Motion Parameter Regression for fMRI Denoising: A Comprehensive Guide for Researchers and Clinicians

Abstract

Motion artifact remains a significant impediment to robust functional Magnetic Resonance Imaging (fMRI) analysis, particularly in clinical and developmental populations. This article provides a contemporary, comprehensive guide to motion parameter regression, a cornerstone denoising technique. We explore the foundational principles of how motion corrupts the BOLD signal and systematically review established and emerging denoising strategies, from basic parameter regression to advanced automated methods like ICA-AROMA and CICADA. A strong emphasis is placed on practical troubleshooting, pipeline optimization for specific populations (e.g., TBI, older adults), and rigorous validation using a multi-metric framework. Designed for researchers, scientists, and drug development professionals, this review synthesizes recent large-scale benchmarking studies to offer evidence-based recommendations for improving the reliability and reproducibility of fMRI findings in biomedical research.

The Motion Problem: Foundations and Impact on fMRI Signal Fidelity

Characteristics and Impact of Motion Artifacts

Motion artifacts represent a major methodological challenge in functional magnetic resonance imaging (fMRI), particularly in studies examining blood oxygenation level-dependent (BOLD) signal correlations and functional connectivity [1]. When subjects move during scanning, the resulting artifacts can systematically bias functional connectivity measures, potentially leading to false scientific conclusions [1] [2]. These artifacts are especially problematic because in-scanner motion frequently correlates with variables of scientific interest such as age, clinical status, cognitive ability, and symptom severity [1].

The spatial distribution of motion artifacts follows a predictable pattern, with minimal movement near the atlas vertebrae (where the skull attaches to the neck) and increasing motion with distance from this anchor point [1]. Frontal cortex regions typically exhibit particularly high motion, likely due to the prevalence of y-axis rotation associated with nodding movements [1]. The temporal properties of motion artifacts include both immediate, large-amplitude signal changes following movement events and longer-duration artifacts that may persist for 8-10 seconds, potentially due to motion-related changes in CO₂ from yawning or deep breathing [1].

Motion artifacts introduce distance-dependent biases in inferred signal correlations, where even small residual motion artifacts continue to corrupt BOLD signal correlations throughout the brain after standard correction approaches [2]. These artifacts manifest through several mechanisms: changes in tissue composition within voxels, distortions of the magnetic field, disruption of steady-state magnetization recovery in slices that have moved, signal dropouts, and artifactual amplitude changes across brain regions [2].

Measuring and Quantifying Motion Artifacts

In-scanner motion is typically estimated from the functional time series itself during preprocessing. Each volume in the time series is rigidly realigned to a reference volume, producing six realignment parameters (RPs) that describe how much a given volume must be moved relative to the reference [1]. These parameters are commonly summarized as frame displacement (FD), which computes relative movement from one volume to the next, providing a concise index of volume-to-volume motion [1].

Table 1: Common Metrics for Quantifying Head Motion in fMRI

Metric	Calculation	Interpretation	Limitations
Frame Displacement (FD) [1]	Derived from 6 realignment parameters (3 translations + 3 rotations); summarizes volume-to-volume movement	Higher FD values indicate greater motion; typically thresholded at 0.2-0.5mm for censoring	Difficult to compare across studies with different TRs; limited temporal resolution
Voxel-specific FD [1]	Computed directly from image header for specific voxels	Accounts for spatial variation in motion effects; maximal in frontal regions	Highly correlated with global FD measures (r ≈ 0.89)
Standardized FD (e.g., mm/minute) [1]	Normalizes FD by acquisition time	Enables comparison across studies with different repetition times (TR)	Not yet widely adopted in literature

Different methods for calculating FD exist, with various formulations showing high correlations but different scaling properties [1]. The advent of multiband imaging with shorter repetition times (TR) has complicated direct comparison of FD values across studies, prompting suggestions to convert FD into standardized measures such as millimeters of RMS displacement per minute [1].

Experimental Protocols for Motion Artifact Characterization

Protocol: Assessing Spatial Distribution of Motion Artifacts

Purpose: To quantify the spatial distribution of motion artifacts and their relationship to anatomical constraints.

Materials and Methods:

• Acquire resting-state fMRI data using standard parameters (e.g., TR=2000ms, TE=30ms, 30 slices)
• Extract voxel-specific frame displacement measures directly from image headers [1]
• Coregister motion estimates with anatomical reference scans
• Calculate correlation between global and voxel-specific FD measures
• Analyze motion gradients relative to distance from atlas vertebrae

Expected Outcomes: Motion is typically minimal near the atlas vertebrae and increases with distance from this anchor point, with particularly high values in frontal regions due to nodding movements [1]. The correlation between voxel-specific and global FD measures is typically high (approximately r=0.89) [1].

Protocol: Temporal Characteristics of Motion Artifacts

Purpose: To characterize immediate and prolonged effects of motion on BOLD signal.

Materials and Methods:

• Acquire fMRI data with simultaneous monitoring of movement parameters
• Identify motion events exceeding threshold (e.g., FD > 0.2mm)
• Analyze signal trajectories for 10 seconds preceding and 20 seconds following motion events
• Quantify signal drop magnitude as a function of motion magnitude
• Document duration of signal disruptions following motion events

Expected Outcomes: Motion typically results in substantial, immediate signal drops that scale with motion magnitude, with maximal effect at the volume acquired immediately after movement [1]. Longer-duration artifacts (8-10 seconds) occur sporadically, potentially due to motion-related physiological changes [1].

Diagram 1: Temporal effects of motion on BOLD signal.

Denoising Pipelines and Methodological Comparisons

Multiple denoising approaches have been developed to mitigate motion artifacts in fMRI data, each with relative strengths and weaknesses. No single pipeline universally excels across all datasets and research objectives, requiring researchers to select methods based on their specific needs [3].

Table 2: Comparison of Common fMRI Denoising Pipelines

Method	Mechanism	Advantages	Disadvantages	Impact on BWAS
Volume Censoring ("Scrubbing") [2] [4]	Removes motion-corrupted volumes from analysis	Effectively removes severe motion artifacts; simple implementation	Creates discontinuities; reduces temporal degrees of freedom	Mixed effects on brain-behavior association strength
ICA-AROMA [3] [5]	Data-driven classification and removal of motion-related independent components	Preserves temporal continuity; automated classification	May remove neural signal in aggressive mode; computational intensity	Reasonable trade-off between motion reduction and behavioral prediction
Global Signal Regression (GSR) [1] [3]	Regresses whole-brain average signal from time series	Effectively reduces motion-related variance	Controversial - may remove neural signal; alters correlation structure	Can enhance behavioral prediction in combination with other methods
Structured Low-Rank Matrix Completion [2] [6]	Recovers censored data using matrix completion algorithms	Maintains data continuity; provides slice-time correction	High computational demand; memory intensive	Not fully established for BWAS
Short Echo Time Regression [7]	Uses short TE data to regress noise from BOLD-weighted time series	Effectively removes physiological noise; acquisition is "free"	Potential BOLD contamination in short TE data	Not widely evaluated for BWAS

Recent comprehensive evaluations of denoising pipelines reveal that combinations of methods often provide the most balanced approach. Pipelines combining ICA-FIX and GSR demonstrate a reasonable trade-off between motion reduction and behavioral prediction performance, though inter-pipeline variations in predictive performance are generally modest [3]. For older adult populations (60-85 years), aggressive ICA-AROMA has been identified as particularly effective, considering reproducibility as the most important factor for longitudinal studies [5].

Protocol for Pipeline Comparison and Optimization

Purpose: To systematically evaluate denoising pipeline performance for specific research contexts and datasets.

Materials and Methods:

• Select multiple denoising pipelines representing diverse approaches (e.g., censoring, ICA-AROMA, GSR, aCompCor)
• Process identical dataset through each pipeline
• Evaluate performance using multiple metrics:
  • QC-FC correlations (correlation between motion and functional connectivity)
  • Network reproducibility and identifiability
  • Temporal degrees of freedom (tDOF) loss
  • Edge activity and spatial smoothness
  • Brain-behavior association strength [3]
• For volume censoring, determine dataset-specific optimal parameters (FD threshold, contiguous frames retained) [4]

Expected Outcomes: Pipeline performance varies across datasets and research objectives. Censoring-based pipelines often show strong motion reduction but substantial data loss, while ICA-based approaches provide more balanced performance [3] [5]. Quantitative metrics should guide pipeline selection rather than relying on default parameters.

Diagram 2: Denoising pipeline selection and evaluation workflow.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Tools for Motion Artifact Research in fMRI

Tool/Resource	Function	Example Implementation	Considerations
Frame Displacement Calculators	Quantifies volume-to-volume head motion	FSL (Jenkinson et al.), Power et al. implementation	Different formulations correlate highly but scale differently
Volume Censoring Tools	Identifies and removes motion-corrupted volumes	"Scrubbing" with FD/DVARS thresholds	Optimize threshold for specific datasets; typical FD threshold: 0.2-0.5mm
ICA-Based Denoising Packages	Automated identification of motion components	ICA-AROMA, ICA-FIX, SOCK	ICA-AROMA preferred for multi-site studies without retraining
Structured Matrix Completion Algorithms	Recovers missing data after censoring	Low-rank Hankel matrix completion	High memory demand; provides slice-time correction
Prospective Motion Correction	Real-time motion tracking and correction	Fetal head tracking with U-Net segmentation	23% increase in temporal SNR demonstrated [8]
Dual-Echo Sequences	Simultaneous acquisition of BOLD and short-TE data	TE = 3.3ms (short) and 35ms (BOLD)	Effectively removes physiological noise without additional scan time [7]
Multi-Band Acquisition	Simultaneous multi-slice imaging	HCP-style protocols	Improved temporal resolution but poses unique denoising challenges [4]

Emerging Approaches and Future Directions

Prospective Motion Correction

Emerging approaches focus on preventing motion artifacts rather than removing them during post-processing. Real-time fetal head motion tracking represents an advanced prospective motion correction (PMC) system that integrates U-Net-based segmentation and rigid registration to track head motion and adjust slice positioning in real-time [8]. This approach has demonstrated a 23% increase in temporal signal-to-noise ratio and a 22% increase in Dice similarity index in fMRI time series compared to uncorrected data [8].

Structured Matrix Completion

Novel reconstruction-based approaches address the limitations of censoring by recovering missing entries using structured low-rank matrix completion [2] [6]. This method formulates the artifact-reduction problem as recovery of a super-resolved matrix from unprocessed fMRI measurements, enforcing a low-rank prior on a large structured matrix formed from time series samples [2]. This approach not only compensates for motion but also provides slice-time correction at fine temporal resolution.

Dataset-Specific Pipeline Optimization

Growing evidence suggests that optimal denoising approaches vary across datasets and populations [4] [5]. Quantitative methods for determining dataset-specific optimal parameters prior to final analysis are emerging, with recommendations for tailored application to specific RSFC datasets [4]. This recognizes that motion and physiological noise characteristics differ substantially across populations, such as between healthy young adults and older adults at risk for Alzheimer's disease [5].

Consensus Recommendations and Best Practices

Current consensus recommendations emphasize that clinical fMRI applications require special consideration of motion artifact mitigation [9]. For clinical language mapping, ensuring data quality through effective motion correction is essential for valid surgical planning [9]. Different paradigms elicit varying degrees of motion, with sensory stimulation causing fewer artifacts than motor tasks, though with reduced sensitivity [10].

The field is moving toward greater transparency and reporting of processing pipelines, with the strongest recommendation being for detailed documentation of methods and outcomes [9]. This facilitates comparison across studies and enables much-needed evaluation of ultimate clinical goals, including minimization of postoperative deficits through accurate functional mapping [9].

Resting-state functional connectivity magnetic resonance imaging (rs-fcMRI) provides critical insights into the brain's intrinsic functional architecture by measuring temporal correlations in blood oxygen level-dependent (BOLD) signals between different brain regions. However, this powerful technique remains vulnerable to a pervasive confound: in-scanner head motion. Even sub-millimeter movements systematically alter fMRI data, introducing spurious correlation structures that can profoundly distort connectivity findings [11] [12]. Despite spatial registration and regression of motion parameters, motion-related artifacts persist in processed data, masquerading as biologically plausible connectivity patterns [11].

This Application Note details the mechanisms through which motion induces systematic bias in functional connectivity analyses, provides quantitative frameworks for assessing these artifacts, and outlines robust methodological protocols to mitigate their impact. Understanding these biases is particularly crucial for studies involving populations prone to increased movement (e.g., children, older adults, or individuals with neurological disorders), where motion can create spurious group differences [11] [12].

Quantitative Evidence of Systematic Motion Effects

Characteristic Patterns of Motion-Induced Connectivity

Head motion does not introduce random noise but produces highly structured artifacts with predictable spatial patterns. Quantitative analyses reveal that motion systematically alters correlation structures throughout the brain:

Table 1: Characteristic Effects of Motion on Functional Connectivity

Connectivity Type	Effect of Motion	Representative Change	Network Impact
Long-Distance Correlations	Substantial decrease	Reduced inter-hemispheric connectivity [11]	Default Mode Network disruption [12]
Short-Distance Correlations	Significant increase	Elevated local correlations [11]	Altered local network topology
Default Mode Network	Selective decrease	Reduced PCC-mPFC connectivity [12]	Impaired network integration

The motion-FC effect matrix demonstrates a strong negative correlation (Spearman ρ = -0.58) with the average FC matrix, indicating that participants who move more show systematically weaker connections across the brain [12]. This bias arises because head movement fundamentally disrupts the spin history assumptions underlying BOLD signal acquisition, creating intensity changes that spatial realignment cannot fully correct [11].

Efficacy of Denoising Pipelines

Multiple denoising strategies have been developed to mitigate motion artifacts, yet none completely eliminate the problem. Recent comprehensive evaluations demonstrate the variable efficacy of different approaches:

Table 2: Performance of fMRI Denoising Pipelines for Motion Artifact Removal

Denoising Method	Key Mechanism	Impact on Motion Artifacts	Effect on Brain-Behavior Correlations
Motion Parameter Regression	Regresses out head position estimates	Partial artifact reduction [11]	Variable effects on validity [3]
Global Signal Regression (GSR)	Removes global signal average	Reduces distance-dependent artifacts [3]	Can attenuate behavioral correlations [3]
ICA-Based Cleaning (e.g., ICA-FIX)	Identifies and removes noise components	Effective for structured noise [3]	Preserves behavioral prediction [3]
Volume Censoring ("Scrubbing")	Removes high-motion timepoints	Dramatically reduces spurious correlations [11] [12]	Risk of biasing sample distributions [12]
DiCER	Diffuse cluster estimation and regression	Targets widespread motion effects	Moderate performance for behavioral prediction [3]

No single pipeline universally excels at both motion mitigation and preservation of biological signals. Pipelines combining ICA-FIX and GSR often represent a reasonable trade-off, though inter-pipeline variations in predictive performance remain modest [3].

Experimental Protocols for Motion Artifact Assessment

Protocol 1: Framewise Displacement and DVARS Calculation

Purpose: To quantify head motion at each timepoint (frame) of the fMRI acquisition.

Materials: Preprocessed fMRI time series; head realignment parameters (3 translational, 3 rotational).

Procedure:

• Compute Framewise Displacement (FD):
  • Calculate the absolute differences in each of the 6 realignment parameters between consecutive timepoints: ΔX, ΔY, ΔZ, Δα, Δβ, Δγ.
  • Convert rotational displacements from radians to millimeters by assuming a brain radius of 50 mm (or mean head radius of your cohort): Δα_mm = 50 * Δα, etc.
  • Sum the absolute values: FD = |ΔX| + |ΔY| + |ΔZ| + |Δα_mm| + |Δβ_mm| + |Δγ_mm|.

• Compute DVARS:

  • Calculate the root mean square of the voxel-wise differentiated time series (from timepoint t to t+1).
  • DVARS represents the rate of change of BOLD signal across the entire brain.

• Identify High-Motion Volumes:

  • Flag timepoints where FD exceeds a predetermined threshold (e.g., 0.2-0.5 mm) [12] or where DVARS values are outliers.

Interpretation: FD provides a scalar summary of head movement between volumes. DVARS captures signal changes potentially induced by motion. These metrics form the basis for volume censoring.

Protocol 2: Volume Censoring (Scrubbing)

Purpose: To exclude high-motion timepoints from functional connectivity analysis.

Materials: fMRI time series; computed FD and DVARS values.

Procedure:

• Define Censoring Threshold:
  • Set an FD threshold (e.g., 0.2 mm) based on data quality and research questions [12].
  • Consider also using a normalized DVARS threshold.

• Flag Volumes for Censoring:

  • Identify volumes where FD exceeds the threshold.
  • Extend censoring to include one volume before and two volumes after each high-motion volume to account for hemodynamic response lag.

• Compute Data Retention Metrics:

  • Calculate the percentage of volumes retained for each participant.
  • Exclude participants with excessive data loss (e.g., <50% volumes remaining) from analysis.

• Conduct Connectivity Analysis:

  • Compute correlation matrices using only retained volumes.
  • Account for unequal number of timepoints across participants in group analyses.

Interpretation: Censoring significantly reduces spurious motion-related correlations. However, aggressive censoring may bias participant inclusion and alter sample characteristics [12].

Protocol 3: SHAMAN for Trait-Specific Motion Impact

Purpose: To quantify how motion impacts specific brain-behavior relationships using Split Half Analysis of Motion Associated Networks [12].

Materials: Resting-state fMRI data; trait measures of interest; motion parameters.

Procedure:

• Split fMRI Timeseries:
  • Divide each participant's cleaned fMRI data into high-motion and low-motion halves based on median FD.

• Compute Trait-FC Effects:

  • Calculate functional connectivity matrices separately for high-motion and low-motion halves.
  • Compute correlation between trait measures and FC for each half.

• Calculate Motion Impact Score:

  • Test for significant differences between trait-FC effects in high-motion vs. low-motion halves.
  • A positive score aligned with the trait-FC effect indicates motion overestimation.
  • A negative score opposite the trait-FC effect indicates motion underestimation.

• Statistical Testing:

  • Use permutation testing (e.g., 1000 permutations) to assess significance of motion impact scores.
  • Apply false discovery rate correction for multiple comparisons.

Interpretation: SHAMAN specifically evaluates whether trait-FC relationships are confounded by motion, helping prevent false positive and false negative conclusions [12].

Visualization of Motion Artifact Mechanisms

Figure 1: Pathway of Motion-Induced Artifacts in fMRI

Figure 2: Volume Censoring Workflow for Motion Mitigation

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Tools for Motion Artifact Management in fcMRI

Tool/Category	Specific Examples	Function	Implementation Considerations
Motion Quantification	Framewise Displacement (FD), DVARS	Quantifies frame-by-frame head movement	FD threshold of 0.2 mm recommended for censoring [12]
Real-Time Monitoring	FIRMM, FIXEL	Monitors motion during scanning	Enables prospective intervention
Denoising Algorithms	ICA-AROMA, FIX, ABCD-BIDS	Removes motion-related variance from data	ABCD-BIDS includes respiratory filtering, GSR, despiking [12]
Statistical Tools	SHAMAN, QC-FC correlations	Quantifies motion impact on specific findings	SHAMAN distinguishes over/underestimation [12]
Data Censoring	"Scrubbing", Spike Regression	Removes high-motion timepoints	Balances artifact reduction with data retention [11]

Motion induces systematic spatial patterns of spurious functional connectivity that cannot be eliminated by standard preprocessing alone. Robust mitigation requires a multi-pronged approach:

• Implement rigorous volume censoring with an FD threshold of 0.2 mm, while monitoring data retention rates to avoid exclusion biases [12].
• Apply validated denoising pipelines such as those combining ICA-FIX with global signal regression to balance artifact removal and signal preservation [3].
• Quantify trait-specific motion impacts using methods like SHAMAN, particularly for traits correlated with movement propensity (e.g., psychiatric symptoms) [12].
• Report motion metrics comprehensively including mean FD, data retention rates, and methods used for motion correction to enable evaluation and replication.

These protocols provide a framework for identifying and addressing motion-related confounds, strengthening the validity of functional connectivity findings in basic neuroscience and drug development research.

In-scanner head motion represents the most substantial source of artifact in functional magnetic resonance imaging (fMRI) signals, introducing systematic bias into resting-state functional connectivity (FC) measurements that cannot be completely eliminated by standard denoising algorithms [12]. This technical challenge is particularly acute when studying clinical and developmental populations who exhibit elevated motion characteristics due to their underlying conditions. Research has demonstrated that participants with neurological, psychiatric, or developmental conditions consistently display higher in-scanner head motion compared to neurotypical participants, creating a systematic confounding variable that can generate spurious brain-behavior associations [12] [13]. For example, early studies erroneously concluded that autism decreases long-distance functional connectivity when these findings were actually attributable to increased head motion in autistic study participants [12]. This confounding effect extends to numerous clinical populations, including traumatic brain injury (TBI) patients who present frequent abnormal movements such as posturing, shivering, tremors, dystonia, and seizures [13]. Understanding these motion correlates across populations is therefore essential for ensuring the validity of fMRI research in clinical neuroscience and drug development.

Quantitative Data on Motion Correlates Across Populations

Motion Prevalence in High-Risk Cohorts

Table 1: Motion Characteristics Across Clinical and Developmental Populations

Population	Motion Correlates	Impact on FC	Data Source
Traumatic Brain Injury (TBI)	Frequent abnormal movements (posturing, shivering, seizures, tremor, dystonia)	Significant artifact in functional connectivity; extensive brain pathology leads to suboptimal performance of conventional software	EpiBioS4Rx Study (n=88) [13]
Developmental Disorders (ADHD, Autism)	Significantly higher in-scanner head motion than neurotypical participants	Spurious decreases in long-distance connectivity; systematic bias in trait-FC relationships	ABCD Study [12]
Psychiatric Populations	Increased motion associated with symptom severity	Altered connectivity measures that may reflect motion artifacts rather than neural correlates	ds000030 Dataset [14]
Pediatric Cohorts	Age inversely related to motion; children show higher motion	Artifactual correlations across brain regions; requires specialized denoising approaches	ABCD Study (n=11,874) [15] [12]

Efficacy of Denoising Strategies

Table 2: Performance of Denoising Pipelines Across High-Motion Populations

Denoising Strategy	Residual Motion Artifact	Data Loss	Recommended Use Cases
Spike Regression + Physiological Regressors	Relatively effective performance	Moderate	TBI populations after high-motion participant exclusion [13]
Volume Censoring (FD < 0.2 mm)	Reduces motion overestimation to 2% of traits	Substantial, may bias sample distribution	When continuous sampling not required [12] [16]
ICA-AROMA	Performs well across benchmarks	Low cost in terms of data loss	General use, balanced approach [16]
Global Signal Regression	Improves performance of most pipelines	Low	Can be combined with other methods despite distance-dependence exacerbation [16]
aCompCor	Only viable in low-motion data	Low	Limited utility for high-risk populations [16]
ABCD-BIDS (Standard Pipeline)	23% of signal variance explained by motion after denoising (vs. 73% before)	Moderate	Large-scale studies like ABCD [12]

Experimental Protocols for Motion Mitigation

Protocol 1: Trait-Specific Motion Impact Assessment Using SHAMAN

Purpose: To quantify motion impact on specific trait-FC relationships using Split Half Analysis of Motion Associated Networks (SHAMAN) [12].

