Riba:
- Ability don raba sassa da ayyukan rarrabuwa kuma zaɓi kayan aikin da aka shirya da ya dace (Celpose, StarDist, MegaDetector) ga kowane
- Ikon cire ingantattun ma'auni daga hotuna ta amfani da daidaita ma'auni da tabbatarwa ta hannu
- Fahimtar buƙatun don jagorantar ƙididdige ƙima mara ƙarfi ga ingancin ɗan adam kuma tabbatar da ƙirar a waje da rarraba horo.
Ilimin halitta kimiyya ne na gani: sel a ƙarƙashin na'urar hangen nesa, mazauna a cikin abincin petri, dabbobi a cikin kyamarar gandun daji, alamun cututtuka akan ganye. Ƙididdigewa, aunawa da rarraba waɗannan hotuna sun kasance aiki ne mai ban gajiya da ra'ayi wanda ya ɗauki sa'o'i. Hankali na wucin gadi, musamman zurfafa ilmantarwa (fitowar ƙira tare da cibiyoyin sadarwa na wucin gadi na wucin gadi) tushen ƙirar hoto, ya haɓaka da daidaita wannan aikin. A cikin wannan rukunin, za mu tattauna yadda ake amfani da basirar ɗan adam a cikin nazarin hoton halitta, shirye-shiryen kayan aikin da buƙatun tabbatarwa.
Gargaɗi daga farko: samfurin na iya "gane" tantanin halitta ko jinsin, amma kuma yana iya kuskure; Idanun ɗan adam da tabbatarwa ta ainihi suna da mahimmanci don yanke shawara mai mahimmanci.
Nau'i biyu na ayyukan nuni
- Rabewa: Warewa da kirga abubuwa (kwayoyin, nuclei) a cikin hoton pixel ta pixel. Misali kayan aikin: Cellpose, StarDist. "Yawancin sel nawa ne a wannan hoton microscope kuma kowannensu nawa ne?"
- Rabewa/ganewa: Faɗin abin da ke cikin hoton ko gano wurinsa. Misali: gane nau'in a cikin hoton tarkon kyamara (iNaturalist, MegaDetector), rarraba ko ganye yana da cuta ko a'a.
Waɗannan ayyuka guda biyu suna buƙatar kayan aiki daban-daban. Hankali na wucin gadi (LLM) yana taimaka muku zaɓi kayan aiki da ya dace, rubuta lambar shigarwa da kira, da fassara fitarwa.
Tukwici: A cikin nazarin hoto, yawancin lokaci ba kwa buƙatar horar da samfuri daga karce. Kayan aikin da aka riga aka horar kamar Cellpose suna aiki kai tsaye akan nau'ikan tantanin halitta da yawa. Gwada kayan aikin da aka shirya da farko; amma tabbatar da sakamakon da hannu a hotuna da yawa.
Mataki-mataki: ƙidayar sel akan hotunan microscope
- Shirya hotuna: Ƙimar haɓakawa, haske; lura da tsarin fayil (TIFF, da dai sauransu).
- Daidaita sikelin: San adadin micrometers akan pixel (mahimmanci don auna girman).
- Zaɓi kayan aikin yanki: StarDist idan ainihin tabo; Cellpose ga cytoplasm.
- Gudu: Gwada kayan aikin akan hoto.
- Tabbatar da hannu: Kwatanta sel da samfurin ya samo tare da ainihin hoton; Ƙididdigar sel da aka haɗa/ɓace.
- Batch: Idan inganci ya gamsar, yi amfani da duk saitin.
- Rahoton: Yawan sel, rarraba girman; Hanyar daftarin aiki da daidaito.
Samfuran gaggawar da za a iya kwafi
Matsayi: Kai ne mataimaki na nazarin halittu. Aiki: Rubuta Python code wanda ya raba kuma yana kirga sel a cikin hoton microscope tare da Cellpose. Shigarwa: image.tif, tashar guda ɗaya. Fitowa: ƙidaya tantanin halitta da hangen nesa. Ba da lambar aiki tare da sharhi kuma tabbatar da zaɓin ƙira.
Ta yaya zan tabbatar da fitarwa na kashi na? Ba da shawarar hanya da awo (daidaici, tunowa, F1) don kwatanta adadin ƙwayoyin da aka samo ta samfurin tare da samfurin da na ƙidaya da hannu. Rubuta code kuma.
Bayyana aikin MegaDetector don gano dabbobi a cikin hotunan tarkon kyamara na. Yi bayanin tanadin lokaci da haɗarin ɓarna a cikin kawar da firam ɗin da ba komai.
Rubuta lambar da ke yin jujjuyawar pixel-to-micrometer tare da bayanin ma'auni kuma yana ƙididdige yankin kowane tantanin halitta a cikin mitoci masu murabba'ai. Daidaitawa: [X] pixel = [Y] micrometer.
