MICCAI 2018 Daily - Monday

Adrian’s team have code available for people to play with all of these aspects. It’s a neural network-based code that has all of the classical methods, so it gives you diffeomorphisms and probabilities and uncertainties. Finally, Adrian tells us that things didn’t always go to plan: “ We actually had something really funny happen when we were trying to build an atlas as part of this. We let our atlas be completely free. We said design any atlas that seems like it’s a good mix for all these images. What happened is the algorithm designed this checkerboard that was one voxel black, one voxel white. It turns out, that’s a very good atlas, because you can basically morph that just the right way to create any image you want. It’s like a palette. It wasn’t exactly what we wanted and so we had to be a little smarter about it. In the end, we got the atlas to work with a bi-directional cost function, because diffeomorphisms are invertible .” Adrian’s oral presentation is at 16:45 today and you can visit his poster session [M-83] at 18:00-19:30. Adrian Dalca 15 Monday “… that’s a very good atlas, because you can basically morph that just the right way to create any image you want. ”