Learning Latent Spaces for Individualized Functional Neuroimaging with Variational Autoencoders

Kajal Singla, Pierre-Louis Bazin, Nico Scherf

Abstract: Functional Magnetic Resonance Imaging (fMRI) studies often use dimensionality reduction methods like independent component analysis or diffusion map embedding to identify group-level brain networks and dynamics. These approaches struggle to capture individual-specific differences. To address this gap, we explore the use of variational autoencoders (VAEs) to model Blood Oxygen Level Dependent (BOLD) signals in a subject-specific latent space. Our approach effectively denoises fMRI data using a compressed, low-dimensional latent representation, enhancing the separation of signals from distinct functional networks without directly aligning them to specific latent axes. While direct alignment of latent dimensions across subjects is not straightforward, we observe shared geometric patterns across subjects’ latent spaces, enabling meaningful cross-subject comparisons. Deep latent modeling offers a promising avenue for individualized fMRI analysis, providing new insights into the brain’s complex functional architecture.