High Resolution fMRI using Compressed Sensing

This page illustrates some material that is challenging to convey in PDF format and gives a telegraphic summary of the work.
For everything else, please refer to the report.

It is well known that ''compressed sensing'' can be used to reconstruct MRI images from undersampled data. This undersampling can be used to improve temporal resolution (reduce scan time). It can also be used to improve spatial resolution (by focusing sampling on the high spatial frequencies). This is what we did in this project, with functional MRI (fMRI) as an application.
To perform the reconstruction, we used a simple TV regularization (across all four dimensions):

Experiments

In these experiments we acquired the same number of samples (same scan time), but focused sampling on higher spatial frequencies. Each example shows ''acquisition'' -- raw aliased image (a least-l2-norm reconstruction), ''reconstruction'' -- result of our sparse-reconstruction-method, and the ''activation map'' - the image that the fMRI setting is actually interested in.

2x undersampling (2.4 matrix size increase)

3x undersampling (5.7 matrix size increase)