Rauni mai ƙarfi / Ƙarfi mai ƙarfi
Rauni: "Kidaya sel a wannan hoton."
Ƙarfafa: "Ina da hoton nuclei mai launin DAPI (TIFF, tashar guda ɗaya, 2048x2048, sikelin 0.32 µm/pixel). Rubuta lambar cewa sassan da ƙidaya nuclei tare da samfurin 2D na StarDist wanda aka rigaya ya horar da shi, yana ba da yanki na kowane tsakiya a cikin ƙananan micrometers tare da rahoton ƙididdiga tare da ƙididdigewa.
Bambanci: Ƙarfin faɗakarwa yana da nau'in hoto, zanen, ƙuduri, ma'auni da tsarin tabbatarwa. Samfurin yana amfani da kayan aiki daidai da ma'auni daidai.
uku mini lokuta
Case 1 - Kwayoyin da suka haɗa: Wani ɗalibi ya ƙidaya sel 400 a cikin hoto mai yawa tare da Cellpose; Lokacin da na ƙidaya da hannu, ya zama 560. Samfurin ya haɗu da kwayoyin halitta suna taɓa juna a matsayin kwayar halitta guda ɗaya. AI ta ba da shawarar daidaita ma'aunin diamita ta tantanin halitta da canza madaidaicin magudanar ruwa; Adadin ya tashi zuwa 545. Darasi: daidaita sigogi bisa ga ainihin hoton.
Case 2 — Species confusion: In an ecology project, automatic species recognition repeatedly tagged a fox as “cat” in nighttime images. Samfurin ya bayyana cewa ƙananan haske da rashin daidaituwa a cikin bayanan horo na iya haifar da wannan kuskure; Tawagar ta aika da alamun da ba su da tabbas ga ingancin ɗan adam. Darasi: idan ƙimar amincewar ƙirar ta yi ƙasa, tabbatar da ɗan adam ya zama tilas.
Case 3 - Kuskuren Sikeli: Wani mai bincike ya ba da rahoton wuraren tantanin halitta a cikin pixels amma ya manta don canzawa zuwa micrometers; Sakamakon bai dace da wallafe-wallafe ba. Lokacin da hankali na wucin gadi ya ƙara matakin daidaitawa, ƙimar sun faɗi cikin kewayon da ya dace. Darasi: Juya juzu'in jiki yana da mahimmanci.
kwatanta jadawalin
nema
abin hawa mai dacewa
ma'aunin tabbatarwa
Mahimman kashi
StarDist
F1 tare da kirgawa da hannu
Cytoplasm/ iyakoki cell
cellpose
IoU (yawan daidaitawa)
Gane nau'ikan (namun daji)
MegaDetector/iNaturalist
Makin amana + amincewar ɗan adam
rarraba cuta
Samfurin horo na musamman
Rudani matrix
Kuskuren gama gari
- Sarrafa tsari ba tare da ingantaccen aikin hannu ba: Yada kuskuren samfurin a cikin dukkan bayanan.
- Tsallake daidaita ma'auni: Baya canza pixel-unit zuwa naúrar jiki.
- Karɓar tsinkaya tare da ƙarancin ƙarfin gwiwa: Ba magana ga mutane marasa tabbas ba.
- Barin sigogi azaman tsoho: Rashin daidaita saitunan kamar diamita tantanin halitta zuwa hoton.
- Hoto a waje da rarraba horo: Makafi ta amfani da samfurin a cikin yanayin da bai taɓa gani ba (launi daban-daban, nau'in).
Tsanaki: Samfuran hoto na iya yin aiki da kyau akan hotuna masu kama da bayanan horo, amma ba zato ba tsammani a ƙarƙashin yanayi daban-daban (sabon tabo, sabon nau'in, nau'in microscope daban-daban). Lokacin matsawa zuwa sabon saitin bayanai, kar a amince da ƙirar ba tare da inganta shi da hannu akan hotuna da yawa ba. Yardar ɗan adam yana da mahimmanci don yanke shawara mai girma kamar kiyaye nau'in nau'in ko ganewar asali.
Daga kashi zuwa aunawa: bayan lambobi
Ƙididdigar sel sau da yawa shine mataki na farko; Ƙimar cikin ciki tana fitar da ma'auni game da kowane abu. Dubban fasali a kowace tantanin halitta, kamar yanki, kewaye, zagaye, ƙarfin haske (nawa aka bayyana alamar), ana iya ƙididdige su daga abin rufe fuska. Ayyukan yanki a cikin ɗakin karatu na scikit-image yana yin wannan. AI yana rubuta lambar da ke fitar da waɗannan ma'auni kuma ta zuba sakamakon a cikin tebur; Hakanan zaka iya kwatanta ƙungiyoyi (misali "waɗanda aka yi maganin sun fi ƙanƙanta da sarrafawa?").
