6% ya Akuda, 30.8% ya Hispanics, 4% ya Asiya, ndi 1.7% ya mafuko ena (Table 1).).Kugawa kwa zaka za odwala omwe ali ndi OCTA kunali kosiyana kwambiri malinga ndi mtundu wa chithunzicho (p <0.001).Chiwerengero cha zithunzi zamtengo wapatali kwa odwala aang'ono a zaka za 18-45 zinali 33.8% poyerekeza ndi 12.2% ya zithunzi zotsika (Table 1).Kugawidwa kwa matenda a shuga a retinopathy kunasiyananso kwambiri pazithunzi (p <0.017).Pakati pa zithunzi zonse zapamwamba, chiwerengero cha odwala omwe ali ndi PDR chinali 18.8% poyerekeza ndi 38.8% ya zithunzi zotsika kwambiri (Table 1).
Ziwerengero zapayekha pazithunzi zonse zidawonetsa kudalirika kwapakati mpaka kolimba pakati pa anthu omwe amawerenga zithunzizo (Cohen's weighted kappa = 0.79, 95% CI: 0.76-0.82), ndipo panalibe zithunzi zomwe ovotera amasiyana kuposa 1 (Mkuyu. 2A)..Kukula kwa ma signature kumagwirizana kwambiri ndi zolemba zamanja (Pearson product moment correlation = 0.58, 95% CI 0.51-0.65, p<0.001), koma zithunzi zambiri zidadziwika kuti zili ndi mphamvu yayikulu koma yotsika pamanja (mkuyu .2B).
Pakuphunzitsidwa kwa zomangamanga za ResNet152 ndi AlexNet, kutayika kwa cross-entropy pa kutsimikizika ndi maphunziro kumagwera pa 50 epochs (Chithunzi 3B, C).Kulondola kotsimikizirika mu nthawi yomaliza yophunzitsira ndi yoposa 90% pazochitika zonse zapamwamba komanso zotsika.
Ma curve performance curves akuwonetsa kuti ResNet152 model imaposa mphamvu yazizindikiro yomwe imanenedwa ndi makina pamachitidwe otsika komanso apamwamba kwambiri (p <0.001).Mtundu wa ResNet152 umachitanso bwino kwambiri kuposa zomangamanga za AlexNet (p = 0.005 ndi p = 0.014 pamilandu yotsika komanso yapamwamba, motsatana).The zitsanzo chifukwa aliyense wa ntchito zimenezi anatha kukwaniritsa AUC mfundo za 0,99 ndi 0,97 motero, amene ali bwino kwambiri kuposa lolingana AUC mfundo za 0,82 ndi 0,78 kwa makina chizindikiro mphamvu index kapena 0,97 ndi 0,94 kwa AlexNet ..(mkuyu 3).Kusiyana pakati pa ResNet ndi AUC mu mphamvu ya siginecha ndikokwera kwambiri pozindikira zithunzi zapamwamba, kuwonetsa maubwino owonjezera ogwiritsira ntchito ResNet pa ntchitoyi.
Ma grafu amawonetsa kuthekera kwa aliyense wodziyimira pawokha kugoletsa ndikuyerekeza ndi mphamvu ya siginecha yomwe imanenedwa ndi makina.(A) Chiwerengero cha mfundo zomwe ziyenera kuyesedwa zimagwiritsidwa ntchito kupanga chiwerengero cha mfundo zomwe ziyenera kuyesedwa.Zithunzi zokhala ndi ma scalability a 4 zimaperekedwa zamtundu wapamwamba, pomwe zithunzi zokhala ndi 1 kapena kuchepera zimapatsidwa mawonekedwe otsika.(B) Kulimba kwa siginecha kumagwirizana ndi kuyerekezera kwapamanja, koma zithunzi zokhala ndi mphamvu yazizindikiro zitha kukhala zotsika kwambiri.Mzere wamadontho ofiyira ukuwonetsa momwe wopanga amapangira kutengera mphamvu ya siginecha (mphamvu ya siginecha \(\ge\)6).
Kuphunzira kusamutsa kwa ResNet kumapereka kusintha kwakukulu pakuzindikiritsa zamtundu wazithunzi pamachitidwe otsika komanso ogwiritsira ntchito apamwamba poyerekeza ndi milingo yamasigino onenedwa ndi makina.(A) Zojambula zophweka za zomangamanga zophunzitsidwa kale (i) ResNet152 ndi (ii) AlexNet architectures.(B) Mbiri yophunzitsa ndi ma curves olandila olandila a ResNet152 poyerekeza ndi mphamvu zamakina zomwe zidanenedwa ndi njira zotsika za AlexNet.(C) Mbiri ya ResNet152 yophunzitsira olandila ndi ma curve ogwirira ntchito poyerekeza ndi mphamvu zamakina zomwe zidanenedwa komanso njira zapamwamba za AlexNet.
