Konvolutsionaalsed tehisnärvivõrgud rakupiltide segmenteerimiseks
Failid
Kuupäev
2018
Autorid
Ajakirja pealkiri
Ajakirja ISSN
Köite pealkiri
Kirjastaja
Abstrakt
Üha enam lülituvad algoritmid töö tegemisel väärtuslikeks abimeesteks. Tänapäevase tehnoloogia toel on võimalik inimesed vabastada lihtsamatest ülesannetest, et nad saaksid keskenduda teistele töödele, mis on arvuti jaoks keerulised. Üks abistavatest tehnoloogiatest on süvaõpe. Selle abil suudavad arvutid lahendada ülesandeid, mida varem peeti arvutite jaoks raskeks või koguni võimatuks.Üheks selliseks tööks on erevälja rakupiltide segmenteerimine. Seda on tarvis eelkõige biomeditsiinilaborites ning ravimifirmades, mis peavad suurt hulka mikroskoobipilte analüüsima ja kvantifitseerima. Praegused tööprotsessid väldivad ereväljapiltide kasutust, kuna nende segmenteerimiseks pole tööstuslikke lahendusi ning käsitsi töötlemine on keerukas ja aeganõudev.Magistritöö eesmärgiks on tõestada, et masinõpe suudab lahendada seni masinatele raskete ereväljapiltide segmenteerimise ülesande. Loodud lahendus aitab teadlastel üle maailma katsetada teisi uurimismeetodeid ja säästa palju aega.
There is a persistent demand for work-assisting algorithms in industry. Using present-day technology, it is possible to free people from mundane tasks so they can concentrate on work that requires human skills and flexibility. Deep learning methods can complete tasks that were previously considered hard or even impossible for machines.One example of this kind of task is segmenting brightfield microscopy images of cells. This work is needed mostly in biomedical laboratories and pharmaceutical companies that must analyse and quantify vast amounts of image data. Current workflows avoid useful brightfield imagery because automatic industrial solutions for segmentation do not exist. Manual annotation is very challenging and time consuming, even for experienced professionals.The goal of the thesis is to demonstrate that deep learning can solve the task of segmenting challenging brightfield images. The developed solution opens new experimental approaches, saving time and resources for biomedical scientists across the globe.
There is a persistent demand for work-assisting algorithms in industry. Using present-day technology, it is possible to free people from mundane tasks so they can concentrate on work that requires human skills and flexibility. Deep learning methods can complete tasks that were previously considered hard or even impossible for machines.One example of this kind of task is segmenting brightfield microscopy images of cells. This work is needed mostly in biomedical laboratories and pharmaceutical companies that must analyse and quantify vast amounts of image data. Current workflows avoid useful brightfield imagery because automatic industrial solutions for segmentation do not exist. Manual annotation is very challenging and time consuming, even for experienced professionals.The goal of the thesis is to demonstrate that deep learning can solve the task of segmenting challenging brightfield images. The developed solution opens new experimental approaches, saving time and resources for biomedical scientists across the globe.