Stabilizing and cleaning functional connectivity measures via native eigenspace denoising of resting state fMRI data

Resting state functional magnetic resonance imaging (rs-fMRI) signals are sensitive to artifacts caused by head motion and non-neural physiological noise, complicating its use to investigate brain function. These effects contaminate rs-fMRI signal timeseries, confounding the calculation and analysis of functional connectivity measures and degrading the interpretation of brain function or changes due to neurological and psychiatric disorders. rs-fMRI denoising strategies play an essential role in addressing motion and non-neural noise and greatly enhance the interpretability of connectivity measures, yet this is still a highly active area of research. We propose an automated denoising method that performs data-driven noise estimation and suppression for rs-fMRI. The method is based on sliding window segmentation and nuisance regression in eigenspace for temporal and spatial eigenvectors, respectively. We show that efficient noise identification/rejection produces not only improved denoising but also enhances the reliability of functional connectivity. Without removing the global signal, the proposed method achieves denoising performance comparable to global signal regression, with trade-offs in different quality metrics. NESD shows advantages in motion and temporal noise suppression, while GSR excels in signal amplitude. Both methods produce similar negative connectivity correlations. We provide data quality visualization tools for automated assessment of noise contamination including time, space, frequency, and connectivity indicators. Our findings demonstrate that denoising is critical for processing rs-fMRI signals for connectivity analyses and that NESD offers a practical alternative to existing approaches, with trade-offs that should be considered based on specific study goals.