Resolution Enhancement and Anti-aliasing in 3D and 2D MRI

Resolution in MRI is often sacrificed in magnetic resonance imaging (MRI)
for faster imaging or higher signal to noise ratios. A common tradeoff is
to acquire data with thicker through-plane resolution than in-plane
resolution. In addition to the introduction of poor resolution in one
orientation, this strategy also tends to introduce aliasing and its
accompanying high-frequency artifacts. In this talk, a new approach that
uses the presence of both low-res and high-res information as well as
aliased and non-aliased information in these types of acquisitions to
enhance resolution and reduce aliasing is described. The approach, based
on the use of a fully convolutional deep neural network trained on patches
from the image itself, does not require external atlases or the assumption
of self-similarity across spatial scales or across tissue contrasts. This
talk provides some background and history on super-resolution
reconstruction and then presents the new algorithm, SMORE, along with
several experiments to illustrate and quantify its performance. Overall,
this approach opens up new opportunities for improved medical image
analysis using existing data as well as new opportunities for faster
imaging.