DeepCor: Denoising fMRI data using Contrastive Variational Autoencoders

Functional magnetic resonance imaging (fMRI) is widely used in neuroscience research to measure neural activity non-invasively with high spatial resolution. However, fMRI data is affected by noise that hinders researchers from making novel discoveries about the brain. In consideration of the complexity of noise sources and their interactions, we introduce and evaluate a denoising method which utilizes adversarial or deep generative models to disentangle and remove noise (DeepCor). The method is applicable to data from single participants, without requiring datasets with large numbers of individuals. DeepCor outperforms other denoising approaches on a variety of real datasets (StudyForrest, Adolescent Brain Cognitive Development, and THINGS-fMRI), more effectively enhancing BOLD signal responses to face selectivity in face selective regions, and place selectivity in place selective regions.