Pambuyo pokonza malire a chigamulo, kulondola kwakukulu kwa chitsanzo cha ResNet152 ndi 95.3% pamtundu wochepa kwambiri ndi 93.5% pazochitika zapamwamba (Table 2).Kulondola kwakukulu kwachidziwitso cha chitsanzo cha AlexNet ndi 91.0% pamtundu wochepa kwambiri ndi 90.1% pamtundu wapamwamba (Table 2).Kulondola kwamphamvu kwamphamvu kwazizindikiro ndi 76.1% pakugwiritsa ntchito kwapamwamba kwambiri ndi 77.8% pamagwiritsidwe apamwamba kwambiri.Malinga ndi Cohen's kappa (\(\kappa\)), mgwirizano pakati pa mtundu wa ResNet152 ndi oyerekeza ndi 0.90 pamilandu yotsika kwambiri ndi 0.81 pamilandu yapamwamba kwambiri.Cohen's AlexNet kappa ndi 0.82 ndi 0.71 pamachitidwe otsika komanso apamwamba kwambiri, motsatana.Chizindikiro champhamvu cha Cohen kappa ndi 0.52 ndi 0.27 pamagwiritsidwe otsika komanso apamwamba kwambiri, motsatana.
Kutsimikizira kwa zitsanzo zozindikiritsa zapamwamba ndi zotsika pazithunzi za 6\(\x\) za mbale yafulati ya 6 mm zimasonyeza kuthekera kwa chitsanzo chophunzitsidwa kudziŵa khalidwe lachithunzi pazithunzi zosiyanasiyana.Mukamagwiritsa ntchito 6\(\x\) 6 mm ma slabs osaya pamtundu wojambula, mtundu wocheperako unali ndi AUC ya 0.83 (95% CI: 0.69-0.98) ndipo mtundu wapamwamba kwambiri unali ndi AUC ya 0.85.(95% CI: 0.55-1.00) (Table 2).
Kuyang'ana kowonekera kwa mamapu otsegulira kalasi yolowetsa kunawonetsa kuti ma neural network onse ophunzitsidwa bwino amagwiritsa ntchito mawonekedwe azithunzi panthawi yamagulu azithunzi (mkuyu 4A, B).Kwa zithunzi za 8 \(\ times \) 8 mm ndi 6 \ (\times \) 6 mm zithunzi, zithunzi zoyambitsa ResNet zimatsata kwambiri mitsempha ya retinal.Mamapu otsegulira a AlexNet amatsatiranso ziwiya za retina, koma ndikusintha kokulirapo.
Mamapu otsegulira kalasi amitundu ya ResNet152 ndi AlexNet amawunikira zinthu zokhudzana ndi mtundu wazithunzi.(A) Mapu otsegulira kalasi omwe akuwonetsa kutsegulira kogwirizana pambuyo pa minyewa yapakhungu pazithunzi zotsimikizira za 8 \(\times \) 8 mm ndi (B) kukula pazithunzi zazing'ono 6 \(\nthawi \) 6 mm zotsimikizira.Mtundu wa LQ wophunzitsidwa pazikhalidwe zotsika, mtundu wa HQ wophunzitsidwa pamakhalidwe apamwamba.
Zawonetsedwa kale kuti mawonekedwe azithunzi amatha kukhudza kwambiri kuchuluka kwa zithunzi za OCTA.Kuphatikiza apo, kukhalapo kwa retinopathy kumawonjezera kuchuluka kwa zinthu zopangidwa ndi zithunzi7,26.Ndipotu, mu deta yathu, mogwirizana ndi maphunziro apitalo, tinapeza mgwirizano waukulu pakati pa kukula kwa msinkhu ndi kuopsa kwa matenda a retinal ndi kuwonongeka kwa chithunzithunzi (p <0.001, p = 0.017 kwa zaka ndi DR, motsatira; Table 1) 27 . Chifukwa chake, ndikofunikira kuwunika mtundu wazithunzi musanapange kusanthula kulikonse kwa zithunzi za OCTA.Maphunziro ambiri omwe amasanthula zithunzi za OCTA amagwiritsa ntchito zidziwitso zamphamvu zamakina zomwe zimawonetsedwa ndi makina kuti aletse zithunzi zotsika.Ngakhale kuti mphamvu yazizindikiro yasonyezedwa kuti ikukhudza kuchuluka kwa magawo a OCTA, mphamvu yapamwamba yokhayokhayo sikungakhale yokwanira kutulutsa zithunzi ndi zithunzi zojambula2,3,28,29.Choncho, m'pofunika kupanga njira yodalirika yoyendetsera khalidwe lachifanizo.Kuti izi zitheke, tikuwunika momwe makina amagwiritsidwira ntchito pophunzira mozama motsutsana ndi mphamvu ya siginecha yomwe imanenedwa ndi makina.
