Yin Rarraba Hotunan Huhu ta Amfani da Cibiyoyin Sadarwar Adawa na Halitta (GANs): Nazarin Fasaha

1. Gabatarwa

Rarraba hotunan huhu wani muhimmin mataki ne na farko a cikin tsarin bincike na kwamfuta (CAD) don cututtukan huhu, kamar ciwon daji na huhu, COPD, da COVID-19. Rarraba daidai na filayen huhu da nodules daga hotunan CT ko X-ray yana da mahimmanci don nazarin ƙididdiga, sa ido kan cuta, da tsarin magani. Hanyoyin rarraba na gargajiya, ciki har da ƙididdiga, girma yanki, da saitin matakin, sau da yawa suna fuskantar ƙalubalen da ke tattare da hotunan likitanci: amo, ƙarancin bambanci, da bambancin tsarin jiki.

Wannan takarda ta gabatar da wata sabuwar hanya ta hanyar tsara aikin rarraba a matsayin matsalar fassarar hotuna zuwa hotuna ta amfani da Cibiyoyin Sadarwar Adawa na Halitta (GANs). Musamman, tana amfani da tsarin Pix2Pix don fassara hoton huhu na danye zuwa abin rufe fuska mai dacewa. Wannan sauyin tsari daga rarraba pixel zuwa samar da hoto mai sharadi yana nufin samar da sakamakon rarraba mai daidaituwa da cikakkun bayanai, musamman ga lokuta masu ƙalubala kamar ƙananan nodules ko ɓoyayyun.

2. Hanya

Hanyar jigo ta ƙunshi amfani da tsarin GAN mai sharadi don koyon taswirar daga hoton huhun da aka shigar zuwa taswirar rarraba da aka fitar.

2.1 Cibiyoyin Sadarwar Adawa na Halitta (GAN)

GAN ya ƙunshi cibiyoyin sadarwar jijiyoyi guda biyu, Mai Samarwa ($G$) da Mai Rarraba ($D$), waɗanda aka horar da su lokaci guda a cikin wasan minimax. Mai samarwa yana koyon samar da samfuran bayanai na gaske daga vector amo ko, a cikin GANs masu sharadi, daga hoton shigarwa. Mai rarraba yana koyon bambancewa tsakanin samfuran gaske (masks na rarraba na gaskiya) da samfuran karya (masks da aka samar). Aikin manufa don GAN na yau da kullun shine:

$\min_G \max_D V(D, G) = \mathbb{E}_{x \sim p_{data}(x)}[\log D(x)] + \mathbb{E}_{z \sim p_z(z)}[\log(1 - D(G(z)))]$

Inda $x$ bayanai na gaske ne kuma $z$ amo shigarwa ne. A cikin saitin sharadi (cGAN), duka $G$ da $D$ suna karɓar ƙarin bayani, kamar hoton shigarwa.

2.2 Pix2Pix don Fassarar Hotuna

Takardar tana amfani da ƙirar Pix2Pix, wani tsari na farko na cGAN wanda Isola et al. (2017) suka gabatar. Pix2Pix yana amfani da mai samarwa na tushen U-Net don daidaitawar daidai da mai rarraba PatchGAN wanda ke rarraba facin hoto na gida a matsayin na gaske ko na karya, yana ƙarfafa cikakkun bayanai na mitar girma. Aikin asara ya haɗu da asarar adawa na GAN na yau da kullun tare da asarar sake gina L1:

$\mathcal{L}_{cGAN}(G, D) = \mathbb{E}_{x,y}[\log D(x, y)] + \mathbb{E}_{x,z}[\log(1 - D(x, G(x, z)))]$

$\mathcal{L}_{L1}(G) = \mathbb{E}_{x,y,z}[\|y - G(x, z)\|_1]$

$G^* = \arg \min_G \max_D \mathcal{L}_{cGAN}(G, D) + \lambda \mathcal{L}_{L1}(G)$

A nan, $x$ shine hoton huhun da aka shigar, $y$ shine abin rufe fuska na rarraba da aka yi niyya, $z$ amo ne, kuma $\lambda$ yana sarrafa nauyin asarar L1.

