VOXEL-Level FMRI Analysis By Representation Learning And Deep Clustering for Alzheimer'S Disease

In functional magnetic resonance imaging (fMRI) analysis, functional connectivity between brain regions can be measured by Blood oxygen level dependent (BOLD) signals locally. Most studies assume that BOLD signals are homogeneous within each brain region, which ignore voxel-level connectivity changes. In this paper, we propose a novel framework for voxel-based feature extraction and recollection to characterize BOLD signal. Specifically, weakly-supervised learning strategy is adopted to extract discriminative representation from original voxel-wise BOLD signal. Considering the heterogeneity of BOLD signals within brain region of interests (ROIs), an unsupervised-based deep clustering method is employed to recollect features to different clusters automatically. Experiments on Alzheimer's Disease (AD) identification using Graph neural network (GNN) validate the effectiveness of proposed framework. To the best of our knowledge, this is the first work to take BOLD signal heterogeneity into account for feature extraction, which also provides a voxel-level scenario that can be migrated to other fMRI based tasks.