MICCAI 2018 Daily - Monday

Adrian Dalca is a postdoc at MGH and MIT. He speaks to us ahead of his oral and poster today, where he is presenting his work called VoxelMorph. Adrian tells us that it is a registration paper. The idea is that there are two camps of registration. One is the classical methods that have very nice technical development but are slow in practice, because they have to optimise a deformation field between two images. Then there are learning- based methods, which are more popular nowadays. The idea there is that we learn a function that takes us from the images to the deformation very quickly. His project combines those two camps. It starts from these principled models from classical methods, and then ends up with a neural network that registers everything fast. Adrian says that nowadays, when people are developing learning models, they’re coming from a very different camp. For this work, their question was, can we take the understanding that we’ve developed over decades and combine it with all this new, exciting stuff? Their focus was on whether they could find a bridge between the two camps. Adrian will talk in his oral about some of the things they have been doing recently, and at the poster, they will have experiments to show you on recent work that they’ve done. He points out that registration now is both really fast and has those nice guarantees from the classical work, so they’re able to do things like building an atlas as part of the training process, or using auxiliary information when registering, so if you have segmentation at training time or genetics or something like that. 14 Oral Presentation With Adrian Dalca “… it is a registration paper… ”