Welcome back to Beyond the Patterns. So today I have the great pleasure to introduce Fabian

Isensee. He is working at the German Cancer Research Center and he is working on medical

image segmentation. In a recent paper he introduced a method called NN-UNet and this builds upon

UNet and suggests several strategies how to quickly make it feasible onto new segmentation tasks

in domains where only limited training data is available. So he kind of has a auto ML strategy

to make the method applicable to new tasks. The paper was published in Nature Methods and the

interesting side story is when he first submitted this paper to Mikai it actually was rejected

because he is essentially using UNet and demonstrates that he can outperform many

modifications of UNet just by setting up the training setup correctly and choosing the

parameters right. And the reason for rejection was actually that it's not new. So I'm very happy that

his paper is now accepted in such a prestigious journal such as Nature Methods. Also Fabian now

graduated from his PhD. He is now running a junior research group in the German Center for Cancer

Research and by now he has won numerous medical image segmentation challenges and he also won the

award of the German Conference on Medical Image Processing for his achievements. So it's a great

pleasure to introduce him here and his presentation today will be entitled NN-UNet, a self-configuring

method for deep learning based biomedical image segmentation. And without further ado I can only

say Fabian the stage is yours.

Thank you so much Andreas for the introduction and also for inviting me to present this work

in front of your PhD students and your lab. I'm very happy to be here. Again thank you so much

also for your warm words. So I'd like to talk about NN-UNet today. This was as Andreas mentioned a

relatively recent Nature Methods publication but if you've been doing semantic segmentation in the

past years you may already have stumbled upon preliminary releases of it. We started working

on it in 2018 and finally got it into a proper journal publication in late 2020 was the acceptance.

So yeah and NN-UNet is a method for semantic segmentation. So what is semantic segmentation?

I'm sure all of you are very familiar with this type of image analysis problem. In semantic

segmentation we are interested in classifying each individual pixel of an image into one of

several predefined classes. This kind of image analysis problem is very common. It's used in

natural image processing and biological image analysis and of course also in medical image

analysis which is what my personal background is. As a side note I should mention that NN-UNet is

not necessarily restricted to the use in medical images but because it was developed in the medical

image analysis area all of the most of the examples you will be seeing throughout this talk

are from the medical domain. So semantic segmentation especially in the biomedical domain

is one of the problems that is among like among those problems that are most often addressed in

the literature and also in competitions and this kind of highlights how many people are actually

working on solutions to this problem. So if we are looking into a survey of competitions in the

biomedical domain we can actually see that 70 percent of competitions in this particular survey

that I'm citing down here were dealing with segmentation problems and this kind of highlights

the popularity of these kinds of things. So I'm going to go back to the last slide.

Naturally being a part of a group that does medical image analysis we also have been working

on multiple segmentation approaches in the past with multiple different data sets and

we've been coming up with all kinds of different UNet inspired architectures to solve all these

problems. But in an ideal world what we would like to have is given some annotated training data set

some button that we can just press and then we press the button and what happens is that

we magically obtain a fully trained fully configured UNet pipeline that can then be applied

to segment previously unseen images at state-of-the-art quality. Unfortunately the real world

is a little bit different than that and what is essentially the current quote-unquote state-of-the-art

in semantic segmentation in not just the biomedical domain but in many other domains as well is that

we are given some training data set and then we have some poor PhD student intern postdoc whoever

that does a manual configuration of a segmentation method. And the way this works is that this person

needs to think about all the different parts of the pipeline for example pre-processing, resampling,