FMRI Data Analysis Preserving Map Variability Via Unsupervised Object-Centric Learning

A novel data-driven functional magnetic resonance imaging (fMRI) data analysis method is proposed using a deep object-centric learning paradigm. The method can faithfully estimate the variabilities in the spatial neural activation maps, which capture functional interconnections in the brain, over fMRI volumes. The key idea is to treat the component maps composing individual fMRI volumes as "objects," whose latent representations are separately learned by a set of autoencoders. Numerical tests using synthetic and real data sets verify the advantages of the proposed method compared to existing matrix factorization-based approaches.