Signal Loss and Artifacts in Neuroimaging Denoising

25 concepts19 papers

When recording brain activity, neuroscientists face a fundamental challenge: the signals of interest are buried in a sea of noise. Whether it is the rhythmic throbbing of a heartbeat, the expansion of the chest during respiration, or subtle head movements inside the scanner, non-neural physiological processes heavily contaminate the data.

To isolate true neural communication, researchers rely on denoising pipelines. However, a growing body of literature reveals that aggressive data cleaning is a double-edged sword. Because noise and legitimate neural signals overlap in time, space, and frequency, preprocessing choices can inadvertently introduce mathematical artifacts, destroy meaningful biological variance, and alter the conclusions of a study.

The Mathematical Trap of Global Signal Regression

One of the most common and controversial tools in functional magnetic resonance imaging (fMRI) preprocessing is Global Signal Regression (GSR). The intuition behind GSR is straightforward: calculate the average time-series of all voxels across the entire brain (the "global signal") to serve as a proxy for widespread, non-neural noise—such as respiratory fluctuations—and regress it out of the data.

While GSR is highly effective at removing motion and breathing artifacts, it introduces a severe mathematical distortion. Mathematically, subtracting the global average from every voxel forces the sum of all correlations across the brain to equal zero. This shifts the entire distribution of correlation coefficients downward.

[ True Positive Correlations ]  ---> GSR Shifts Distribution --->  [ Spurious Negative Correlations ]
       (All regions sync'd)              (Mean forced to 0)             (Artificial anti-correlations)

As a result, regions that were originally uncorrelated or weakly positively correlated are mechanically driven into negative territory, creating "spurious anti-correlations" that do not exist biologically.

Chong-Yaw Wee et al. highlighted this exact tradeoff in their work on Resting-State Multi-Spectrum Functional Connectivity Networks for Identification of MCI Patients, noting that the introduction of artificial negative correlations complicates the interpretation of functional networks. The real-world impact of this shift was demonstrated by Csaba Orbán et al. in Correction: Time of day is associated with paradoxical reductions in global signal fluctuation and functional connectivity, where the authors showed that applying GSR fundamentally alters the observed relationship between scan timing and brain connectivity, suppressing certain correlations while manufacturing entirely new ones.

When Noise Carries Meaning: The Biological Cost

The danger of over-cleaning extends beyond mathematical artifacts; it also risks discarding the very biology researchers want to study. For years, the global signal was dismissed as a nuisance parameter. However, macro-scale brain activity is not entirely noise.

In Topography and behavioral relevance of the global signal in the human brain, Jingwei Li et al. demonstrated that the spatial distribution of the global signal tracks major functional networks and significantly predicts an individual's cognitive abilities and life outcomes. When preprocessing pipelines aggressively strip away this global variance, they inadvertently eliminate the neural correlates of individual behavioral and cognitive traits.

This loss of diagnostic information is particularly problematic when studying psychiatric and neurological conditions:

The Over-Cleaning Trap in Machine Learning

The impact of preprocessing choices is magnified when neuroimaging data is fed into predictive models. If a data-cleaning step removes global metrics to isolate specific local features, it can inadvertently break the model's predictive power.

This phenomenon was clearly illustrated in the ABCD Neurocognitive Prediction Challenge 2019: Predicting Individual Residual Fluid Intelligence Scores from Cortical Grey Matter Morphology. To isolate fine-grained structural predictors of intelligence, Neil P. Oxtoby et al. regressed out global brain metrics and demographic factors from cognitive scores. However, because intelligence is intrinsically linked to global brain properties, this residualization process stripped away the primary biological signals required for prediction, leaving a target variable that lacked any detectable relationship with the remaining structural features.

In fact, preprocessing choices can have a larger footprint on the final data than the biological variables under investigation. Jae Wook Cho et al. showed in Impact of Concatenating fMRI Data on Reliability for Functional Connectomics that denoising strategies like GSR exert a more dominant influence on functional connectivity patterns than the actual brain state or task performed during the scan. When the cleaning process introduces more variance than the experimental condition, the validity of the downstream analysis is compromised.

Overlapping Domains: The Spectral and Temporal Challenge

Why is it so difficult to separate noise from signal? The core issue is that physiological artifacts do not neatly occupy a separate domain from neural activity; they share the same temporal and spectral frequencies.

In electrophysiological recordings like electroencephalography (EEG) performed inside an MRI scanner, the heartbeat induces massive ballistocardiogram (BCG) and pulse artifacts. As Eleonora Maggioni et al. observed in Removal of Pulse Artefact from EEG Data Recorded in MR Environment at 3T. Setting of ICA Parameters for Marking Artefactual Components: Application to Resting-State Data, these artifacts overlap directly with genuine neural oscillations like alpha rhythms. Aggressive filtering or poorly calibrated independent component analysis (ICA) can easily delete these critical brain rhythms.

Similarly, when processing magnetoencephalography (MEG) or EEG data, standard high-pass filtering meant to eliminate slow scanner drift can distort or suppress late-occurring, low-frequency cognitive neural components, as detailed by Mainak Jas et al. in A Reproducible MEG/EEG Group Study With the MNE Software: Recommendations, Quality Assessments, and Good Practices.

Frequency Spectrum:
|-------------------|-------------------|-------------------|
0 Hz             0.1 Hz              0.5 Hz              1.0 Hz
[  Scanner Drift  ] [  Neural Signals ] [   Respiration   ]
                    [ <--- Overlap ---> ]

Towards Adaptive and Selective Denoising

To overcome the blunt-force limitations of traditional preprocessing, modern neuroimaging is moving toward adaptive, data-driven filtering techniques that isolate noise without destroying signal architecture. This is especially vital in low signal-to-noise environments, such as in utero imaging, where extreme fetal motion requires highly selective filtering pipelines (Vyacheslav Karolis et al. in The developing Human Connectome Project fetal functional MRI release: Methods and data structures).

Rather than applying uniform regressions, researchers are developing methods that adapt to individual physiology:

When carefully calibrated, standard workflows can still preserve the underlying stability of the brain's architecture. For instance, Linlin Zhu et al. demonstrated in Temporal Reliability and Lateralization of the Resting-State Language Network that the core functional architecture of the language network remains highly reliable across sessions regardless of whether GSR is applied.

The consensus across recent neuroimaging literature is clear: denoising is not a neutral administrative step. Every filter, regression, and artifact rejection tool alters the underlying geometry of the data. Building reliable models of the human brain requires moving away from rigid, aggressive cleaning recipes and toward adaptive pipelines that respect the thin, overlapping boundary between noise and signal.

Key Foundational Work

Resting-State Multi-Spectrum Functional Connectivity Networks for Identification of MCI Patients (2012) — contributed 2 concept elements to this direction.
Convergent and Divergent Functional Connectivity Patterns in Schizophrenia and Depression (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_global_signal_regression_artifacts_0f219e07,
  title = {global signal regression artifacts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/global-signal-regression-artifacts/rdr_0f219e079df942199598ab00370f91e7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_global_signal_regression_artifacts_0f219e07_v1,
  title = {global signal regression artifacts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/global-signal-regression-artifacts/rdr_0f219e079df942199598ab00370f91e7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake