The following content has been provided by the University of Erlangen-Nürnberg.

Okay, so welcome to the Tuesday lecture in interventional medical image processing, the

second guest lecture within two days. My name is Sebastian Bauer. I am a PhD student at

the Pattern Recognition Lab and today we're going to learn about feature detectors, feature

descriptors and its application in medical image processing. So first of all, an overview

about the next 45 minutes of this lecture. First I'm going to motivate you. So why do

we need features? What is important for features? And then we have a look at feature detectors

and descriptors. First some basic aspects. We have a look at the requirements. So what

is a feature required to be good and to be suitable for matching? We have a look at the

invariances that we want to address. And then I give you the big picture of the feature

matching pipeline. So how or what has to be done in order to match points. Then, and this

will be the main part of this lecture, we have a look at the feature detectors and descriptors.

In particular we give you some examples and you will also learn how to match those descriptors

or feature vectors at the end in order to find correspondences or address your application.

I will conclude this lecture with some practical applications, a short insight in current research

at our lab where you apply features in different applications. Okay, so let us start with the

motivation from the field of stereo vision and in particular stereo vision in radiation

therapy. So when a patient has cancer, basically it gets into the hospital and you perform

a CT scan of the patient. Then you find the tumor and the CT data and the physician is

labeling the tumor contours. Based on the tumor contours a treatment plan is generated.

And then after this initial setup where you scan the patient, the patient is coming several

times to the hospital in order to get treatment. And in treatment the patient has to be aligned

very precisely with respect to this planning data in order to get a precise alignment and

to get a nice and accurate dose delivery. And here is where stereo vision comes into

play and here you see a commercial system from a British company that is basically using

two stereo pots that acquire images from different viewpoints. Based on those images from different

viewpoints you can acquire and reconstruct the 3D surface of the patient that is currently

lying on the table. And using the planning data or the surface that is extracted from

planning data that we have scanned in the initial phase, we can register those two surfaces

and align the patient very precisely in order to get a successful treatment.

So this is an application of stereo vision. I know you already learned about stereo vision

in one of the last lectures so we can skip this part. Basically here is another sketch

of the epipolar geometry. But what is important in order to be able to use stereo vision you

need to identify pairs of corresponding points in those two images. And only when you have

those pairs of corresponding points you are able to reconstruct depth by triangulation.

So features are important not only for reconstruction techniques such as stereo vision or multiple

view geometry such as structure of motion etc. We have different applications where

we need features. For example if you want to recognise an object, imagine you have a

CT volume and you want to find lymph nodes in the CT volume. So typically you compute

features in the intensity of the volume in order to find those lymph nodes. Then we have

applications for example in image alignment, in image registration. And at the end of this

lecture I will show you some examples from our lab. So first let us have a look at the

requirements. What makes a feature suitable for matching? And let us discuss now five

points. The first one is locality. So imagine we have one image of this hand, we have another

image of this hand. If you want to find the alignment of those two images basically you

can have an initialisation of the alignment, you are a little bit to find the biggest overlap,

the largest overlap and this is basically the solution you want to find. So if you have

this global registration alignment problem that is rather easy. So when I have in one

hand my presenter and you have the image of the hand on the other side, you want to do

the registration again and you do it in a global manner you will get run into problems.