Disentangle the group and individual components of functional connectome with autoencoders.

One of the central goals of neuroscience is to understand the group commonality and individual variability in functional connectome. However, the entangled nature of the group and individual components in functional connectome poses challenges. Some methods have attempted to disentangle these group and individual components, typically using functional connectivity (FC). Among them, some first compute FC from BOLD signals and then disentangle group and individual components with FC; these approaches are termed FC-level methods. In contrast, some methods first disentangle group and individual components at the BOLD level and then compute FC; these techniques are termed BOLD-level methods. BOLD-level research has demonstrated that directly modeling BOLD signals enables the capture of novel aspects of group and individual components and achieves a better disentangling effect. To this end, we propose a novel network framework, termed BRAin Signal DEcoupling (BRASDE), to disentangle group and individual components from BOLD signals, as well as complementary inductive biases that serves as disentangling strategies. Here, we assume that group components are consistent across different subjects and sessions in BOLD signals; individual components are consistent across different sessions within the same subject but variable across different subjects in BOLD signals. Utilizing the multiple sessions of fMRI data from the Human Connectome Project (HCP), we demonstrate that compared to the existing methods, BRASDE yields enhanced consistency across subjects for group components. At the same time, BRASDE amplifies the differentiation among individual components across subjects, and provides enhanced consistency within the same subjects across various sessions. Moreover, the superior performance achieved on novel sessions and subjects demonstrates the excellent generalization of BRASDE. Our methods also reveal significantly higher individual differences in the right hemisphere than in the left hemisphere. In addition, experiments validate the associations between individual components and cognitive behaviors. Overall, we propose an effective approach for disentangling group and individual components, which will facilitate further investigation into the general principles and neural mechanisms underlying individual variability in the human brain. The code can be found at https://github.com/PeiKeepMoving/BRASDE.