Learning and guidance approaches for generative and physics-driven models in computational MRI

Abstract: Lengthy data acquisition remains a major bottleneck in magnetic resonance imaging (MRI), often necessitating tradeoffs in resolution and signal-to-noise ratio. Thus, computational MRI techniques for reconstruction from sub-sampled noisy data have received great interest. In this talk, we will describe recent advances for improving computational MRI using both generative and physics-driven deep learning models. We will cover improved self-supervised learning methods for physics-driven deep learning that allows training without reference data, zero-shot guidance techniques for diffusion model based MRI reconstruction, and recent work for faster MRI reconstruction with generative priors using consistency models.

Bio: Mehmet Akçakaya is the Jim and Sara Anderson Chair Professor of Electrical Engineering at the University of Minnesota. He received the Bachelor's degree with great distinction from McGill University, Montreal, QC; and the S.M. and Ph.D. degrees from Harvard University, Cambridge, MA. His work on computational imaging and accelerated MRI has received a number of international recognitions and best paper awards. He was the recipient of a Trailblazer Award from NIH and a CAREER Award from NSF. His research interests include image processing, artificial intelligence, MRI and inverse problems.

Zoom link: https://utexas.zoom.us/j/84254847215