2.3 Aikace-aikacen ga Rarraba Hotunan Huhu

A cikin wannan mahallin, shigarwar $x$ ita ce yankin CT na huhu na grayscale na asali. Manufar $y$ ita ce abin rufe fuska na binary inda aka yiwa alamar pixels na cikin parenchyma na huhu (da yuwuwar nodules). Mai samarwa $G$ yana koyon taswirar $G: x \rightarrow y$. Horon adawa yana tilasta $G$ don samar da abin rufe fuska wanda ba kawai daidai pixel ba (ta hanyar asarar L1) amma kuma mai yuwuwar tsari kuma ba za a iya bambanta shi da abin rufe fuska na gaske ba (ta hanyar mai rarraba).

3. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Nasarar ta dogara ne da ikon mai samarwa na U-Net na ɗaukar mahallin da daidaitawar daidai ta hanyar tsarin mai ɗaukar bayanai-mai fitar da bayanai tare da haɗin tsalle. Mai da hankali na mai rarraba PatchGAN akan rubutun gida yana hana mai samarwa samar da sakamako mai duhu da aka saba da shi tare da asarar L1/L2 mai tsafta. Haɗaɗɗen aikin asara yana da mahimmanci:

Asarar Adawa ($\mathcal{L}_{cGAN}$): Yana tabbatar da gaskiyar tsarin duniya na abin rufe fuska da aka samar.
Asarar L1 ($\mathcal{L}_{L1}$): Yana tilasta daidaiton mitar ƙasa, yana tabbatar da cewa abin rufe fuska ya yi daidai da gaskiyar gaskiya a matakin pixel.

Tsarin horo a zahiri bai da ƙarfi, yana buƙatar daidaita ma'auni na hyperparameters, daidaitawar rukuni, da dabaru kamar daidaitawar misali don hana rugujewar yanayi.

4. Sakamakon Gwaji & Nazari

Takardar ta ba da rahoton gwada hanyar da aka gabatar ta tushen Pix2Pix akan ainihin bayanan hoton huhu. Duk da yake cikakkun bayanai game da bayanan (misali, LIDC-IDRI, LUNA16) da ma'auni na ƙididdiga (misali, Coefficient Dice, Index Jaccard, Hankali) ba a cika bayyana su ba a cikin ɓangaren da aka bayar, marubutan sun yi iƙirarin cewa hanyar tana "da tasiri kuma ta fi dacewa da hanyar da ta fi dacewa."

Sakamakon da aka ƙaddara & Bayanin Ginshiƙi: Wani yanki na sakamako na yau da kullun don irin wannan aikin zai haɗa da:

Kwatanta Halayya: Ganin gani na gefe-da-gefe na yankunan CT na shigarwa, abin rufe fuska na gaskiya, da tsinkaya daga hanyar GAN da aka gabatar sabanin ma'auni (misali, U-Net, FCN). Fitowar GAN zai iya nuna iyakoki masu kaifi a kusa da lobes na huhu da kuma ɗaukar ƙananan kwane-kwane fiye da yuwuwar fitarwar CNN mai duhu.
Tebur na Ma'auni na Ƙididdiga: Tebur da ke kwatanta Maki Dice, Daidaito, Tunawa, da Nisan Hausdorff a cikin hanyoyi daban-daban. Hanyar da ta dogara da GAN za ta yi jagora a tebur, musamman akan ma'auni masu hankali ga daidaiton iyaka.
Nazarin Matsalar Kasawa: Tattaunawa kan iyakoki, kamar raguwar aiki akan hotuna tare da cututtuka masu tsanani (babban ƙarfafawa) ko amo mai tsanani, inda mai samarwa zai iya yin hasashe na tsarin da ba daidai ba.

5. Tsarin Nazari: Fahimtar Jigo & Zargi

Fahimtar Jigo: Shawarar asali na wannan takarda tana da ƙarfi amma mai ma'ana: kula da rarraba hoton likitanci ba a matsayin aikin rarrabuwa ba, amma a matsayin matsala ta canja salo. Ainihin fahimtar ba kawai amfani da GAN ba ne, amma sanin cewa abin rufe fuska na rarraba mai inganci shine sigar "salo" na ainihin hoton inda "salo" shine gaskiyar tsarin jiki. Wannan sake tsarawa yana ba da damar ƙirar yin amfani da ƙa'idodin haɗa hoto masu ƙarfi da aka koya daga bayanai, yana iya ƙetare buƙatar ayyukan asara na hannu don santsin iyaka ko haɗin kai.

