Fusion of Fmri Multisource Data Using a Two-stream Separable 3D Network for the Diagnosis of Autism Spectrum Disorder

Renzhou Gui,Shuai Liu,Aobo Zhang,Mei Song Tong
DOI: https://doi.org/10.1109/piers59004.2023.10221333
2023-01-01
Abstract:With the continuous development of neuroscience, the computer-aided diagnosis of autism spectrum disorder (ASD) has become a widely studied topic. An increasing number of research results have shown that functional connectivity networks (FCNs) based on resting-state functional magnetic resonance imaging (rs-fMRI) are effective in aiding the diagnosis of various neurological disorders including ASD. However, most conventional FCNs are static low-level networks based on pairwise correlations between brain regions, ignoring the time-varying nature of brain activity and high-level interactions across multiple brain regions of interest (ROIs), which limits application to brain disease diagnosis. In response to the above problems, we propose to fuse features from low-order and high-order dynamic functional connectivity data to train an ensemble classifier to improve classification performance. We first obtained dynamic functional connectivity networks (D-FCNs) based on the sliding window method, and then constructed new high-order dynamic functional connectivity networks based on the principle of “correlation of correlations” to further explore the higher-level and more complex interactions between multiple brain regions of interest. We then normalized and thresholded the dynamic functional connectivity data for both modalities to reduce the effects of noise and weak connectivity. In addition, considering that the fusion of multimodal data by supervised deep learning methods has shown more and more promising application prospects in the field of neuroimaging, we extracted the spatio- temporal features of low-order D-FCNs and high-order D-FCNs respectively through a Two- Stream separable 3D network and performed fusion classification for the diagnosis of ASD and normal control subjects. Our method achieved good ASD classification accuracy and verified that high-order dynamic functional connectivity data could provide additional information for ASD diagnosis.
What problem does this paper attempt to address?