Title: Computationally efficient methods for magnetic resonance image reconstruction

Abstract: Medical image reconstruction is the process of recovering clinically useful images from noisy and often undersampled measurements. In MRI, the combination of advanced acquisition strategies with sophisticated reconstruction algorithms has dramatically improved both imaging speed and quality. However, these advances have also led to a steep rise in computational demands, an issue accentuated by the adoption of deep learning methods and comprehensive high-dimensional acquisitions. This computational bottleneck remains a major barrier to clinical translation.

In this thesis, I present a set of methods aimed at improving the computational efficiency of MRI reconstruction across different acquisition and reconstruction settings. For three-dimensional non-Cartesian imaging, I introduce Coil Sketching, a technique inspired by randomized dimensionality reduction that reduces computational complexity without sacrificing accuracy. For dynamic imaging, I develop a deep learning–based method for phase-contrast MRI that exploits the redundancy across multiple acquisitions required to encode velocity vectors, enabling faster acquisitions while maintaining accuracy. Finally, for generalized image reconstruction with generative diffusion priors, I propose a framework for designing rapid inverse problem solvers, including principled approaches for noise schedule design and hyperparameter selection tailored to inverse problem setups.

Zoom: Please contact Leyre Caracuel for the zoom link.