Tapanga mitundu ingapo yowunikira mtundu wazithunzi chifukwa mawonekedwe osiyanasiyana ogwiritsira ntchito OCTA amatha kukhala ndi zofunikira zosiyanasiyana zazithunzi.Mwachitsanzo, zithunzi ziyenera kukhala zapamwamba kwambiri.Kuonjezera apo, magawo enieni a chiwongoladzanja ndi ofunikanso.Mwachitsanzo, dera la foveal avascular zone silitengera turbidity ya sing'anga yapakati, koma imakhudza kachulukidwe ka zombo.Ngakhale kuti kafukufuku wathu akupitirizabe kuyang'ana pa njira yodziwika bwino ya chithunzithunzi, osati yogwirizana ndi zofunikira za mayesero enaake, koma cholinga chofuna kusintha mwachindunji mphamvu ya siginecha yomwe imanenedwa ndi makina, tikuyembekeza kupatsa ogwiritsa ntchito mphamvu zambiri kuti azitha kulamulira. ikhoza kusankha metric yomwe imakonda kwa wogwiritsa ntchito.sankhani chitsanzo chomwe chimagwirizana ndi kuchuluka kwa zinthu zopangidwa ndi zithunzi zomwe zimavomerezedwa.
Pazithunzi zotsika komanso zapamwamba kwambiri, tikuwonetsa magwiridwe antchito abwino kwambiri a maukonde osokonekera azama neural, okhala ndi ma AUC a 0.97 ndi 0.99 ndi mitundu yotsika kwambiri, motsatana.Tikuwonetsanso magwiridwe antchito apamwamba a njira yathu yophunzirira mwakuya poyerekeza ndi ma siginecha omwe amanenedwa ndi makina okha.Lumphani maulumikizidwe amalola ma neural network kuti aphunzire zinthu mwatsatanetsatane, kujambula mawonekedwe abwino kwambiri azithunzi (monga kusiyanitsa) komanso mawonekedwe ake onse (monga chithunzi centering30,31).Popeza kuti zithunzi zomwe zimakhudza mtundu wa zithunzi zimazindikirika bwino pamitundu yambiri, ma neural network architecture omwe ali ndi zolumikizira zosowa amatha kuwonetsa magwiridwe antchito abwino kuposa omwe alibe ntchito zowunikira mtundu.
Poyesa chitsanzo chathu pazithunzi za OCTA za 6 \ (\ × 6mm) , tinawona kuchepa kwa ntchito zamagulu amitundu yonse yapamwamba komanso yotsika kwambiri (mkuyu 2), mosiyana ndi kukula kwa chitsanzo chomwe chimaphunzitsidwa kuti chikhale chamagulu.Poyerekeza ndi chitsanzo cha ResNet, chitsanzo cha AlexNet chili ndi kugwa kwakukulu.Kuchita bwino kwambiri kwa ResNet kungakhale chifukwa cha kuthekera kwa maulumikizidwe otsalira kuti atumize zidziwitso pamiyeso ingapo, zomwe zimapangitsa kuti chitsanzocho chikhale cholimba kwambiri pakuyika zithunzi zojambulidwa pamasikelo osiyanasiyana ndi / kapena kukulitsa.
