De l'échantillonnage compressé à l'apprentissage profond pour la reconstruction d'Imagerie par Résonance Magnétique à très haut champ

Functional Magnetic Resonance Imaging (fMRI) is a widely used brain imaging technique based on the Blood Oxygenation Level Dependent (BOLD) effect, a proxy for neural activity.fMRI faces a trade-off between spatial resolution, temporal resolution, signal-to-noise ratio, and brain coverage. Achieving whole-brain coverage at high spatial and temporal resolution remains a challenge. This PhD thesis aims to improve the acquisition and reconstruction stages of the fMRI pipeline to reach higher resolution, ultimately optimizing both processes jointly. To this end, we developed "SNAKE" an open-source fMRI simulator from neuronal activation and BOLD responses to raw data acquisition. This simulator enables the generation of high spatio-temporal resolution data, facilitating the development of optimized acquisition and reconstruction methods. Using SNAKE, we demonstrated the effectiveness of a warm-start strategy for fMRI reconstruction with dynamic acquisition patterns. Finally, we also explored integrating denoising methods into the reconstruction using a "Plug-and-Play" method. Preliminary results are obtained in the context of anatomical MRI, setting the stage for future work extending it to fMRI.