Chii chaitika
Iyo itsva arXiv preprint ongororo kana yakaomeswa nheyo-modhi yekumisikidza inoramba yakavimbika kana chena-yeropa-sero mifananidzo inosiyana pama scanner, masaiti, mavara uye mapaipi ekugadzirira. Vanyori vakaongorora maencoder gumi neshanu munzvimbo ina dzeruzhinji dzekuwana-sero rekutora, vachiyera zvese zviri zviviri kurongeka uye . Vanoshuma kuti mamodheru ane padhuze-akazara mu-domain mhinduro anogona kuita zvakanyanya kuipa pane yakachinjika data.
Vanyori vanoongorora gumi neshanu maencoders akatorwa kubva kuhematology, pathology uye general-vision modhi mapoka. Pane kudzidzisazve maencoder, vanoisa maprobes ekudzika kune yavo yekumisikidza uye kuyedza kuita munzvimbo ina yeruzhinji yeruzhinji-sero yekutora madomasi inosanganisira chena-yeropa-sero. Iro bepa rinogadzira dambudziko repakati sekuchinja kwekutora: iyo data yemufananidzo inogona kuchinja nekuda kwema scanner, masaiti, mavara kana mapaipi ekugadzirira, kunyangwe kana basa repasi richiramba rakafanana. Basa rakaendeswa kuarXiv musi waNyamavhuvhu 25, 2026, uye chidimbu chinoti chakagamuchirwa kuti chitaurirwe nemuromo pamusangano weHemaRAI 2026, chiitiko chesetiraiti yeMICCAI 2026.
Iyo yakashumwa mu-domain mhedzisiro yakanyatso kuunganidzwa uye yakakwirira kwazvo. Linear-probe macro-F1 inotangira pa0.98 kusvika 0.997, inotsanangurwa nevanyori sekuita kwakazara pane sosi yekumisikidza. Iko kurongeka hakubate kana mamodheru akaedzwa pane data kubva kune imwe nzvimbo yekutora. Muchinjikwa-dataset macro-F1 inodonha ne34% kusvika 72%, maererano neabstract. DinoBloom-L, iyo yakanakisa modhi mu-domain, inodonha kusvika yechigumi pakati pegumi neshanu encoder pane yakanyanya kuchinjika tarisiro, inozivikanwa seMLL23, pabenchmark yakagovaniswa 224-pixel yekupinda saizi. RedDino uye akati wandei-yekuona uye pathology encoders inomira pamberi payo mune iyo mamiriro.
Bepa racho rinoshumawo kuti kusimba kuri pachena kunoenderana nekuti iyo chando chinomiririra chinoshandiswa sei. Imwe-yepedyo-yemuvakidzani kudzoreredza yakanyanya kugadzikana paavhareji kupfuura sosi-yakakodzera mutsara musoro: iyo yepakati renji kuwirirana pakati pekwakabva uye chinangwa chekuita ndeye 0.65 ye1-NN yekudzosa maringe ne0.45 yemutsara probe. Hapana imwe nzira, zvisinei, inofanotaura pasi rose kuti ndeipi encoder ichave yakasimba pane yainotarirwa domain. Izvi zvinoreva kuti mhedzisiro yakanaka kubva kune imwe nzira yekuongorora yepasi haifanirwe kungotorwa seumboo hwekuti chinomiririra chichaendeswa zvakavimbika.
Ruvimbo fungidziro inotowedzera kuipa. Source-akadzidziswa probes akada kuenzaniswa mu-domain, paine inotarisirwa kukanganisa kukanganisa kwe0.004, asi iyo yakashumwa off-domain ECE inokwira kusvika 0.35, zvichiratidza kurongeka kwakanyanya pakati pekuvimba uye kurongeka pasi pekuedzwa kwakaedzwa. Source-fitted tembiricha scaling inofambiswa zvisina kunaka. Vanyori vanozivisazve MLL23 seDinoBloom yemukati cohort. Nekuti iyo chete yakachengetedzwa-yakabuda dataset yeDinoBloom zvakare ndiyo bhenji tsime dura, odhiyo haigone kupatsanura zvinobvira kudzidziswa kuratidzwa kubva kune scanner-yakabatana shift. Kuganhurirwa ikoko kuri pakati pekududzira chiyero uye haigadziriswe neyakashumwa mibairo yechokwadi.
