fMRI  denoising and Quality Control

Default denoising pipeline

CONN's default denoising pipeline combines two general steps: linear regression of potential confounding effects in the BOLD signal, and temporal band-pass filtering

Temporal band-pass filtering

Temporal frequencies below 0.008 Hz or above 0.09 Hz are removed from the BOLD signal in order to focus on slow-frequency fluctuations while minimizing the influence of physiological, head-motion and other noise sources. Filtering is implemented using a discrete cosine transform windowing operation to minimize border effects, and performed after regression to avoid any frequency mismatch in the nuisance regression procedure (Hallquist et al. 2013)

Quality Control of fMRI data after denoising

The effect of denoising, acting to minimize the influence of artifactual factors on functional connectivity measures, can be exemplified by estimating the distribution of functional connectivity values (FC) between randomly-selected pairs of points within the brain before and after denoising (available as part of CONN's Quality Control plots). Considering the BOLD signal after a standard minimal preprocessing pipeline (but before denoising), FC distributions show extremely large inter-session and inter-subject variability, and skewed distributions with varying degrees of positive biases, consistent with the influence of global or large-scale physiological and subject-motion effects. After denoising FC distributions show approximately centered distributions, with small but noticeable larger tails in the positive side, and considerably reduced inter-session and inter-subject heterogeneity.

Another useful way to evaluate the quality of the fMRI data after denoising include computing QC-FC correlations (Ciric et al. 2017). QC-FC correlations look, again, at the connectivity values (FC) between randomly-selected pairs of points within the brain separately for each subject, but instead of simply displaying the distribution of these values, this method evaluates whether these FC values are correlated, across subjects, with individual Quality Control ( QC) measures (e.g. subject-motion indicators). The method computes a QC-FC correlation value for each randomly-selected pair of points within the brain, and then it displays the distribution of the resulting QC-FC correlation values (shown in gray before and after denoising in the plots below). The observed QC-FC distribution shape f(r) can then be directly compared to a permutation-derived null distribution f0(r) (shown as red dashed lines in the plots below) in order to evaluate the potential influence of motion-related or other confounds in functional connectivity measures.  The difference between these two distributions is described by their overlap coefficient (Ihman and Bradley 1989), the integrated minimum of two probability density functions.

Alternatives/additions to the default denoising pipeline

While the above denoising pipeline is designed to remove most outlier, physiological, including respiratory and cardiac effects, and residual subject-motion effects from the BOLD signal, researchers are strongly encouraged to evaluate the quality of their data after denoising and consider modifying these or using additional denoising steps as necessary. In addition to simple variations of the default denoising strategy, such as using different scrubbing options (e.g. identifying outliers based on a lower framewise displacement threshold) or selecting a higher number of aCompCor noise components, some useful additional denoising approaches include:

1. ICA denoising is a data-driven approach where Independent Component Analyses are used to identify potential noise-related temporal components  either manually or semi-automatically (Griffanti et al. 2017) which can then be entered in CONN as additional potential confounding effects in the standard Linear Regression denoising step

2. Retroicor (Glover et al 2000) is a popular technique to  use cardiac and respiratory state information recorded during the scanning session to build a series of predicted sine and cosine components of the respiratory and cardiac effects, which can then be used in CONN as additional potential confounding effects in the standard Linear Regression denoising step

3. Simult (Hallquist et al. 2013) is an alternative implementation of the standard sequential regression followed by filtering approach, where both regression and filtering are implemented simultaneously as a single regression step (either by explicitly adding low- and high- frequency regressors or by pre-filtering the matrix of linear regressors). This approach can be used in CONN directly (e.g. switching the default 'RegBP' option to 'Simult'), or it can be used selectively over a subset of regressors only (by selecting the 'Filtered' option on the selected subset of potential confounding effects). Since Simult is equivalent to removing the frequency-specific effect of potential confounder variables (the effect estimated only within the frequency window defined by the band-pass filter) it is recommended to not apply this filtering to confounders such as outliers which are broadband and not frequency-modulated

