Computer Vision News - March 2020

2 Summary AI in Medical Imaging 1 Even i f you are talking about using the same architecture for solving your problem, there are so many degrees of freedom in your pipel ine, so many architectural choices, and these choices are not made in a systematic way at the moment. The team are trying to systemize this now and come up with methods that have al l these choices included, and that see a new dataset or a new segmentation problem and then adapt themselves. They don’t focus on the core architecture, but on adapting the whole set of parameters to a new dataset. “This is real ly one of my favorite topics at the moment,” Klaus tel ls us enthusiastical ly. “We could show that i f you actual ly systematical ly look at that, you can gain so much. We are currently leading so many international segmentation chal lenges which other groups have worked on for years. We apply a black-box method which adapts itsel f and are performing real ly wel l .” Fabian Isensee is the PhD student who has been working on this, and the method he is talking about is cal led the nnU-Net. “Do you know why it ’s cal led nnU-Net?” Klaus asks. “It ’s an abbreviation for no-new-Net. The principle was basical ly we are doing nothing new in terms of architectures, so i f you asked people, they would say it ’s just a new Net. We were real ly focusing on how to apply that new Net in an optimal way. We are showing that we can bui ld many, many fancy methods just by doing very optimized appl ication of that basic methodology.” He points out that at the Kidney Tumor Segmentation (KiTS) chal lenge at MICCAI 2019 there were more than 100 teams and you would imagine that by looking at the leaderboard as a PhD student you could get some kind of a signal e.g. which method should I use in order to solve such a segmentation problem? But it ’s very hard, because i f you look at architecture type A, the performance is scattered from rank 1 to 100. Then when you look at architecture type B, it ’s the same, and it ’s very hard to gain any signal that wi l l tel l you what is best. Klaus explains: “There is so much you can do wrong and so many choices that can be good or bad that basical ly the architecture is not important anymore. That is something that the community has overlooked in the past and that we are currently focusing on. I think that is going to be a very important development and is