Welcome back everybody to Beyond the Patterns.

So today I really have an exciting guest for you and another presentation here.

Our guest speaker today will be Adrian Dalke.

So he is an assistant professor at Harvard Medical School and research scientist at the

Massachusetts Institute of Technology.

He obtained his PhD from CSAIL MIT and his research focuses on probabilistic models and

machine learning techniques to capture relationships between medical images, clinical diagnosis

and other complex medical data.

His work spans medical image analysis, computer vision, machine learning and computational

biology.

He received his bachelor's and master's in computer science from the University of Toronto.

So today I have the great pleasure to announce his presentation entitled Unsupervised Learning

of Medical Image Correspondencies in Medical Image Analysis.

Adrian, I'm very, very happy to have you here as a guest and I'm very much looking forward

to your talk.

So the stage is yours.

Thank you very much.

Thanks for the intro.

Thank you for having me here virtually.

So I'm going to talk about sort of image registration and a lot of our recent work in machine learning

for it.

I'm going to focus on brain MRIs throughout the talk, but really anything I say applies

more broadly.

It's just that we have a lot of brain MRI data that we're really interested in.

And to anyone who's here live, please do feel free to interrupt.

It makes it a bit easier to know there's people on the other side.

So I'm going to tell a little bit of a story of how I got into image registration and bring

everyone on the same page for the first 15 to 20 minutes.

And then I'll talk about sort of the most recent work that we're tackling.

So back in my PhD, which is not that long ago, I worked very closely with stroke neurologist

collaborators who had all these interesting data and problems that came from stroke patients.

And we had all kinds of interesting models we wanted to play with.

We wanted to segment strokes, segment various diseases, predict how a disease is going to

change, analyze various progressions of the disease in the brain, even sort of predict

how a brain itself is going to change with the disease.

So if I have a brain and I have the genetics of a person, maybe some clinical variables,

how are all these going to affect the brain in the future?

Can we predict how it's all going to shape up later on?

So we have all these models in our head.

Some of them we managed to do.

But there was this one fundamental process that we kept hitting against, which was this

image alignment or image registration.

And the problem at the time was that there was an awful lot of research into it and it

was fundamental to everything we did, but it was quite slow.

And so we had to work with thousands of images and we didn't just want to align these images

once, but we wanted to build models that involve the lining the images over and over and over,

either to compare the images one to another, to put them all in a common reference frame,

to predict how they're going to spatially change, all this sort of stuff required registration.

The top algorithm at the time, once we kind of modified it to work with these images,

required a lot of CPU runtime, maybe a couple hours per image.