Welcome everybody to another episode of Beyond the Patterns.

So today I have the great pleasure to introduce Luis Pineda.

He is a researcher at Facebook AI Research Fair in Montreal.

He obtained his PhD from the University of Massachusetts Amherst in 2018, advised by

Professor Schlomo Silberstein. During his PhD he focused on developing heuristic search

algorithms for probabilistic planning and their applications to robotic problems.

At FAIR his focus has been on studying deep reinforcement learning and its applications.

His recent work includes exploring the use of deep learning for active MRI acquisition

and developing novel reinforcement learning based methods for multi-agent collaboration

in Hanabi.

So today we have the great pleasure to have Luis here as a speaker and his presentation

will be entitled Active MR Case-based Sampling with Reinforcement Learning.

Luis, great pleasure to have you here, at least virtually, and I'm very much looking

forward to your presentation.

So the stage is yours.

First of all, thanks a lot for the invitation.

I'm really happy to present their work here today.

Well, the virtual here, right?

And yeah, let me get started.

So this was work done, we collaborated with FAIR and also from McGill University.

Suman Abbasu is a student at McGill and Adriana Romero, Roberto Calandra and Michal Drossel

are from FAIR.

Yes, so let's get started.

So magnetic resonance imaging is a powerful imaging technique.

Some of its advantages over other modalities like computational tomography include or might

include superior image quality and zero radiation exposure.

But unfortunately, MRI acquisition is very slow.

Sometimes it can take an hour or more.

And then this can lead to patient discomfort and artifacts due to patient motion.

From personal experience, the MRI, being an MRI machine is a pretty stressful process.

You have to be sitting still or lying still for quite a while.

So anything to reduce this time will be of great benefit.

So I think I'm guessing a lot of people here know this much better than I am, but just

as a sort of like for context, MRI works is that it scans what's so called raw case-based

measurements, which are a 2D frequency representation of an image.

And from this collected case-based data, an image can be reconstructed using an inverse

Fourier transform.

So as the video shows, every time we acquire some frequencies, so in this 2D plane, they

say we are acquiring rows of frequency measurements, then each additional row gets acquired, the

images keeps getting sharper and sharper.

And typically it's not that the frequencies are acquired in just kind of completely free

order where there is some sub-sampling pattern, for example, rows as in the video or columns

or some other type of pattern of acquisition.

So this means that an obvious way to accelerate this process would be to use a sub-sampling

pattern that instead of acquiring all the possible frequencies for reconstructed image,

you only acquire a partial number of them and then apply some reconstruction technique

to do that.

So here is an example of possible acquisition patterns or also called random masks, different

acceleration factors.

So you can have some random pattern or some equi-space pattern of acquisition and different