Materials:

• Resting-state fMRI data (minimum 8 minutes per participant)
• Framewise displacement (FD) calculations
• Behavioral/cognitive trait measures
• Computing environment with SHAMAN implementation

Procedure:

• Data Acquisition: Acquire rs-fMRI data using standardized protocols (e.g., ABCD Study parameters: 2-4 runs, 20 minutes total) [12].
• Motion Quantification: Calculate framewise displacement for each participant across all timepoints.
• Data Splitting: For each participant, split the fMRI timeseries into high-motion and low-motion halves based on median FD.
• Connectivity Calculation: Compute separate FC matrices for high-motion and low-motion halves.
• Trait-FC Effect Estimation: Calculate correlation between trait measures and FC for both halves.
• Impact Score Calculation:
  • Compute difference in trait-FC effects between high and low-motion halves
  • Aligned direction = motion overestimation score
  • Opposite direction = motion underestimation score
• Statistical Testing: Use permutation testing (e.g., 1000 permutations) and non-parametric combining to derive p-values.

Validation: Apply to negative control traits (those theoretically unrelated to motion) to confirm specificity.

Protocol 2: Denoising Pipeline Evaluation for Clinical Populations

Purpose: To evaluate and select optimal denoising strategies for clinical populations with high motion [13] [14].

Materials:

• fMRIPrep software (version 1.4+)
• Nilearn Python library (version 0.9.0+)
• Custom code for quality control metrics
• Clinical population data (e.g., TBI, psychiatric cohorts)

Procedure:

• Data Preprocessing: Process all data through fMRIPrep with standardized settings.
• Pipeline Implementation: Apply multiple denoising pipelines including:
  • 24HMP (Friston-24): 6 motion parameters + derivatives + squares [13]
  • Spike regression (FDJenk >0.25mm) + physiological regressors [13]
  • ICA-AROMA [16]
  • aCompCor [16]
  • Global signal regression + censoring [14]
• Quality Control Assessment: Evaluate each pipeline using:
  • Residual relationship between motion and FC
  • Distance-dependent effects of motion on FC
  • Test-retest reliability of FC estimates
  • Group differences between high- and low-motion participants
• Performance Benchmarking: Rank pipelines based on ability to mitigate motion artifacts while preserving neural signal.
• Strategy Selection: Choose optimal pipeline based on specific research question and participant characteristics.

Signaling Pathways and Workflows

Diagram 1: Motion Artifact Assessment Workflow for High-Risk Populations

Diagram 2: Denoising Pipeline Implementation Workflow

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Tools for Motion Correction in fMRI Research

Tool/Software	Function	Application Context
fMRIPrep (v1.4+)	Standardized fMRI preprocessing	Generates consistent confound variables across studies; foundation for denoising pipelines [14]
Nilearn (v0.9.0+)	Python library for neuroimaging analysis	Implements load_confounds API for flexible denoising strategy application [14]
SHAMAN Algorithm	Trait-specific motion impact quantification	Determines whether motion causes overestimation or underestimation of trait-FC effects [12]
Framewise Displacement (FD)	Head motion quantification	Standard metric for identifying high-motion timepoints and participants [12]
ICA-AROMA	Data-driven noise component identification	Automatically identifies and removes motion-related independent components [16]
ABCD-BIDS Pipeline	Standardized denoising for large datasets	Incorporates global signal regression, respiratory filtering, and motion parameter regression [12]

Head motion is a significant and pervasive challenge in functional magnetic resonance imaging (fMRI) studies, particularly in resting-state functional connectivity MRI (rs-fcMRI). Even small, transient subject movements can introduce systematic but spurious correlation structures throughout the brain, fundamentally altering the interpretation of functional connectivity [11]. These motion-induced signal changes are problematic because they are often complex and variable waveforms that can persist for more than 10 seconds after the physical movement has ceased, and are frequently shared across nearly all brain voxels [17]. The impact is not merely cosmetic; motion artifacts can increase observed rs-fcMRI correlations in a distance-dependent manner, spuriously strengthening short-distance correlations while weakening long-distance connections [11]. This poses particular challenges for studies involving populations prone to movement, such as pediatric, elderly, or clinical cohorts, potentially creating spurious group differences that confound scientific conclusions [18] [11].

Key Motion Metrics and Their Calculations

Framewise Displacement (FD)

Framewise displacement (FD) is a scalar quantity that quantifies the total head movement between consecutive fMRI volumes [19]. It is derived from the six rigid-body realignment parameters (three translations: X, Y, Z; three rotations: pitch, yaw, roll) obtained during volume registration. FD provides a comprehensive measure of volume-to-volume head movement by calculating the sum of the absolute values of the derivatives of these six parameters [20] [21].

The standard calculation for FD at time point t is:

FD_t = |ΔX_t| + |ΔY_t| + |ΔZ_t| + |Δα_t| + |Δβ_t| + |Δγ_t|

Where rotational displacements (Δα, Δβ, Δγ) are converted from degrees to millimeters by calculating the arc length on a sphere of radius 50 mm, effectively approximating the typical distance from the cerebral cortex to the center of the head [20] [17].

Complementary Motion Metrics

While FD is a crucial metric, comprehensive motion assessment requires additional measures:

• Relative Root Mean Square (RMS): Measures intensity differences between consecutive volumes, providing a complementary measure of signal disruption beyond physical head movement [18].

• Delta Variation Signal (DVARS): Quantifies the rate of change of the BOLD signal across the entire brain at each time point, calculated as the root mean square of the spatial difference of volume It from volume I(t-1) [17].

Table 1: Key Motion Metrics in fMRI Quality Control

Metric	Calculation	Interpretation	Primary Utility
Framewise Displacement (FD)	Sum of absolute derivatives of 6 motion parameters	Total head movement between volumes	Identifying volumes with excessive movement
Relative RMS	RMS of intensity differences between consecutive volumes	Signal disruption from movement	Detecting signal changes independent of physical motion
DVARS	RMS of spatial difference of volume It from I(t-1)	Rate of BOLD signal change across brain	Identifying rapid global signal changes
RMS Movement	Root mean squared head position change	Summary of overall subject motion	Subject-level inclusion/exclusion

Experimental Protocols for Motion Metric Implementation

Real-Time Motion Monitoring Protocol

For challenging populations such as pediatric cohorts, real-time motion monitoring can significantly improve data quality:

• Setup: Utilize real-time monitoring software such as Framewise Integrated Real-time MRI Monitoring (FIRMM) during scanning sessions [18].

• Acquisition: Continue scanning until achieving at least 4 minutes of total data comprised of frames with FD less than 0.4 mm, or until the subject requests to end the session [18].

• Thresholding: Apply a framewise displacement threshold of 0.3 mm for volume censoring in preprocessing pipelines for pediatric populations [18].

Post-Acquisition Quality Control Protocol

Implement a systematic quality control protocol using statistical parametric mapping (SPM) and MATLAB:

• Initial Data Check: Verify consistency of imaging parameters across participants (number of volumes, TR, voxel sizes) and inspect image quality, coverage, and orientations [20].

• Functional Image Realignment: Align all functional images to the first image using rigid-body transformation to obtain motion parameters [20].

• FD Calculation: Compute framewise displacement from motion parameters using the standard formula [20].

• Visualization and Exclusion: Plot FD distributions across the sample and exclude participants with excessive head motions based on predetermined thresholds [20].

Volume Censoring (Scrubbing) Protocol

For datasets with significant motion, implement a censoring approach:

• Identification: Flag volumes exceeding specific FD thresholds (typically 0.2-0.5 mm) as potentially contaminated [18] [17].

• Removal: Remove flagged volumes from subsequent analyses, taking care to maintain temporal structure when possible [18].

• Validation: Ensure sufficient data remains after censoring (e.g., >4 minutes of clean data) for reliable connectivity estimates [18].

Diagram 1: Framewise Displacement Quality Control Workflow. This diagram illustrates the standard processing pipeline for incorporating FD metrics into fMRI quality control, showing decision points for volume censoring and nuisance regression.

Research Reagents and Computational Tools

Table 2: Essential Tools for Motion Metric Calculation and Analysis

Tool/Software	Function	Implementation
SPM (Statistical Parametric Mapping)	Realignment and motion parameter estimation	MATLAB-based; calculates 6 motion parameters
FSL (FMRIB Software Library)	Volume registration and motion correction	MCFLIRT for realignment; FSL motion outliers for FD
AFNI (Analysis of Functional NeuroImages)	Comprehensive fMRI processing	Afni_proc.py script for FD calculation
FIRMM (Framewise Integrated Real-time MRI Monitoring)	Real-time motion tracking	Enables adaptive scanning based on motion thresholds
BRAMILA Toolbox	Framewise displacement calculation	MATLAB package for computing FD from motion parameters
ICA-AROMA	ICA-based automatic removal of motion artifacts	Classifies and removes motion-related components

Threshold Determination and Interpretation

Establishing appropriate FD thresholds is critical for effective motion correction:

• FD > 0.5 mm: Volumes show marked correlation changes and should be rigorously censored [17].

• FD = 0.15-0.2 mm: Significant correlation changes begin to be observed, suggesting a potential lower threshold for stringent analyses [17].

• FD = 0.3 mm: Effective threshold for pediatric cohorts, balancing data quality and retention of subjects (83% participant retention in one study) [18].

It is important to note that QC measure "improvement" during processing may be partially cosmetic - volumes with initially "bad" FD values that become "good" after processing may still harbor residual motion artifact [17].

Integration with Denoising Pipelines

FD metrics are most effective when integrated into comprehensive denoising strategies:

Nuisance Regression Approaches

Traditional nuisance regression incorporates the 6 motion parameters as regressors, sometimes expanded to 12, 24, or 36 regressors by including temporal derivatives and squared terms [22]. However, motion-related signal changes are not completely removed by a variety of motion-based regressors alone [17]. Global signal regression (GSR) has been shown to markedly reduce motion-related variance, though it remains controversial due to potential introduction of artificial anti-correlations [22] [17].

ICA-Based Approaches

ICA-AROMA (Independent Component Analysis-based Automatic Removal Of Motion Artifacts) provides a data-driven approach to motion correction by identifying and removing motion-related components [5]. For older adult populations, aggressive ICA-AROMA has been identified as particularly effective, showing high reproducibility and better preservation of temporal structure [5].

Advanced Regression Models

Emerging approaches include convolutional neural network (CNN) models that derive optimized motion regressors from the basic motion parameters. These models can non-parametrically model the prolonged effects of head motion, potentially outperforming traditional regression approaches [22].

Impact on Functional Connectivity and Statistical Inference

The effects of motion on functional connectivity are systematic and spatially structured:

• Motion generally increases short-distance correlations while decreasing long-distance correlations [11].

• These effects create distance-dependent artifacts that persist after standard motion correction approaches [17].

• Residual motion artifacts can lead to spurious group differences in studies comparing populations with different movement characteristics (e.g., children vs. adults, patients vs. controls) [11].

The most effective approaches for eliminating motion-related artifacts combine volume censoring based on FD thresholds with global signal regression, which together can reduce motion-related group differences to chance levels [17]. However, censoring approaches must be carefully implemented to maintain sufficient temporal data for reliable connectivity estimates, particularly for populations with high motion [18].

Denoising in Practice: From Basic Regression to Advanced Automated Pipelines

Head motion is the largest source of artifact in functional magnetic resonance imaging (fMRI) data, particularly for resting-state functional connectivity (FC) analyses where the timing of underlying neural processes is unknown [12]. Motion artifact systematically alters the blood oxygenation level-dependent (BOLD) signal, causing decreased long-distance connectivity and increased short-range connectivity, most notably in the default mode network [12]. These motion-induced artifacts can lead to false positive findings in brain-behavior association studies (BWAS), especially when investigating traits that correlate with motion propensity, such as psychiatric disorders [12].

Motion parameter regression represents a fundamental denoising approach that uses the estimated head movement parameters as nuisance regressors in a general linear model (GLM) to remove motion-related variance from fMRI time series [23] [16]. This technique operates on the principle that head movement causes systematic signal changes that can be modeled and removed statistically. The efficacy of motion regression depends critically on the completeness of the motion model and proper accounting for the statistical properties of fMRI data [24].

Table 1: Fundamental Concepts of Motion Regression in fMRI

Concept	Description	Impact on fMRI Data
Motion Artifact	Signal changes induced by head movement	Spurious functional connectivity; systematic bias in brain-behavior associations [12]
Nuisance Regression	Statistical removal of unwanted variance using a noise model	Reduced motion contamination; improved validity of connectivity estimates [24]
Framewise Displacement (FD)	Quantitative measure of head motion between volumes	Used to identify high-motion timepoints for censoring; quality metric [12]
Residual Motion	Motion-related signal remaining after denoising	Can still significantly impact trait-FC relationships after standard processing [12]

Motion Regression Parameters: Basic and Expanded

Basic Motion Parameters (6-Parameter Model)

The most fundamental motion regression approach includes six rigid-body head motion parameters estimated during volume realignment: three translational (x, y, z) and three rotational (pitch, roll, yaw) parameters [23] [16]. These parameters capture the bulk head movement between consecutive volumes and represent the minimal motion correction that should be applied to fMRI data.

Expanded Motion Parameters (24-Parameter Model)

The expanded 24-parameter model provides a more comprehensive motion model that accounts for more complex motion-related signal changes [23] [16]. This approach includes:

• The 6 basic motion parameters
• Their first-order temporal derivatives (6 parameters)
• The squared basic motion parameters (6 parameters)
• The squared first-order derivatives (6 parameters)

This expanded set forms a Taylor expansion of motion-related signal changes, better capturing nonlinear relationships and dynamic motion effects that the basic 6-parameter model misses [23].

Table 2: Composition of 24-Parameter Motion Model

Parameter Group	Components	Purpose	Count
Basic Parameters	Translations (X, Y, Z); Rotations (pitch, roll, yaw)	Model bulk head movement between volumes [23]	6
Temporal Derivatives	Derivatives of all 6 basic parameters	Model gradual, continuous motion effects [16]	6
Squared Basic Parameters	Squares of all 6 basic parameters	Capture nonlinear motion effects [23]	6
Squared Derivatives	Squares of all 6 derivative parameters	Model nonlinear aspects of continuous motion [23]	6

Figure 1: Workflow for implementing motion parameter regression in fMRI preprocessing.

Comparative Efficacy of Motion Regression Approaches

Performance Against Quality Control Benchmarks

Evaluations of denoising pipelines across multiple datasets indicate that simple linear regression of motion parameters alone is insufficient to completely remove motion artifacts [16]. The 24-parameter model demonstrates superior performance compared to the basic 6-parameter approach, but still leaves significant residual motion contamination, particularly in high-motion datasets [16].

The relationship between motion and functional connectivity exhibits strong distance dependence, with motion artifacts particularly affecting long-distance connections [12]. Even after denoising with standard approaches like ABCD-BIDS (which includes motion parameter regression), a strong negative correlation (Spearman ρ = -0.58) persists between motion-FC effects and the average FC matrix [12].

Complementary Denoising Techniques

Motion parameter regression is typically combined with other denoising strategies for improved efficacy:

• Global Signal Regression (GSR): Improves performance of most pipelines but exacerbates distance-dependence of motion-connectivity correlations [16]
• Volume Censoring (Scrubbing): Effectively removes high-motion timepoints but reduces temporal degrees of freedom and can introduce bias [12] [16]
• ICA-AROMA: Data-driven approach that identifies motion-related components without requiring parameter expansion [16] [5]
• aCompCor: Anatomical component-based noise correction method that may only be viable in low-motion data [16]

Table 3: Performance Comparison of Denoising Pipelines Including Motion Regression

Pipeline Components	Residual Motion Artifact	Data Loss	Best Use Case	Key Limitations
6-Parameter Motion Regression	High	None	Minimal preprocessing; initial quality assessment [16]	Limited efficacy; leaves significant motion artifact [16]
24-Parameter Motion Regression	Moderate	None	Standard preprocessing where data retention is critical [23]	Does not eliminate motion-connectivity relationships [16]
24-Parameter + GSR	Low-Moderate	None	Studies where global signal regression is appropriate [16]	Alters connectivity interpretation; enhances distance-dependence [16]
24-Parameter + Censoring (FD < 0.2 mm)	Low	High (timepoints)	High-motion data; critical brain-behavior analyses [12]	Reduces degrees of freedom; may bias sample distribution [12]
ICA-AROMA (with motion parameters)	Low	Moderate (components)	General purpose; balanced approach [16] [5]	Requires specialized implementation; classification errors possible [5]

Experimental Protocol: Implementation and Validation

Protocol: Implementing 24-Parameter Motion Regression

Software Requirements: AFNI [23] [16], FSL [25], HALFpipe [25], or SPM with in-house scripting

Step-by-Step Procedure:

• Volume Realignment: Perform rigid-body registration of all functional volumes to a reference volume (typically the first or middle volume, or a mean volume) to generate the 6 basic motion parameters [23].

• Parameter Expansion: Calculate the additional parameters needed for the 24-parameter model:

  • Compute first-order temporal derivatives of each basic parameter (forward differences)
  • Square each of the 6 basic parameters
  • Square each of the 6 derivative parameters [23]

• Nuisance Regression Implementation: Incorporate the 24-parameter set as regressors of no interest in a General Linear Model (GLM). Critical implementation considerations include:

  • Pre-whitening: Account for temporal autocorrelation in fMRI data [24]
  • Temporal Filtering: Incorporate bandpass filtering (typically 0.01-0.1 Hz) directly into the GLM to properly account for degrees of freedom [24]
  • Temporal Shifting: Consider optimal temporal shifts of regressors to account for delayed motion effects [24]

• Residual Extraction: Save the residuals of the GLM fit as the "cleaned" fMRI data for subsequent functional connectivity analysis [24].

Protocol: Validation and Quality Control

Validation Metrics:

• Motion-Connectivity Correlation: After denoising, compute the correlation between participants' mean framewise displacement (FD) and their functional connectivity matrices. Significant residual correlations indicate incomplete motion correction [12] [16].

• Distance-Dependence Analysis: Examine whether residual motion effects show the characteristic distance-dependent pattern (stronger effects on long-range connections) [12] [16].

• High-Motion vs Low-Motion Comparison: Compare functional connectivity between high-motion and low-motion participants after denoising. Effective pipelines should minimize systematic differences [16].

• Temporal Degrees of Freedom (tDOF): Account for the reduction in tDOF due to nuisance regression, as this affects statistical inference in downstream analyses [16] [5].

Figure 2: Quality control workflow for validating motion regression efficacy.

Advanced Implementation Considerations

Integration with Complementary Techniques

For comprehensive motion denoising, 24-parameter motion regression should be combined with additional strategies:

• Physiological Noise Modeling: Include mean signals from white matter and cerebrospinal fluid (CSF) masks to account for physiological fluctuations [16] [5]
• Edge Voxel Regression: Incorporate signals from the edge of the brain, which are most affected by motion, as additional nuisance regressors [23]
• Motion Censoring: Identify and remove or interpolate high-motion timepoints (typically FD > 0.2-0.5 mm) [12] [16]
• Data-Driven Cleaning: Apply ICA-based approaches (ICA-AROMA) to identify and remove motion-related components not captured by motion parameters [16] [5]

Trait-Specific Motion Impact Assessment

For brain-behavior association studies, implement trait-specific motion impact analyses such as SHAMAN (Split Half Analysis of Motion Associated Networks) [12]. This approach:

• Quantifies whether motion causes overestimation or underestimation of specific trait-FC relationships
• Capitalizes on the stability of traits versus the state-dependent nature of motion
• Provides motion impact scores with significance values for specific trait-FC associations [12]

Dataset-Specific Optimization

Motion regression efficacy varies across datasets with different acquisition parameters and participant populations [3] [5]. Critical factors requiring adjustment include:

• Repetition Time (TR): Affects optimal temporal shifting of regressors [24]
• Participant Population: Motion characteristics differ across age groups and clinical populations [5]
• Motion Severity: High-motion datasets may require more aggressive approaches (censoring) [12] [16]
• Multi-Site Studies: Account for site-specific noise characteristics when implementing motion regression [5]

The Scientist's Toolkit: Research Reagent Solutions

Table 4: Essential Resources for Implementing Motion Regression

Resource Category	Specific Tools	Function	Implementation Considerations
Software Packages	AFNI [23], FSL [25], HALFpipe [25], fMRIPrep [3]	Provides volume realignment, parameter estimation, and GLM implementation	HALFpipe offers containerized standardization; AFNI provides flexible regression options
Motion Quantification	Framewise Displacement (FD) [12], DVARS [16]	Quantifies head motion for censoring and quality assessment	FD threshold of 0.2 mm effectively reduces motion overestimation [12]
Quality Control Metrics	Motion-connectivity correlation [12], Distance-dependence [16], tDOF calculation [5]	Validates denoising efficacy and identifies residual artifacts	Should be reported in all studies to demonstrate motion control
Data-Driven Supplements	ICA-AROMA [16] [5], aCompCor [16], Edge Voxel Regression [23]	Complementary approaches to capture motion not modeled by parameters	ICA-AROMA performs well across benchmarks with moderate data loss [16]
Validation Frameworks	SHAMAN [12], Benchmarking pipelines [16]	Quantifies trait-specific motion impacts and compares method efficacy	Essential for brain-behavior association studies to avoid false positives [12]

Functional magnetic resonance imaging (fMRI) has become an indispensable tool for studying human brain function, yet its signal is notoriously contaminated by various noise sources. Physiological noise arising from cardiac pulsations and respiratory cycles represents a dominant confounding factor, particularly at higher field strengths [26] [27]. This noise increases signal variance, effectively decreasing detection power for neural activation, and compromises the statistical assumptions underlying most fMRI analyses [26]. Additionally, head motion introduces complex artifacts that persist even after standard image realignment [28]. In the context of resting-state fMRI (rs-fMRI), where the goal is to measure intrinsic functional connectivity through temporal correlations between brain regions, these non-neural fluctuations can mimic or mask true functional connections, leading to both false positives and false negatives [28] [29].

The integration of multiple denoising strategies has emerged as a powerful approach to mitigate these diverse noise sources. This application note provides a comprehensive framework for combining three principal denoising methodologies: model-based physiological noise correction (e.g., RETROICOR), data-driven component-based correction (CompCor), and global signal regression (GSR). When strategically integrated within a motion parameter regression framework, these techniques can significantly enhance the sensitivity and specificity of fMRI analyses for both task-based and resting-state paradigms.

