Manifold Learning in Detecting the Transitions of Dynamic Functional Connectivities Boosts Brain State-Specific Recognition

Abstract

Exploring functional dynamics, especially with regard to the topology of functional networks, evolves into the forefront in neuroscience. Despite recent advances in identifying the transitions of functional connectivities (FCs), recognizing accurately the specific brain cognitive states along time series is few reported. Direct classification of time-varying brain data often produces sub-optimal recognition results that do not adhere to the principle of the quasi-stationary functional state. On account of the predicted brain states in such manner will be disorderly change along time. To overcome this challenge, we exploit a novel state recognition network (SR-Net) guided by the detection for transitions of dynamic FCs on Riemannian manifold. To do so, we regard the temporal evolution of functional brain networks as a set of landmarks residing on a Riemannian manifold. Accounting for high-dimensional properties of the brain networks, we elaborate a feature distillation network to capture low-dimensional FC signatures with symmetric positive definite (SPD) geometry properties. Stratifying the distribution of functional networks is devised to detect cognition state changes, which can be well solved by identifying latent modes through mean shift on the Riemannian manifold. Since functional dynamic recognition is implicated in cognitive state changes, we propose to classify these latent modes from the stratified time-varying data. Empirical results show that our SR-Net has achieved favorable state recognition results than other state-of-the-art methods on the simulated and task functional neuroimaging data from Human Connectome Project (HCP).