Kusiyana kwina pakati pa zithunzi za 8 \(\×\) 8 mm ndi 6 \(\×\) zithunzi za 6 mm zingayambitse kusakhazikika bwino, kuphatikizapo kuchuluka kwa zithunzi zomwe zili ndi madera a foveal avascular, kusintha kwa mawonekedwe, ma arcades a mitsempha, ndi palibe mitsempha yamawonedwe pa chithunzi 6 × 6 mm.Ngakhale izi, mtundu wathu wapamwamba kwambiri wa ResNet udatha kukwaniritsa AUC ya 85% pazithunzi za 6 \(\x\) 6 mm, kasinthidwe komwe mtunduwo sunaphunzitsidwe, kutanthauza kuti chidziwitso chamtundu wa chithunzicho chimasungidwa mu neural network. ndi oyenera.kwa kukula kwa chithunzi chimodzi kapena kasinthidwe ka makina kunja kwa maphunziro ake (Table 2).Zolimbikitsa, ResNet- ndi AlexNet-monga ma activation mapu a 8 \ (\ nthawi \) 8 mm ndi 6 \ (\ nthawi \) zithunzi za 6 mm zinatha kuwonetsa ziwiya za retinal muzochitika zonsezi, kutanthauza kuti chitsanzocho chili ndi chidziwitso chofunikira.zikugwiritsidwa ntchito pogawa mitundu yonse ya zithunzi za OCTA (mkuyu 4).
Lauerman et al.Kuwunika kwamtundu wazithunzi pazithunzi za OCTA kudachitidwanso pogwiritsa ntchito zomangamanga za Inception, skip-connection convolutional neural network6,32 pogwiritsa ntchito njira zophunzirira mozama.Anachepetsanso kafukufukuyu ku zithunzi za capillary plexus, koma kungogwiritsa ntchito zithunzi zazing'ono za 3 × 3 mm zochokera ku Optovue AngioVue, ngakhale odwala omwe ali ndi matenda osiyanasiyana a chorioretinal adaphatikizidwanso.Ntchito yathu imamanga pamaziko awo, kuphatikiza mitundu ingapo kuti athetse mawonekedwe osiyanasiyana azithunzi ndikutsimikizira zotsatira zazithunzi zamitundu yosiyanasiyana.Timaperekanso malipoti a AUC amitundu yophunzirira makina ndikuwonjezera kulondola kwawoko kale (90%)6 pamitundu yonse yotsika (96%) ndi mitundu yapamwamba (95.7%)6.
Maphunzirowa ali ndi malire angapo.Choyamba, zithunzizo zinapezedwa ndi makina amodzi a OCTA, kuphatikizapo zithunzi zokha za capillary plexus yapamwamba pa 8 \ (\ nthawi \) 8 mm ndi 6 \ (\ nthawi \) 6 mm.Chifukwa chochotsa zithunzi kuchokera m'magawo akuya ndikuti zinthu zakale zimatha kupangitsa kuwunika pamanja kwazithunzi kukhala kovuta komanso kosasinthasintha.Komanso, zithunzi zangopezeka mwa odwala matenda a shuga, omwe OCTA ikuwonekera ngati chida chofunikira chodziwiratu komanso chodziwikiratu33,34.Ngakhale kuti tinatha kuyesa chitsanzo chathu pazithunzi zamitundu yosiyanasiyana kuti tiwonetsetse kuti zotsatira zake zinali zolimba, sitinathe kuzindikira ma dataset oyenerera ochokera kumadera osiyanasiyana, zomwe zimalepheretsa kuwunika kwathu kwa generalizability ya chitsanzo.Ngakhale kuti zithunzizo zinapezedwa kuchokera ku malo amodzi okha, zinapezedwa kuchokera kwa odwala amitundu ndi mafuko osiyanasiyana, omwe ndi mphamvu yapadera ya phunziro lathu.Pophatikiza kusiyanasiyana pamaphunziro athu, tikukhulupirira kuti zotsatira zathu zidzasinthidwa mokulirapo, komanso kuti tidzapewa kuyika tsankho mumitundu yomwe timaphunzitsa.
Kafukufuku wathu akuwonetsa kuti kulumikizana-kudumpha ma neural network kumatha kuphunzitsidwa kuti akwaniritse magwiridwe antchito apamwamba pakuzindikira mtundu wazithunzi za OCTA.Timapereka zitsanzozi ngati zida zopititsira patsogolo kafukufuku.Chifukwa ma metric osiyanasiyana amatha kukhala ndi zofunikira zosiyanasiyana zamtundu wazithunzi, mtundu wowongolera upangiri pawokha utha kupangidwa pa metric iliyonse pogwiritsa ntchito zomwe zakhazikitsidwa pano.
