WhoIsWho: exploiting Google Colab to classify cells based on features related to their nuclei and their neighbours
WhoIsWho is a collection of self-explanatory Jupyter Notebooks for Google Colab that features an easy-to-use graphical user interface. They are meant to quickly get you started on learning to use deep learning for microscopy. Google Colab itself provides the computations resources needed at no-cost. WhoIsWho is designed for researchers that have little or no coding expertise to quickly train and test models.
This repository is part of the bachelor's thesis (undergraduate project) conducted by Jan Ribas-Garriga under the frame of Erasmus+ Traineeship in the Toledo Group at the Center for Chromosome Stability from the Department of Cellular and Molecular Medicine from October 2020 to June 2021.
Cell lines have long served as tools for understanding cancer at the molecular level. However, since most studies are carried out using a handful of cell lines, our ability to entirely understand the genetic heterogeneity of cancer might be reaching a plateau. A change of approach must take place in order to routinely carry out studies in a larger number of cancer cell lines, with the goal of finding new medicines and identifying genotype-specific cancer vulnerabilities. Nevertheless, replicating costly mechanistic studies or large screenings in multiple cell lines using classical imaging analysis techniques is currently unaffordable. One straightforward solution would be to pool many cell lines in a co-culture and devise a way to differentiate them from one another. Here we report a novel deep learning-based method that allows identifying cells according to the cell line they belong to in a co-culture using structural information obtained by transmitted light microscopy. This work presents a successful multi-stage pipeline consisting of a (i) segmentation of cell nuclei, (ii) localization of these nuclei by generating a bounding box around them, and (iii) classification of each nucleus. Moreover, we show the best cells' culture conditions for the pipeline's training dataset which boosts its performance yielding up to 78% classification accuracy.
Running a WhoIsWho notebook |
---|
![]() |
Any researcher interested in microscopy, independent of their background training. WhoIsWho is designed for anyone with little or no coding expertise to quickly train, test and use popular deep learning tools used to process microscopy data.
I would like to thank my co-supervisors Luis Toledo from the University of Copenhagen and Christian Mata i Miquel from the Polytechnic University of Catalonia for their help and guidance throughout the project. I am also vastly grateful for the endless patience and wise bits of advice from H. Martin Kjer. I would also like to show my appreciation to Chris Dinant for such useful discussions and feedback.