Identifying Functional Brain Networks under Naturalistic Paradigm via A Three-Dimensional Spatial Attention Convolution Autoencoder

Functional Magnetic Resonance Imaging under the naturalistic paradigm (NfMRI) has great advantages in triggering complex and interactive functional brain networks (FBNs) due to its dynamic and multimodal nature. Various deep learning (DL) models have become the dominant means of analyzing NfMRI data. However, many studies adopt two-dimensional (2D) approaches to extract spatio-temporal information from NfMRI data, which capture limited spatial information. Moreover, these models overlook the brain's attentional mechanisms, which hinders the accurate and comprehensive understanding of neural activities. To address this, we propose a three-dimensional convolutional autoencoder network incorporating the Convolutional Block Attention Module (SA-3DCAE) to efficiently identify complex FBNs from 3D NfMRI volumes. By comparisons with the state-of-art (SOTA) methods, the proposed SA-3DCAE is more effective and reliable in characterizing FBNs, indicating the effectiveness and feasibility of incorporating spatial information of fMRI data.