Kwararar Ma'ana: Hujjar tana da daidaituwa. 1) Hanyoyin gargajiya da koyon zurfin (U-Net) suna da aibobi da aka sani (iyakoki masu duhu, ƙarancin aiki akan sifofi masu sauƙi). 2) GANs, musamman Pix2Pix, suna ƙware wajen koyon wuraren fitarwa masu tsari da kiyaye cikakkun bayanai. 3) Don haka, amfani da Pix2Pix akan hotunan huhu yakamata ya haifar da rarraba mafi girma, musamman ga ƙananan nodules masu ƙalubala. Ma'ana tana da inganci, kodayake tana ɗauka cewa fa'idodin horon adawa sun fi rikitarwarsa.

Ƙarfi & Aibobi:
Ƙarfi: Hanyar tana da kyau a ka'idar. Asarar adawa ma'auni ne mai ƙarfi na kamanceceniya da aka koya wanda zai iya ɗaukar dangantaka masu rikitarwa, waɗanda ba na gida ba fiye da asarar pixel. Yana da babban yuwuwar samar da rarraba mai yuwuwar tsarin jiki ko da tare da shigarwa mara tabbas, kamar yadda aka lura a cikin aikin da ke da alaƙa kamar "CycleGAN: Fassarar Hotuna zuwa Hotuna mara Haɗin kai" (Zhu et al., 2017) wanda ke nuna ikon GANs na koyon sifofi marasa bambanci na yanki.
Aibobi Masu Muhimmanci: Takardar, kamar yadda aka gabatar, tana fama da rashin zurfi. Da'awar fiye da hanyoyin da suka fi dacewa tana da ƙarfi amma ba ta da goyan baya a nan ta takamaiman ma'auni ko masu fafatawa da aka ambata. GANs sanannen su ne masu wahala kuma marasa ƙarfi don horarwa—suna buƙatar bayanai masu yawa, daidaitawa mai kyau, da albarkatun lissafi. Tsarin yanke shawara na ƙirar "akwatin baƙar fata" ne, yana haifar da damuwa mai mahimmanci ga turawa na asibiti inda bayyanawa ke da mahimmanci. Hakanan akwai haɗarin mai samarwa "yin fentin" tsarin da ya dace amma ba daidai ba a cikin lokuta masu tsanani na cuta, wata matsala da aka sani da samfuran halitta.

Fahimtar Aiki: Ga masu bincike: Kada ku ɗauki wannan a matsayin mafita mai toshewa-da-kunnawa. Ainihin aikin yana farawa bayan zaɓin Pix2Pix. Mayar da hankali kan:

Haɗaɗɗun Asara: Haɗa asarar musamman na aiki (misali, asarar Dice) tare da asarar adawa don horo mai ƙarfi da ingantaccen ma'auni.
Ƙwararrun Tabbatarwa: Yi kwatankwacin ma'auni ba kawai da tsofaffin hanyoyi ba amma da ma'auni masu ƙarfi na zamani kamar nnU-Net (Isensee et al., 2021), daidaitaccen ma'auni na yanzu a cikin rarraba likitanci.
Bayyanawa: Yi amfani da dabaru kamar Grad-CAM ko taswirar hankali don fassara wane yanki na hoto mai rarraba ya mai da hankali akai, gina amincewa.
Jirgin Sama na Asibiti: Matsa zuwa bayan ma'auni na bayanai zuwa tabbatar da duniyar gaske tare da likitocin rediyo, auna lokacin da aka ajiye da daidaiton bincike.

Ga masu aiki: Ku tuntuɓe da kyakkyawan fata mai hankali. Dabarar tana da ban sha'awa don ayyuka ƙanana kamar gyara rarraba mai kauri ko sarrafa takamaiman hanyoyi masu ƙalubala, amma har yanzu ba ta maye gurbin ƙirar ƙwaƙƙwaran, masu fassara kamar U-Net a cikin hanyoyin samarwa ba.

6. Misalin Tsarin Nazari

Yanayi: Kimanta aikin ƙirar GAN akan rarraba nodules na juxtapleural—nodules da ke manne da bangon huhu, waɗanda suke da wahala ga algorithms na gargajiya su raba.

