MICCAI 2018 Daily - Wednesday

Gustav Bredell 11 Wednesday The novelty is that the network is trained in an interactive fashion. The user gives scribbles, then it will update, you’ll see if it’s good enough, you’ll make an update again, and that’s exactly how they trained their network to keep it as close as possible to real world. They also did it on multi-class data, which has not been done before. Gustav explains further: “ Our training is split into two parts. In the first part, everything is as usual. We have automatic segmentation during training. First, we sample a lot of these automatic segmentation images and then we use a robot user to make scribbles. Normally during the training of a neural network, you need a lot of images – a lot of scribbles – and we want to keep this close to real world, but it’s not feasible to have a real person do this. We had a robot user make all these scribbles, and then we used these automated predictions and scribbles from the user to iteratively train our neural network .” Christine Tanner , Gustav’s co- supervisor, says that he has done great work and she will be very happy if he stays with them. She tells us that beyond the work they presented at the workshop, they have now applied it on different data, and in the beginning it didn’t work so well. She teases: “ Gustav has now found a new way, which we will hopefully publish soon, to get it all working on a different dataset. It’s on micro-CT images, and that’s a global change which is quite different from doing single object segmentation prostate images. We are quite excited about that work, but for now, we can’t reveal too much! ” We hope to see that at the next MICCAI! For now, Christine has her own poster – [W-33] Framework for Fusion of Data- and Model-Based Approaches for Ultrasound Simulation – which you can see today at 11:30- 12:30. Christine was Woman in Science at MICCAI 2017 , where she received the MICCAI IJCARS Best Paper Runner-up Award