Theoretical Foundations of Individual Methods

Physiological Noise Correction (RETROICOR)

RETROICOR is a model-based physiological noise correction technique that uses external measurements of cardiac and respiratory cycles to model signal fluctuations via Fourier series [26]. The method calculates the phase of the cardiac cycle based on the timing between heartbeats relative to image acquisition and the respiratory phase based on the depth of breathing relative to a histogram of respiratory depth across the entire imaging run [26]. These phase values are then used to create Fourier basis functions that model the physiological noise, which are subsequently regressed out from the fMRI data.

The fundamental equations governing RETROICOR are:

• Cardiac/Respiratory Fluctuation Model: ( y{c/r}(x,t) = \sum{m=1}^{M} [a{c/r}(x)\cos(m\varphi{c/r}(t)) + b{c/r}(x)\sin(m\varphi{c/r}(t)) ] )
• Cardiac Phase: ( \varphic(t) = 2\pi(t - t1)/(t2 - t1) )
• Respiratory Phase: ( \varphir(t) = \pi \frac{ \sum{b=1}^{ \lfloor 100 \cdot \text{rnd}[R(t)/R{\text{max}}] \rfloor } H(b) }{ \sum{1}^{100} H(b) } )

Where (y{c/r}(x,t)) represents the cardiac or respiratory-induced signal fluctuation, (\varphi{c/r}(t)) is the phase of the cardiac or respiratory cycle at image acquisition time, (M) is the Fourier fit order, and (a{c/r}), (b{c/r}) are Fourier coefficients determined through regression analysis [26].

CompCor (Component-Based Noise Correction)

CompCor is a data-driven approach that operates on the principle that signals from regions unlikely to contain neural activity (e.g., white matter and cerebrospinal fluid) can be used to model physiological noise in gray matter [27]. The method employs principal component analysis (PCA) on time-series data from defined "noise regions-of-interest" to characterize physiological noise compactly. The significant principal components are then included as nuisance regressors in a general linear model (GLM) to remove noise from the fMRI data.

Two primary variants of CompCor exist:

• Anatomical CompCor (aCompCor): Uses anatomically defined masks of white matter and cerebrospinal fluid to identify noise regions [27] [28].
• Temporal CompCor (tCompCor): Identifies noise regions based on high temporal standard deviation of voxel time series, irrespective of anatomical location [30].

A key advantage of CompCor over simple mean signal regression from noise regions is its ability to capture multiple spatial patterns of noise, as physiological noise manifestations can vary across different brain regions [27] [28].

Global Signal Regression (GSR)

Global Signal Regression involves removing the global mean signal (average of all voxels within the brain) via linear regression from the fMRI data [31]. The global signal represents a "catch-all" signal that contains contributions from various sources, including physiological noise, motion artifacts, scanner drift, and potentially widespread neural activity [31].

The computation is straightforward: for each time point, the global signal ( GS(t) ) is calculated as the average of all voxel values within the brain mask at that time point. This time course is then included as a nuisance regressor in the general linear model. Despite ongoing controversy about its potential removal of neural information and induction of artificial anti-correlations [31] [30], GSR has been shown to effectively reduce motion-related and respiratory-related artifacts [32] and strengthen associations between functional connectivity and behavior [32].

Integrated Protocol: Strategic Combination of Denoising Methods

Optimal Processing Order and Workflow

Research indicates that the sequence of denoising operations significantly impacts their efficacy. The following workflow represents an optimized processing stream based on empirical evidence:

Rationale for Processing Order:

• Volume registration should precede RETROICOR because motion affects the amplitude of physiological fluctuations, and volume registration can distort timing information if performed after physiological correction [26].
• RETROICOR should be applied before slice-time correction to preserve accurate cardiac timing information, as interpolating data to a common time grid may corrupt the cardiac phase information critical for RETROICOR's efficacy [26].
• CompCor should be implemented after RETROICOR to address residual physiological noise not fully captured by the model-based approach.
• Global Signal Regression is optimally positioned after CompCor, as both methods target spatially widespread signals but through different mechanisms.
• Slice-time correction should be applied last, after physiological noise correction, to avoid corrupting timing information needed for RETROICOR [26].

Motion-Modified RETROICOR Protocol

Traditional RETROICOR does not account for timing errors introduced by subject motion. A motion-modified RETROICOR approach has been developed to address this limitation:

Procedure:

• Acquire physiological recordings: Collect cardiac data via pulse oximeter and respiratory data via respiration belt throughout the fMRI acquisition.
• Extract physiological phases: Calculate cardiac and respiratory phases using standard RETROICOR methods [26].
• Incorporate motion parameters: Modify the physiological noise model to account for motion-induced timing discrepancies by incorporating volume registration parameters.
• Generate nuisance regressors: Create Fourier basis functions based on the motion-modified physiological phases.
• Regression: Include these regressors in the GLM alongside traditional motion parameters.

Performance: Simulations indicate that motion-modified RETROICOR reduces temporal standard deviation by up to 36% compared to traditional RETROICOR, with demonstrated efficacy in both high- and low-resolution fMRI data [26].

CompCor Implementation Guidelines

aCompCor Protocol:

• Tissue segmentation: Use high-resolution anatomical data to create masks of white matter (WM) and cerebrospinal fluid (CSF), excluding voxels near gray matter boundaries to minimize partial volume effects [27] [28].
• Noise ROI definition: Combine WM and CSF masks to define the noise region of interest.
• Principal component extraction: Perform PCA on the time series from all voxels within the noise ROI.
• Component selection: Select the top 5-10 principal components based on variance explanation (typically components explaining >1-2% of variance each) [28].
• Regression: Include selected components as nuisance regressors in the GLM.

tCompCor Protocol:

• Temporal standard deviation calculation: Compute the temporal standard deviation (tSTD) for each voxel's time series.
• Noise voxel identification: Identify voxels with the highest tSTD (typically top 2-5% of voxels) [27].
• Principal component extraction: Perform PCA on the time series from these high-tSTD voxels.
• Component selection: Select components based on variance explanation criteria similar to aCompCor.
• Regression: Include selected components as nuisance regressors.

Global Signal Regression with Nuance

Implementation Considerations:

• Computation: Calculate the global signal as the mean of all voxels within the brain mask at each time point.
• Normalization: Consider implementing grand-mean scaling or per-voxel percent signal change normalization before GSR [31].
• Controversy management: Acknowledge that GSR may remove neural information along with noise [31] [30]. For studies focusing on absolute connectivity strengths or between-network anti-correlations, consider alternative approaches or interpret results with appropriate caution.

Enhanced GSR Approach: For studies where GSR is deemed appropriate, an enhanced approach involves:

• Comprehensive nuisance regression prior to GSR computation, including motion parameters, physiological noise, and tissue-specific signals [31].
• Visual inspection of carpet plots before and after GSR to verify removal of widespread signal deflections [33].
• Comparison with results from non-GSR pipelines to ensure robustness of findings.

Quantitative Comparison of Denoising Efficacy

Performance Metrics Across Methods

Table 1: Comparative Performance of Denoising Methods

Method	Noise Reduction Efficacy	Impact on Functional Connectivity	Test-Retest Reliability	Key Limitations
RETROICOR	Reduces cardiac/respiratory spectral bands; Motion-modified version reduces temporal SD by up to 36% [26]	Improves specificity by reducing non-neural correlations [26]	Improves intra-subject variability but reduces inter-subject variability [29]	Requires external physiological recordings; Sensitive to timing errors from motion [26]
CompCor (aCompCor)	Effectively removes motion artifacts; Superior to mean signal regression for motion reduction [28]	Preserves known network anatomy; Maintains age-related connectivity differences better than GSR [30] [34]	Similar to RETROICOR for physiological noise components [27]	May remove neural signal if noise ROIs contain gray matter; Anatomical segmentation required [27]
Global Signal Regression	Highly effective for motion/respiratory artifacts; Explains ~48% variance with motion regressors, additional 31% with physiological regressors [31]	Increases behavioral variance explained by 40-47%; Induces negative correlations [32]	Reduces both intra- and inter-subject variability [29]	Removes potential neural information; Controversial for resting-state studies [31] [30]

Frequency Domain Characteristics

Table 2: Spectral Impact of Denoising Methods

Method	Cardiac/Respiratory Noise Reduction	Low-Frequency Signal Preservation	Impact on Age-Related Connectivity Differences
No Correction	No removal of physiological noise	Full preservation of low-frequency signals	High similarity to pseudo-ground truth (reference) [34]
ICA-AROMA	Effective removal of heartbeat/respiration frequencies [30]	Removes the most low-frequency signals [30] [34]	Reduces detection of age-related differences [34]
Global Signal Regression	Effective removal of respiratory-related artifacts [30] [32]	Removes significant low-frequency content [30]	Diminished similarity to pseudo-ground truth [34]
aCompCor/tCompCor	Effective for high-frequency physiological signals [30]	Better preservation of low-frequency signals [30]	Highest similarity to pseudo-ground truth [34]

Experimental Protocols for Method Validation

Protocol 1: Integrated Pipeline Validation

Purpose: To validate the efficacy of the combined denoising approach in restoring known functional network architecture.

Materials:

• Resting-state fMRI data (≥10 minutes acquisition)
• High-resolution T1-weighted anatomical scan
• Physiological recordings (cardiac and respiratory)
• Processing software: AFNI, FSL, or SPM with in-house scripts

Procedure:

• Preprocessing: Apply standard volume registration, non-brain removal, and spatial smoothing.
• Implement integrated pipeline following the workflow in Section 3.1.
• Quality assessment:
  • Calculate DVARS (root mean square variance over voxels) and Framewise Displacement (FD) to quantify motion artifacts.
  • Generate correlation matrices between known functional networks (default mode, frontoparietal, somatomotor).
  • Compare network segregation (within-network vs. between-network correlation differences) before and after denoising.
• Validation: Assess the anatomical specificity of functional connectivity maps using known network templates.

Expected Outcomes: The integrated pipeline should yield reduced correlation between FD and DVARS, increased network segregation scores, and improved specificity of functional connectivity maps to canonical network boundaries [28].

Protocol 2: Method Comparison Study

Purpose: To quantitatively compare individual and combined denoising methods using objective metrics.

Materials:

• Task-based fMRI data with known activation patterns (e.g., motor or visual task)
• Resting-state fMRI data from the same participants
• Physiological recordings

Procedure:

• Process data through multiple pipelines:
  • Pipeline A: Motion regression only
  • Pipeline B: Motion regression + RETROICOR
  • Pipeline C: Motion regression + CompCor
  • Pipeline D: Motion regression + GSR
  • Pipeline E: Full integrated approach
• For task data:
  • Calculate t-values for activation in expected regions
  • Compute contrast-to-noise ratios
  • Compare activation extent and effect sizes
• For resting-state data:
  • Calculate functional connectivity within and between known networks
  • Assess network modularity
  • Measure motion-connectivity correlations
• Statistical comparison: Use repeated measures ANOVA to compare pipeline performance across metrics.

Expected Outcomes: The integrated pipeline should optimize multiple metrics simultaneously, demonstrating superior noise reduction while preserving neural signals [30] [32].

The Scientist's Toolkit: Essential Research Reagents

Table 3: Essential Tools for Integrated fMRI Denoising

Tool/Reagent	Function	Implementation Notes
Physiological Monitoring Equipment (Pulse oximeter, respiratory belt)	Captures cardiac and respiratory waveforms for RETROICOR	Essential for model-based physiological noise correction; Synchronize with scanner triggers [26]
High-resolution T1-weighted Anatomical Sequence	Provides structural reference for tissue segmentation	Enables accurate WM/CSF masks for aCompCor; ≤1mm isotropic resolution recommended [27] [28]
Volume Registration Algorithm (3dvolreg in AFNI, MCFLIRT in FSL)	Corrects for head motion between volumes	Include derivatives and squares of motion parameters for improved motion correction [28]
Tissue Segmentation Tool (FSL FAST, Freesurfer, SPM)	Identifies white matter and CSF compartments	Crucial for aCompCor; Use conservative thresholds to avoid gray matter contamination [27] [28]
Principal Component Analysis Software (Implementation in AFNI, FSL, or CONN)	Extracts noise components from noise ROIs	For CompCor; Select components explaining >1-2% variance each [27] [28]
Global Signal Computation	Calculates whole-brain average signal	Simple mean of all brain voxels at each time point; Can be computed after minimal preprocessing [31]

Advanced Integration: DiCER as a GSR Alternative

For researchers concerned about the potential biases introduced by GSR, the Diffuse Cluster Estimation and Regression (DiCER) method offers a promising alternative. DiCER identifies and removes widespread signal deflections (WSDs) by:

• Clustering voxels based on similarity in signal time courses
• Identifying large clusters exhibiting widespread, temporally coherent signals
• Regressing out representative signals from these clusters

DiCER has been shown to effectively remove diverse WSDs while better preserving the spatial structure of task-related activation patterns compared to GSR [33]. Implementation code is publicly available, making it a viable alternative worth considering in the integrated denoising pipeline.

The strategic integration of physiological, CompCor, and global signal regression methods within a motion parameter regression framework represents a robust approach to fMRI denoising. Each method targets distinct noise components: RETROICOR addresses phase-locked physiological fluctuations, CompCor captures data-driven noise components from non-neural tissues, and GSR removes whole-brain artifacts. When combined in the recommended sequence, these methods synergistically enhance data quality while mitigating their individual limitations. Researchers should validate their specific pipeline using the provided protocols and consider their experimental goals when making decisions about method inclusion, particularly regarding the controversial but effective GSR technique.

In the field of functional magnetic resonance imaging (fMRI) research, motion-induced artifacts represent a significant confound that can induce spurious findings and obscure true neural effects. While traditional motion parameter regression has been widely used for denoising, this approach removes motion-related signal variations at the cost of destroying the autocorrelation structure of fMRI time-series and reducing temporal degrees of freedom. Within this context, Independent Component Analysis (ICA)-based strategies have emerged as powerful alternatives that overcome these limitations. This application note provides a comprehensive technical overview of two advanced data-driven denoising methods: ICA-AROMA (Automatic Removal of Motion Artifacts) and CICADA (Comprehensive Independent Component Analysis Denoising Assistant). Framed within the broader thesis of motion parameter regression for denoising fMRI research, we detail their methodologies, performance characteristics, and implementation protocols to guide researchers, scientists, and drug development professionals in selecting and applying these tools to enhance the validity and reliability of their fMRI findings.

Fundamental Principles of ICA-based Denoising

ICA-based denoising strategies for fMRI data leverage spatial independence to separate neuronally-driven BOLD signals from noise sources. These methods decompose the 4D fMRI data into spatially independent components (ICs), each characterized by a spatial map and associated time course. The fundamental premise is that physiological noise and motion artifacts exhibit spatial and temporal characteristics distinct from neural signals, enabling their identification and removal. Unlike model-driven approaches that require a priori specification of noise regressors, ICA methods are data-driven, automatically adapting to the unique noise characteristics present in each dataset. This flexibility is particularly valuable for addressing the complex, multi-source noise contamination that plagues fMRI data, especially in clinical populations with elevated motion.

The Evolution from Manual to Automated ICA Denoising

Early ICA denoising implementations required manual classification of components as noise or signal—a process that demanded extensive training, introduced subjectivity, and was prohibitively time-consuming for large datasets. This limitation spurred the development of automated classifiers, beginning with ICA-FIX, which required dataset-specific classifier training. ICA-AROMA advanced the field by eliminating the need for re-training through its use of a robust set of theoretically-motivated features. Most recently, CICADA has emerged, designed to fully automate the manual ICA denoising gold-standard while offering unprecedented accuracy and flexibility across diverse data types and populations [35] [36].

Methodological Deep Dive: ICA-AROMA vs. CICADA

ICA-AROMA: Core Architecture and Algorithm

ICA-AROMA employs a streamlined, transparent classification framework based on a small (n=4) but robust set of theoretically motivated features designed to capture the distinctive characteristics of motion-related components [37]. The algorithm operates through four sequential steps:

• Feature Extraction: For each IC, four specific features are calculated:

  • Maximum CSF Correlation: The highest correlation between the component's time course and the average cerebrospinal fluid (CSF) signal.
  • High-Frequency Content: The proportion of the component's power spectral density above 0.5 Hz.
  • Spatial Edge Fraction: The fraction of component voxels located at the edges of the brain.
  • Spatial CSF Fraction: The fraction of component voxels located within the CSF.

• Component Classification: Each IC is evaluated against pre-defined thresholds for these features. Components exceeding thresholds for one or more features are classified as noise.

• Noise Removal: Denoising is performed using FSL's fsl_regfilt, which applies non-aggressive regression to remove the variance associated with the noise components' time courses from the original data.

• Output: The final output is a denoised 4D fMRI dataset that retains the data's autocorrelation structure and preserves temporal degrees of freedom to a greater extent than scrubbing or spike regression methods [37].

CICADA: Advanced Multi-Stage Denoising Framework

CICADA introduces a more comprehensive, multi-stage architecture designed to capture a wider range of common fMRI noise sources with high accuracy [35] [36]. Its operation consists of three configurable modules:

• Automatic CICADA:

  • Regional Masking: Creates detailed regional masks of the functional data.
  • Component Generation: Uses FSL's MELODIC to generate ICs.
  • Feature Calculation: Computes extensive features from regional masks, time series, and power spectra.
  • Intelligent Sorting: Resorts ICs based on relative neural signal probability.
  • Classification: Employs 3-group k-means clustering on the calculated features to automatically tag ICs into various signal-like and noise-like categories.
  • Denoising & QC: Performs non-aggressive denoising with fsl_regfilt and conducts comprehensive quality control analyses.

• Manual CICADA: Allows for user inspection and adjustment of the automatically classified ICs. A key efficiency is that users only need to examine a small subset (approximately 25%) of the total components, as the majority are classified with high confidence automatically [35].

• Group CICADA: Performs group-level quality control, identifies outliers, and prepares data for group analyses.

Table 1: Core Algorithmic Comparison of ICA-AROMA and CICADA

Feature	ICA-AROMA	CICADA
Classification Basis	4 robust temporal/spatial features [37]	Extensive features analyzed via 3-group k-means clustering [36]
Noise Scope	Primarily motion artifacts [37]	Multiple common fMRI noise sources [35]
Automation Level	Full automation	Full automation with optional manual refinement [36]
Manual Effort	None required	Reduces manual component inspection by ~75% [35]
Core Dependencies	FSL	FSL, MATLAB [36]
Output Strategy	Non-aggressive regression	Non-aggressive regression

Performance and Efficacy Analysis

Quantitative Performance Benchmarks

Recent validation studies provide direct comparisons of the classification accuracy and denoising efficacy of these tools against manual classification and other established methods.

Table 2: Quantitative Performance Comparison Across Denoising Methods

Method	Classification Accuracy	Motion Reduction	Effect on Behavioral Prediction	Data Type Suitability
CICADA	97.9% (vs. manual) [35]	Matches or outperforms FIX & AROMA in QC metrics [35]	Preserves behavioral correlations	Resting-state & task-fMRI; high & low motion [36]
ICA-AROMA	83.8% (vs. manual) [35]	Superior to 24-parameter regression & spike regression [37]	Increases group-level activation sensitivity [37]	Resting-state & task-fMRI [37]
ICA-FIX	92.9% (vs. manual) [35]	Effective, but requires training [35]	N/A	Requires classifier training
Manual ICA	Gold Standard (100%)	Gold Standard	Varies with expertise	All data types, but time-intensive

Contextual Performance in fMRI Research Applications

The performance of denoising methods must be evaluated in the context of specific research goals. Studies comparing multiple pipelines reveal that:

• ICA-AROMA effectively reduces motion-induced signal variations and increases sensitivity to group-level activation [37]. However, it may remove more low-frequency signals along with physiological noise, which can subsequently diminish the detection of age-related functional connectivity differences [38] [39].

• CICADA performs nearly identically to the labor-intensive manual IC classification gold-standard across diverse datasets, including high-motion data from clinical populations such as individuals with schizophrenia [35]. Its high accuracy minimizes both Type I and Type II errors in component classification.

• Multi-echo fMRI applications at ultra-high field (7T) demonstrate that ME-ICA combined with aCompCor may preserve more signal-of-interest compared to the aggressive option of ICA-AROMA, highlighting that the optimal tool can depend on acquisition parameters [40] [41].

Experimental Protocols and Implementation

Protocol for Implementing ICA-AROMA

Required Software and Materials:

• FSL (version 5.0.1 or higher)
• 4D preprocessed fMRI data (in MNI space, recommended without smoothing)
• Structural data (for improved registration)

Step-by-Step Procedure:

• Data Preparation: Ensure your fMRI data is preprocessed with standard steps (realignment, slice-time correction, normalization to standard space). fMRIPrep is highly recommended for this purpose.
• Feature Extraction and Classification: Execute the core AROMA classification script. This step performs MELODIC ICA and classifies components based on the four key features. python ICA_AROMA.py -in <input_feat_directory> -out <output_directory> -affmat <mat_file> -mc <mc_file>
• Noise Regression: Apply non-aggressive regression to remove the identified noise components from your original data. python ICA_AROMA.py -in <input_feat_directory> -out <output_directory> -classify <classified_motion_ICs.txt> -den
• Output Verification: Inspect the classified components file and the denoised data using standard QC metrics (e.g., DVARS, FD plots) to confirm artifact reduction.

Protocol for Implementing CICADA

Required Software and Materials:

• FSL (tested on 6.0)
• MATLAB (tested on 2022a)
• Preprocessed fMRI data (ideally fMRIPrep outputs)
• Subject-specific segmented tissue probability masks (GM, WM, CSF)

Step-by-Step Procedure:

• Environment Setup: Clone the CICADA repository and add it to your MATLAB path. git clone https://github.com/keithcdodd/CICADA.git addpath(genpath('CICADA'));
• Automatic Denoising: For fMRIPrep-processed data, use the dedicated wrapper function. fmriprep_auto_CICADA(fmriprep_dir, output_dir, subject_id);
• Manual Verification (Optional): If manual refinement is desired, open the generated "IC_auto_checker.csv" file, adjust the "Signal Label Column" (1 for signal, 0 for noise), save as "IC_manual_checker.csv", and run the manual module. fmriprep_manual_CICADA(fmriprep_dir, output_dir, subject_id);
• Group-Level Analysis: For cohort-level processing and QC, use the group function. cicada_group_qc(project_directory);

Visualization of Workflows

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Software and Computational Tools for ICA Denoising

Tool/Resource	Function/Purpose	Implementation Notes
FSL	Provides core MELODICA ICA and fsl_regfilt utilities	Required for both ICA-AROMA and CICADA [37] [36]
MATLAB	Computational engine for CICADA's classification algorithms	Required for CICADA, version 2022a tested [36]
fMRIPrep	Standardized fMRI preprocessing	Provides optimized inputs for both tools, enabling robust denoising [36]
Subject-specific tissue masks	Improve accuracy of noise feature calculation	From fMRIPrep/FreeSurfer; critical for CICADA's regional analysis [36]
High-performance computing	Handles computational load of ICA	ICA is memory and computationally intensive for large datasets

Within the evolving landscape of fMRI denoising that moves beyond traditional motion parameter regression, both ICA-AROMA and CICADA represent significant advancements for addressing motion and physiological artifacts. ICA-AROMA stands out for its computational efficiency, theoretical transparency, and robust performance for standard motion removal without requiring classifier training. CICADA distinguishes itself through its exceptional classification accuracy that nears manual inspection, its comprehensive approach to multiple noise sources, and its unique flexibility in supporting full automation or manual refinement.

