Thank you very much for this introduction. Actually I am almost blushing after hearing

all of this and I'm also amazed to see so many of you here listening to this presentation.

Wow, thank you very much for coming. I hope you enjoy the presentation so I'm trying to

do my best here and I want to show you a bit about the research that we've been

doing in the last couple of years and in particular I want to introduce the

technique of known operator learning because sometimes there's things that

are being tackled right now in machine learning where we already know more

about the problem that can be integrated and sometimes it makes sense actually to

reuse that prior knowledge. So I'll give a short introduction if possible, if not.

I'll give a short introduction. I'll not remote control but I'll press buttons

instead. So the well in deep learning we see quite a few things happening and

there's many many many things that are being tackled and most of the things

that are being tackled are perceptual problems. Things that you can see and

typically you need humans to do it and what people are trying to solve in these

new deep machine learning algorithms is try to process data and then extract

some information from that and I want to show some example. This is based on a

reinforcement learning technique and here a colleague, so it's a collaboration

with Siemens actually, they tried to detect anatomical landmarks in whole

body CT datasets and the idea that they used is that they formulated as a kind

of game problem where you look at some patch of a volume and try to decide to

go into the right direction in order to detect that anatomical landmark. Somehow

motivated by how a radiologist looks at an image, you know, he looks for

anatomical structures that he knows and then follows towards the network he's

interested in and here we try to model this as a game approach and we use an

agent that is kind of playing this game to detect the landmarks. We're using a

multi-resolution scheme here so we gradually zoom into the image in order

to refine finding that landmark and the nice idea about this whole process is

that it's really quick. You don't have to process the entire volume but you only

process relevant parts towards finding that certain landmark. So this approach

is very fast. We can detect 200 anatomical landmarks in approximately two

seconds on a full body CT. So that's pretty quick. The other thing is it's not

just finding landmarks but we also get some kind of well interpretation because

we find the path towards that anatomical landmark and we can see that it follows

anatomical structures that make sense in order to figure out where that landmark

is located. A cool thing about this is if you're missing a part of the anatomy

then this agent will try to leave the volume so there's no hip bone in this

volume and the agent tries to run out of the volume at the very bottom. So this is

a nice approach and you know it has a couple of advantages compared to other

approaches if you consider traditional liver segmentation tasks they would

segment the liver everywhere also in a head volume. With this one we get some

additional trust but we don't really understand how the agent makes the

decisions. We have this deep network that is taking some input and is deriving

some reaction from it. So it's very hard to interpret this and we see that these

deep learning techniques they're not only used for these perceptual problems

but more and more they're being applied towards all kinds of problems. One

particular one that I want to show here is an application where they try to

complete CT reconstruction images. So what you see here is an incomplete CT scan

and it's a it's you could say a partial reconstruction. The image here on the

left-hand side has been reconstructed only from 120 degrees of rotation which

means it's an incomplete scan and now what people have been investigating is