AI DA KE KARANTA HATIMIN ASIBITI MAIMAKON HUHU

Mai rarraba hotunan X-ray na ƙirji (classifier) zai iya koyon karanta lambar da asibiti ke buga hatiminta a kusurwar kowane hoto maimakon cutar da ke cikin huhu. Sannan yana aiki da kyau a kan bayanai daga wannan asibiti kuma yana gazawa a ɓoye a na gaba. Binciken da aka yi a baya ya rubuta wannan ga masu rarraba COVID-19 da ciwon huhu (pneumonia). Abin da ba su iya faɗi ba shi ne wane sashe na samfurin ke yin karatun, da ko cire wannan sashe zai taimaka.

Kayan aikin bayani da aka saba — taswirorin zafi kamar Grad-CAM, attention rollout, LIME — suna nuna inda samfuri ya duba. Ba sa gano sashen da ke da alhaki, kuma ba za su iya tabbatar da sanadi ba. Joshua Muhumuza da Cornelius Kalenzi, a Jami’ar Makerere, maimakon haka sun aro kayan aikin fassarar injiniya (mechanistic interpretability), wanda aka ƙirƙira don rarrabe samfuran harshe, suka yi amfani da shi a kan Vision Transformer.

Daga asibitin koyarwa zuwa asibitin gunduma

Marubutan sun tsara abin da ke cikin haɗari a zahiri. Ana ƙara amfani da masu rarraba X-ray masu tushe a Transformer don tantance tarin fuka (tuberculosis) da ciwon huhu a wurare marasa isassun kayan aiki, sau da yawa an horar da su a babban asibitin karɓar marasa lafiya kuma aka mayar da su, ba tare da sake bincike ba, zuwa asibitocin gunduma ko na’urorin tafi-da-gidanka masu injuna daban, manhaja daban da al’adun lakabi daban. Samfurin da ya koyi hatimin asibitinsa na asali zai lalace da zarar ya haɗu da asibitin da ba ya buga hatimi, ko yana buga wani abu daban, a wani wuri daban.

Head ɗaya daga cikin 144

Samfurin da aka yi nazari, ViT-Base/16 (checkpoint na jama’a google/vit-base-patch16-224, wanda aka riga aka horar a kan ImageNet-21k), yana yanka kowane hoto zuwa faci 196 kuma yana sarrafa su ta matakai (layers) 12 masu attention heads 12 kowanne. Ƙungiyar ta ba dukkan heads 144 maki kan yadda hankalinsu ya taru da yadda ya yi nisa.

Tattarawar hankali tana bayyana farat ɗaya a matakai na 7 zuwa 9. Head ɗaya ya fice: L8H10, head na 10 na mataki na 8. Yana sanya kashi 34.6% na hankalinsa a kan faci ɗaya — ninki 2.8 fiye da kowane head. A kan hoton X-ray da aka rubuta “D CLINO” a kusurwar sama a hagu, yana sanya 0.346 na hankalinsa a kan wannan faci, ninki 68 na matakin daidai na 0.0051, yayin da huhu ke samun kusan babu komai.

Head L8H10 a kan hoton X-ray na farko da aka bincika: wuri ɗaya mai haske a kan rubutun kusurwa, kusan baƙi a kan huhu. — Hoto na 4, Muhumuza da Kalenzi (2026), arXiv:2610.21638.

A cikin hotunan X-ray 200 na COVID-19, head ɗin yana yin haka a cikin 123 (kashi 62%). A cikin waɗannan hotunan, hankali a kan kusurwa a matsakaici ninki 57.5 ne na hankali a kan huhu; a cikin sauran, ninki 0.77. Ƙungiyoyin biyu ba sa haɗuwa. Abin da ke hasashen wannan hali ba haske ba ne amma bambancin launi na gida (local contrast) a kusurwa — daidai abin da baƙaƙen haruffa a kan bango mai haske ke samarwa.

Mafi girman hankali a kan faci ɗaya ga kowane head: L8H10 (ja) yana a 0.346, na gaba a 0.124. — Hoto na 3, Muhumuza da Kalenzi (2026), arXiv:2610.21638.

Hujja ta hanyar gogewa

Rufe kusurwar da launin toka mai shimfiɗe ba gwaji mai tsabta ba ne: murabba’in launin toka iri ɗaya shi kansa yana da babban bambancin launi kuma yana jan hankalin head ɗin. Don haka marubutan sun cike (inpainted) kusurwar maimakon haka, suna cika ta da siffar nama ta gaskiya. Hankali a kan kusurwa sai ya ragu a cikin hotuna 158 daga cikin 200 (kashi 79%). A cikin hotunan da abin ya shafa, hankali ga huhu ya ƙaru da ninki 1.52; a cikin sauran, bai motsa ba (ninki 0.99). Rubutun kansa, ba haske ko gefuna ba, ke tura head ɗin.

Tsarin yana maimaituwa a wasu wurare: a cikin kashi 79% na hotunan X-ray 200 daga rukunin bayanai na NIH ChestX-ray14, da kuma a cikin wasu tsare-tsare biyu — DeiT-Base (head L9H2, kashi 88%) da ViT-Large/16 (head L8H8, kashi 80%).

Da’irar da gwaje-gwajen kowa ba sa iya gani

Babban sakamakon mara kyau ne. A cikin mai rarraba X-ray mai rukunai biyar da aka gina a kan tsari ɗaya, kashe L8H10 ya canza kashi 5% na hasashe — amma sau da yawa a kan hotuna na al’ada kamar a kan waɗanda abin ya shafa (p = 0.340). Mai rarraba ciwon huhu na biyu, wanda aka horar daban, ya ba da sakamako ɗaya mara tasiri. Kashe dukkan heads uku da ake zargi a lokaci guda (L8H10, L7H2, L9H3) ya ninka canje-canjen zuwa kashi 10%, har yanzu ba tare da tasiri na musamman ba (p = 0.532).

Marubutan sun fassara wannan a matsayin yawaitar ajiya (redundancy): bayanin ya bazu a kan sassa da yawa, don haka cire masu ɗaukarsa mafi bayyana yana bar sauran su yi aikin. Sake horarwa ma ba ya taimakawa. Bayan cikakken gyara (fine-tuning) a kan hotunan X-ray, 121 daga cikin hotuna 123 da abin ya shafa (kashi 98.4%) har yanzu sun kunna head ɗin, domin hatimin yana ci gaba da zama mai hasashe a cikin bayanai inda al’adun lakabi ke bin asibitin da hoto ya fito — da cututtukan da ke tare da shi.

Duba ciki kafin ka tura

Haɗarin, marubutan sun rubuta, ba ganewar cuta ɗaya mara daidai ba ne amma raunin ɓoye wanda ke kunnawa lokacin da samfuri ya canza asibiti. Shawararsu: a binciki attention heads, a gano masu ɗaukar bayanin, a tabbatar da su ta hanyar sanadi, kuma a yi haka kafin turawa — musamman tsakanin cibiyoyi. Gyare-gyare kawai masu inganci da suka sani suna a matakin bayanai: a horar a kan hotunan da aka cire rubuce-rubucensu, ko a hukunta heads ɗin da aka gano a lokacin horo.

An kuma bayyana iyakokin. An saita yankin kusurwa da iyakarsa don waɗannan rukunonin bayanai; gwaje-gwajen gyara sun yi amfani da hotuna 200 kawai; kuma gwajin yanke hukunci — horo a kan hotuna masu rubutu idan aka kwatanta da waɗanda aka cire musu, masu girma ɗaya — har yanzu ba a yi shi ba.