Computational Learning for Efficient and Accurate Reconstruction in Microscopy – CLEAR-Microscopy

Microscopic imaging is a key technology for our understanding of molecular and cellular processes. In recent years, artificial intelligence (AI) has emerged as a promising tool to revolutionize this field. It currently dominates the academic image reconstruction challenges and begins to appear in commercial solutions. However, significant problems remain. In particular, the lack of theoretical guarantees and reliability of AI models, the insufficient integration of physics models, the difficulty of scaling up to handle large multi-dimensional data. Overall, the adoption of AI in daily practice is still limited. Our project promises to address these challenges.
On the theoretical side, we plan to develop a rigorous mathematical framework to understand and guarantee the statistical properties of AI algorithms applied to inverse problems. We also plan to develop scalable numerical methods for advanced computational microscopy techniques.
On the applied side, we will develop innovative imaging solutions by focusing on three applications: i) the development of numerical tools to characterize the transfer function of fluorescence microscopes, thereby allowing us to design smart self-calibrated microscopes, ii) the quantitative analysis of molecular dynamics in condensates by FRAP (Fluorescence Recovery After Photo-bleaching) and FCS (Fluorescence Correlation Spetroscopy) and iii) the exploration of cryo-ET (Cryogenic Electron Tomography) for the study of intrinsically disordered proteins, of chromatin and ribosomes.
By advancing optical super-resolution microscopy, developing novel methods to quantify molecular dynamics in FRAP, and unraveling the conformational ensembles of IDPs with Cryo-ET, this project bridges key scales of biological organization. In particular, it promises to advance the understanding of the nucleolar structure and to establish a strong computational microscopy pole in Toulouse.