Graph Representation Neural Architecture Search for Optimal Spatial/Temporal Functional Brain Network Decomposition

Decomposing the spatial/temporal functional brain networks from 4D functional magnetic resonance imaging (fMRI) data has attracted extensive attention. Among all these efforts, deep neural network-based methods have shown significant advantages due to their powerful hierarchical representation ability. However, the network architectures of those deep learning models are manually crafted, which is time consuming and non-optimal. This paper presents a novel graph representation neural architecture search (GR-NAS) method based on graph representation to optimize the vanilla RNN cell structure for decomposing spatial/temporal brain networks. The core idea is to embed the discrete search space of the RNN cell into a continuous domain that preserves the topological information. After that, popular search algorithms, e.g., reinforcement learning (RL) and Bayesian optimization (BO), can be employed to find the optimal architecture in this continuous space. The proposed method was evaluated on the Human Connectome Project (HCP) task fMRI datasets. Extensive experiments demonstrated the superiority of the proposed model in brain network decomposition both spatially and temporally. To our best knowledge, the proposed model is among the early efforts using NAS strategy to optimally decompose spatial/temporal functional brain networks from fMRI data.

Automated summary

A graph representation neural architecture search method is proposed in this paper to optimize the RNN cell structure for decomposing spatial/temporal brain networks from fMRI data.