I-DAMO yase-Alibaba, ngokubambisana neSibhedlela i-Shengjing esixhumene ne-China Medical University nezinye izikhungo, ithuthukise imodeli ye-AI yokuxilonga umdlavuza wesibindi i-DAMO LiON, ekwazi ukubona izilonda ezincane zomdlavuza esibindini ngezithombe ze-CT.
Esivivinyweni somtholampilo esibheke phambili somhlaba wangempela esathatha izinyanga ezimbili, le modeli ye-AI yathola izimila eziyingozi ezingu-15 ezazingatholwanga ngaphambilini, eziningi zazo ziyizilonda ezingaba ngu-1 cm, yasiza iziguli ukuba zithole ukuhlinzwa noma ukwelashwa ngemithi ngesikhathi. Ucwaningo oluhlobene nalo lushicilelwe ephephabhukwini lamazwe ngamazwe elihamba phambili, i-Nature Medicine.
Nansi incazelo eningiliziwe esemthethweni:
Isibindi singesinye sezitho ezivame ukuhlaselwa yizimila eziyingozi. Kukhona umdlavuza wesibindi oyinhloko oqala esibindini, ovame ukubizwa ngokuthi “umdlavuza wesibindi”, kanye “nokusabalala komdlavuza esibindini” okuvela emdlavuzeni wamathumbu, wamanyikwe nakwezinye izindawo, futhi lokhu kwakamuva kuvame kakhulu. Kuzo zombili lezi zimo, ukutholakala kusenesikhathi kuthuthukisa umphumela wokwelashwa. Nokho, ukutholakala kusenesikhathi akulula. Ngenxa yokuthi izilonda zincane futhi zingafihlwa yisifo sokuqina kwesibindi, isibindi esinamafutha kanye nesakhiwo esiyinkimbinkimbi sesibindi, odokotela bangase bangaziboni ngisho naku-CT enokuthuthukiswa komfanekiso. Lokhu kwenzeka ikakhulukazi ekusabalaleni komdlavuza esibindini, lapho ukunaka kukadokotela kungadonselwa esilondeni sokuqala bese kunganakwa izilonda ezincane esibindini.
Ngenxa yalokho, i-DAMO yasebenzisa ulwazi lwayo lweminyaka eminingi ku-AI yezithombe zezokwelapha ukuze ithuthukise imodeli ye-DAMO LiON, esebenza “njengesikhulu sokuphepha se-AI” esisiza odokotela ekuhloleni izithombe. Ikwazi ukubona ngokunembile umdlavuza wesibindi oyinhloko, futhi inekhono elikhethekile lokuthola ukusabalala komdlavuza esibindini okuvame ukunganakwa. Imiphumela yokuhlola ikhombise ukuthi imodeli ye-AI inembayo kakhulu kunodokotela bemisebe ekuboneni izimila eziyingozi. Lapho odokotela besebenzisa i-AI ukubasiza ekuhloleni izithombe, isikhathi esisetshenziswayo sehla ngo-27%, kanti ukuzwela ekuboneni izimila eziyingozi kwenyuka ngo-11.5%, ngaleyo ndlela kuncipha izimo ezingatholwanga. Ngosizo lwe-AI, odokotela abasaqala bakwazi ukufinyelela ezingeni lodokotela abanolwazi.
Ngemva kwalokho, ithimba locwaningo lafaka imodeli ye-AI esibhedlela ukuze isetshenziswe ekuhloleni izithombe kwansuku zonke. Lapho isiphetho sokuqala se-AI singahambisani nesikadokotela, icala lalinikezwa udokotela wemisebe onolwazi ukuze alibuyekeze, futhi uma kudingeka lidluliselwe ethimbeni lezinhlaka eziningi (MDT) ukuze kuxoxwe ngalo. Ezinyangeni ezimbili, i-AI yafunda izithombe ze-CT ezithuthukisiwe zeziguli ezingaphezu kuka-10 000, yasiza odokotela ukuthola izimo ezingu-15 zokusabalala komdlavuza esibindini ezazinganakwa ngaphambilini, kwashintshwa izinhlelo zokwelashwa zalezi ziguli. Phakathi kwazo kwakukhona isiguli esineminyaka engu-65 esake saba nomdlavuza wesinye. Ngesikhathi sibuyela ukuyohlolwa, izinkomba zokusebenza kwesibindi nezimpawu zomdlavuza zazijwayelekile. Udokotela wathola kuphela indawo enokubalwa esibindini ekuhloleni kokuqala, kodwa i-AI yakhomba ngokunembile ukusabalala komdlavuza esibindini, yakhumbuza udokotela ukuba alungise umbiko wokuhlolwa futhi anikeze iseluleko sokwelashwa ngamakhemikhali.
U-Yan Ke, uchwepheshe wama-algorithm e-DAMO, uthe izilonda eziyingozi ezatholwa yi-AI ngokuvamile “zincane, azigqami futhi zisendaweni engajwayelekile”: ububanzi bazo obumaphakathi bungaba ngu-1 cm, umehluko wazo nezicubu zesibindi uphansi, noma zitholakala ezindaweni ezingavamile zokwakheka komzimba. Ngenxa yalokho, imodeli ye-DAMO LiON isebenzisa ukwakheka kwenethiwekhi okuthuthukisiwe, okukwazi ukubamba ubudlelwano phakathi kwesilonda nesibindi sonke, kuyilapho kugcinwa kahle ukuthungwa nemingcele yendawo, ngaleyo ndlela kuthuthukiswe ukusebenza ezimweni ezinzima njengesibindi esinamafutha, isifo sokuqina kwesibindi kanye nesibindi sangemva kokuhlinzwa. Ngaphezu kwalokho, i-AI ingahlanganisa futhi ihlanganise ngokuphindaphindiwe izithombe zezigaba ezahlukene, ibambe kahle umehluko wamaphikseli phakathi kwezigaba ezahlukene, futhi ithole ngokunembile izilonda ezincane “ezicwayiza zidlule” ekuhlolweni kwe-CT enokuthuthukiswa komfanekiso.
Kubikwa ukuthi kusukela yasungulwa ngo-2017, i-DAMO ibitshala emkhakheni we-AI yezokwelapha. Yaba ngenye yezokuqala embonini ukuphakamisa ukusebenzisa i-AI ukubona izilonda ezincane ezithombeni zezokwelapha ezingabonwa kalula ngeso lomuntu. Ithuthukise amamodeli e-AI okuhlola umdlavuza, okuhlanganisa i-DAMO PANDA yomdlavuza wamanyikwe esebenzisa i-CT evamile, i-DAMO GRAPE yomdlavuza wesisu kanye ne-DAMO COCA yomdlavuza wamathumbu. Imiphumela ehlobene nayo ishicilelwe izikhathi ezintathu ku-Nature Medicine, yangena ohlelweni lokusungula lwe-National Medical Products Administration, futhi yaqashelwa yi-U.S. Food and Drug Administration (FDA) njengedivayisi yezokwelapha “enempumelelo entsha”. Kulokhu, i-DAMO yandisa amakhono ayo e-AI yezokwelapha ukusuka ekuhloleni kuya ekuxilongeni, isize iziguli eziningi ukuba zithole ukwelashwa okufika ngesikhathi nokunembile.