Deep Representations for Time-varying Brain Datasets

Sikun Lin, Shuyun Tang, Ambuj Singh

Keywords: fMRI, graph neural networks, feature attribution

Abstract: Finding an appropriate representation of dynamic activities in the brain is crucial for many downstream applications. Due to its highly dynamic nature, temporally averaged fMRI (functional magnetic resonance imaging) cannot capture the whole picture of underlying brain activities, and previous works lack the ability to learn and interpret the latent dynamics in brain architectures. In this paper, we build an efficient graph neural network model that incorporates both region-mapped fMRI sequences and structural connectivities obtained from DWI (diffusion-weighted imaging) as inputs. Through novel sample-level adaptive adjacency matrix learning and multi-resolution inner cluster smoothing, we find good representations of the latent brain dynamics. We also attribute inputs with integrated gradients, which enables us to infer (1) highly involved brain connections and subnetworks for each task (2) keyframes of imaging sequences along the temporal axis, and (3) subnetworks that discriminate between individual subjects. This ability to identify critical subnetworks that characterize brain states across heterogeneous tasks and individuals is of great importance to neuroscience research. Extensive experiments and ablation studies demonstrate our proposed method's superiority and efficiency in spatial-temporal graph signal modeling with insightful interpretations of brain dynamics.

One-sentence Summary: We build an efficient graph neural network model that incorporates both structural connectivities and dynamic functional brain signals to learn deep representations of brain activities with insightful interpretations.

Community Implementations: [ 3 code implementations](https://www.catalyzex.com/paper/arxiv:2205.11648/code)