Under the direction of : Jacques Pécréaux and Marc Tramier
Optical microscopy systems are complex tools that are essential to the understanding of living organisms through fundamental and applied research in biology. They are also a tool of choice in the search for new drugs by screening in the pharmaceutical industry. Their automation is a rapidly growing field of study. The aim is to analyze images autonomously during their acquisition in order to decide what should be imaged and under what conditions. This will not only speed up and standardize experiments, but will also pave the way for the study of complex biological phenomena that were previously inaccessible. Currently, microscopy is either highly supervised and used by research experts, or restricted to rudimentary approaches during screening.
In this thesis, we present the design and development of a prototype of an embedded system analyzing microscopy images in real-time and performing a feedback loop with a microscope. We designed this system in permanent confrontation with more or less complex theoretical biological applications, which allowed us to obtain a generic and modular theoretical model. We have also started testing this system in real conditions. In particular, our system allows us to modify on-the-fly the image acquisition modalities of a microscope according to the images analyzed on-the-fly. In addition, it sorts these images by returning only those representing objects of interest to biologists.
A critical point in this process is the analysis of the images in real time. We wondered how to achieve accurate and fast classifications. We propose a method for optimizing a generic image classification tool. To do so, we identify relevant characteristics for the classification of images whose extraction time remains low. In this approach, we have highlighted a redundancy of characteristics that allows to exclude the longest to compute even if they are the most discriminating in the classification. Thus, a dozen characteristics thus chosen allow to classify 14 cells per second with a precision higher than 80% with a random forest algorithm or a simple neural network. This work opens perspectives for the optimization of classification systems in machine learning, including deep learning.
In conclusion, we have laid the foundations for an intelligent and autonomous microscope. Beyond accelerating research, it opens the way to a new generation of microscopy, like sequencing, contributing to the emergence of precision medicine based on microscopy.
Composition of the jury :
Loic Le Goff,
Sébastien Le Nours,