Computer Vision News - February 2019

15 Medical Segmentation Decathlon bit special in a way. One of them was relatively simple. It was large organs with a lot of training data. There was another one which was a little bit different because you were segmenting vessel and tumors. Previously, there were no tasks which had to do segmentation of any tumor structures. So that was slightly different. The most impressive was the third task which was on rectal primaries, cancer in the colon. That is truly hard to see. If you open the images yourself, it’s pretty hard to see the areas that need to be segmented. Probably two were leaders in this area. We wanted to push the algorithms to see how they were able to perform. The surprising outcome from all of this was that the results seen in phase 1 translated to phase 2 without almost any changes whatsoever. The scale of the data that we had was very hard for people to fix, even on the training sets, because the data was so large. The data was pretty much the same in terms of proportions between different methods, the training and test sets, even though there was no organization of parameters, obviously, because that was a whole other task. That was exactly the purpose of the challenge. Again, we saw the same behavior. The team from NVIDIA forfeited their award because NVIDIA was sponsoring the event: they decided to reward their own Titan V GPU card, giving it away to someone in the public. We did a really funny thing where we put a sticker underneath the chairs, and the person who found the sticker won the GPU. We have a picture of pretty much everyone in the room trying to search under their chairs. From a scientific point-of-view, what do we learn from the challenge? That’s the interesting part. We’re going through quite an expansive exploration of the data itself. The aim of the challenge was to ask the question: can algorithms learn to segment images without human interaction? Can we just develop a learner that can automatically solve a very large variety of tasks without a human? That question is very important because, in the medical domain, we have a huge number of tasks to solve. All of them have relatively small sometimes large datasets. In the medical domain, there are many, many tasks that we could possibly solve. We at least need an image segmentation context. You can segment any organ or different types of tumors. It’s very hard to develop an Challenge Computer Vision News