For researchers selecting between these tools, we recommend:

• Choose ICA-AROMA for studies prioritizing a well-established, computationally efficient method for robust motion artifact removal in standard population cohorts, particularly when manual inspection is not feasible.
• Choose CICADA for studies requiring the highest possible classification accuracy, for data with complex noise profiles (e.g., clinical populations with high motion), or when investigating subtle neural effects where minimizing both false positive and false negative noise classification is paramount.

The integration of these advanced data-driven methods into the fMRI processing pipeline significantly enhances data quality, ultimately leading to more reliable and valid findings in basic neuroscience and drug development research.

In-scanner head motion remains the largest source of artifact in functional magnetic resonance imaging (fMRI) signals, introducing systematic biases that can lead to both false positive and false negative findings in brain-behavior association studies [12]. While numerous denoising techniques exist—including global signal regression, motion parameter regression, and independent component analysis—the practice of scrubbing, or censoring high-motion volumes from analysis, continues to play a crucial role in comprehensive motion mitigation pipelines [42]. The fundamental challenge lies in the non-linear characteristics of MRI physics, which make complete removal of motion artifact during post-processing exceptionally difficult [12]. This application note examines the current debate surrounding scrubbing protocols, provides evidence-based recommendations for implementation, and situates these protocols within the broader context of motion parameter regression for denoising fMRI research.

The tension inherent in scrubbing practices revolves around a critical trade-off: removing sufficient motion-contaminated data to reduce spurious findings while retaining enough data to preserve statistical power and avoid biasing sample distributions [12]. This balance is particularly crucial when studying populations that tend to exhibit higher motion, such as children, older adults, or patients with neurological or psychiatric conditions, as overly aggressive censoring may systematically exclude these very participants from analysis [43]. Recent methodological advances, including the development of motion impact scores and data-driven scrubbing approaches, offer new pathways for optimizing this balance [12] [44].

Quantitative Landscape: Assessing Scrubbing Efficacy

Motion Impact on Functional Connectivity

Recent large-scale studies have quantified the substantial impact of residual head motion on functional connectivity (FC) metrics even after extensive denoising. Analysis of the Adolescent Brain Cognitive Development (ABCD) Study dataset (n = 7,270) revealed that the motion-FC effect matrix exhibited a strong negative correlation (Spearman ρ = -0.58) with the average FC matrix, indicating that connection strength tended to be systematically weaker in participants who moved more [12]. This pervasive effect persisted even after motion censoring at framewise displacement (FD) < 0.2 mm (Spearman ρ = -0.51), highlighting the tenacity of motion artifacts [12].

Table 1: Prevalence of Motion-Related Trait-FC Distortions Before and After Scrubbing in ABCD Study Data (n=7,270)

Motion Impact Type	Prevalence Before Censoring (FD < 0.2 mm)	Prevalence After Censoring (FD < 0.2 mm)
Significant Motion Overestimation	42% (19/45 traits)	2% (1/45 traits)
Significant Motion Underestimation	38% (17/45 traits)	38% (17/45 traits)

Data derived from SHAMAN analysis of 45 behavioral traits [12]

Notably, Table 1 demonstrates that while stringent censoring effectively addresses motion-induced overestimation of trait-FC effects, it does not similarly reduce underestimation effects [12]. This differential impact underscores the complexity of motion artifact influences on functional connectivity measures.

Comparative Performance of Denoising Pipelines

Research evaluating multiple denoising approaches has revealed that no single pipeline universally excels at both motion mitigation and behavioral prediction. Pipelines combining ICA-FIX with global signal regression (GSR) appear to offer a reasonable trade-off between these competing objectives [45] [46]. Importantly, the efficacy of volume censoring appears to be enhanced when integrated with complementary denoising techniques rather than applied in isolation [42].

Table 2: Efficacy of Scrubbing Thresholds Across Different Research Contexts

Application Context	Recommended Scrubbing Threshold	Key Outcomes and Considerations
General Adult Populations	FD < 0.2 - 0.5 mm	Balances noise reduction with data retention; improves validity and identifiability of functional connectivity [47]
Pediatric Populations (High-Motion)	FD < 0.3 mm	Retained 83% of participants in first-grade children while meeting rigorous quality standards [48]
Data-Driven Projection Scrubbing	ICA-based outlier detection	Improves fingerprinting without worsening validity/reliability; excludes fewer volumes than motion scrubbing [44]
HCP Data Optimization	Varied thresholds with low-pass filtering	Small improvement with scrubbing; largest gains from GS and WM component removal [42]

Experimental Protocols: Implementing Scrubbing Frameworks

Protocol 1: Standardized Motion Scrubbing with Framewise Displacement

Purpose: To identify and censor individual volumes contaminated by excessive head motion using framewise displacement (FD) calculations.

Materials and Reagents:

• Preprocessed fMRI time series data (slice-time corrected, realigned)
• Head motion parameters (3 translation, 3 rotation) from realignment
• Computing environment with MATLAB, Python, or similar analytical platform

Procedure:

• Calculate Framewise Displacement: Compute FD for each time point using the formula:
  where Δ represents the volume-to-volume changes in translation (x, y, z) and rotation (α, β, γ) parameters [20]. For rotational displacements, convert to millimeters by assuming a brain radius of 50 mm (or population-specific mean).

• Identify Censoring Targets: Flag all time points where FD exceeds a predetermined threshold (e.g., 0.2-0.5 mm, depending on research context and population).

• Expand Censoring Frame: Include one additional time point before and two time points after each flagged volume to account for the temporal spread of motion artifacts [47].

• Apply Censoring: Remove identified volumes from functional connectivity analysis. For seed-based correlation analysis, use pairwise approaches that only exclude censored time points from specific correlation calculations rather than deleting them from the entire time series.

• Quality Assessment: Calculate the percentage of censored volumes per participant. Establish a data retention threshold (e.g., ≥ 5 minutes of clean data) for inclusion in final analyses [48].

Protocol 2: Data-Driven Projection Scrubbing

Purpose: To identify artifactual volumes using a statistical outlier detection framework that isolates noise through strategic dimension reduction.

Materials and Reagents:

• Preprocessed fMRI time series data
• Independent Component Analysis (ICA) software (e.g., FSL MELODIC)
• Computing environment with statistical processing capabilities

Procedure:

• Data Decomposition: Perform ICA on the concatenated fMRI data across all participants to identify spatially independent components.

• Component Classification: Classify components as neural signals or noise using automated classifiers (e.g., FSL FIX) or manual inspection based on spatial patterns and frequency content.

• Noise Time Course Extraction: Extract the time courses associated with noise components identified in Step 2.

• Multivariate Outlier Detection: For each time point, calculate the squared Mahalanobis distance across all noise component time courses to identify volumes exhibiting distinctive noise characteristics.

• Statistical Thresholding: Establish significance thresholds for outlier detection using median absolute deviation or similar robust statistical measures [44].

• Targeted Censoring: Censor identified outlier volumes following the same temporal expansion principles described in Protocol 1.

Protocol 3: Integrated Scrubbing with SHAMAN Motion Impact Assessment

Purpose: To implement scrubbing within a framework that quantitatively evaluates trait-specific motion impacts on functional connectivity.

Materials and Reagents:

• Resting-state fMRI data with minimal preprocessing
• Behavioral or trait measures of interest
• SHAMAN (Split Half Analysis of Motion Associated Networks) computational tools

Procedure:

• Data Preparation: Process resting-state fMRI data using standard pipelines (e.g., ABCD-BIDS) without initial motion censoring.

• Split-Half Analysis: Divide each participant's time series into high-motion and low-motion halves based on median Framewise Displacement.

• Trait-FC Estimation: Calculate trait-functional connectivity relationships separately for high-motion and low-motion partitions.

• Motion Impact Score Calculation: Compute motion impact scores by comparing trait-FC effects between high-motion and low-motion partitions:

  • Positive scores aligned with trait-FC effect direction indicate motion overestimation
  • Negative scores opposite trait-FC effect direction indicate motion underestimation

• Informed Censoring Implementation: Apply scrubbing thresholds specifically tuned to address the identified motion impact profile (overestimation vs. underestimation bias) [12].

• Validation: Recalculate motion impact scores post-censoring to verify reduction of spurious motion-trait associations.

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Analytical Tools and Resources for Scrubbing Implementation

Tool Name	Function	Application Context
Framewise Displacement (FD)	Quantifies volume-to-volume head movement	Universal metric for motion scrubbing thresholding [20]
SHAMAN Framework	Assigns motion impact scores to specific trait-FC relationships	Identifying motion-induced over/underestimation of brain-behavior associations [12]
Projection Scrubbing	Data-driven outlier detection for artifactual volumes	Alternative to motion-based scrubbing; improved data retention [44]
ICA-FIX	Classifies noise components from independent component analysis	Automated noise component identification for data-driven scrubbing [45] [46]
ABCD-BIDS Pipeline	Integrated denoising with respiratory filtering, motion regression, and despiking	Large-scale dataset processing; benchmark for evaluating additional scrubbing [12]

Decision Framework: Strategic Scrubbing Implementation

The following workflow diagram illustrates a systematic approach for implementing scrubbing protocols based on research objectives, sample characteristics, and data quality considerations:

Diagram 1: Scrubbing Protocol Decision Workflow. This framework guides researchers in selecting appropriate scrubbing strategies based on study-specific considerations.

The evidence reviewed in this application note supports several strategic recommendations for researchers implementing scrubbing protocols:

First, scrubbing should be implemented as part of a comprehensive denoising strategy rather than as a standalone solution. The most effective approaches combine motion censoring with other techniques such as global signal regression, ICA-based artifact removal, and physiological noise modeling [45] [42].

Second, scrubbing thresholds must be calibrated to specific research contexts. While FD < 0.2 mm effectively controls false positives in general adult populations, more lenient thresholds (e.g., FD < 0.3 mm) may be necessary when studying high-motion populations to avoid systematic exclusion biases [12] [48] [43].

Third, emerging data-driven approaches offer promising alternatives to traditional motion-based scrubbing. Projection scrubbing and similar techniques can improve data retention while maintaining connectivity measurement quality, particularly valuable for large-scale population neuroscience [44].

Finally, researchers should explicitly account for potential selection biases introduced by scrubbing-related exclusions. Statistical approaches such as multiple imputation or inclusion of quality metrics as covariates can help address the non-random missingness patterns created when participants with higher motion are excluded from analysis [43].

As the field continues to develop more sophisticated methods for quantifying and addressing motion artifacts, scrubbing remains an essential component of the fMRI preprocessing toolkit. By implementing context-appropriate scrubbing protocols and transparently reporting their procedures, researchers can strengthen the validity and reproducibility of functional connectivity findings across basic neuroscience and clinical drug development applications.

Functional magnetic resonance imaging (fMRI) research fundamentally relies on detecting subtle blood-oxygenation-level-dependent (BOLD) signal fluctuations that reflect neural activity. These signals are notoriously contaminated by multiple non-neuronal sources of noise, including head motion, cardiac cycles, respiratory variations, and other physiological processes [3] [49]. Motion artifacts present a particularly significant challenge, as they can introduce spurious correlations in functional connectivity (FC) estimates, potentially leading to false positives in brain-behavior association studies (BWAS) [3] [42]. The imperative for effective denoising is especially pronounced in clinical and drug development contexts, where population comparisons often involve groups with inherent motion differences (e.g., patients versus controls, or pediatric versus adult cohorts) [42].

In this landscape, standardized and reproducible preprocessing workflows have emerged as essential tools for ensuring the validity and reliability of fMRI findings. The integration of fMRIPrep for robust data preprocessing and Nilearn for subsequent denoising and connectivity analysis represents a powerful, state-of-the-art approach [50] [51]. This application note details protocols for implementing denoising strategies within this framework, providing researchers with practical methodologies to enhance data quality for motion parameter regression and other denoising techniques.

Table 1: Essential Research Reagent Solutions for fMRI Denoising

Tool Name	Primary Function	Role in Denoising Workflow	Key Features
fMRIPrep [51]	Automated fMRI Preprocessing	Generates preprocessed BOLD time series and comprehensive confound matrices.	Robust, BIDS-compliant, instrument-agnostic, generates QC reports.
Nilearn [50]	Python-based fMRI Analysis	Implements denoising strategies on fMRIPrep outputs; calculates functional connectivity.	Integrates with fMRIPrep outputs; offers multiple denoising algorithms and atlases.
BIDS Validator	Data Standardization Check	Ensures input dataset is properly formatted for fMRIPrep.	Validates compliance with Brain Imaging Data Structure (BIDS) standard.
Confound Matrix [50]	Structured Noise Data	fMRIPrep output file (e.g., *_confounds.tsv) containing motion parameters and tissue signals.	Provides nuisance regressors for denoising in Nilearn.

Integrated Denoising Workflow: From Raw Data to Clean Signals

The following diagram illustrates the complete pathway from raw data to denoised functional connectivity matrices, integrating fMRIPrep and Nilearn with decision points for different denoising strategies.

Experimental Protocols and Implementation

Protocol 1: fMRIPrep Preprocessing and Confound Extraction

Objective: To generate standardized preprocessed functional MRI data and a comprehensive confound matrix for subsequent denoising.

• Data Preparation: Ensure your dataset is organized according to the Brain Imaging Data Structure (BIDS) standard. Validate it using the BIDS Validator [52].
• fMRIPrep Execution: Run fMRIPrep (version 25.1.0 or later is recommended for latest fixes and features [53]) via Docker or Singularity. A basic command structure is:

• Output Acquisition: Upon successful completion, locate the preprocessed BOLD file (e.g., *_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz) and the corresponding confounds file ( *_confounds.tsv) in the fMRIPrep output directory. The confounds.tsv file contains the nuisance regressors critical for denoising [50].

Protocol 2: Denoising Strategy Implementation in Nilearn

Objective: To remove motion and physiological artifacts from the preprocessed BOLD signal using selected strategies in Nilearn.

• Data Loading: Use Nilearn's image module to load the preprocessed BOLD file and its mask. Use pandas to load the confounds.tsv file.
• Confound Selection: Create a strategy-specific list of confound regressors from the TSV file. The required confounds vary by strategy, as detailed in Table 2.
• Nuisance Regression: Use Nilearn's clean_img function to perform linear regression and remove the selected confounds from the BOLD time series.

Quantitative Comparison of Denoising Pipelines

Benchmarking studies are crucial for selecting an appropriate denoising strategy, as no single pipeline universally excels across all datasets and objectives [3]. Performance is typically evaluated based on two key, and sometimes competing, metrics: the efficacy in mitigating motion-related artifacts and the capacity to enhance the detection of brain-behavior relationships.

Table 2: Performance Comparison of Common Denoising Pipelines

Denoising Strategy	Key Confound Regressors	Motion Reduction Efficacy	Impact on Brain-Behavior Prediction	Best-Suited Applications
Global Signal Regression (GSR) [3] [50]	Global signal, 6 Motion Parameters (MP), derivatives.	High efficacy in reducing motion artifacts.	Shows a reasonable trade-off, but can remove neural signal of interest.	Studies where motion is the primary concern and global signal is considered a confound.
aCompCor [50] [42]	6 MP, average WM/CSF signals, top 5 Principal Components (PCs) from WM and CSF.	Moderate to high motion reduction.	Preserves neural variability, potentially enhancing behavioral prediction.	General-purpose resting-state studies; preferred when avoiding GSR.
ICA-AROMA & GSR [3]	ICA-based noise components, Global signal.	High motion reduction, effective for removing various structured noises.	Demonstrates a reasonable trade-off between motion reduction and behavioral prediction performance.	Pipelines requiring robust, automated noise component removal.
Scrubbing & GSR [50]	6 MP, Global signal, and removal of high-motion volumes (FD > 0.5mm).	Highest efficacy in reducing motion artifacts.	Incompatible with analyses requiring continuous sampling; can reduce statistical power.	Datasets with severe motion contamination, where data loss is acceptable.

Selecting an optimal denoising protocol requires balancing the trade-offs inherent in each method. The following decision diagram synthesizes evidence from recent benchmarks to guide researchers in choosing a strategy aligned with their study goals and data characteristics.

In conclusion, the integrated use of fMRIPrep and Nilearn establishes a robust, standardized foundation for fMRI denoising. The protocols and data-driven comparisons provided here empower researchers to make informed decisions, thereby enhancing the validity and reproducibility of their functional connectivity findings, which is paramount for both basic neuroscience and applied drug development.

Optimizing Your Pipeline: Strategies for Challenging Data and Populations

In-scanner head motion represents a major confounding factor in functional connectivity (FC) studies employing the blood oxygenation level dependent (BOLD) signal [54]. Despite substantial efforts to develop denoising strategies, residual motion-related artifacts persist in functional magnetic resonance imaging (fMRI) data, complicating interpretation of results, particularly in studies where motion correlates with the experimental conditions of interest [54] [55]. This application note examines the technical challenges underlying incomplete motion denoising and provides evidence-based protocols for optimizing denoising pipelines in both resting-state and task-based fMRI paradigms.

The fundamental challenge in motion denoising stems from two key issues: first, motion introduces both global and spatially-dependent variance that mimics true functional connections [54]; second, investigators lack a priori information regarding the exact temporal characteristics of neuronal-related variance, making it difficult to distinguish signal from noise [54]. Furthermore, the problem is particularly acute in study designs comparing different cognitive states, as subjects tend to move less during engaging tasks compared to unconstrained rest conditions, creating systematic biases in connectivity measures [54].

Performance Comparison of Denoising Pipelines

Efficacy of Common Denoising Strategies

Table 1: Performance benchmarks of common denoising pipelines across multiple studies

Denoising Pipeline	Residual Motion Artifacts	Network Identifiability	Distance-Dependent Artifacts	Best Application Context
aCompCor	Moderate reduction [54]	High [54]	Limited efficacy [54]	Task-based fMRI with optimized noise prediction [54]
Global Signal Regression (GSR)	Effective reduction [54]	High [54]	Can unmask or exacerbate [55]	When motion-connectivity relationship must be minimized [55]
ICA-AROMA (aggressive)	Effective reduction [5]	Moderate [5]	Moderate control [5]	Resting-state fMRI in older adults [5]
ICA-AROMA (non-aggressive)	Less effective [5]	Moderate [5]	Moderate control [5]	Standard resting-state fMRI [5]
FIX	Effective reduction, conserves signal [56]	High [56]	Better balance for task fMRI [56]	Task-fMRI with physiological changes [56]
Censoring/Scrubbing	Substantial reduction [54]	Reduced [54]	Substantially reduces [54]	When sudden motion bursts are present [54]

Quantitative Trade-offs in Denoising Performance

Table 2: Trade-offs in denoising pipeline performance metrics

Performance Metric	Best Performing Pipeline(s)	Key Trade-off
Minimizing motion-connectivity relationships	GSR, aCompCor [54] [55]	GSR unmasks distance-dependent artifacts [55]
Reducing distance-dependent artifacts	Censoring techniques [54]	Reduced network identifiability and temporal degrees of freedom [54] [55]
Network identifiability	aCompCor, GSR [54]	Less effective at balancing artifacts between conditions [54]
Conserving signal of interest	FIX [56]	Requires manual training; workload intensive [56]
Reproducibility in longitudinal studies	Aggressive ICA-AROMA [5]	Less effective for motion spike reduction [5]

Experimental Protocols for Denoising Pipeline Implementation

Protocol 1: Optimized aCompCor for Task-Based fMRI

Application Context: Task-based fMRI studies with sustained cognitive engagement where motion systematically differs between conditions [54].

Step-by-Step Methodology:

• Data Acquisition: Acquire BOLD data using block-design fMRI with prolonged epochs (e.g., 24-minute runs with alternating rest and task conditions) [54].
• Initial Preprocessing: Perform standard steps including motion correction, slice timing correction, and spatial smoothing [56].
• Noise Component Identification:
  • Extract signals from white matter and cerebrospinal fluid (CSF) masks [56].
  • Perform principal component analysis (PCA) on these noise region signals [54] [56].
  • Select principal components that explain the maximum variance in noise regions [54].
• Noise Regression: Include the noise components as regressors in a general linear model (GLM) to remove noise signals from the BOLD data [56].
• Optimization: Increase the noise prediction power by optimizing the number of components and variance explained thresholds [54].

Quality Control Metrics:

• Calculate framewise displacement to quantify motion [54].
• Assess residual relationship between motion and connectivity [55].
• Evaluate network identifiability using established benchmarks [54].

Protocol 2: FIX for Task-fMRI with Physiological Changes

Application Context: Task-based fMRI involving substantial physiological responses or movements (e.g., pain research, breathing alterations) [56].

Step-by-Step Methodology:

• Classifier Training (Dataset-Specific):
  • Use 30-40 participants from your dataset for manual training [56].
  • Perform independent component analysis (ICA) to decompose data into components [56].
  • Manually classify components as signal or noise based on spatial and temporal features [56].
  • Train FIX classifier using these labeled components [56].
• Data Preprocessing: Apply standard preprocessing including motion correction and high-pass filtering [56].
• ICA Decomposition: Run ICA on preprocessed data to obtain spatial components and time courses [56].
• Component Classification: Apply the trained FIX classifier to automatically identify noise components [56].
• Noise Regression: Remove identified noise components from the data [56].

Quality Control Metrics:

• Compare signal conservation in task-activated regions [56].
• Quantify noise removal in white matter and CSF regions [56].
• Assess false positive rates in group-level analyses [56].

Protocol 3: Aggressive ICA-AROMA for Resting-State fMRI in Special Populations

Application Context: Resting-state fMRI in older adults or clinical populations where motion characteristics may differ from healthy young adults [5].

Step-by-Step Methodology:

• Data Preprocessing: Perform standard volume realignment, slice timing correction, and brain extraction [5].
• ICA-AROMA Application:
  • Run ICA-AROMA in aggressive mode to completely remove motion-related components [5].
  • The algorithm automatically identifies motion components based on four spatial and temporal features [5].
• Component Removal: Completely regress out identified motion components (aggressive approach) [5].
• Additional Filtering: Apply high-pass filtering (>0.01 Hz) to remove low-frequency drifts [5].

Quality Control Metrics:

• Evaluate network reproducibility across sessions [5].
• Assess loss of temporal degrees of freedom [5].
• Measure edge activity and spatial smoothness [5].

