Recurrent Multigraph Integrator Network for Predicting the Evolution of Population-Driven Brain Connectivity Templates

Learning how to estimate a connectional brain template (CBT) from population of multigraphs, where each graph (e.g., functional) quantifies particular relationship between pairs regions interest (ROIs), allows pin down the unique connectivity patterns shared across individuals. Specifically, CBT is viewed as an integral representation set highly heterogeneous graphs and ideally meeting centeredness (i.e., minimum distance all in population) discriminativeness distinguishes healthy disordered criteria. So far, existing works have been limited only integrating fusing multigraphs acquired at single timepoint. In this paper, we unprecedentedly tackle question: “Given baseline multigraph population, can learn integrate forecast its representations follow-up timepoints?” Addressing such question paramount predicting common alternations populations. To fill gap, propose Recurrent Multigraph Integrator Network (ReMI-Net), first recurrent neural network which infers input \(t_1\) predicts longitudinal evolution over time (\(t_i > t_1\)). Our ReMI-Net composed blocks with convolutional layers using cross-node message passing hidden-states embeddings node region interest) then predict consecutive Moreover, design novel time-dependent loss regularize trajectory further introduce cyclic recursion learnable normalization layer generate well-centered CBTs hidden-state embeddings. Finally, derive adjacency matrix learned hidden state representation. significantly outperformed benchmark methods both discriminative biomarker discovery criteria demented patients. GitHub code available https://github.com/basiralab/ReMI-Net.