MIDL Vision 2020

and to encode the images as shapes rather than textures. Having shape filters instead of the textural map to see if the performance is as good, to give a nice comparison between purely textural and purely shape information. Another thought is that this area of work relates to a new area in deep learning called representation disentanglement. It is a similar conceptwhere researcherswould like to find out more about what is in the black box, but in a slightly different way. Using variational autoencoders, for example. This work relates to that as it does the work before to train the network, while they are working on the actual architecture. He thinks it would be interesting to explore if they could contribute to that kind of research. Finally, we ask Ahmed what it is like to work with Daniel Rueckert: “He is very kind. He gives you a lot of freedom to explore your intellectual curiosity. In addition to doing what you are hired to do, you have got a lot of freedom to research other interesting things. He is very excited about discovering newthings or going out of the way to research new things that are not necessarily a core part of the project. This paper is an example of that, because my main work is on the Developing Human Connectome Project (dHCP).” To learn more about Ahmed’s work [P230], you are invited to visit Poster Session #5 at 09:30-11:00 today. Ahmed E. Fetit 29 “The great thing about them is that they are explicitly textured representations of the image”