Nei zvichikosha
Izvo zvakawanikwa zvinopokana nekushandiswa kwe-in-domain kurongeka seyedzo yakakwana yezvokurapa AI modhi. Iyo sisitimu inogona kutaridzika chaizvo uye yakanyatso kurongeka pane yayo sosi data uku ichiva nechivimbo kukanganisa pamifananidzo yakaunganidzwa pasi pemamiriro akasiyana. Izvo zvinogadzira dambudziko rekuongorora rinoshanda kune marabhoritari achifunga anogona kushandiswazve nheyo-modhi remienzaniso.
Yambiro inoshanda ndeye mukaha uripo pakati pekuziva patani uye kuziva apo kuzivikanwa ikoko kunogona kuvimbwa. Mukuongorora uku, modhi inogona kuwana yakakwira zvakanyanya padhata rakafanana nekumisikidzwa kwayo uku ichirasikirwa nechikamu chikuru cheiyo cross-dataset macro-F1. Kana marabhoritari ichiongorora chete nzvimbo yakajairika, inogona kupotsa mamiriro ekuti chinomiririra chinoshaya basa. Iyo bepa saka inobata kusimba sechinhu chinodiwa chekuendesa kwete chechipiri chekutsvagisa metric.
ine basa nekuti fungidziro isiriyo inounzwa uine chivimbo chepamusoro inogona kukanganisa mashandisiro anoita vanhu kuburitsa kweAI. Iyo abstract haitauri kuendeswa kwekiriniki kana ongororo yemurwere, saka hairatidze kuti yakayerwa shanduko yeECE ingakanganisa sei kuongororwa, triage kana kurapwa sarudzo mukuita. Zvinoratidza kuti hunhu hwekuvimba hunogona kuchinja zvakanyanya pasi pekuyedzwa kwekutora mashifiti. Izvi zvinoita kuti kuongororwa kwechivimbo kuve kwakakodzera pese panoshandiswa modhi inobuda kuisa pamberi pekuongorora kana kukanganisa sarudzo yemunhu.
Chidzidzo ichi zvakare chinoomesa pfungwa yekuti mukurumbira wenheyo modhi kana mu-domain chinzvimbo chinogona kumira mukusimbisa-chaiyo saiti. DinoBloom-L's kudzoserwa kubva pekutanga-mu-domain kuenda kune yechigumi pane yakanyanya kuchinjika tarisiro imhaka kubva pabhenjimark iyi, kwete chiyero chepasi rose chemhando dzehematology. Ongororo yevanyori yekufumurwa inoenderera mberi ichiratidza kuti nei modhi yemhando uye data inopindirana nyaya: mukana unooneka unogona kuratidza kujairana necohort kana kutora maitiro kwete kusimba zvakazara. Ongororo dzinosiya kutariswa kwekutarisa dzinogona saka kuwedzeredza zvakadzidzwa nemuenzaniso.
Bepa rinoshuma kuti label-isina adapta uye marginal-entropy-yakavakirwa modhi sarudzo inotaridzika yakachengeteka pasi pekuongororwa kwakaringana asi inotadza kana kirasi yeWBC isati yachinja mune imwe mamiriro echokwadi. Mhedzisiro iyoyo inobatanidza sarudzo dzekuongorora dzehunyanzvi kumamiriro ekushanda: kugoverwa kwemakirasi mumasampuri anouya anogona kusaenderana nefungidziro yakaenzana inoshandiswa mubhenji. Vanyori vanokurudzira Class-Balanced Re-standardization, yekudzidzira-isina pseudo-label-yakaenzana chimiro-yakajairwa nzira, uye inoshuma kuti inovandudza ese akaongororwa chinangwa-yepamberi mamiriro ekureva uku ichivandudza zvishoma. Iyo abstract inotiwo zvisizvo uye zvakasara zvisizvo zvinoramba zviripo, saka nzira yacho mhedzisiro yekutsvagisa kwete mhinduro yakaratidza yekushandiswa kwekiriniki.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Ongororo dzenguva yemberi dzinozoda kupatsanura shanduko-inosanganisirwa yekugovera shanduko kubva painobvira kudzidzira kuratidzwa, kuyedza zvese chokwadi uye chivimbo, uye kuongorora shanduko muWBC kirasi zviyero. Bepa rinoshuma zvinovimbisa asi zvisina kukwana mhedzisiro yekudzidzira-isina Kirasi-Yakaenzana Re-standardization nzira; kuita kwayo mukati meencoders, datasets uye chaiyo yekiriniki workflows inoramba isina kugadziriswa.