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key software tools and resources for fMRI denoising

Tool/Resource	Function	Application Context
FSL FIX	Classifier-based ICA component removal [56]	Task-based fMRI with physiological changes; high-quality datasets [56]
ICA-AROMA	Automatic motion component identification and removal [5]	Multi-site studies; when classifier training is not feasible [5]
aCompCor	PCA-based noise signal regression from WM/CSF [54] [56]	Task-based fMRI where motion differs between conditions [54]
Global Signal Regression	Removal of whole-brain average signal [54] [55]	When motion-connectivity relationships must be minimized [55]
Censoring/Scrubbing	Removal of motion-contaminated volumes [54]	Datasets with sudden motion bursts; when distance-dependent artifacts are primary concern [54]
fMRIPrep	Standardized preprocessing pipeline [3]	Ensuring consistent initial preprocessing across studies [3]

Visual Workflows for Denoising Pipelines

Decision Framework for fMRI Denoising Pipeline Selection

Optimization Pathway for Task-Based fMRI Denoising

The challenge of incomplete denoising in fMRI studies stems from fundamental trade-offs between different performance metrics, with no single pipeline achieving optimal results across all benchmarks. The selection of an appropriate denoising strategy must be guided by the specific research context, including the experimental paradigm, participant population, and primary analytical goals. For task-based fMRI where motion systematically differs between conditions, optimized aCompCor or FIX provide the best balance between noise removal and signal conservation. In resting-state studies of special populations, aggressive ICA-AROMA offers advantages for reproducibility. Critically, researchers should implement quality control metrics specific to their denoising approach and acknowledge the residual limitations in their interpretations of functional connectivity findings.

Functional magnetic resonance imaging (fMRI) has emerged as a pivotal tool for understanding brain function in traumatic brain injury (TBI) and psychiatric disorders. However, the blood-oxygen-level-dependent (BOLD) signal measured by fMRI is contaminated by substantial non-neural noise, with in-scanner head motion representing the most significant source of artifact [13] [12]. This challenge is particularly acute in clinical populations, where patients may exhibit abnormal movements such as posturing, tremor, dystonia, or restlessness due to their underlying conditions [13]. Motion artifacts systematically generate artifactual correlations across brain regions, leading to spurious functional connectivity (FC) results that can misinterpret the very neuropathology researchers seek to understand [13] [12].

The problem extends beyond mere data quality to fundamental questions of validity and reliability in clinical inference. As resting-state fMRI (rs-fMRI) becomes incorporated into some diagnostic guidelines for severe brain injury patients, ensuring the validity of findings through appropriate denoising becomes a clinical imperative [13]. This application note provides a structured framework for selecting and implementing denoising pipelines specifically tailored to the challenges of TBI and psychiatric disorders research, with particular emphasis on motion parameter regression strategies within a broader thesis on fMRI denoising.

Evaluating Denoising Pipeline Efficacy: Metrics and Comparative Performance

Quality Control Metrics for Pipeline Assessment

Selecting an optimal denoising pipeline requires evaluation against multiple quality control metrics that reflect the dual goals of noise removal and neural signal preservation. Based on large-scale evaluations across diverse populations, three primary classes of assessment metrics have emerged:

• Motion-FC Association: Quantifies the residual relationship between head motion (typically framewise displacement) and functional connectivity after denoising. Effective pipelines minimize this association, reducing distance-dependent biases where motion artificially decreases long-distance connectivity while increasing short-range connections [12].
• Brain-Behavior Correlation: Measures how denoising impacts the association between functional connectivity and behavioral or clinical variables. Pipelines should maximize valid brain-behavior relationships while minimizing spurious associations driven by motion [3].
• Trait-Specific Motion Impact: Novel methods like SHAMAN (Split Half Analysis of Motion Associated Networks) assign motion impact scores to specific trait-FC relationships, distinguishing between motion causing overestimation or underestimation of effects [12].

Comparative Performance of Denoising Strategies

Table 1: Efficacy of Denoising Pipelines Across Clinical and Healthy Populations

Pipeline	Key Components	Performance in TBI	Performance in Psychiatric Populations	Impact on Brain-Behavior Correlations
Spike Regression + Physiological Regressors	Framewise displacement censoring, physiological noise removal	Best performer in TBI populations [13]	Not specifically evaluated	Maintains biological interpretability while reducing motion artifacts
ICA-AROMA	Automatic ICA-based component classification, aggressive/non-aggressive options	Not specifically evaluated	Effective for motion-correlated traits [12]	Removes substantial physiological noise but may also remove low-frequency neural signals [38]
Global Signal Regression (GSR)	Regression of global brain signal	Use with caution in TBI with extensive lesions	Controversial due to potential neural signal removal	Can improve behavioral correlations but may introduce anticorrelations [3] [38]
aCompCor/tCompCor	PCA-based noise components from WM/CSF or high-variance voxels	Not the best performer in TBI [13]	Better at removing high-frequency physiological signals [38]	Associated with relatively higher age-related FC differences [38]
Structured Matrix Completion	Low-rank matrix completion to recover censored data	Not evaluated in clinical populations	Not evaluated in clinical populations	Improved delineation of networks like DMN with lower correlation errors [57] [2]

Large-scale evaluations in challenging clinical populations reveal that no single pipeline eliminates noise effects on functional connectivity completely [13]. The performance of denoising strategies varies significantly based on the exclusion regime, participant motion profiles, and the specific clinical population under investigation. For TBI patients, pipelines combining spike regression with physiological regressors have demonstrated superior performance, whereas automated data-driven methods performed comparatively worse in this population [13].

Experimental Protocols for Denoising Pipeline Implementation

Protocol 1: Comprehensive Denoising for TBI Populations

This protocol is adapted from the EpiBioS4Rx clinical trial for TBI patients, which conducted one of the largest evaluations of denoising strategies in this population [13].

Preprocessing Steps (Common to All Pipelines):

• Remove first four volumes to eliminate T1 equilibrium effects using FSL's fslroi
• Realignment of volumes to acquire raw motion parameters
• Slice-time correction
• Two-pass realignment (first to first volume, then to mean volume)
• Co-registration of EPI to native T1-weighted image via rigid-body registration
• Application of inverse nonlinear transform to EPI data
• Linear detrending of spatially normalized BOLD time series
• Intensity normalization to mode 1000 units
• Application of specific denoising pipeline
• Band-pass filtering (0.008-0.08 Hz)
• Spatial smoothing with 6mm FWHM kernel (except for ICA-AROMA)

Denoising Strategy Implementation:

• Spike Regression: Identify volumes with framewise displacement (FDJenk) >0.25mm and generate separate regressors for each contaminated volume [13]
• Physiological Regressors: Extract signals from eroded white matter (5 erosion cycles) and cerebrospinal fluid (2 erosion cycles) masks, with ventricles extracted for CSF
• Motion Parameter Regression: Include 24 motion parameters (6 rigid-body parameters + temporal derivatives + quadratic terms) to account for delayed and nonlinear motion-induced spin history effects [13]

Quality Control Checkpoints:

• Visual confirmation of accurate alignment after each registration step
• Framewise displacement and DVARS calculation for each participant
• Evaluation of residual motion-FC correlations after denoising

Protocol 2: Structured Matrix Completion for Data Recovery

For studies where censoring would result in excessive data loss, structured matrix completion approaches offer an alternative for recovering motion-corrupted volumes [57] [2].

Matrix Formation and Recovery:

• Excise fMRI volumes with elevated motion (FD > 0.2-0.3mm)
• Model remaining unprocessed fMRI data using motion parameters and slice-timing information
• Formulate artifact-reduction as recovery of a super-resolved matrix from sparse measurements
• Enforce low-rank prior on large structured Hankel matrix formed from time series samples
• Employ variable splitting strategy to solve optimization problem with reduced memory demand
• Recover complete time series with inherent slice-time correction

Validation Steps:

• Compare connectivity matrices from recovered data against censored and non-censored approaches
• Evaluate seed-based correlation analyses for improved network delineation (e.g., default mode network)
• Quantify errors in pair-wise correlation relative to ground truth when available

Integration Framework and Decision Pathway

The following workflow provides a systematic approach for selecting and implementing denoising pipelines based on study-specific characteristics:

Table 2: Essential Tools for fMRI Denoising in Clinical Research

Tool/Resource	Function	Application Context
fMRIPrep	Robust, analysis-agnostic preprocessing pipeline	Standardized preprocessing across diverse datasets; provides consistent motion correction, normalization, and confound estimation [51] [58]
ICA-AROMA	Automatic Removal of Motion Artifacts using ICA	Automatic identification and removal of motion-related components without requiring training data [38]
FSL	FMRIB Software Library (mcflirt, FIX)	Motion correction (mcflirt) and ICA-based denoising (FIX) [58]
ANTs	Advanced Normalization Tools	Sophisticated image registration and normalization, particularly valuable for brains with structural abnormalities [13]
CompCor	Anatomical and Temporal Component Noise Correction	PCA-based noise estimation from noise ROIs (aCompCor) or high-variance voxels (tCompCor) [38]
SHAMAN	Split Half Analysis of Motion Associated Networks	Quantifying trait-specific motion impact scores for specific brain-behavior relationships [12]
Structured Matrix Completion	Advanced matrix recovery for censored data	Recovery of motion-corrupted volumes without discarding data [57] [2]
ABCD-BIDS Pipeline	Integrated denoising for large-scale studies	Combines global signal regression, respiratory filtering, motion regression, and despiking [12]

Selecting optimal denoising pipelines for clinical fMRI research in TBI and psychiatric disorders requires careful consideration of population-specific characteristics, motion profiles, and research objectives. The evidence suggests that a one-size-fits-all approach to denoising is inadequate, with pipeline performance varying substantially across clinical populations. For TBI research, combined approaches incorporating spike regression and physiological regressors currently demonstrate superior performance, while data-driven approaches like ICA-AROMA may offer advantages in psychiatric populations where motion is correlated with traits of interest.

Future directions in the field include the development of trait-specific motion impact assessments that move beyond general motion metrics to evaluate how residual artifacts affect specific research questions [12]. Additionally, structured matrix completion methods show promise for addressing the fundamental tension between data retention and artifact removal, particularly in clinical populations where data loss from censoring may exclude precisely those patients most critical to study [57] [2]. As the field advances, the integration of multiple denoising approaches with careful attention to population-specific validation will be essential for ensuring the validity and reproducibility of clinical fMRI findings.

In functional magnetic resonance imaging (fMRI) research, scrubbing is a widely used technique to mitigate the impact of motion artifacts on functional connectivity measures. However, this method necessitates the exclusion of contaminated volumes, leading to a direct reduction in temporal degrees of freedom (tDOF), which can compromise the statistical power and validity of downstream analyses [55]. This application note examines the critical balance between effective noise removal through scrubbing and the preservation of data integrity. We present a systematic comparison of scrubbing methodologies, provide protocols for their implementation, and introduce data-driven alternatives that minimize data loss while effectively controlling for motion artifacts, framed within the broader context of motion parameter regression for denoising fMRI data.

Quantitative Comparison of Denoising and Scrubbing Strategies

The table below summarizes the performance of various denoising strategies, including scrubbing and its combinations with other methods, based on key quality metrics such as tDOF loss and effectiveness in motion artifact reduction.

Table 1: Performance Comparison of fMRI Denoising and Scrubbing Pipelines

Pipeline Number & Name	Key Components	tDOF Loss	Residual Motion Artifact	Best-Suited Population	Key Trade-offs
Motion Scrubbing [5] [55]	Removal of volumes with high Framewise Displacement (FD)	High (Direct loss of volumes)	Low	General populations with low-motion data	High data loss, altered temporal structure [5]
Spike Regression (SpikeReg) [59] [55]	Regression of spike regressors for high-motion volumes	Moderate	Low	Populations with sporadic motion	Less data loss than scrubbing, but uses regressors
ICA-AROMA (non-aggressive) [5]	ICA-based noise component classification and regression	Low	Moderate	Healthy adults & teenagers [5]	Less effective for high-motion data [5]
ICA-AROMA (aggressive) [5]	ICA-based noise component removal	Low	Low	Older adults [5]	Potential slight signal alteration
Projection Scrubbing [60]	Data-driven outlier detection via ICA/dimension reduction	Low (Targeted removal)	Low	Broad, including high-motion studies	Superior balance of retention & denoising [60]
Combination: ICA-AROMA + Scrubbing [59]	Pipeline #9, #14 in comparative studies	Moderate to High	Very Low	Non-lesional encephalopathic conditions [59]	Enhanced denoising at cost of higher tDOF loss
Combination: aCompCor + Scrubbing [59]	Pipeline #12, #16 in comparative studies	Moderate to High	Very Low	Lesional conditions (e.g., glioma) [59]	Enhanced denoising at cost of higher tDOF loss

The selection of an optimal pipeline is highly dependent on the specific study population. Research indicates that at comparable motion levels, combinations involving ICA-AROMA are most effective for non-lesional conditions (e.g., encephalopathy), whereas combinations including Anatomical Component Correction (aCompCor) yield the best results for lesional conditions (e.g., glioma, meningioma) [59]. Furthermore, the characteristics of motion and physiological noise can vary with age, necessitating age-specific optimization; for instance, aggressive ICA-AROMA has been identified as particularly suitable for older adult populations [5].

Experimental Protocols for Evaluating Scrubbing Strategies

Protocol 1: Benchmarking Scrubbing Pipelines Using QC-FC Correlation

Objective: To evaluate the efficacy of scrubbing and other denoising pipelines in removing motion artifacts by assessing the residual relationship between head motion and functional connectivity (QC-FC correlation) [59] [55].

Materials:

• Preprocessed rs-fMRI data (e.g., via fMRIPrep)
• Participant head motion parameters (Framewise Displacement)
• Computing environment (e.g., HALFpipe toolbox, FSL, AFNI)

Procedure:

• Data Preparation: Extract mean Framewise Displacement (FD) for each subject as the quantitative measure of head motion.
• Pipeline Application: Apply multiple denoising pipelines to the same dataset. Essential pipelines to compare include:
  • Motion scrubbing (e.g., using a threshold of FD > 0.2 mm)
  • Spike Regression
  • ICA-AROMA (aggressive and non-aggressive)
  • Projection scrubbing
  • Combinations (e.g., ICA-AROMA + scrubbing)
• Functional Connectivity Calculation: For each cleaned dataset, compute a whole-brain functional connectivity matrix (e.g., using Pearson correlation between region time series from a chosen atlas).
• QC-FC Correlation: For each pipeline, correlate the subject-level mean FD with each edge in the functional connectivity matrix across subjects.
• Evaluation: The performance of a pipeline is inversely related to the strength of the observed QC-FC correlations. Pipelines that successfully mitigate motion artifact will show weaker QC-FC correlations [55].

Protocol 2: Assessing Impact on Temporal Degrees of Freedom and Network Identifiability

Objective: To quantify the trade-off between data retention (tDOF) and the quality of denoising, measured by the identifiability of resting-state networks (RSNs) [59] [5].

Materials:

• Denoised fMRI data from Protocol 1
• Standard RSN templates (e.g., from Smith et al.)

Procedure:

• Calculate tDOF Loss: For each subject and pipeline, compute the effective tDOF after denoising. For scrubbing, this is simply the number of volumes retained. For regression-based strategies, calculate as tDOF = N - k, where N is the number of time points and k is the number of regressors used.
• RSN Identifiability: For each pipeline, perform group-level Independent Component Analysis (ICA) to extract RSNs. Alternatively, use dual regression to back-project standard RSN maps into each subject's data.
• Quantification: Calculate the spatial correlation between the derived RSNs and standard template maps. Higher spatial correlation indicates better network identifiability.
• Balance Analysis: Plot tDOF against RSN identifiability for each pipeline. The optimal pipeline achieves high RSN identifiability with minimal tDOF loss.

Visualizing the Scrubbing Strategy Selection Workflow

The following diagram outlines a systematic decision process for selecting an appropriate scrubbing strategy, balancing data retention with denoising efficacy.

Table 2: Key Software Tools and Resources for fMRI Denoising

Tool/Resource Name	Type/Brief Description	Primary Function in Denoising
HALFpipe [25]	Harmonized AnaLysis of Functional MRI pipeline	Standardized workflow from raw data to group stats; integrates multiple denoising options.
fMRIPrep [25]	Robust fMRI preprocessing pipeline	Provides standardized preprocessed data, a crucial starting point for all denoising strategies.
ICA-AROMA [5]	ICA-based Automatic Removal Of Motion Artifacts	Classifies and removes motion-related ICA components without requiring manual training.
FSL [25]	FMRIB Software Library	A comprehensive library of MRI analysis tools, including MELODIC for ICA and FEAT for model regression.
AFNI [25]	Analysis of Functional NeuroImages	Provides a suite for analyzing and visualizing functional MRI data, including 3dTproject for confound regression.
Quality Control Metrics (QC-FC, tDOF, RSN-ID) [59] [55]	Quantitative benchmarking metrics	Used to objectively evaluate and compare the performance of different denoising pipelines.

Effectively mitigating data loss in fMRI analysis requires a carefully considered balance. While scrubbing is a powerful tool, its aggressive application can severely diminish temporal degrees of freedom. The emergence of data-driven methods like projection scrubbing offers a promising path forward, providing robust denoising while maximizing data retention [60]. Furthermore, the optimal denoising strategy is not universal; it must be tailored to the specific clinical population [59] and age group [5] under study. By adopting the systematic evaluation protocols and decision framework outlined in this document, researchers can make informed, evidence-based choices in their motion parameter regression pipelines, enhancing the reliability and reproducibility of their functional connectomics findings.

Resting-state functional magnetic resonance imaging (rs-fMRI) is a pivotal tool for mapping the brain's intrinsic functional organization by measuring spontaneous blood-oxygenation-level-dependent (BOLD) signal fluctuations [3]. However, the BOLD signal is notoriously susceptible to contamination by multiple noise sources, particularly subject motion, which can severely compromise the reliability and validity of functional connectivity (FC) estimates [3] [4]. While motion denoising is a universal challenge in fMRI research, it presents unique considerations in older adult and pediatric populations due to distinct physiological and behavioral characteristics. Effective denoising is especially critical for Brain-Wide Association Studies (BWAS) seeking to link FC with behavior, as noise can attenuate or artificially inflate these correlations [3]. This Application Note provides population-specific tuning guidelines for motion parameter regression and other denoising techniques, framed within the broader context of optimizing fMRI preprocessing pipelines for reliable neuroimaging research across the lifespan.

Population-Specific Considerations

Older Adults (Aged 60+ Years)

The BOLD signal in older adults is a complex convolution of neural and vascular factors. Advancing age affects cerebral vascular reactivity and neurovascular coupling, meaning that age-related differences in BOLD signal may reflect vascular rather than neural changes [61]. The resting-state fluctuation amplitude (RSFA) in the fMRI signal has been validated as an index of vascular reactivity and is particularly useful for large-scale studies of aging where alternative measures like breath-hold are impractical [61]. Physiological noise characteristics also differ substantially in older adults, necessitating specialized noise regression techniques [5].

Pediatric Populations (Aged 5-12 Years)

Childhood development features unique neuromaturational processes that influence fMRI data quality and interpretation. While basic retinotopic organization and population receptive field (pRF) properties in early visual cortex are largely adult-like by middle childhood [62] [63], subtle developmental changes continue through adolescence. Notably, children exhibit similar visual performance at the lower and upper vertical meridian, unlike adults who show performance asymmetries—a difference mirrored in the distribution of V1 cortical surface area [63]. Children also typically exhibit greater motion during scanning, requiring robust denoising approaches that accommodate this expected behavior.

Denoising Pipeline Efficacy and Recommendations

Comparative Performance of Denoising Techniques

Table 1: Denoising Pipeline Efficacy Across Populations

Pipeline	Older Adults	Pediatric Cohorts	Key Advantages	Key Limitations
Aggressive ICA-AROMA	Recommended for longitudinal studies; high reproducibility [5]	Not specifically tested in found studies	Preserves temporal structure; pre-trained classifier works across sites [5]	May remove neural signal of interest
Censoring (scrubbing)	Performs well but alters temporal autocorrelation [5]	Effective for high-motion subjects; requires parameter optimization [4]	Directly removes motion-contaminated volumes	Reduces temporal degrees of freedom; creates non-continuous data [3]
Global Signal Regression (GSR)	Combined with censoring for effective denoising [5]	Shows variable effects on behavioral prediction [3]	Reduces widespread motion-related artifacts	Removes potential biologically relevant signal; controversial interpretation [3]
Non-aggressive ICA-AROMA	Lower performance compared to aggressive variant [5]	Recommended for general use in younger populations [5]	Less removal of potential neural signal	Lower efficacy at removing motion artifacts [5]

Optimized Pipeline Recommendations

For older adult populations, aggressive ICA-AROMA is currently the most suitable noise regression technique, demonstrating optimal reproducibility for longitudinal studies with low false-positive rates and better preservation of temporal structure [5]. This recommendation is based on comprehensive evaluation using data from 434 older adults (60-84 years) in the Risk Reduction for Alzheimer's Disease trial, which considered network reproducibility, identifiability, edge activity, spatial smoothness, and loss of temporal degrees of freedom [5].

For pediatric populations, while specific studies in the search results didn't evaluate denoising efficacy, general principles for high-motion populations apply. Censoring-based approaches effectively address motion spikes common in children, though they should be carefully tuned to minimize data loss [4]. Non-aggressive ICA-AROMA has been recommended for younger populations based on studies in teenagers and adults [5].

Experimental Protocols for Population-Specific Tuning

RSFA Scaling Protocol for Older Adults

Purpose: To correct for age-related changes in vascular reactivity that confound neural BOLD signal interpretation [61].

Materials:

• Resting-state fMRI data (6-10 minutes)
• Physiological monitoring (cardiac, respiratory) if available
• Processing software with bandpass filtering capabilities

Procedure:

• Acquire resting-state fMRI using standard parameters (e.g., TR=2-3s, 6-10 minutes)
• Preprocess with standard motion correction, spatial normalization, and minimal smoothing
• Bandpass filter BOLD signal (typically 0.01-0.1 Hz)
• Calculate RSFA as standard deviation of filtered BOLD timeseries within gray matter masks
• Apply RSFA scaling to task-based BOLD data using regression approaches
• Verify reduced age-effects on vascular components post-scaling [61]

Validation: Mediation analysis should confirm that age effects on RSFA are significantly mediated by vascular factors (e.g., heart rate variability) but not by neural activity variability [61].

Optimized Volume Censoring Protocol for High-Motion Populations

Purpose: To remove motion-contaminated volumes while preserving data integrity in pediatric or clinical populations with elevated motion.

