Speaker: Mehmet Akçakaya
From: University of Minnesota
Abstract
Lengthy data acquisition times remain a bottleneck in magnetic resonance imaging (MRI).
Thus accelerated imaging methodologies have received great interest over the last three dec- ades.
Recently, deep learning (DL) techniques have gathered interest as a means to improve
reconstruction quality for accelerated MRI. DL-based reconstruction techniques can be broadly
divided into two categories, data-driven and physics-driven. The former methods learn a map- ping
from aliased images/k-space to artifact-free images/k-space. In the physics-driven ap- proaches,
the knowledge of the forward encoding operator is taken into account in solving an inverse problem.
In this talk, we will concentrate on physics-driven approaches. We will specifically focus on novel
self-supervised training strategies for such reconstruction algorithms when ground-truth data is
not available, which is a common problem in MRI. We will also discuss how we can use
insights from optimization theory to design better neural network structures.
For more info, please follow this link.
Read More