Vision par ordinateur pour le profilage phénotypique du cycle cellulaire

Cells are the fundamental units of life. However, despite extensive research over several decades, our understanding of their functions remains incomplete. Microscopy plays a crucial role in addressing this scientific challenge, as recent advancements in microscopy techniques have enabled researchers to visualize cellular structures at both genetic and phenotypic levels. Additionally, modern microscopes allow for conducting experiments under varied conditions and obtaining digitized results that greatly facilitate the creation of large-scale datasets. Such datasets enable the application of deep computer vision algorithms to microscopy data, paving the way for future discoveries.In this thesis, we adapted state-of-the-art algorithms from the the field of computer vision to address biological challenges linked to the cell cycle. First, we aimed at automating tedious and time-consuming image analysis tasks to alleviate the work of biologists. Second, we performed prediction tasks that are inaccessible to biologists, reducing the need for unnecessary additional fluorescent markers.We first introduced a new methodogical strategy to facilitate nucleus and cell segmentation. We proposed to leverage In-Silico Labeling (ISL), a deep learning method that generates fluorescent-like images from label-free microscopy inputs. We use ISL as a pretraining scheme for segmentation tasks, showing that it significantly reduces the number of required annotations. We later developed Cut-Detector, an end-to-end pipeline for the automatic analysis of cytokinesis timing from time-lapse microscopy images. Cut-Detector is implemented as a Napari plugin, enabling immediate user adoption for performing timing comparisons between experimental and control conditions, in order to investigate the role of specific genes in cytokinesis. Finally, we predicted the cell cycle phase - G1, S or G2/M - of a given cell from a DNA marker microscopy image. This prediction allows biologists to get rid of dedicated cell cycle fluorescent indicators, reducing cell toxicity and freeing up fluorescence channels for other markers. Our self-supervised method is based on the Variational Auto-Encoder (VAE) architecture to extract a meaningful latent representation of cell images, which is subsequently used to predict the cell cycle phase.