Computer Vision News - November 2020

3 Lu Zhang 2 Best of MICCAI 2020 between structural and functional connectivity is not known. The MSE loss helps to control the predicted structural and real connectivity at the elementary level. The PCC loss controls the predicted and real structure at the structure level. “The predicted structure needs to be close to the real one because, overall, the structure is very similar in different people,” Lu tells us. “However, there are some subtle differences between functional connectivities . The predicted structure in our model can catch these differences, which is important because we need to capture the relationship between the function and structure at an individual level and avoid tracking the predicted structure into a common pattern.” To evaluate their model, they explore the similarities between the predicted structure and the real structure for one person and then for another person. If the predicted structure is a bit different for each person, it means it captured the subtle differences . So far, their model has produced some very promising results. Lu tells us she chose this field of work because of her curiosity about the brain: “If I can study the brain, I am very excited!” she grins. Ø Brain structure and function 1. Functional connectivity (FC) (rs-fMRI) 2. Structural connectivity (SC) (DTI) • Multiple multi-layer GCN based generator • Single multi-layer GCN based discriminator Ø Multi-GCN based generative adversarial network (MGCN-GAN) Different GCN components are designed for different feature space and each of them will learn a latent mapping from individual FC to its corresponding SC. Ø Structure-Preserving (SP) Loss Function GAN: adversarial training MSE: element-level magnitude PCC: overall pattern (structure)