Cire yanki, kewaye da matsakaicin ƙarfin haske na kowane tantanin halitta daga abin rufe fuska na Cellpose tare da scikit-image regionprops. Rubuta sakamakon zuwa CSV; Kwatanta ƙungiyoyin kulawa da kulawa a cikin makircin akwatin. Aiwatar da ma'auni (µm/pixel).
Tukwici: Kar a manta cire bango lokacin da ake auna yawa; Ƙididdiga marasa ƙima sun dogara da saitin microscope kuma ana iya kwatanta su ba cikin cikakkun sharuɗɗa ba, amma a cikin sharuddan dangi tsakanin ƙungiyoyin da aka ɗauka ƙarƙashin yanayi ɗaya.
Lokacin horar da samfur: ƙananan bayanai, lakabi a hankali
Wasu lokuta kayan aikin kashe-kashe ba su isa ba kuma kuna buƙatar horar da ƙirar ƙirar ku (misali takamaiman alamar cuta). Tarkuna biyu sun tsaya a nan. Na farko shine zubar da bayanai: Hotunan mutum ɗaya / samfurin da aka haɗa a cikin duka horo da gwajin gwaji ya sa samfurin ya zama mai nasara fiye da yadda yake; yi rabo a daidai matakin. Na biyu shine aji mara daidaituwa: idan samfuran marasa lafiya kaɗan ne, ƙirar tana nuna daidaitattun daidaito ta hanyar cewa "koyaushe lafiya" amma ba shi da amfani; Dubi hankali (tunawa) da matrix ruɗani maimakon daidaito. AI ta rubuta wannan lambar horarwa, amma aikinku ne don kewaya waɗannan ramukan hanyoyin.
Ingancin hoto da preprocessing
Nasarar samfurin nuni ya dogara kai tsaye akan ingancin hoton da kuka bayar. Hotunan da ba a mayar da hankali ba, da yawa, ƙananan bambanci, ko ɗauka a ma'auni daban-daban za su rikita har ma da mafi kyawun samfurin. Shi ya sa sauƙaƙan matakan aiwatarwa kafin rarrabuwa galibi suna haɓaka sakamako sosai: gyare-gyaren bango (cire rashin daidaituwar haske), daidaita daidaituwa da raguwar amo. AI ta rubuta wannan lambar aiwatarwa (tare da scikit-image ko OpenCV); Amma ka yanke shawarar abin da gyara ya zama dole ta hanyar kallon hoton. Tsaftacewa da yawa kuma yana da haɗari: "tsaftacewa" hoto da yawa na iya shafe ainihin tsarin halitta kuma ya zama ƙawata bayanai. Ƙa'idar ita ce: preprocessing yakamata a yi amfani da shi a kan duk hotuna iri ɗaya, ƙayyadaddun tsari, ba zaɓaɓɓu ɗaya bayan ɗaya don ƙawata sakamakon ba.
Aiwatar da daidaitattun preprocessing zuwa hotuna na microscope kafin rarrabuwa: gyare-gyaren bango, daidaita daidaituwa, raguwa kaɗan. Aiwatar da sigogi iri ɗaya zuwa DUKAN hotuna (ba zaɓaɓɓu ba). Nuna kafin/bayan kwatanta. Yi amfani da scikit-image.
a takaice
Hankali na wucin gadi yana adana lokaci mai yawa a cikin nazarin hoto na halitta (rarrabuwar tantanin halitta, nau'in / gano cuta). Kayan aikin kashe-kashe kamar Cellpose, StarDist, MegaDetector suna magance yawancin ayyuka ba tare da horo daga karce ba. Duk da haka, dole ne a tabbatar da sakamakon da hannu, dole ne a daidaita ma'auni, kuma ƙananan ƙididdiga dole ne a kai ga mutane. Samfurin ba shi da tabbas a waje da rarraba horo; Tabbatarwa yana da mahimmanci lokacin matsawa zuwa sabbin bayanai.
Aikin aikace-aikace
Ɗauki hoton microscope (ko samfurin bayanai). Nemi basirar wucin gadi ta rubuta kuma ta gudanar da lambar da ta raba da kirga sel tare da Cellpose ko StarDist. Kwatanta da hannu kirga sel da samfurin ya samo a aƙalla hoto ɗaya; Lura da yawan sel da aka rasa/haɗe. Ba da rahoton wuraren tantanin halitta a cikin murabba'in mitoci ta ƙara daidaita ma'auni.
jerin abubuwan dubawa
- [ ] Na ƙara nau'in hoton, zanen da ƙuduri zuwa faɗakarwa.
- [ ] Na yi amfani da ma'auni (pixel-micrometer).
- [ ] Na tabbatar da sakamakon da hannu akan aƙalla hoto ɗaya.
- [ ] Na saita sigogin rarrabuwa bisa ga hoton.
- [ ] Na gabatar da tsinkaya tare da ƙarancin amincewa ga ingancin ɗan adam.
- [ ] Ban amince da samfurin ba tare da tabbatar da shi a cikin sabon saitin bayanai ba.