Materials:

• fMRI timeseries data
• Framewise displacement (FD) calculations
• Dataset-specific quality control metrics [4]

Procedure:

• Compute framewise displacement (FD) from motion parameters
• Determine optimal FD threshold (e.g., 0.2-0.5mm) using quantitative methods [4]
• Identify contaminated volumes exceeding FD threshold
• Apply censoring by excluding contaminated volumes from functional connectivity calculations
• Combine with interpolation or structured covariance approaches for severe motion
• Validate using quantitative metrics agnostic to QC-FC correlations [4]

Optimization: Develop dataset-specific optimal censoring parameters using quantitative methods that evaluate motion denoising efficacy independent of assumptions about true motion-FC relationships [4].

Quality Control Protocol for Multi-Population Studies

Purpose: To ensure consistent data quality across populations with different motion and physiological characteristics.

Materials:

• AFNI software package for QC metrics [64]
• FreeSurfer for anatomical processing
• Multi-modal data (T1-weighted, resting-state fMRI)

Procedure:

• Implement "getting to know your data" (GTKYD) stage to understand acquisition properties [64]
• Apply quantitative measures with thresholds (APQUANT) including:
  • Temporal signal-to-noise ratio (TSNR) maps
  • Framewise displacement calculations
  • Anatomical coverage assessment
• Conduct qualitative image assessment (APQUAL) via systematic HTML reports
• Perform interactive graphical user interface (GUI) checks for subset verification
• For task data: verify stimulus event timing statistics (STIM stage) [64]

The Scientist's Toolkit: Essential Research Reagents

Table 2: Key Research Reagents and Computational Tools

Item	Function/Application	Population-Specific Considerations
ICA-AROMA	ICA-based Automatic Removal of Motion Artifacts; classifies and removes motion-related components [5]	Use aggressive variant for older adults; non-aggressive for general population [5]
fMRIPrep	Standardized preprocessing pipeline; ensures consistent initial processing [3]	Compatible with population-specific tuning in downstream analysis
AFNI QC Tools	Comprehensive quality control metrics and visualization [64]	Essential for verifying data quality across diverse populations
RSFA Scaling Scripts	Implements resting-state fluctuation amplitude correction for vascular effects [61]	Critical for older adult studies to account for neurovascular uncoupling
Censoring Optimization Code	Determines dataset-specific optimal censoring parameters [4]	Particularly valuable for pediatric and clinical populations with elevated motion
Population Receptive Field (pRF) Modeling	Characterizes spatial tuning properties in visual cortex [63]	Useful for developmental studies to establish neural maturity benchmarks

Workflow and Decision Pathways

Diagram 1: Population-Specific fMRI Denoising Decision Pathway

Implementation Guidelines and Best Practices

Multi-Site and Longitudinal Considerations

When implementing these population-specific approaches in multi-site studies or longitudinal designs, additional standardization is crucial. Both ICA-AROMA and SOCK contain pre-trained classifiers that do not require retraining on new data, making them suitable for studies aggregating data from multiple scanners and sites [5]. For longitudinal studies of aging, reproducibility should be prioritized as the most important factor in pipeline selection [5].

Quantitative Validation Metrics

Implement denoising efficacy evaluation using metrics that are agnostic to quality control-functional connectivity (QC-FC) correlations, as assumptions underlying common metrics (especially those based on QC-FC correlations and differences between high- and low-motion participants) are problematic and may be inappropriate as indicators of comparative pipeline performance [4]. Instead, develop and utilize quantitative methods that determine dataset-specific optimal parameters prior to final analysis [4].

Population-specific tuning of motion parameter regression and denoising pipelines is essential for valid and reliable fMRI research across the lifespan. For older adults, accounting for age-related vascular changes through RSFA scaling and utilizing aggressive ICA-AROMA provides optimal correction for confounding factors. For pediatric populations, optimized volume censoring approaches address characteristically higher motion while preserving neural signals of interest. Implementation of these specialized protocols, combined with rigorous population-specific quality control, will enhance the fidelity of functional connectivity measurements and strengthen conclusions in developmental and aging research.

Functional magnetic resonance imaging (fMRI) has become an indispensable tool for studying brain function in both research and clinical settings. However, the blood oxygenation level-dependent (BOLD) signal of interest in fMRI is notoriously small, typically representing only 1-5% of the total measured signal variability [65]. This modest signal is easily obscured by various sources of noise, among which subject motion stands as one of the most significant confounding factors [66]. Motion artifacts reduce statistical significance in activation maps, increase the prevalence of false activations, and substantially diminish the reliability and reproducibility of functional connectivity (FC) estimates [66] [3] [4]. While traditional retrospective motion correction (RMC) techniques have been widely adopted, they possess inherent limitations in addressing the full spectrum of motion-induced artifacts. This creates a critical need for proactive measures that address motion at the time of acquisition, positioning Prospective Motion Correction (PMC) as an essential advancement for robust denoising in fMRI research.

Theoretical Foundations: Prospective versus Retrospective Correction

The Fundamental Limitations of Retrospective Approaches

Retrospective motion correction algorithms, such as those implemented in widely used tools like FSL's MCFLIRT and SPM's realign, operate by applying rigid-body transformations to each volume during post-processing to align all acquired volumes to a reference volume [67] [65]. While these methods are convenient and have been successfully applied to correct for slow motion between acquisitions, they suffer from several fundamental limitations:

• Inability to correct for spin history effects: RMC cannot address the changes in longitudinal magnetization that occur when a brain region moves through different slice positions before the magnetization has fully recovered [66] [68].
• No correction for intra-volume motion: Conventional RMC implementations typically correct for inter-volume motion but neglect intra-volume movement between slices, which is particularly problematic for multi-slice acquisitions [66] [65].
• Voxel loss in partial brain acquisitions: As the acquisition box is not coupled to the brain, edge voxels can be lost due to subject motion [68].
• Failure to address k-space inconsistencies: RMC cannot correct for distortions in k-space data caused by motion during the acquisition process [68].

The Prospective Correction Paradigm

In contrast, Prospective Motion Correction utilizes real-time tracking of head movement to dynamically update the imaging field-of-view, keeping the scan plane orientation and position constant with respect to the head throughout acquisition [68] [69]. By updating the imaging gradients, radiofrequency (RF) frequency, and phase each repetition time (TR), PMC effectively couples the acquisition box to the participant's brain, thereby addressing both intra- and inter-volume motion at the source [68]. This fundamental difference in approach allows PMC to overcome the primary limitations of RMC, particularly for high-resolution protocols and studies involving populations prone to movement.

Table 1: Comparative Analysis of Motion Correction Techniques

Feature	Retrospective Motion Correction (RMC)	Prospective Motion Correction (PMC)
Correction principle	Post-acquisition image realignment	Real-time adjustment of scan plane during acquisition
Spin history effects	Unable to correct	Effectively addresses
Intra-volume motion	Limited correction	Slice-wise realignment capability
Hardware requirements	None (software-only)	Optical tracking system + marker attachment
Impact on acquisition	None	Requires sequence support for real-time updates
Computational load	Post-processing	Real-time during acquisition
Edge voxel preservation	Limited	Excellent

Quantitative Efficacy of PMC

Enhanced Statistical Power in Task-Based fMRI

Empirical evidence demonstrates that PMC significantly enhances the statistical power of BOLD fMRI measurements. In a comprehensive evaluation using a prospective active marker motion correction (PRAMMO) system, researchers observed substantial increases in both the spatial extent and statistical significance of task-specific BOLD signals [65]. The system utilized three active radiofrequency markers integrated into a rigid plastic headband to track head motion in real-time, enabling online slice plane correction. When applied to visual and motor paradigms, this approach resulted in:

• Increased activation cluster sizes by at least 10% compared to conventional retrospective correction [65]
• Significantly higher z-scores in activated regions at the group level [65]
• Reduction in residual variance without decreasing the signal component of the BOLD response [65]

Notably, these improvements were achieved under typical experimental conditions with naturalistic levels of subject motion, rather than exaggerated deliberate motions that might overestimate benefits [68] [65].

Benefits in High-Resolution and Quantitative Imaging

The advantages of PMC become particularly pronounced in high-resolution imaging protocols, where even minute movements can profoundly degrade data quality. In an evaluation of PMC for high-resolution (800 μm isotropic) multi-parameter mapping (MPM), researchers found that PMC considerably improved map quality in the presence of head motion, reflected by fewer visible artifacts and improved consistency [70] [69]. The precision of quantitative maps, parameterized through the coefficient of variation in cortical sub-regions, showed improvements of 11-25% in the presence of deliberate head motion [70] [69]. Crucially, in the absence of motion, the PMC system did not introduce extraneous artifacts into the quantitative maps, demonstrating its safety for general application [70] [69].

Table 2: Quantitative Benefits of PMC Across Imaging Modalities

Imaging Modality	Resolution	Key Improvement Metric	% Improvement with PMC
Task-based fMRI [65]	3-4 mm	Activation cluster size	≥10%
Multi-parameter Mapping [70] [69]	800 μm isotropic	Coefficient of variation in cortical regions	11-25%
Multivoxel Pattern Analysis [68]	1.5 mm isotropic	Pattern decoding accuracy	Most apparent at higher resolutions
Resting-state fMRI [4]	Multiband EPI	Functional connectivity reliability	Significant improvement in motion-affected data

Impact on Multivoxel Pattern Analysis and Decoding

The benefits of PMC extend beyond conventional univariate analyses to more sophisticated multivariate approaches. Research has demonstrated that the advantage of PMC is most apparent for multi-voxel pattern decoding at higher resolutions [68]. In studies examining visual cortical response patterns, PMC enhanced the accuracy of multivariate decoding in primary visual cortex (V1), particularly for analyses that require accurate voxel registration across time [68]. This suggests that PMC is increasingly important for advanced fMRI analyses that leverage distributed pattern information rather than isolated activation foci.

Implementation Protocols and Methodologies

Optical Tracking Systems and Marker Attachment

The most widely evaluated PMC approach for fMRI utilizes optical tracking systems with specialized marker technology. One implemented system employs an optical camera mounted inside the scanner bore that tracks the motion of a passive Moiré phase marker at a frame rate of 80 Hz [69]. The gratings and patterns on the marker enable measurement of all three translational and three rotational degrees of freedom with precision on the order of tens of microns for translations and fractions of degrees for rotations [69].

A critical implementation challenge is ensuring rigid coupling between the tracking marker and the participant's head. Research consensus indicates that skin attachment is insufficiently rigid [68]. The most effective solutions utilize:

• Custom-molded mouthpieces: Created using medical-grade hydroplastic molded to the participant's upper front teeth [69]
• Mini bite-bars: That remain securely in place without requiring continuous biting pressure [69]
• Lightweight mounting systems: Made from plastic components that allow flexible positioning within the camera's field of view [69]

Comparative testing has demonstrated that inexpensive, commercially available mouthpiece solutions can produce comparable results to dentist-molded alternatives, improving practicality for widespread implementation [68].

Integration with fMRI Acquisition Sequences

For 2D echo-planar imaging (EPI) sequences typical of fMRI, PMC implementation requires pulse sequence modifications to accept real-time position updates. The tracking information is transformed from camera to scanner coordinates using a pre-calibrated transformation matrix and used to dynamically update the imaging field-of-view (FOV) such that it tracks the movement of the marker [69]. The imaging gradients, RF frequency, and phase are updated each TR, enabling slice-wise realignment that corrects for both intra- and inter-volume motion [68].

This integration preserves edge voxels that might otherwise be lost in partial brain acquisitions and maintains consistent voxel-wise registration across volumes, which is particularly crucial for high-resolution studies and multivariate pattern analysis [68].

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Components for PMC Implementation

Component	Function	Implementation Example
Optical Tracking Camera	Records marker position in real-time	In-bore camera tracking at 80 Hz frame rate [69]
Motion Marker	Provides visual pattern for tracking	Passive Moiré phase marker [69] or active RF markers [65]
Marker Attachment System	Ensures rigid head-marker coupling	Custom-molded mouthpiece [69] or active marker headband [65]
Real-Time Interface	Communicates position data to scanner	Direct link to scanner host computer [69]
Modified Pulse Sequences	Enables dynamic FOV adjustment	EPI sequences accepting real-time position updates [68]

Integration with Denoising Pipelines in Resting-State fMRI

While PMC addresses motion at the acquisition stage, comprehensive denoising for resting-state fMRI (rs-fMRI) typically involves additional post-processing steps. Current research indicates that no single pipeline universally excels at simultaneously mitigating motion-related artifacts and augmenting brain-behaviour associations across different cohorts [3]. Common post-processing approaches include:

• White matter and cerebrospinal fluid regression [3]
• Independent component analysis (ICA)-based artefact removal (e.g., ICA-FIX) [3]
• Volume censoring ("scrubbing") [4]
• Global signal regression (GSR) [3]
• Diffuse cluster estimation and regression (DiCER) [3]

Studies suggest that pipelines combining ICA-FIX and GSR demonstrate a reasonable trade-off between motion reduction and behavioral prediction performance [3]. However, PMC offers the fundamental advantage of reducing the burden on these post-processing techniques by addressing the motion problem at its source, potentially leading to more robust and biologically valid results.

Experimental Workflow for PMC-Enhanced fMRI

The following diagram illustrates the comprehensive workflow for implementing Prospective Motion Correction in fMRI studies, integrating both prospective acquisition and retrospective denoising components:

Prospective Motion Correction represents a significant advancement in the quest for robust motion denoising in fMRI research. By addressing motion artifacts at their source during data acquisition rather than relying solely on post-processing corrections, PMC provides fundamental improvements in data quality that complement existing denoising pipelines. The technique demonstrates particular value for high-resolution imaging, quantitative mapping, multivariate pattern analysis, and studies involving populations prone to movement. While implementation requires specialized hardware and careful attention to marker attachment, the resulting enhancements in statistical power, reliability, and validity position PMC as an essential component of a comprehensive motion denoising strategy in modern fMRI research.

Benchmarking Success: A Multi-Metric Framework for Pipeline Validation

In the domain of functional magnetic resonance imaging (fMRI) research, motion-related artifacts represent a principal confound that significantly diminishes the reliability and validity of functional connectivity (FC) estimates [3] [42]. The evaluation of denoising efficacy, particularly within the context of motion parameter regression, relies critically on two pivotal benchmarks: QC-FC correlations and motion-dependent distance dependence [4] [42]. QC-FC correlation quantifies the association between subject motion and observed functional connectivity, a relationship that effective denoising should minimize [3] [4]. Concurrently, motion-induced artifacts exhibit a characteristic spatial pattern, inflating short-distance correlations more than long-distance ones; thus, the reduction of this distance-dependent bias serves as a second crucial metric for evaluating preprocessing pipelines [42]. This protocol details the application of these benchmarks, providing a standardized framework for assessing the performance of denoising strategies in resting-state fMRI studies, with a specific focus on their integration within a broader thesis on motion parameter regression.

Theoretical Foundation and Significance

The Problem of Motion in fMRI

Resting-state fMRI (rs-fMRI) signals are notoriously contaminated by multiple sources of noise, with head motion being a predominant challenge [3] [42]. Even small movements can introduce spurious correlations that do not reflect true neural synchrony [42]. These motion artifacts are not random; they systematically bias connectivity estimates, often inflating correlations between physically proximate brain regions while having less impact on long-range connections [42]. This creates a distance-dependent confound that can invalidate group comparisons, particularly when comparing populations with different inherent motion characteristics (e.g., patients vs. controls, children vs. adults) [42].

QC-FC Correlations as a Benchmark

The QC-FC correlation is a direct measure of how much a subject's motion level (the Quality Control, or QC, metric) predicts their functional connectivity (FC) [4]. In inadequately denoised data, higher motion is often associated with specific patterns of altered connectivity, leading to strong QC-FC correlations [3] [4]. A successful denoising pipeline must therefore mitigate this relationship, resulting in QC-FC correlations that are weak and non-significant [4]. It is critical to note that recent research has questioned the universal suitability of QC-FC correlations as a sole benchmark, suggesting they can be problematic when motion is genuinely correlated with trait-level neural characteristics [4]. This underscores the necessity of a multi-measure approach to evaluation [42].

Distance Dependence as a Benchmark

Motion artifacts manifest with a spatial fingerprint. Because the signal displacement caused by head movement is more similar for voxels that are close together, motion tends to artificially inflate short-distance correlations [42]. A key signature of effective denoising is, therefore, the reduction or elimination of this distance-dependent bias. The pipeline should produce connectivity matrices where the relationship between physical distance and correlation strength is no longer driven by motion, thereby reflecting more accurately the underlying neurobiology [42].

Experimental Protocols for Benchmarking

Core Workflow for Evaluation

The following workflow provides a standardized procedure for evaluating denoising pipelines using QC-FC and distance-dependence metrics. This process can be applied to compare multiple pipelines or to optimize parameters for a single dataset.

Protocol 1: Quantifying QC-FC Correlations

This protocol measures the residual relationship between motion and connectivity after denoising.

• Inputs: Preprocessed BOLD time series for all subjects; framewise displacement (FD) time series for all subjects.
• Procedure:
  • Compute Summary Motion Metric: For each subject, calculate the mean Framewise Displacement (mean FD) from the six rigid-body motion parameters [4] [5]. This serves as the primary QC variable.
  • Generate Denoised Time Series: Apply the denoising strategy under evaluation (e.g., motion parameter regression, ICA-AROMA, GSR) to the preprocessed BOLD data [3] [50].
  • Estimate Functional Connectivity: From the denoised time series, compute a full correlation (or partial correlation) matrix for each subject using a predefined brain atlas (e.g., Schaefer, Gordon) [71]. The result is an M x M connectivity matrix per subject, where M is the number of brain regions.
  • Calculate QC-FC Correlations: For each unique connection (edge) between regions i and j across all subjects:
    • Extract the Fisher Z-transformed correlation strength for that edge for all N subjects.
    • Extract the mean FD for all N subjects.
    • Compute the Pearson's correlation (across subjects) between the edge strength and mean FD.
  • Evaluate Results: The output is a distribution of correlation coefficients (one for each edge). The efficacy of the denoising pipeline is judged by the central tendency and spread of this distribution. A well-performing pipeline will yield a QC-FC distribution that is centered on zero with a narrow spread [4].

Protocol 2: Quantifying Distance Dependence of Motion Artifacts

This protocol assesses the extent to which the denoising pipeline removes the characteristic spatial pattern of motion artifacts.

• Inputs: Denoised functional connectivity matrices from Protocol 1; Euclidean distance matrix between brain regions.
• Procedure:
  • Compute Euclidean Distance Matrix: Based on the centroid coordinates of the regions in the atlas, calculate an M x M matrix where each element d_ij is the physical distance (in mm) between region i and region j.
  • Calculate Edge-wise Motion Effect: For each subject and each edge, a measure of the motion-related change is needed. This is often derived by comparing a high-motion and a low-motion state within the same subject, but a common proxy is to use the QC-FC correlation value for that edge from Protocol 1 as a group-level measure of its sensitivity to motion [42].
  • Relate Motion Effect to Distance: For each edge, plot its QC-FC value (y-axis) against the Euclidean distance between its two nodes (x-axis). Perform a robust linear regression to quantify the relationship.
  • Evaluate Results: A significant negative slope indicates that motion artifacts are stronger for shorter-distance connections, which is the hallmark of residual motion contamination. An effective denoising pipeline should attenuate this negative slope, ideally rendering it non-significant [42].

Quantitative Benchmark Data

Performance of Common Denoising Pipelines

The following table synthesizes benchmark results from recent large-scale studies, illustrating how different pipeline performances can be quantified using the described metrics.

Table 1: Performance of Common Denoising Pipelines on QC Benchmarks

Denoising Pipeline	QC-FC Correlation (Distribution Mean ± SD)	Distance Dependence (Slope)	Key Trade-offs and Considerations
24-Parameter Model (6 MP + derivatives + squares) [5]	Moderate reduction	Moderate negative slope	Less effective at removing motion-related spikes; a common baseline.
Global Signal Regression (GSR) [3] [42]	Strong reduction	Strong attenuation of negative slope	Controversial due to potential introduction of negative correlations and removal of neural signal [3].
ICA-AROMA (aggressive) [5] [50]	Strong reduction	Strong attenuation	Effective for motion and physiological noise; performs well in older adults [5].
Volume Censoring (Scrubbing) [4] [50]	Very strong reduction	Strong attenuation	Disrupts temporal continuity; high cost in temporal degrees of freedom (tDOF); may not be suitable for all analyses [4] [50].
aCompCor [42] [50]	Moderate reduction	Moderate attenuation	Data-driven; avoids some GSR controversies; performance depends on number of components [42].

Advanced Connectivity Methods

The choice of functional connectivity metric itself can influence benchmark outcomes. While Pearson's correlation is the default, other measures offer different properties.

Table 2: Impact of Functional Connectivity Metric on Benchmarking [71]

FC Metric Family	Representative Method	Structure-Function Coupling (R²)	Note on Benchmarking
Covariance	Pearson's Correlation	~0.15	The default method; baseline for comparison.
Precision	Partial Correlation	~0.25	Attenuates indirect effects; may require regularization for large networks [72].
Distance	Euclidean Distance	~0.18	A dissimilarity measure; expect a positive QC-FC relationship.
Spectral	Coherence	~0.10	Sensitive to frequency-specific interactions.

This section details the essential software tools and data resources required to implement the described benchmarking protocols.

Table 3: Essential Resources for fMRI Denoising Benchmarking

Resource Name	Type	Primary Function	Application in Protocol
fMRIPrep [50]	Software	Standardized, robust fMRI preprocessing	Generates preprocessed BOLD time series and a comprehensive set of noise regressors (motion parameters, tissue masks, etc.). The starting point for denoising.
ICA-AROMA [5] [50]	Software/Algorithm	Data-driven removal of motion artifacts via ICA	A specific, highly effective denoising pipeline to be benchmarked. Classifies and removes motion-related independent components.
Nilearn [50] [71]	Python Library	Statistical learning and analysis of neuroimaging data	Used to compute functional connectivity matrices, perform QC-FC correlations, and implement various denoising strategies (regression, scrubbing).
CONN	Software Toolbox	Functional connectivity analysis	Provides a GUI-based alternative for implementing preprocessing and denoising pipelines, and calculating QC metrics.
HCP Dataset [4] [42] [71]	Data	High-quality, multiband fMRI data from healthy adults	A common benchmark dataset due to its high temporal resolution and low noise, allowing for clear evaluation of denoising methods.
ABCD Dataset [73]	Data	Large-scale developmental fMRI data	Useful for testing pipelines on a large, heterogeneous sample with variable motion characteristics.
pySPI [71]	Python Library	Calculation of 200+ pairwise similarity metrics	Allows for the benchmarking of denoising efficacy across a wide range of functional connectivity measures beyond Pearson's correlation.

Integrated Evaluation Workflow

To ensure a comprehensive assessment, the benchmarks should be applied collectively. The following diagram outlines the decision-making logic for interpreting the combined results.

