This research introduces DeepCor, a novel method for denoising functional magnetic resonance imaging (fMRI) data, which addresses the issue of noise obscuring meaningful neural patterns. DeepCor utilizes a contrastive variational autoencoder (CVAE), a form of deep generative model, to disentangle neural signal features (from regions of interest, ROI) from common noise features (shared with regions of no interest, RONI). Evaluation demonstrated that DeepCor significantly surpassed traditional and other deep-learning denoising techniques, achieving a 339% improvement over the CompCor method in recovering the ground truth in realistic simulated datasets. Furthermore, when applied to real fMRI data, DeepCor boosted the measured BOLD signal response to face stimuli in the fusiform face area by 215% compared to CompCor. The method is notably flexible, as it can be trained efficiently within individual participants and is applicable to both task-based experiments and resting-state data. These characteristics position DeepCor as a powerful advancement for improving the accuracy and clarity of brain activity studies.

References:

• Zhu Y, Aglinskas A, Anzellotti S. DeepCor: denoising fMRI data with contrastive autoencoders[J]. Nature Methods, 2025: 1-4.