DeepCor: Denoising fMRI Data with Contrastive Autoencoders

Abstract

Functional magnetic resonance imaging (fMRI) is widely used in neuroscience research. FMRI data is noisy; improving denoising methods could lead to novel discoveries. Here, we introduce and evaluate a denoising method (DeepCor) which utilizes deep generative models to disentangle and remove noise. DeepCor outperforms CompCor (a state-of-the art denoising approach) on a variety of simulated datasets. In addition, DeepCor enhances differences in connectivity between brain networks in real datasets.