The rigorous evaluation of denoising pipelines via QC-FC correlations and distance dependence is a non-negotiable step in modern rs-fMRI analysis [4] [42]. The protocols outlined herein provide a standardized framework for this critical assessment. Evidence suggests that no single pipeline universally excels, and the optimal choice may depend on the specific dataset and research question [3] [50]. Pipelines combining multiple techniques, such as ICA-based cleaning with GSR, often offer a favorable trade-off, effectively mitigating motion artifacts while preserving behaviorally relevant neural signals [3] [5]. By adopting these benchmarking practices, researchers can make informed, empirically-grounded decisions about denoising strategies, thereby enhancing the reliability and interpretability of their functional connectivity findings.

This application note provides a structured framework for assessing the identifiability and reproducibility of resting-state functional connectivity (RSFC) in clinical and cognitive neuroscience research. With a specific focus on integrating motion parameter regression into denoising pipelines, we present standardized protocols and quantitative benchmarks to enhance the reliability of network integrity metrics across diverse populations, including healthy aging and clinical cohorts such as traumatic brain injury (TBI) and vascular dementia (VD).

Resting-state functional magnetic resonance imaging (rs-fMRI) has become a pivotal tool for mapping the brain's intrinsic functional organization through blood oxygenation level-dependent (BOLD) signal correlations [3]. However, the reproducibility of resting-state network (RSN) metrics remains challenging due to multiple methodological variables, particularly motion artifacts and analytical choices. The reliability of RSFC is especially consequential in clinical populations where neurological compromise may alter the neurovascular coupling underlying BOLD signals [74]. Establishing standardized protocols for assessing network integrity is thus essential for developing robust biomarkers for drug development and clinical applications.

Recent evidence indicates that while RSFC methods can identify robust canonical networks in healthy and clinical samples, findings have been challenging to reproduce, especially in populations with significant neurological disruption such as TBI [74]. This application note addresses these challenges by providing a comprehensive framework for evaluating RSFC reliability, with particular emphasis on motion denoising strategies and their impact on network identifiability metrics.

Quantitative Foundations of Network Integrity

Reliability Benchmarks for Graph Metrics

Extensive reliability testing using intraclass correlation coefficients (ICCs) has established benchmarks for various graph theory metrics in both healthy aging and clinical populations. These benchmarks provide critical reference points for evaluating network integrity in research applications.

Table 1: Test-Retest Reliability of Global Graph Metrics (ICC Values)

Graph Metric	Reliability Classification	Representative ICC Range	Key Applications
Within-Network Connectivity	Excellent	0.75-0.90	Tracking system-level plasticity post-TBI
Network Segregation	Good-Excellent	0.70-0.85	Assessing network specialization in aging
Clustering Coefficient	Fair-Poor	0.40-0.60	Screening for gross connectivity alterations
Eigenvector Centrality	Fair-Poor	0.45-0.55	Identifying hub disruption in pathologies

Data derived from back-to-back 10-minute resting-state scans in healthy aging (n=41) and TBI (n=45) samples [74].

Canonical Network Reliability Profiles

Different resting-state networks demonstrate varying levels of intrinsic reliability, which must be considered when selecting networks for longitudinal intervention studies or clinical trial endpoints.

Table 2: Network-Specific Reliability and Sensitivity Profiles

Canonical Network	Reliability Ranking	Motion Sensitivity	Clinical Utility
Default Mode Network	Highest	Moderate	Alzheimer's disease, vascular dementia
Salience Network	High	Moderate-High	TBI, psychiatric disorders
Dorsal Attention Network	Moderate	Low-Moderate	Aging, attention deficits
Sensorimotor Network	Moderate	Low	Parkinson's disease, stroke
Frontoparietal Network	Moderate-Low	High	Executive function assessment

Research indicates the default mode and salience networks demonstrate the highest test-retest reliability, making them particularly suitable for longitudinal studies [74]. The sensorimotor network shows enhanced negative connectivity strength with adult age, providing a potential biomarker for aging studies [75].

Experimental Protocols for Network Integrity Assessment

Protocol 1: Basic RSFC Acquisition and Denoising

Purpose: To acquire resting-state fMRI data with optimized parameters for network identifiability while integrating comprehensive motion parameter regression.

Materials:

• 3T MRI scanner with phased-array head coil
• Gradient-echo EPI sequence (TR=720-2000ms, TE=30-33ms, flip angle=52°-90°)
• T1-weighted anatomical scan (MPRAGE, 1mm³ resolution)
• Pulse oximeter and respiratory bellows for physiological monitoring

Procedure:

• Participant Preparation: Instruct participants to keep eyes open or closed (standardize across study), remain still, and let mind wander without structured thinking.
• Structural Acquisition: Acquire high-resolution T1-weighted anatomical scan for spatial normalization.
• Functional Acquisition: Obtain 10-15 minutes of resting-state BOLD data (minimum 10 minutes for reliability).
• Motion Parameter Integration: Include rigid-body transformation parameters (3 translation, 3 rotation) plus their temporal derivatives in nuisance regression [3].
• Data Denoising: Apply integrated denoising pipeline combining:
  • Motion parameter regression (6 parameters + derivatives)
  • Tissue-based nuisance regression (white matter, CSF signals)
  • Global signal regression (context-dependent) [3]
  • Motion scrubbing (FD threshold = 0.5mm) [74]
• Temporal Filtering: Apply bandpass filter (0.01-0.1 Hz) to isolate low-frequency fluctuations.

Validation Metrics: Frame-wise displacement (FD) should not exceed 0.5mm mean; signal-to-noise ratio (SNR) >100; visual inspection for artifacts.

Protocol 2: Multiverse Reliability Analysis

Purpose: To evaluate the robustness of network integrity metrics across multiple processing pipelines, explicitly testing the impact of motion correction strategies.

Materials:

• Back-to-back resting-state scans (same session)
• Multiple brain atlases (e.g., AAL, Yeo-17, Brainnetome)
• Computational resources for parallel processing

Procedure:

• Data Acquisition: Collect consecutive resting-state scans (10 minutes each) with identical parameters.
• Pipeline Implementation: Execute 5 distinct processing pipelines varying in:
  • Motion correction approach (volume censoring, global signal regression, etc.)
  • Brain parcellation (8 different atlases)
  • Temporal filtering parameters
• Network Metric Calculation: Compute graph theory metrics for each pipeline combination.
• Reliability Assessment: Calculate intraclass correlation coefficients (ICCs) between back-to-back scans for each pipeline.
• Motion Sensitivity Analysis: Quantify the relationship between motion scrubbing intensity and ICC values.

Analysis: Identify pipelines producing optimal reliability (ICC>0.75) while minimizing motion artifacts. Research demonstrates a notable influence of motion scrubbing on ICCs, with diminished reliability proportional to the number of volumes removed [74].

Protocol 3: Dynamic Functional Network Connectivity (dFNC)

Purpose: To capture time-varying properties of network interactions and their modulation by motion parameters.

Materials:

• High-temporal resolution rs-fMRI (TR<1s)
• Sliding window analysis toolbox
• k-means clustering algorithms

Procedure:

• Data Acquisition: Obtain resting-state data with high temporal resolution (TR=720ms).
• Windowed Analysis: Apply sliding window (e.g., 30-60s) to compute dynamic FC matrices.
• State Identification: Cluster windowed FC matrices into discrete brain states.
• Temporal Characterization: Calculate dwell times, fractional windows, and transition probabilities.
• Motion Covariate Analysis: Regress motion parameters against state expression metrics.

Applications: This approach has demonstrated utility in differentiating vascular dementia from Alzheimer's disease, where static FC approaches show overlapping findings [76].

Visualizing Analytical Frameworks

Network Integrity Assessment Workflow

Diagram 1: Network integrity assessment workflow integrating motion parameter regression throughout the analytical pipeline.

Multiverse Reliability Analysis Framework

Diagram 2: Multiverse analysis framework evaluating reliability across multiple processing decisions.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Critical Resources for Network Integrity Research

Resource Category	Specific Tools/Reagents	Function/Application	Implementation Notes
Analysis Software	fMRIPrep, AFNI, FSL, CONN	Data preprocessing and denoising	fMRIPrep provides standardized preprocessing [3]
Brain Atlases	AAL, Yeo-17, Brainnetome, Glasser	Brain parcellation for network nodes	Atlas choice modestly affects reliability findings [74]
Motion Metrics	Framewise Displacement (FD), DVARS	Quantifying head motion artifacts	FD threshold of 0.5mm recommended for censoring [77]
Connectivity Metrics	Pearson correlation, distance metrics	Functional connectivity calculation	Correlation metrics outperform partial correlation for detecting neural decline [78]
Reliability Statistics	Intraclass Correlation Coefficients (ICCs)	Test-retest reliability assessment	ICCs >0.75 indicate excellent reliability [74]
Hybrid Decomposition Tools	NeuroMark Pipeline	Individualized network identification	Integrates spatial priors with data-driven refinement [79]

This application note provides a comprehensive framework for assessing resting-state network identifiability and reproducibility, with specific attention to integrating motion parameter regression into analytical pipelines. The protocols and benchmarks presented here enable researchers to quantitatively evaluate network integrity metrics while accounting for the confounding effects of motion.

Implementation of these standardized approaches is particularly crucial in clinical populations such as traumatic brain injury and vascular dementia, where neurological compromise may affect both functional connectivity and the neurovascular coupling underlying BOLD signals. By adopting these reproducible methods, researchers in drug development and clinical neuroscience can enhance the reliability of functional connectivity biomarkers for therapeutic development and patient stratification.

Future directions in the field include the development of dynamic fusion models that incorporate multiple time-resolved data streams and the refinement of generative models that can synthesize multimodal data to address missing data challenges in clinical research [79].

Resting-state functional magnetic resonance imaging (rs-fMRI) is a pivotal tool for mapping the brain's intrinsic functional organization and investigating its relationship to behaviour and cognition [45] [3]. However, the utility of rs-fMRI is compromised by a critical challenge: the measured signals are contaminated by multiple noise sources, particularly motion-induced artefacts, which diminish the reliability and validity of functional connectivity (FC) estimates [3] [4]. These artefacts can spuriously inflate or attenuate brain-behaviour associations in brain-wide association studies (BWAS), threatening the reproducibility of findings [45].

The field has responded by developing numerous data-processing pipelines for noise mitigation. Yet, this abundance creates a "combinatorial explosion problem" [80], making it difficult for researchers to select optimal strategies. This application note systematically evaluates popular denoising pipelines, providing performance comparisons across diverse datasets and offering detailed protocols for implementation, all within the context of advancing motion parameter regression for denoising fMRI research.

Comparative Efficacy in Motion Correction and Behavioural Prediction

A 2025 systematic investigation evaluated 14 distinct denoising pipelines, combining 5 common noise mitigation methods: white matter and cerebrospinal fluid regression, independent component analysis (ICA)-based artefact removal (ICA-FIX), volume censoring, global signal regression (GSR), and diffuse cluster estimation and regression (DiCER) [45] [3]. The study employed three distinct quality control metrics to evaluate motion influence and used kernel ridge regression to predict 81 behavioural variables across two independent datasets—the Genomics Superstruct Project (GSP, N=1,570) and the Human Connectome Project (HCP, N=1,200) [45].

Table 1: Pipeline Performance in Motion Reduction and Behavioural Prediction

Pipeline Category	Key Components	Motion Reduction Efficacy	Behavioural Prediction Performance	Overall Recommendation
Optimal Trade-off	ICA-FIX + GSR	High	Reasonable, modest inter-pipeline variations	Recommended for balanced performance [45]
GSR-Inclusive	Various pipelines incorporating GSR	Effective	Variable, can augment certain brain-behaviour associations [45]	Context-dependent application advised [80]
Volume Censoring	Framewise displacement-based data removal	Effective for multiband fMRI	Requires dataset-specific parameter optimization [4]	Recommended with optimized parameters [4]

A critical finding was that no single pipeline universally excelled at simultaneously achieving both optimal motion reduction and maximal behavioural prediction performance across different cohorts [45]. However, pipelines combining ICA-FIX and GSR demonstrated a reasonable trade-off between these two objectives. Inter-pipeline variations in predictive performance were generally modest [45].

Functional Connectomics and Network Reliability

A separate 2024 large-scale evaluation focused on 768 data-processing pipelines for constructing functional brain networks from preprocessed fMRI data [80]. The study assessed pipelines based on their ability to minimize motion confounds and spurious test-retest discrepancies in network topology while remaining sensitive to inter-subject differences and experimental effects.

Table 2: Network Construction Pipeline Reliability and Sensitivity

Evaluation Criterion	Performance Range Across Pipelines	Key Influencing Factors
Test-Retest Reliability	Vast and systematic variability; majority of pipelines failed at least one criterion [80]	Parcellation scheme, edge definition, global signal regression [80]
Sensitivity to Individual Differences	Highly pipeline-dependent [80]	Node definition (parcellation), edge weighting [80]
Motion Confound Resistance	Systematically variable [80]	Global signal regression, connectivity definition [80]
Generalizability Across Datasets	A subset of pipelines performed consistently across minutes, weeks, and months [80]	End-to-end pipeline configuration [80]

This research revealed that an inappropriate choice of pipeline can produce misleading and systematically replicable results [80]. Despite this variability, a subset of optimal pipelines consistently satisfied all evaluation criteria across different datasets, providing a foundation for robust functional connectomics.

Experimental Protocols

Protocol 1: Benchmarking Denoising Pipelines for BWAS

This protocol outlines the procedure for comparing the efficacy of different denoising pipelines in mitigating motion artefacts and enhancing brain-behaviour association studies [45] [3].

3.1.1 Data Acquisition and Initial Preprocessing

• Datasets: Utilize publicly available datasets such as the Human Connectome Project (HCP) or the Genomics Superstruct Project (GSP). The HCP dataset (N=1,200) was acquired on a customized Siemens 3T Skyra with a multiband sequence (TR=720 ms). The GSP dataset (N=1,570) was acquired on 3T Tim Trio scanners (TR=3000 ms) [3].
• Initial Preprocessing: Process structural and functional data through standard pipelines (e.g., fMRIprep v23.0.2). This includes brain extraction, tissue segmentation, and spatial normalization [3].

3.1.2 Denoising Pipeline Implementation Implement a set of distinct denoising pipelines by combining the following methods in different configurations:

• White Matter and Cerebrospinal Fluid (WM/CSF) Regression: Extract average signals from WM and CSF masks and regress them out from the BOLD signal.
• Independent Component Analysis (ICA-FIX): Use ICA to decompose data and a classifier (e.g., FIX) to identify and remove noise components.
• Global Signal Regression (GSR): Regress out the global mean signal from each voxel's time series.
• Volume Censoring (e.g., "Scrubbing"): Identify and remove volumes with excessive motion (e.g., framewise displacement > 0.2 mm).
• Diffuse Cluster Estimation and Regression (DiCER): A method that regresses out widespread, spatially heterogeneous noise [45] [3].

3.1.3 Outcome Evaluation

• Motion Influence Metrics: Calculate three distinct quality control metrics (e.g., framewise displacement correlation measures) to quantify residual motion artefacts after denoising [45].
• Behavioural Prediction: Use kernel ridge regression to predict a wide array of behavioural variables (e.g., from NIH Toolbox or personality inventories) from the functional connectivity matrices generated by each pipeline. Assess performance via cross-validated prediction accuracy [45] [3].

Protocol 2: Optimized Volume Censoring for Multiband fMRI

This protocol details a method for determining dataset-specific optimal parameters for volume censoring, a critical denoising step for multiband fMRI data [4].

3.2.1 Parameter Optimization

• Define Censoring Parameters: The key parameter is the framewise displacement (FD) threshold. Do not rely on a single universal threshold (e.g., 0.2 mm).
• Generate Metrics: For a range of FD thresholds (e.g., 0.05 mm to 0.3 mm), calculate two novel, quantitative metrics that are agnostic to QC-FC correlations:
  • Distance-Dependent Effects (DDE): Measure how motion artefacts manifest as spurious distance-dependent correlations.
  • Data Loss vs. Noise Retention: Quantify the trade-off between the amount of data censored and the level of noise retained.
• Identify Optimal Threshold: Select the FD threshold that best balances the reduction of motion artefacts with the retention of usable data. This threshold is often dataset-specific [4].

3.2.2 Application and Validation

• Apply Censoring: Apply the optimized volume censoring parameters to the dataset.
• Validate with Quality Metrics: Assess denoising efficacy using the DDE and other quality metrics to ensure artefactual patterns are sufficiently reduced without introducing bias [4].

Workflow and Signaling Pathways

Denoising Pipeline Evaluation Framework

Motion Artefact Mitigation Logic

Research Reagent Solutions

Table 3: Essential Reagents and Tools for fMRI Denoising Research

Reagent/Tool Name	Type/Classification	Primary Function in Research
fMRIprep	Software Pipeline	Standardized automated preprocessing of structural and functional MRI data, ensuring reproducibility and reducing manual intervention [3].
FIX (FMRIB's ICA-based Xnoiseifier)	Software Classifier	Automates the identification of noise components from ICA decompositions, crucial for the ICA-FIX denoising method [80].
Volume Censoring (Scrubbing) Tools	Software Scripts/Metrics	Implements framewise displacement calculation and identifies high-motion volumes for removal, optimizing denoising for multiband fMRI [4].
DiCER (Diffuse Cluster Estimation and Regression)	Software Algorithm	A denoising method that regresses out widespread, spatially heterogeneous noise, particularly effective for certain artefact types [45].
Kernel Ridge Regression	Statistical Tool	A machine learning method used to evaluate pipeline efficacy by predicting behavioural variables from functional connectivity patterns [45].
Portrait Divergence (PDiv)	Network Metric	An information-theoretic measure for quantifying dissimilarity between whole-network topologies, used for test-retest reliability assessment [80].

Head motion is the largest source of artifact in functional MRI, systematically altering functional connectivity (FC) measurements and potentially leading to false positive associations in brain-behavior research. This is particularly problematic when studying traits inherently correlated with motion propensity, such as psychiatric disorders. The Split Half Analysis of Motion Associated Networks (SHAMAN) framework addresses this challenge by providing a quantitative motion impact score for specific trait-FC relationships, distinguishing between effects causing overestimation or underestimation of true effects. Application to 45 traits from n=7,270 participants in the Adolescent Brain Cognitive Development (ABCD) Study revealed that after standard denoising, 42% of traits exhibited significant motion overestimation scores and 38% exhibited significant underestimation scores. Motion censoring at framewise displacement (FD) < 0.2 mm substantially reduced overestimation to just 2% of traits, though it did not mitigate underestimation effects. These findings underscore the critical importance of trait-specific motion impact assessment beyond standard denoising pipelines.

The Problem of Motion Artifact in fMRI

In-scanner head motion introduces systematic bias to resting-state fMRI functional connectivity that is not completely removed by standard denoising algorithms [12]. The technical challenge posed by motion cannot be overstated, as even involuntary sub-millimeter head movements systematically alter fMRI data through non-linear characteristics of MRI physics [12]. Compared to task fMRI, resting-state FC is especially vulnerable to motion artifact because the timing of the underlying neural processes is unknown [12].

The effect of motion on FC is spatially systematic, causing decreased long-distance connectivity and increased short-range connectivity, most notably in the default mode network [12]. This spatial pattern has led to spurious findings in studies of children, older adults, and patients with neurological or psychiatric disorders, where investigators have mistakenly attributed motion artifacts to genuine neurobiological differences [12].

Limitations of Current Denoising Approaches

Numerous approaches have been developed to mitigate motion artifact, including global signal regression, motion parameter regression, spectral filtering, respiratory filtering, principal component analysis, independent component analysis, multi-echo pulse sequences, despiking of high-motion frames, and combinations thereof [12]. However, given the complexity of these approaches, it is difficult to be certain that enough motion artifact has been removed to avoid over- or underestimating trait-FC effects [12].

A particular challenge arises from the natural tension between removing motion-contaminated volumes to reduce spurious findings while not systematically excluding individuals with high motion who may exhibit important variance in the trait of interest [12]. This difficulty in censoring threshold selection arises because most approaches for quantifying motion are agnostic to the specific hypothesis under study [12].

Quantitative Evidence of Motion Impact

Efficacy of Standard Denoising

The ABCD-BIDS denoising algorithm, which includes global signal regression, respiratory filtering, spectral filtering, despiking, and motion parameter regression, achieves a significant but incomplete reduction in motion-related variance. After minimal processing (motion-correction by frame realignment only), 73% of signal variance is explained by head motion. After ABCD-BIDS denoising, this reduces to 23%, representing a 69% relative reduction compared to minimal processing alone [12].

Despite this improvement, substantial motion-related effects persist. The motion-FC effect matrix shows a strong negative correlation (Spearman ρ = -0.58) with the average FC matrix, indicating that connection strength tends to be weaker in participants who moved more. This strong negative correlation persists even after motion censoring at FD < 0.2 mm (Spearman ρ = -0.51) [12].

Table 1: Motion Impact on Traits After Standard Denoising (ABCD Study, n=7,270)

Impact Type	Percentage of Traits	Number of Traits	Description
Motion Overestimation	42%	19/45	Motion artifact inflates apparent trait-FC effect sizes
Motion Underestimation	38%	17/45	Motion artifact diminishes apparent trait-FC effect sizes
No Significant Impact	20%	9/45	Trait-FC relationships not significantly affected by residual motion

Impact of Motion Censoring

Motion censoring (scrubbing) at different framewise displacement thresholds has differential effects on overestimation versus underestimation:

Table 2: Effect of Motion Censoring on Trait-FC Relationships

Censoring Threshold	Overestimation Impact	Underestimation Impact	Data Retention
No censoring	42% of traits affected	38% of traits affected	Maximum data
FD < 0.2 mm	2% of traits affected	No reduction	Reduced data, especially for high-motion subjects
FD < 0.1 mm	Not reported	Not reported	Substantially reduced data

Censoring at FD < 0.2 mm dramatically reduces significant overestimation from 42% to just 2% of traits. However, it does not decrease the number of traits with significant motion underestimation scores, highlighting the complex relationship between motion artifact and trait-FC effect directions [12] [81].

The SHAMAN Framework: Protocol and Implementation

Theoretical Foundation

SHAMAN capitalizes on the observation that traits (e.g., weight, intelligence) are stable over the timescale of an MRI scan whereas motion is a state that varies from second to second [12]. The method measures differences in the correlation structure between split high- and low-motion halves of each participant's fMRI timeseries. When trait-FC effects are independent of motion, the difference between halves will be non-significant because traits are stable over time. A significant difference indicates that state-dependent differences in motion impact the trait's connectivity [12].

