Computer Vision News - April 2021

Emma, tell us about your work! My work is focused on the development of new machine learning and advanced image processing techniques for analyzing medical imaging data of the brain. Nowadays, I’m much more focused on cortical modeling, which is trying to understand the outer surface of the brain: the area of the brain which is responsible for complex thought and cognitive processes that are implicated in a wide range of neurological psychiatric disorders. I am also starting to be more interested in cognitive neuroscience, modeling of how the brain works. How are things going? We’ve had some success. My biggest area of focus, at the moment, is the fact that traditional approaches foranalyzing this data make simplified assumptions, which make it very difficult to pick up on subtle differences between individual datasets. Historically, people make inferences at the population level: “On average, people with Alzheimer’s have the following brain properties” and “On average, people with schizophrenia have the following brain properties”. But a lot of these diseases that affect cognition originate in highly complex parts of the brain and are highly variable across individuals. So, they tend to be confounded by individual variability in brain shape and brain organization, which nobody really ever takes into account. We had a recent NeurIPS paper about deep generative modeling on MRI brain datasets, which very specifically looked to disentangle this issue. It has one latent space which models natural brain variability and one latent space that models the variability that is specific to the disease or the phenotype. So that works quite well actually. I have interviewed a couple of scientists who work in that area in Canada: Sandrine de Ribaupierre and Marta Kersten-Oertel. I know you worked with Carol Sudre. Yes, Carol is on that paper. She helped us with all the Alzheimer’s data. Marta taught me that some areas of the brain can be touched by the surgeon, while other areas shouldn’t. I don’t know those specific areas as well as she would. A lot of the data stuff that I’ve done has been on modeling healthy brains to develop new methods which we can use for disease. We’re starting to look at trying to detect the signs, the origins, the seizure point for epileptics, things like that. Do I know which areas of the brain are more vulnerable than others? Not so much... [ laughs ] It’s a difficult question! 341 Emma Robinson “You cannot treat brain data in a data-agnostic way!”