Computer Vision News - October 2022

59 Nicola Dinsdale two barriers to enable us to answer the question of interest while removing scanner effects and protecting individual privacy,” Nicola tells us. “The idea is that you should be able to take data from different hospital sites and learn from it without ever having to move it. This method should allow us to combine data across different institutions.” The paper proposes a federated framework in which the data doesn’t move. However, for harmonization, data must be compared between sites to be able to remove the scanner information. To solve this, it proposes domainadaptation ina federated setting while reducing the amount of shared information . “We reduce the amount of shared information to just a mean and standard deviation per feature per site,” she reveals. “What thismeans is that rather than sharing the whole amount of data, or an entire feature embedding, you’re sharing about 96 pieces of information.” Encoding the information while ensuring privacy was challenging because standard approaches do not protect privacy. However, by encoding it as the mean and standard deviation, this method could share just that information and then create example features by pulling them from a Gaussian distribution. “We’re modeling features as a Gaussian distribution,” Nicola explains. “It’s a deep learning-based approach. We use an iterative framework that allows us to remove the information adversarially. We do the task we’re interested in while removing the scanner information, but we use the Gaussian distribution to generate