Denoising fMRI Message on Population Graph for Multi-site Disease Prediction

In general, large-scale fMRI analysis helps to uncover functional biomarkers and diagnose neuropsychiatric disorders. However, the existence of multi-site problem caused by inter-site variation hinders the full exploitation of fMRI data from multiple sites. To address the heterogeneity across sites, we propose a novel end-to-end framework for multi-site disease prediction, which aims to build a robust population graph and denoise the message passing on it. Specifically, we decompose the fMRI feature into site-invariant and site-specific embeddings through representation disentanglement, and construct the edge of population graph through the site-specific embedding and represent each subject using its site-invariant embedding, followed by the feature propagation and transformation over the constructed population graph via graph convolutional networks. Compared to the state-of-the-art methods, we have demonstrated its superior performance of our framework on the challenging ABIDE dataset.