4. Friston24 (Friston et al. 1996) is a richer set of motion-related regressors designed to remove an autoregressive-moving-average model of the effects of subject motion on the BOLD signal. In CONN this can be used simply selecting a  polynomial expansion (quadratic effects) for the realignment parameters, in addition to the temporal expansion (first-order derivatives) that is already part of the default denoising strategy

5. Global Signal Regression (GSR) is an alternative approach which uses the average BOLD signal (across the entire brain) as a potential confounding effect. This approach is generally not recommended, since it can introduce artifactual biases (Murphy et al. 2009), remove potentially meaningful neural components (Chai et al. 2012), and introduce confounding effects across populations. It is nevertheless possible to use as a last resort or reference alternative approach in CONN simply defining  a new ROI encompassing the entire brain (e.g. a subject-specific brain mask resulting from the outlier detection step during preprocessing, or gray matter mask resulting from the segmentation step) and using this ROI as an additional potential confounding effect in the standard Linear Regression denoising step

References

Behzadi, Y., Restom, K., Liau, J., & Liu, T. T. (2007). A component based noise correction method (CompCor) for BOLD and perfusion based fMRI. Neuroimage, 37(1), 90-101

Chai, X. J., Castañón, A. N., Öngür, D., & Whitfield-Gabrieli, S. (2012). Anticorrelations in resting state networks without global signal regression. Neuroimage, 59(2), 1420-1428

Morfini F, Whitfield-Gabrieli S and Nieto-Castañón A (2023) Functional connectivity MRI quality control procedures in CONN. Front. Neurosci. 17:1092125.

Ciric, R., Wolf, D. H., Power, J. D., Roalf, D. R., Baum, G. L., Ruparel, K., ... & Gur, R. C. (2017). Benchmarking of participant-level confound regression strategies for the control of motion artifact in studies of functional connectivity. Neuroimage, 154, 174-187

Friston, K. J., Williams, S., Howard, R., Frackowiak, R. S., & Turner, R. (1996). Movement‐related effects in fMRI time‐series. Magnetic resonance in medicine, 35(3), 346-355.

Glover, G. H., Li, T. Q., & Ress, D. (2000). Image‐based method for retrospective correction of physiological motion effects in fMRI: RETROICOR. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, 44(1), 162-167.

Nieto-Castanon, A. (2025). Preparing fMRI data for statistical analysis. In fMRI techniques and protocols (pp. 163-191). New York, NY: Springer US.

Inman, H. F., & Bradley Jr, E. L. (1989). The overlapping coefficient as a measure of agreement between probability distributions and point estimation of the overlap of two normal densities. Communications in Statistics-theory and Methods, 18(10), 3851-3874.

Satterthwaite, F. E. (1946). An approximate distribution of estimates of variance components. Biometrics bulletin, 2(6), 110-114.

Griffanti, L., Douaud, G., Bijsterbosch, J., Evangelisti, S., Alfaro-Almagro, F., Glasser, M. F., ... & Beckmann, C. F. (2017). Hand classification of fMRI ICA noise components. Neuroimage, 154, 188-205.

Hallquist, M. N., Hwang, K., & Luna, B. (2013). The nuisance of nuisance regression: spectral misspecification in a common approach to resting-state fMRI preprocessing reintroduces noise and obscures functional connectivity. Neuroimage, 82, 208-225

Murphy, K., Birn, R. M., Handwerker, D. A., Jones, T. B., & Bandettini, P. A. (2009). The impact of global signal regression on resting state correlations: are anti-correlated networks introduced?. Neuroimage, 44(3), 893-905.

Power, J. D., Mitra, A., Laumann, T. O., Snyder, A. Z., Schlaggar, B. L., & Petersen, S. E. (2014). Methods to detect, characterize, and remove motion artifact in resting state fMRI. Neuroimage, 84, 320-341

Whitfield-Gabrieli, S., & Nieto-Castanon, A. (2012). Conn: a functional connectivity toolbox for correlated and anticorrelated brain networks. Brain connectivity, 2(3), 125-141