ResGAT: Embedding Adjacent Connectivity of Brain Regions from fMRI for Accurate Parkinson's Disease Recognition

Shuyang Li, Yu Yang, Yichen Lv, Jiangnan Xia, Xin Wang

Abstract: Despite extensive research, the exact mechanisms of Parkinson’s disease (PD) remain elusive, hindering the development of effective treatments. With the development of deep learning, particularly graph neural networks, identifying and learning biomarkers from Functional Magnetic Resonance Imaging (fMRI) is increasingly employed for accurate PD recognition. However, existing studies often model the brain regions as a fully connected network across all fMRI slices regardless of their geo-neighborhood locations, introducing biases to the recognition models since blood and electrical signals primarily flow to adjacent brain areas only. In this study, we model the 3D adjacent connectivity of brain regions and construct the corresponding brain network from fMRI data. Then, we innovatively propose ResGAT, a Residual Graph Attention Network, to learn the connectivity patterns of brain regions in each fMRI slice and its adjacent upper and lower layers by graph attention network in a residual architecture. ResGAT is highly effective in exploring valuable correlations among different functional areas in the brain, closely aligning with its real structure. Extensive experiments have been conducted on a real-world dataset. The results demonstrate the superior performance of ResGAT over existing methods in accurate PD identification, highlighting its ability to capture complex functional correlations and achieve high accuracy in PD recognition.

External IDs:dblp:conf/adma/LiYLXW24