Okay, so welcome back to interventional medical image processing.

And so far, we have started to talk about like image analysis.

So we were interested in investigating different structures in images.

And in the last lecture, we've seen that we can, for example, use epipolar geometry to

establish correspondence between two views and can use multiple views in order to get

a 3D impression of the scene.

And we also seen in the previous lecture that we were talking about factorization, where

we had the structure-for-motion approach, and we were able to reconstruct 3D point clouds

from a static scene and as well the camera motion just from a factorization of the tracked

points, essentially.

At least for the orthogonal case, in the perspective case, it was slightly more challenging.

Okay, so based on that, I want to go to a slightly different direction, where we also

do some analysis on the image.

We want to extract some information from the image.

And the information that we are now interested in is segmenting a particular structure.

So segmentation is the process where you try to find the outline or the area of a specific

object in an image.

And you can apply this to 2D images or 3D images or even 4D images.

So you want to find the outline, the boundary of a specific object.

And if you have the boundary, of course, you can also find the area that is within the

boundary.

And for example, you can use that to do measurements.

And of course, you don't want to click at every pixel individually, so you want to find

some smart methods that allow you to do a fast segmentation of a specific structure.

And one approach we will look at today is a, let's say, bottom-up approach.

Here we want to do some small sets of labels, so we only want to label a few pixels and

use this information to derive the interior and the outside of the object.

Or you could even use multiple markers, so you could even segment different organs in

the body.

And then you want to assign one label to each pixel describing this is the liver, this is

the lung, and this is background.

So this is what we want to do.

And this is essentially a segmentation problem.

So we want to find labels for every pixel and we want to label as few pixels as possible.

And the nice thing with the random marker here is that we can also do this in a semi-automatic

way.

So we can first label a couple of pixels, then run the segmentation algorithm.

It's very quick.

You will see that.

It's essentially all matrix calculus that we will be doing, and based on that we can

get a segmentation of the full image.

And if we spot then a couple of problems in the segmentation, we want to interactively

edit them and then update the segmentation.

In the next lecture, we will also discuss segmentation, but we'll look at it from a

very different perspective.

We will essentially follow a top-down approach, and there we want to follow the idea that

we use prior information about the shape and the structure of the object that we seek to

segment.

And we will look into statistical shape models for that.

That's what we'll do in the next lecture.