Semi-supervised real-time classification for microscopy: Towards a universal cell cycle phase classifier
Current optical microscopes are now fully motorised, calling for making it smart and automatically adjusting the acquisition modalities to objects and events detected in the image, in real-time. We set to design a real-time deep-learning analysis of cells images to support autonomous microscopy. Specifically, biology requires training on small datasets due to costly sample preparation, prohibiting large CNN training from scratch. We propose a low-depth semi-supervised technique using a generative adversarial network GAN. It enabled us to improve training by supplementing the labelled training set with non-labelled images, decreasing experts' burden of image annotating. Our algorithm generic, i.e. easy to tweak to account for different labellings, classes (question of interest), cell lines, or even microscopy modalities, by partially retrained the pretrained model. The proposed approach, fast, generic and with a reduced-training-set, will nicely integrate an automated microscope.
Perrine Paul-Gilloteau :
Correlative microscopies: image registration for data fusion and for deep learning
Correlative microscopies is a set of methods allowing to fuse the information from complementary imaging modalities, usually also at different scales, on the same sample. In addition to the experimental elements helping to bridge different imaging modalities acquisitions, automated software solutions to correlate multiscale, multimodal and volumetric image data are an essential pillar of correlative microscopies workflows. I will introduce through biological examples the interest of correlative microscopies, and discuss methods and interest of the statistical estimation of registration errors.