MICCAI 2018 Daily - Wednesday

Gustav Bredell is a Master’s student at ETH Zurich under the supervision of Christine Tanner and Professor Ender Konukoglu. He speaks to us about his work on iterative interaction training for segmentation editing networks, which he presented at the Machine Learning in Medical Imaging workshop on Sunday, where his Iterative Interaction Training for Segmentation Editing Networks won best paper! Gustav says they thought about the work very practically. There’s a lot of very good automatic segmentation algorithms out there, but what often happens is that they’re not perfect and then the user has to segment from scratch whatever they want to segment. In most cases automatic segmentation works, but he says they thought it would be great if when you have to segment, you can just change the automatic segmentation directly. They trained a network that takes the automatic segmentation on the one hand, and on the other hand user input, and then combines them to give an updated prediction. 10 Winning Presentation With Gustav Bredell “ We are quite excited about that work, but for now, we can’t reveal too much! ” Two examples of T2-weighted MRI images of the prostate. The yellow and red represents the central- and peripheral zone, respectively. The left column shows the ground truth and the column next to it (autoCNN) shows the output of a standard automatic segmentation algorithm (here U-Net). The next two columns show how the prediction is updated after the first and fifth iteration with a user. Wednesday