Mubvunzo unotevera wakakosha ndewekuti shanduko dzakataurwa dzinoenderana nemhando yekusiyana yakasangana mumarabhoritari chaiwo ehematology. Iri bepa rinoshandisa mana eruzhinji e-single-cell acquisition domains, asi iyo abstract hairatidze masangano avo, scanner modhi, staining protocol, mavhoriyamu emuenzaniso kana makiriniki vanhu. Izvo zvakare hazvitauri zvinotarisirwa kukiriniki kusimbiswa, mhedzisiro yemurwere, kana kuenzanisa neanogara achishandiswa diagnostic workflows. Izvo zvakasiiwa zvinodzikamisa kuti zvakananga mhedzisiro inogona kududzirwa sei kuita sarudzo dzekuendesa.
Nyaya yechipiri ndeyekuparadzanisa kuchinja kwekutora kubva kune pretraining exposure. Vanyori vanoti MLL23 yaive yemukati meboka reDinoBloom uye kuti bhenji haigone kupatsanura kuratidzwa kubva kune scanner-inosanganisirwa shanduko nekuti iyo chete yakabatwa-kunze dhatabheti zvakare inzvimbo yekubva. Rimwe basa raizoda kupatsanurwa kwakachena pakati pekudzidziswa kana kufanodzidzira dhata uye data yekuongorora, pamwe nemagwaro enzvimbo dzekuunganidzira, zvishandiso, mavara uye mapaipi ekugadzirira. Pasina kupatsanurwa ikoko, zvinoramba zvakaoma kuziva kana modhi yakasimba pakuchinja kwekutora kana kuti yakabatsirwa kubva mukujairana nerimwe cohort.
Iko kukundikana kwakashumwa pasi peWBC kirasi-yepamberi shanduko yakakodzera kuyedzwa kwepedyo. Kuongorora kwakadzikama kunogona kuita kuti kuchinjika kana nzira dzekusarudza-modhi dziratidzike dzakachengeteka kupfuura zvadziri kana musanganiswa wemhando dzemasero uchichinja. Zvidzidzo zvenguva yemberi zvinofanirwa kuongorora kana patani imwe chete ichienderera mberi mukugoverwa kwekirasi kwakasiyana, huwandu hwevanhu vanotarisirwa uye marongero ekutora, uye kana kuvimba kunoramba kuchishandiswa nepo ma frequency ekirasi achichinja. Iyo abstract inogadza mamiriro evanyori akaedzwa, asi hairatidze saizi kana kuwanda kwekuchinja kwakadaro mukuita kwekiriniki.
Class-Balanced Re-standardization ndiyo bepa rinonyanya kuita hurongwa hwekupindira, asi miganhu yaro iri pachena. Iyo abstract mishumo kuvandudzwa mune zvese zvakaongororwa zvakanangwa-yepamberi mamiriro ezvinhu uye chikamu checalibration zvakawanikwa, ukuwo uchicherekedza encoder-level yekusarudzika uye yasara miscalibration. Izvo zvisati zvanyatsojeka ndezvekuti nzira yacho inoita sei kunze kweakaongororwa encoders uye veruzhinji madomasi, ingave pseudo-label zvikanganiso zvinogona kusanganiswa panguva ye normalization, uye kuti yaizokwana sei mukufambiswa kwemabasa. Iyo yakafara yekutora yekutarisa ndeyekuti ramangwana hematology AI mabhenji akabatana anorondedzera chokwadi, , kuratidzwa uye kusimba kwekirasi-yepamberi pane kubata chero metric imwe seyakakwana.