A Spatial-Temporal Integration Analysis to Classify Dynamic Functional Connectivity for Brain Disease Diagnosis

Abstract

The analysis of functional connectivity (FC) has become the major method in recent functional magnetic resonance imaging (fMRI) research for brain disease diagnosis. Most of the present FC classification methods were established based on the static FC patterns. However, an increasing number of studies have shown that dynamic FC (DFC) patterns contain abundant spatial and temporal information, which may further promote the classification. In this study, we constructed the DFC patterns and proposed a novel DFC spatial-temporal integration analysis (DFC-ST) to classify DFC patterns for the brain disease diagnosis. This model extracted the abstract spatial FC features and retained the time dependence of DFC by adopting a two-stage configuration, which separately centered on the spatial and temporal property. In the spatial analysis stage, we designed multiple feature extractors based on autoencoders to extract the abstract feature representations for the FC patterns at each time point. In the temporal analysis stage, the learnt abstract features were gathered chronologically, and a separate deep neural network (DNN) was trained to classify the fused features and obtain the prediction labels. To validate the model performance, we conducted experiments on the Autism Brain Imaging Data Exchange (ABIDE) dataset. Results demonstrate that the proposed model can more accurately distinguish the autism groups from healthy controls and the fused DFC features contain more decoding information. Our study provides an effective approach to analyze and classify the DFC patterns and further promote the classification performance for the FC-based brain diseases diagnosis.