Kafukufuku wam'tsogolo ayenera kukhala ndi zithunzi zamitundu yosiyanasiyana kuchokera kuya kosiyana ndi makina osiyanasiyana a OCTA kuti apeze njira yowunikira yakuya yazithunzi zomwe zitha kusinthidwa kukhala nsanja za OCTA ndi ma protocol oyerekeza.Kafukufuku wamakono amachokeranso ku njira zoyang'aniridwa zozama zomwe zimafuna kuunika kwaumunthu ndi kuwunika kwazithunzi, zomwe zingakhale zogwira ntchito komanso zowononga nthawi pamagulu akuluakulu.Zikuwonekerabe ngati njira zophunzirira mozama zosayang'aniridwa zingathe kusiyanitsa mokwanira pakati pa zithunzi zotsika ndi zithunzi zapamwamba.
Pamene ukadaulo wa OCTA ukupitilirabe kusinthika komanso kuthamanga kwa sikani kumachulukira, kuchuluka kwa zithunzi zakale ndi zithunzi zopanda pake zitha kuchepa.Kuwongolera kwa mapulogalamu, monga mawonekedwe omwe angotulutsidwa kumene, kungathenso kuchepetsa malirewa.Komabe, mavuto ambiri amakhalabe ngati kuyerekezera kwa odwala omwe ali ndi vuto losakonza bwino kapena kusokonezeka kwapa media nthawi zonse kumabweretsa zojambula.Pamene OCTA ikugwiritsidwa ntchito kwambiri m'mayesero achipatala, kuganiziridwa mozama kumafunika kuti pakhale ndondomeko zomveka bwino za milingo yovomerezeka ya zithunzi zojambula pazithunzi.Kugwiritsa ntchito njira zophunzirira mozama pazithunzi za OCTA kumakhala ndi lonjezo lalikulu ndipo kufufuza kwina kumafunika m'derali kuti apange njira yolimba yowongolera mawonekedwe azithunzi.
Khodi yomwe yagwiritsidwa ntchito mu kafukufuku wapano ikupezeka mu octa-qc repository, https://github.com/rahuldhodapkar/octa-qc.Ma data opangidwa ndi/kapena owunikidwa pa kafukufuku wapano akupezeka kuchokera kwa olemba omwe afunsidwa pazofunikira.
Spaide, RF, Fujimoto, JG & Waheed, NK Zithunzi zojambula mu optical coherence angiography.Retina 35, 2163-2180 (2015).
Fenner, BJ et al.Kuzindikiritsa mawonekedwe ojambulira omwe amatsimikizira mtundu ndi kubalana kwa retinal capillary plexus density miyeso mu OCT angiography.BR.J. Ophthalmol.102, 509–514 (2018).
Lauerman, JL et al.Chikoka chaukadaulo wotsata maso pamtundu wa chithunzi cha OCT angiography pakuwonongeka kokhudzana ndi zaka.Manda arch.zachipatala.Exp.ophthalmology.255, 1535-1542 (2017).
Babyuch AS et al.Miyezo ya OCTA capillary perfusion density imagwiritsidwa ntchito pozindikira ndikuwunika macular ischemia.opaleshoni ya ophthalmic.Retinal Laser Imaging 51, S30–S36 (2020).
Iye, K., Zhang, X., Ren, S., ndi Sun, J. Deep Residual Learning for Image Recognition.Mu 2016 pa Msonkhano wa IEEE pa Computer Vision and Pattern Recognition (2016).
Lauerman, JL et al.Kuwunika kwazithunzi za OCT angiographic pogwiritsa ntchito njira zophunzirira mozama.Manda arch.zachipatala.Exp.ophthalmology.257, 1641–1648 (2019).
Lauermann, J. et al.Kuchuluka kwa zolakwika zamagawo ndi zinthu zoyenda mu OCT angiography zimatengera matenda a retina.Manda arch.zachipatala.Exp.ophthalmology.256, 1807-1816 (2018).
Pask, Adam et al.Pytorch: Laibulale Yofunikira, Yopambana Kwambiri Yophunzira.Kukonzekera kwapamwamba kwa chidziwitso cha neural.dongosolo.32, 8026–8037 (2019).
Deng, J. et al.ImageNet: Malo Akuluakulu Azithunzi Zakale.Msonkhano wa IEEE wa 2009 pa Computer Vision ndi Kuzindikira Kwachitsanzo.248–255.(2009).
Krizhevsky A., Sutzkever I. ndi Hinton GE Imagenet gulu pogwiritsa ntchito deep convolutional neural network.Kukonzekera kwapamwamba kwa chidziwitso cha neural.dongosolo.25, 1 (2012).