The direction of the motion impact score relative to the trait-FC effect direction indicates the nature of the bias:

• Motion overestimation: Impact score aligned with trait-FC effect direction
• Motion underestimation: Impact score opposite to trait-FC effect direction

Experimental Protocol

Step 1: Data Acquisition and Preprocessing

• Acquire resting-state fMRI data using standard parameters
• Apply minimal preprocessing including motion correction
• Optional: Apply comprehensive denoising pipeline (e.g., ABCD-BIDS, fMRIPrep)

Step 2: Motion Quantification

• Calculate framewise displacement (FD) for each timepoint
• Generate motion-FC effect matrix by regressing FD against FC

Step 3: Data Splitting

• For each participant, split fMRI timeseries into high-motion and low-motion halves based on FD
• Ensure adequate data retention in each half for reliable connectivity estimation

Step 4: Trait-FC Effect Calculation

• Calculate trait-FC effects separately for high-motion and low-motion halves
• Use appropriate statistical models incorporating relevant covariates

Step 5: Motion Impact Scoring

• Compare trait-FC effects between high-motion and low-motion halves
• Calculate motion impact score and significance using permutation testing
• Classify as overestimation or underestimation based on directionality

Step 6: Validation and Interpretation

• Apply non-parametric combining across pairwise connections
• Interpret results in context of specific research question and participant characteristics

Research Reagents and Tools

Table 3: Essential Research Tools for Motion Impact Assessment

Tool Category	Specific Examples	Function	Implementation Considerations
Data Processing	fMRIPrep, ABCD-BIDS pipeline	Standardized preprocessing	Ensure compatibility with SHAMAN framework
Motion Quantification	Framewise Displacement (FD), DVARS	Quantify head motion	Standardize calculation methods across studies
Denoising Strategies	ICA-AROMA, GSR, aCompCor, Scrubbing	Remove motion artifacts	Consider differential effects on over/under-estimation
Statistical Analysis	Permutation testing, Non-parametric combining	Calculate significance	Account for multiple comparisons
Quality Metrics	Network reproducibility, Identifiability, tDOF loss	Evaluate denoising efficacy	Balance noise removal with signal preservation

Integration with Denoising Pipelines

Comparative Performance of Denoising Methods

Different denoising strategies offer varying tradeoffs between motion reduction and preservation of neural signal. Comprehensive comparisons show that no single pipeline universally excels across all objectives and datasets [3]. Key findings include:

• ICA-AROMA (particularly aggressive version) shows strong performance for older adult populations, with good reproducibility and low false-positive rates [5]
• Scrubbing combined with global signal regression is generally effective at noise removal but alters temporal autocorrelation structure [50]
• Simpler strategies using motion parameters, compartment signals, and global signal regression may be preferred for analyses requiring continuous sampling [50]

Contextual Considerations

The optimal denoising approach and motion impact assessment strategy depends on multiple factors:

• Population characteristics: Motion and physiological noise characteristics differ substantially across age groups (e.g., children vs. older adults) [5]
• Trait properties: Traits strongly correlated with motion propensity (e.g., ADHD symptoms) require more rigorous motion impact assessment [12]
• Analysis goals: Studies focusing on specific neural circuits vs. brain-wide associations may tolerate different tradeoffs
• Data quality: Acquisition parameters, scanner stability, and session length influence optimal strategy selection

Applications in Drug Development

Regulatory Context and Biomarker Qualification

Functional MRI holds potential to enhance CNS drug development by providing objective data on drug effects in the living brain [82]. Regulatory agencies recognize the need for novel technologies that facilitate drug development but require rigorous validation for specific contexts of use [82].

No requests have been made to qualify fMRI as a drug development tool, though the European Medicines Agency has issued a letter of support for exploring fMRI biomarkers in autism spectrum disorder [82]. The SHAMAN framework could contribute to biomarker qualification by providing standardized assessment of motion impact, addressing a major potential confound in multi-site clinical trials.

Phase-Specific Applications

• Phase I: Establish CNS penetration and target engagement while controlling for motion-related confounds
• Phase II: Differentiate true drug effects from motion-related artifacts in patient populations
• Phase III: Confirm efficacy with reduced variance through rigorous motion artifact control
• Phase IV: Post-marketing studies of disease modification with enhanced sensitivity

The SHAMAN framework represents a significant advancement in detecting and quantifying motion impact on trait-FC associations. By providing trait-specific motion impact scores that distinguish between overestimation and underestimation biases, it addresses a critical limitation of standard denoising approaches. The high prevalence of significant motion impact scores (80% of traits affected) after standard denoising highlights the necessity of incorporating motion impact assessment into the analytical pipeline for brain-behavior association studies.

Implementation of this approach requires careful consideration of population characteristics, trait properties, and analytical goals. When properly implemented, motion impact assessment can enhance the validity of trait-FC findings across basic neuroscience and applied drug development contexts.

Resting-state functional magnetic resonance imaging (rs-fMRI) is a pivotal tool for mapping the brain's functional organization and its relation to individual differences in behaviour [3]. However, the measured signal is contaminated by multiple sources of noise, including those from head motion, cardiac cycles, and respiratory variations, which can severely impact the reliability and validity of functional connectivity (FC) estimates [3] [25]. These artefacts reduce effect sizes in brain-wide association studies (BWAS) and can induce spurious brain-behaviour associations [3] [12].

While numerous denoising strategies exist—such as white matter and cerebrospinal fluid regression, global signal regression (GSR), independent component analysis (ICA), and motion censoring—a critical challenge persists. The field lacks a single, universally optimal pipeline; performance varies across datasets, and benchmarks can quickly become obsolete as software and methods evolve [3] [83]. This article outlines a paradigm shift from static, one-time pipeline selection towards continuous, framework-based evaluation of denoising strategies, ensuring robust and reproducible fMRI research.

The Need for Continuous Evaluation

Traditional, static benchmarking studies provide valuable snapshots but cannot keep pace with the rapidly evolving neuroimaging software ecosystem. A 2024 study demonstrated that certain denoising strategies behave inconsistently across different datasets and, crucially, across different versions of the same preprocessing software (e.g., fMRIPrep) [83]. This version-dependent behaviour means that a pipeline validated in a past study may perform differently when applied with updated software tools, potentially compromising result reproducibility.

Furthermore, the performance of denoising pipelines involves inherent trade-offs. A strategy effective at mitigating motion-related artefacts may not be the best for augmenting brain-behaviour associations [3] [84]. Similarly, pipelines that excel in one quality metric (e.g., resting-state network identifiability) may perform poorly on another (e.g., reducing motion artefacts) [25]. This complex landscape necessitates evaluation frameworks that are as dynamic and multi-faceted as the data and questions they address.

A Multi-Metric Framework for Pipeline Evaluation

A continuous evaluation framework requires a comprehensive set of benchmarks that reflect the diverse goals of fMRI research. The following table summarizes key metrics derived from recent literature.

Table 1: Key Metrics for Evaluating Denoising Pipeline Efficacy

Metric Category	Specific Metric	What it Quantifies	Interpretation
Data Quality	Temporal Signal-to-Noise Ratio (tSNR) [85] [41]	Stability of the BOLD signal over time	Higher values indicate a cleaner signal.
	DVARS [41]	Rate of change of BOLD signal frame-to-frame	Lower values indicate less introduced noise.
	Variance of Residuals [85]	Unexplained variance after denoising	Lower values suggest better noise removal.
Motion Impact	Motion-FC Effect Correlation [12]	Spatial similarity between motion and trait-FC maps	Strong correlation suggests residual motion contamination.
	Motion Impact Score (SHAMAN) [12]	Trait-specific over/under-estimation due to motion	Scores significant from zero indicate motion bias.
	QC-FC Distance Dependence [12]	Correlation between motion and connectivity, stratified by distance	Reduced dependence indicates better motion mitigation.
Functional Connectivity reliability	Test-Retest Reliability [80]	Consistency of network topology across repeated scans	Higher reliability is crucial for individual differences research.
	Portrait Divergence (PDiv) [80]	Dissimilarity between whole-network organizations	Lower values between test-retest scans indicate higher reliability.
Biological Validity	Behavioural Prediction Accuracy [3] [84]	Cross-validated prediction of behavioural measures from FC	Higher accuracy suggests preserved behaviourally-relevant signal.
	Resting-State Network Identifiability [25]	Sharpness and contrast of known functional networks	Better identifiability suggests preserved neurological signal.

The SHAMAN Protocol for Trait-Specific Motion Impact

The Split Half Analysis of Motion Associated Networks (SHAMAN) is a novel method for assigning a motion impact score to specific trait-FC relationships [12]. This is critical for traits correlated with motion (e.g., in psychiatric disorders) to avoid false positives.

Experimental Protocol:

• Data Requirements: One or more rs-fMRI scans per participant with associated framewise displacement (FD) timeseries and a trait of interest.
• Split Timeseries: For each participant, split the denoised fMRI timeseries into high-motion and low-motion halves based on the median FD.
• Compute Correlation Differences: For each functional connection, calculate the difference in correlation with the trait between the high-motion and low-motion halves.
• Generate Null Distribution: Permute the high/low motion labels many times (e.g., 10,000) to create a null distribution of correlation differences.
• Calculate Significance: Compare the observed correlation difference to the null distribution to obtain a p-value. A significant positive score indicates motion overestimation of the trait-FC effect, while a significant negative score indicates underestimation [12].

Implementing a Continuous Evaluation Workflow

A continuous evaluation framework integrates automated software, standardized metrics, and regular re-assessment. The following workflow diagram illustrates this process.

Protocol for Workflow Implementation:

• Establish a Pipeline Repository: Use containerized software (e.g., fMRIPrep, HALFpipe) to ensure reproducible execution. The repository should include a diverse set of strategies [25] [83]:

  • Base Pipelines: GSR, aCompCor, ICA-AROMA (aggressive/non-aggressive).
  • Advanced Pipelines: Multi-echo ICA (ME-ICA), ME-ICA combined with aCompCor, pipelines incorporating RIPTiDe [84] [41] [86].
  • Censoring Variants: Pipelines with different framewise displacement thresholds (e.g., FD < 0.2 mm).

• Automate Metric Computation: Leverage open-source tools like Nilearn [83] to compute the suite of metrics from Table 1 programmatically. This should be integrated into a continuous integration (CI) environment.

• Maintain a Results Dashboard: A living database that tracks pipeline performance across all metrics and dataset versions. This dashboard should highlight trade-offs and allow researchers to select the optimal pipeline for their specific goal (e.g., maximizing behavioural prediction vs. minimizing motion bias).

• Schedule Re-Evaluation: Automatically trigger the evaluation workflow upon changes to key software dependencies (e.g., new fMRIPrep release) or when a new dataset of interest is acquired.

The Scientist's Toolkit: Essential Research Reagents

Table 2: Key Software and Data Resources for Continuous Evaluation

Category	Item	Function in Evaluation	Example/Reference
Software & Pipelines	fMRIPrep	Standardized, containerized automated preprocessing; foundation for denoising [83].	https://fmriprep.org
	HALFpipe	Standardized workflow from raw data to group stats, built on fMRIPrep [25].	Waller et al., 2022 [25]
	ICA-AROMA	Data-driven method for automatic removal of motion artifacts via ICA [84].	Pruim et al., 2015
	ME-ICA	Denoising method for multi-echo data that separates BOLD from non-BOLD components [41].	Kundu et al., 2012
Evaluation Tools	Nilearn	Python module for statistical learning on neuroimaging data; enables metric computation [83].	https://nilearn.github.io
	SHAMAN	Method to compute trait-specific motion impact scores for brain-behaviour associations [12].	Nature Comm., 2025 [12]
	Portrait Divergence (PDiv)	Information-theoretic measure of whole-network dissimilarity for test-retest reliability [80].	Nature Comm., 2024 [80]
Reference Datasets	HCP, ABCD, CNP, GSP	Public datasets with high-quality data and behavioural measures for benchmarking [3] [12].	HCP Young Adult [3]

Discussion and Future Directions

Adopting a continuous evaluation framework moves the field beyond the quest for a single, universal denoising pipeline and towards a more nuanced, context-aware approach. It formally acknowledges that the optimal pipeline may depend on the specific research question, the population being studied, the acquisition parameters, and the current software landscape.

Future developments will likely involve more sophisticated, automated pipeline optimizers that can dynamically recommend or assemble strategies based on the characteristics of the input data and the desired output metrics. Furthermore, as large-scale models become more prevalent in neuroimaging, the integration of these evaluation frameworks into centralized data repositories will be essential for providing community-wide standards and ensuring the long-term validity of brain-behaviour findings. By making denoising evaluation an ongoing, integrated practice rather than a one-time prelude to analysis, researchers can significantly enhance the robustness, reproducibility, and translational potential of fMRI research.

Conclusion

Motion parameter regression is a critical, yet incomplete, defense against one of fMRI's most pervasive confounds. A successful denoising strategy is not one-size-fits-all; it requires a careful balance between noise removal and signal preservation, tailored to the specific population and research question. The evidence strongly favors a combined approach, often integrating motion regression with physiological noise removal and, where temporally feasible, scrubbing. While automated tools like ICA-AROMA and the novel CICADA offer powerful, standardized solutions, rigorous validation using a multi-metric framework is non-negotiable. For the future, the field must move towards continuous evaluation of denoising software and adopt transparent, reproducible pipelines. This will be paramount for advancing robust biomarker discovery in neurology and psychiatry and for ensuring the reliability of fMRI in clinical trials and drug development.

References

• 1. Motion artifact in studies of functional connectivity [https://pmc.ncbi.nlm.nih.gov/articles/PMC5930165/]
• 2. Reducing the effects of motion artifacts in fMRI - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC8934405/]
• 3. The efficacy of resting-state fMRI denoising pipelines for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12336063/]
• 4. Advancing motion denoising of multiband resting-state ... [https://www.sciencedirect.com/science/article/pii/S1053811922000374]
• 5. Evaluation of noise regression techniques in resting-state ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2022.1006056/full]
• 6. Reducing the Effects of Motion Artifacts in fMRI [https://labs.childrenshospital.org/onurafacan/publications/reducing-effects-motion-artifacts-fmri-structured-matrix-completion-approach]
• 7. Removing motion and physiological artifacts from intrinsic ... [https://www.sciencedirect.com/science/article/pii/S1053811912009536]
• 8. Real-time fetal head motion tracking for prospective ... [https://archive.ismrm.org/2025/3611.html]
• 9. Consensus recommendations for clinical functional MRI ... [https://apertureneuro.org/article/128149-consensus-recommendations-for-clinical-functional-mri-applied-to-language-mapping]
• 10. Movement artefacts and MR BOLD signal increase during ... [https://pubmed.ncbi.nlm.nih.gov/11956941/]
• 11. Spurious but systematic correlations in functional connectivity ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3254728/]
• 12. Motion impact score for detecting spurious brain-behavior ... [https://www.nature.com/articles/s41467-025-63661-2]
• 13. Evaluating denoising strategies in resting‐state functional magnetic resonance in traumatic brain injury (EpiBioS4Rx) - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC9491287/]
• 14. Continuous evaluation of denoising strategies in resting-state fMRI connectivity using fMRIPrep and Nilearn - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10977879/]
• 15. Task and resting state fMRI modelling of brain-behavior ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12593242/]
• 16. An evaluation of the efficacy, reliability, and sensitivity of motion correction strategies for resting-state functional MRI - ScienceDirect [https://www.sciencedirect.com/science/article/pii/S1053811917310972]
• 17. Methods to detect, characterize, and remove motion artifact in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3849338/]
• 18. Can this data be saved? Techniques for high motion in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9694086/]
• 19. Framewise displacement: Significance and symbolism [https://www.wisdomlib.org/concept/framewise-displacement]
• 20. A functional MRI pre-processing and quality control ... [https://www.frontiersin.org/journals/neuroimaging/articles/10.3389/fnimg.2022.1070151/full]
• 21. Beyond motion correction: Framewise displacement [https://crimia.wordpress.com/2016/02/07/beyond-motion-correction-framewise-displacement/]
• 22. Robust Motion Regression of Resting-State Data Using a ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6482337/]
• 23. Using Edge Voxel Information to Improve Motion ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4652211/]
• 24. Potential pitfalls when denoising resting state fMRI data ... [https://www.sciencedirect.com/science/article/pii/S1053811916307480]
• 25. Multi‐Metric Approach for the Comparison of Denoising ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12044599/]
• 26. Integration of motion correction and physiological noise ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC2833099/]
• 27. A Component Based Noise Correction Method (CompCor ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC2214855/]
• 28. Reduction of motion-related artifacts in resting state fMRI ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4043948/]
• 29. The Influence of Physiological Noise Correction on Test– ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4146390/]
• 30. Comparing data-driven physiological denoising ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10876882/]
• 31. The Global Signal in fMRI: Nuisance or Information? - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC5406229/]
• 32. Global signal regression strengthens association between ... [https://pubmed.ncbi.nlm.nih.gov/30974241/]
• 33. Identifying and removing widespread signal deflections ... [https://www.sciencedirect.com/science/article/pii/S1053811920301014]
• 34. Data-driven denoising methods (global-signal regression ... [https://cds.ismrm.org/protected/23MProceedings/PDFfiles/2667_HJIveVQ6z.html]
• 35. An automated and flexible tool for comprehensive fMRI ... [https://pubmed.ncbi.nlm.nih.gov/40851911/]
• 36. keithcdodd/CICADA [https://github.com/keithcdodd/CICADA]
• 37. - ICA : A robust ICA-based strategy for removing motion... AROMA [https://pubmed.ncbi.nlm.nih.gov/25770991/]
• 38. Comparing data-driven physiological denoising ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2024.1223230/full]
• 39. Comparing data-driven physiological denoising ... [https://www.semanticscholar.org/paper/586f33ec539e80ab814b4d3ccda0fdda4d76a339]
• 40. the efficacy of data-driven Comparing methods for... denoising [https://cris.maastrichtuniversity.nl/en/publications/comparing-the-efficacy-of-data-driven-denoising-methods-for-a-mul]
• 41. Comparing the efficacy of data-driven denoising methods ... [https://www.sciencedirect.com/science/article/pii/S1053811923005128]
• 42. A multi-measure approach for assessing the performance ... [https://www.sciencedirect.com/science/article/abs/pii/S0730725X21001958#!]
• 43. Balancing Data Quality and Bias: Investigating Functional ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11717557/]
• 44. Less is more: balancing noise reduction and data retention in ... [https://pubmed.ncbi.nlm.nih.gov/36842522/]
• 45. The efficacy of resting-state fMRI denoising pipelines for ... [https://pubmed.ncbi.nlm.nih.gov/40800742/]
• 46. The efficacy of resting-state fMRI denoising pipelines for ... [https://research.monash.edu/en/publications/the-efficacy-of-resting-state-fmri-denoising-pipelines-for-motion/]
• 47. Quality control procedures and metrics for resting-state ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10406233/]
• 48. Can this data be saved? Techniques for high motion in ... [https://www.sciencedirect.com/science/article/pii/S1878929322001219]
• 49. A Hitchhiker's Guide to Functional Magnetic Resonance ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2016.00515/full]
• 50. Continuous evaluation of denoising strategies in resting-state ... [https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1011942]
• 51. fMRIPrep: A Robust Preprocessing Pipeline for fMRI Data ... [https://fmriprep.org/]
• 52. fMRI pre-processing project using fMRIPrep, Nilearn on a ... [https://github.com/brainhack-school2020/stong3_fMRI_processing]
• 53. What's new — fmriprep version documentation [https://fmriprep.org/en/25.1.0/changes.html]
• 54. Evaluation of denoising strategies for task‐based ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7978116/]
• 55. Benchmarking of participant-level confound regression ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5483393/]
• 56. Denoising task‐related fMRI: Balancing noise reduction ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10619396/]
• 57. Reducing the Effects of Motion Artifacts in fMRI - PubMed - NIH [https://pubmed.ncbi.nlm.nih.gov/34432631/]
• 58. FMRIPrep: a robust preprocessing pipeline for functional MRI [https://pmc.ncbi.nlm.nih.gov/articles/PMC6319393/]
• 59. Denoising strategies of functional connectivity MRI data in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12580809/]
• 60. Less is more: balancing noise reduction and data retention ... [https://www.sciencedirect.com/science/article/pii/S1053811923001180]
• 61. The effect of ageing on fMRI: Correction for the confounding ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4730557/]
• 62. Population receptive field tuning properties of visual cortex ... [https://www.sciencedirect.com/science/article/pii/S1878929318301725]
• 63. Comparing retinotopic maps of children and adults reveals ... [https://www.nature.com/articles/s41467-023-37280-8]
• 64. Quality control practices in FMRI analysis [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2022.1073800/full]
• 65. Prospective active marker motion correction improves ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811912011640]
• 66. Prospective Motion Correction in Functional MRI - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC5427003/]
• 67. Chapter 3: Motion Correction - Andy's Brain Book! [https://andysbrainbook.readthedocs.io/en/latest/fMRI_Short_Course/Preprocessing/Motion_Correction.html]
• 68. Prospective motion correction improves the sensitivity ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6175330/]
• 69. An evaluation of prospective motion correction (PMC) for ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2015.00097/full]
• 70. An evaluation of prospective motion correction (PMC) for ... [https://pubmed.ncbi.nlm.nih.gov/25859178/]
• 71. Benchmarking methods for mapping functional connectivity ... [https://www.nature.com/articles/s41592-025-02704-4]
• 72. An Efficient and Reliable Statistical Method for Estimating ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2016.00123/full]
• 73. Longer scans boost prediction and cut costs in brain-wide ... [https://www.nature.com/articles/s41586-025-09250-1]
• 74. Reproducibility of resting-state functional connectivity in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12548667/]
• 75. A Quantitative Data-Driven Analysis Framework for Resting ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2021.768418/full]
• 76. Perspectives on resting-state functional magnetic ... [https://www.frontiersin.org/journals/aging-neuroscience/articles/10.3389/fnagi.2025.1547965/full]
• 77. The colors of our brain: an integrated approach for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9107439/]
• 78. How to measure functional connectivity using resting-state ... [https://www.sciencedirect.com/science/article/pii/S1053811925001983]
• 79. Data-guided neuroimaging and visualization [https://apertureneuro.org/article/140427-data-guided-neuroimaging-and-visualization-from-functional-decomposition-to-dynamic-fusion]
• 80. Systematic evaluation of fMRI data-processing pipelines for ... [https://www.nature.com/articles/s41467-024-48781-5]
• 81. Motion impact score for detecting spurious brain-behavior associations - PubMed [https://pubmed.ncbi.nlm.nih.gov/41022827/]
• 82. The role of fMRI in drug development - PMC - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC5931333/]
• 83. Continuous evaluation of denoising strategies in resting- ... [https://inria.hal.science/hal-04718572]
• 84. An Evaluation of the Efficacy of Single-Echo and Multi-Echo ... [https://sciety-discovery.elifesciences.org/articles/by?article_doi=10.1101/2025.08.28.672985]
• 85. Correction of Physiological Artifacts in Multi‐Echo fMRI Data ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12183604/]
• 86. Improving delay and strength maps derived from resting ... [https://www.sciencedirect.com/science/article/abs/pii/S0010482525006134]