Sügavate närvivõrkude rakendused segatud rakukultuuride fenotüüpimiseks
Date
2019
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Erinevate rakkude koos kasvatamine on peamine viis neile loomupäraseima keskkonna loomiseks. Koos kasvatatud rakud suhtlevad omavahel, mis aitab kaasa nende pikemale kasvule laboris võrreldes isoleeritud rakkudega. See on oluline näiteks ravimite katsetamisel, sh. maksarakkude vastupidavuse testimiseks. Paraku muudab erinevate rakkude olemasolu pildianalüüsi palju keerukamaks. Igal pildil tuleb sadu rakke automaatselt sildistada, ning treeningandmete annoteerimine võtab kaua aega. Olemasolevad töövood kasutavad rakkude segmenteerimist, millele järgneb raku tunnuste arvutamine või pikslipõhine klassifitseerimine. Viimasel ajal on selleks populaarsust kogunud sügavad neurovõrgud. Selles magistritöös kasutame me konvolutsioonilisi neurovõrke et koos kasvatatud rakke identifitseerida ning kokku lugeda. Me näitame, et ühe andmepunkti põhjal häälestatud mudelid aitavad igavat annoteerimistööd vältida, sealjuures analüüsi kvaliteeti kahjustamata. Rohkemate andmepunktidega töötavad meie mudelid olemasolevatest paremini. Need tulemused võivad olla olulised nii teaduses kui tööstuses, kus kiired klassifitseerimismeetodid võimaldavad rakkude kasvatamist automatiseerida, ning teadusprotsessi kiirendada.
Co-culturing is a widespread method of growing multiple cell lines together to produce an environment that is close to natural conditions. Cells in co-cultures interact with each other which allows them to sustain longer than lonely. It is crucial e.g. in pharmacology to properly measure hepatotoxicity of certain drugs. However, having two or more kinds of cells on the same plate makes downstream image analysis more difficult. Hundreds of cells per image need to be classified automatically, which is a challenging problem. Training a proper classifier requires to spent hours working on annotation. Pipelines that are commonly used for this task employ cell segmentation with subsequent feature extraction or pixel-based classification. Recently, deep learning methods have proven to perform well for image recognition. In this work, we use convolutional neural networks to build a robust algorithm for cell phenotyping and counting in co-cultures. We show that zero-shot transfer learning allows avoiding monotonous annotation and maintaining the classification performance on a baseline level. With retraining, our model is able to outperform all the reference methods. These results are could be extremely important for both research and industry. Out-of-the-box classification models allow to fully automate co-culturing experiments. This would help to diminish scientific routine and speed up pharmacological research.
Co-culturing is a widespread method of growing multiple cell lines together to produce an environment that is close to natural conditions. Cells in co-cultures interact with each other which allows them to sustain longer than lonely. It is crucial e.g. in pharmacology to properly measure hepatotoxicity of certain drugs. However, having two or more kinds of cells on the same plate makes downstream image analysis more difficult. Hundreds of cells per image need to be classified automatically, which is a challenging problem. Training a proper classifier requires to spent hours working on annotation. Pipelines that are commonly used for this task employ cell segmentation with subsequent feature extraction or pixel-based classification. Recently, deep learning methods have proven to perform well for image recognition. In this work, we use convolutional neural networks to build a robust algorithm for cell phenotyping and counting in co-cultures. We show that zero-shot transfer learning allows avoiding monotonous annotation and maintaining the classification performance on a baseline level. With retraining, our model is able to outperform all the reference methods. These results are could be extremely important for both research and industry. Out-of-the-box classification models allow to fully automate co-culturing experiments. This would help to diminish scientific routine and speed up pharmacological research.