Aikace-aikacen Tsarin:

Fahimtar Jigo: Mai rarraba adawa ya kamata ya koyi cewa abin rufe fuska na huhu na gaske yana da iyaka mai santsi, mai ci gaba. Rarraba da kuskure ya yanke nodule na juxtapleural yana haifar da wani abu mara dabi'a a cikin wannan iyaka, wanda mai rarraba zai iya alama a matsayin "karya."
Kwararar Ma'ana: Shigarwa: Yankin CT tare da nodule mai sauƙi da aka manne da bango. U-Net na iya ƙididdige shi saboda raunin gradients na gefe. Mai samarwa na GAN, wanda mai rarraba ya hukunta shi don samar da "maras tsarin jiki" na huhu, ana ƙarfafa shi don haɗa nodule don kiyaye santsin iyaka.
Ƙarfi & Aibobi: Ƙarfi: Yuwuwar mafi girman hankali ga waɗannan takamaiman nodules. Aibi: Haɗarin kuskuren kishiyar—mai samarwa zai iya "yin hasashe" kuma ya sassauta ainihin tsaga ko shiga ciki, yana haɗa nodule da parenchyma ba daidai ba.
Fahimtar Aiki: Don rage aibin, mutum zai iya sanya sharadi ga mai rarraba ba kawai akan abin rufe fuska ba, har ma da taswirar gefen hoton shigarwa, yana kafa "gaskiyar" a cikin sifofin hoto na ƙananan matakan. Kimantawa dole ne ya haɗa da takamaiman "nazarin rukunin nodule na juxtapleural" a cikin sakamakon.

7. Aikace-aikacen Gaba & Hanyoyin Bincike

Tsarin rarraba na tushen GAN yana buɗe hanyoyi masu ban sha'awa da yawa:

Rarraba Multi-modal: Tsawaita tsarin don fassara tsakanin hanyoyin hoto daban-daban (misali, CT zuwa PET) yayin yin rarraba, yana amfani da sifofin tsarin jiki da aka raba.
Koyo mara Kulawa & Semi-kulawa: Yin amfani da tsarin kamar CycleGAN don rarraba a cikin yanayin da bayanan hoto-abin rufe fuska ba su da yawa, amma hotunan da ba a yiwa alama ba suna da yawa.
Rarraba Volumetric 3D: Matsawa daga yankuna 2D zuwa ƙarar 3D ta amfani da tsarin gine-gine kamar 3D Pix2Pix ko Vox2Vox, ɗaukar mahallin sararin samaniya mai mahimmanci don rarraba lobe na huhu da bishiyar jijiyoyi.
Haɗin Rarraba & Rarraba Cututtuka: Horar da GAN mai sharadi guda ɗaya don rarraba huhu da samar da taswirar yuwuwar rauni, kamar yadda aka bincika a cikin ayyukan baya-bayan nan akan "GANs na bincike."
Haɗin Koyo na Tarayya don Lafiya: Haɓaka ka'idojin horar da GAN waɗanda ke kiyaye sirrin marasa lafiya ta hanyar koyo daga bayanan asibiti masu rarrabuwa ba tare da raba ainihin hotuna ba, babban cikas a cikin AI na likitanci.
Haɗin kai tare da Samfuran Yadawa: Bincika tsarar samfuran halitta na gaba, samfuran yadawa, waɗanda ke ba da horo mai ƙarfi da yuwuwar fitarwa mafi girma don cikakken rarraba tsarin jiki.

8. Nassoshi

Goodfellow, I., et al. (2014). Generative Adversarial Nets. Advances in Neural Information Processing Systems (NeurIPS).
Isola, P., Zhu, J.-Y., Zhou, T., & Efros, A. A. (2017). Image-to-Image Translation with Conditional Adversarial Networks. IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI).
Zhu, J.-Y., Park, T., Isola, P., & Efros, A. A. (2017). Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. IEEE International Conference on Computer Vision (ICCV).
Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods.
Litjens, G., et al. (2017). A survey on deep learning in medical image analysis. Medical Image Analysis.
National Cancer Institute. The Cancer Imaging Archive (TCIA). https://www.cancerimagingarchive.net/ (Datasets like LIDC-IDRI).