Okay, so welcome back to interventional medical image processing.

Sorry, I was a bit late today.

So far we have discussed interventional medical image processing and different topics.

So if you remember, you can follow the cloud in your head now.

So this is the summary of our lecture.

So far we have been discussing some tools.

We've been discussing the pre-processing and we were looking into different feature descriptors

and methods how to improve image quality.

So we did edge preserving filtering and we were also interested in performing all of

these steps in real time.

So we also looked into things like the guided filter that could be very efficiently solve

the edge preserving filtering problem using box filters.

And after that we stepped further ahead to a topic that we called image analysis and

we started first with describing epipolar geometry.

We figured out that epipolar geometry is very useful if you're looking for point correspondences

in two camera or x-ray images.

We also looked into the extension that if you use epipolar planes that you essentially

have identical plane integrals in two projections and how this can be used to establish consistency

conditions.

And then we went ahead and talked about the factorization algorithm.

So Tomasi's famous factorization algorithm that allows you to reconstruct 3D point correspondences

from a series of camera images so that you could reconstruct really 3D geometry and camera

parameters from a sequence of tracked images.

Good.

The topic now we will stay in the field of image analysis and what we will be talking

about today is a first approach for image segmentation and what we will discuss today

is data driven image segmentation.

So we will look into a segmentation method that does not require a very sophisticated

modeling of the image itself.

We will do everything with low level descriptors, with low level features for the segmentation

and a very popular algorithm is the random walker.

And the nice thing with the random walker is that you don't have this very sophisticated

model and the other thing is you can compute it very quickly so it's also suitable for

interactive application.

And the way it works is you have an image and you place seed points so you have some

semi-automatic approach so you have a user that is marking some areas in the image as

belonging to one class and to another class.

So you could do foreground, background or you could have heart and lung so these would

be the tissue classes that you assign so it also works with multiple labels and based

on these labels on the input so then you have a couple of points in the image where you

know the class for sure.

And what you do next is you try to derive a segmentation from those initial points.

And this is what we're going to discuss about today.

So the nice thing is if your segmentation fails and sometimes the segmentation fails

and you don't get the right result then if you have this interactive approach that you

can recompute you can place additional labels close to the decision boundary, close to the

segmentation boundary and then you can readjust your segmentation.

And for 2D this works very efficiently and it can be done really in an iterative way

that you update your label masks.

By the way generally when we speak about segmentation we mean the process of delineating a boundary