Noise removal in resting-state and task fMRI: functional connectivity and activation maps

Bianca De Blasi,Anna Barnes,Silvia Francesca Storti,Ilaria Boscolo Galazzo,Lorenzo Caciagli,Marian Galovic,Matthias Koepp,Gloria Menegaz

doi:10.1088/1741-2552/aba5cc

Abstract

Objective. Blood-oxygenated-level dependent (BOLD)-based functional magnetic resonance imaging (fMRI) is a widely used non-invasive tool for mapping brain function and connectivity. However, the BOLD signal is highly affected by non-neuronal contributions arising from head motion, physiological noise and scanner artefacts. Therefore, it is necessary to recover the signal of interest from the other noise-related fluctuations to obtain reliable functional connectivity (FC) results. Several pre-processing pipelines have been developed, mainly based on nuisance regression and independent component analysis (ICA). The aim of this work was to investigate the impact of seven widely used denoising methods on both resting-state and task fMRI. Approach. Task fMRI can provide some ground truth given that the task administered has well established brain activations. The resulting cleaned data were compared using a wide range of measures: motion evaluation and data quality, resting-state networks and task activations, FC. Main results. Improved signal quality and reduced motion artefacts were obtained with all advanced pipelines, compared to the minimally pre-processed data. Larger variability was observed in the case of brain activation and FC estimates, with ICA-based pipelines generally achieving more reliable and accurate results. Significance. This work provides an evidence-based reference for investigators to choose the most appropriate method for their study and data.

Highlights

Functional magnetic resonance imaging, based on the blood-oxygenated-level dependent (BOLD) signal, is a widely used non-invasive tool for mapping brain function and functional connectivity (FC)
The first consists in creating connectivity maps by computing the correlation between the Functional magnetic resonance imaging (fMRI) signal from pairs of regions of interest
Top: In the case of rs-fMRI, the distribution of the mean tSNR values in GM is reported while, for task-fMRI, we report mean values across a representative area known to be active during a verbal fluency (VF) paradigm (L_IFG)

Summary

Introduction

Functional magnetic resonance imaging (fMRI), based on the blood-oxygenated-level dependent (BOLD) signal, is a widely used non-invasive tool for mapping brain function and functional connectivity (FC). Nonneuronal contributions to the BOLD time series arise from several factors including head motion, physiological noise (e.g. cardiac and respiratory) and scanner artefacts (e.g. thermal noise and hardware instability) (Bright, Tench and Murphy, 2017; Caballero-Gaudes and Reynolds, 2017). Several pre-processing pipelines have been developed which are generally based on nuisance regression or ICA (Pruim, Mennes, Buitelaar, et al, 2015) These pipelines result in cleaned up fMRI time series that more accurately reflect the underlying brain fluctuations of interest and reduce possible bias in post-processing analyses due to noise confounds. The second group of pipelines employ ICA, a data-driven method to decompose the fMRI data into signal of interest and structured noise The classification of these independent components (ICs) into physiological signal or noise is usually carried out manually, resulting in a time consuming and user-dependent procedure

Methods

Results

Conclusion