Motion Artifacts in fMRI: Characteristics, Challenges, and Correction Strategies for Resting-State vs. Task-Based Paradigms

Abstract

This article provides a comprehensive analysis of motion artifact characteristics in resting-state and task-based functional magnetic resonance imaging (fMRI) for researchers and drug development professionals. It explores the fundamental spatial, temporal, and spectral properties of motion artifacts, highlighting critical differences between resting-state and task paradigms where motion can become correlated with experimental design. The content details established and emerging methodological approaches for artifact correction, from retrospective preprocessing pipelines to real-time compensation techniques. It further addresses troubleshooting and optimization strategies for challenging populations and study designs, and examines validation frameworks for comparing correction efficacy. This synthesis enables researchers to design more robust fMRI studies and accurately interpret functional connectivity and activation results in both clinical and research settings.

Characterizing Motion Artifacts: Fundamental Differences Between Resting-State and Task fMRI

Motion artifacts represent a significant methodological challenge in functional magnetic resonance imaging (fMRI), systematically biasing connectivity measurements in a spatially dependent manner. This technical review examines the well-established distance-dependent effects of motion artifacts on functional connectivity metrics, distinguishing the specific challenges in both resting-state and task-based fMRI paradigms. We synthesize evidence demonstrating that motion artifacts spuriously inflate short-distance correlations while suppressing long-distance connections, potentially confounding neurodevelopmental, clinical, and pharmacological studies. This review provides a comprehensive overview of quantification methodologies, experimental protocols for artifact characterization, and advanced mitigation strategies essential for robust connectivity analysis in neuroscience research and drug development applications.

Motion artifacts are now recognized as a major methodological challenge for studies of functional connectivity, with particular relevance for research involving pediatric, clinical, or elderly populations where motion may correlate with variables of interest [1] [2]. In-scanner motion systematically alters correlations in resting-state functional connectivity fMRI (RSFC), with potentially severe implications for studies of lifespan development, individual differences, and clinical groups [1] [3]. While motion artifacts affect all fMRI paradigms, their impact manifests differently in resting-state versus task-based designs. In task-based fMRI, motion uncorrelated with task design degrades statistical power by introducing noise, causing false negatives, whereas motion correlated with task timing can lead to false positives [4]. In resting-state fMRI, head motion artifacts introduce colored distance-dependent noise such that covariance is systematically biased between brain regions due to shared proximal artifacts [4].

The spatial distribution of motion artifacts is not uniform across the brain. Due to biomechanical constraints of the neck, motion is minimal near the atlas vertebrae (where the skull attaches to the neck) and increases with distance from the atlas [1]. Furthermore, high motion in frontal cortex is most likely due to the preponderance of y-axis rotation, associated with nodding movement [1]. This spatial heterogeneity has profound implications for connectivity analysis, particularly as it creates predictable patterns of bias that can mimic or obscure genuine neurobiological effects—a critical concern for drug development professionals investigating potential neurotherapeutics.

Characteristics of Distance-Dependent Motion Artifacts

Spatial Distribution Patterns

Motion artifacts exhibit distinctive spatial patterns that reflect both the biomechanics of head movement and the physics of MRI signal acquisition. The artifacts manifest as signal intensity changes across the brain parenchyma, with areas at the edge of the brain demonstrating large increases in signal due to partial volume effects [1]. Motion produces increased image smoothness globally, further complicating accurate connectivity measurement [1].

Critically, motion-related artifacts influence connectivity measures in a systematically distance-dependent manner. Research has consistently demonstrated that motion artifacts spuriously increase correlations between nearby brain regions while decreasing correlations between distant regions [3]. This pattern arises because motion adds spurious variance to 'true' timeseries that is most similar at nearby voxels [3]. Consequently, correlations between BOLD timeseries are spuriously increased across all voxels, but are most increased between nearby voxels [3].

Temporal and Spectral Properties

The temporal characteristics of motion artifacts include both transient components immediately following movement and prolonged signal changes that may persist for 10 seconds or more after motion ceases [1] [3]. Immediately following a movement event, motion typically results in a substantial drop in signal that scales with the magnitude of motion [1]. These signal changes are temporally circumscribed and maximal at the volume acquired immediately after an observed movement. Additionally, longer duration artifacts (up to 8-10 seconds) occur idiosyncratically in individual time series, potentially due to motion-related changes in CO2 that accompany yawning or deep breathing [1].

The spectral properties of motion artifacts further complicate their removal. Motion does not display band-limited frequency content, making frequency filtering (commonly applied in rs-fMRI to isolate the 0.01-0.1 Hz range) ineffective for motion correction [2]. In fact, filtering can potentially smear motion contamination across the entire dataset if not applied carefully [2]. Low-frequency, autocorrelated trends are readily apparent in rs-fMRI data due to motion, necessitating specialized methods beyond frequency filtering to remove these artifacts while retaining true fMRI signal content [2].

Impact on Connectivity Metrics

The distance-dependent nature of motion artifacts systematically biases multiple connectivity metrics. Studies have demonstrated that sub-millimeter motions can distort functional connectivity estimates from approaches including seed correlation analyses, graph theoretic network modularity, dual regression independent component analysis (ICA), and power spectrum methods [2]. The specific direction of motion can also introduce orientation-dependent effects, with increased lateral connectivity at the expense of connectivity in the inferior-superior and anterior-posterior directions [2].

Table 1: Sensitivity of Different Functional Connectivity Measures to Motion Artifacts

Connectivity Measure	Sensitivity to Motion	Residual Distance-Dependent Bias	Notable Characteristics
Full Correlation	High	Pronounced	High test-retest reliability despite motion sensitivity
Partial Correlation	Low	Minimal	Intermediate system identifiability
Coherence	Moderate	Moderate	Frequency-domain approach
Mutual Information	Low	Moderate	Information theory-based
Precision Matrix	Low	Low	Accounts for indirect connections

Different connectivity measures exhibit varying sensitivity to motion artifacts. Recent evaluations of eight different functional connectivity measures found that FC estimated using full correlation maintains a relatively high residual distance-dependent relationship with motion compared to partial correlation, coherence, and information theory-based measures, even after implementing rigorous motion artifact mitigation [5]. This disadvantage of full correlation may be offset by higher test-retest reliability, fingerprinting accuracy, and system identifiability [5]. FC estimated by partial correlation offers a favorable balance with low sensitivity to motion artifact and intermediate system identifiability, though with the caveat of low test-retest reliability and fingerprinting accuracy [5].

Measurement and Quantification Methods

Motion Parameter Estimation

In-scanner motion is typically estimated from the functional time series itself during preprocessing. Each volume of the time series is usually rigidly realigned to a reference volume, producing a set of 6 realignment parameters (RPs; 3 translations, 3 rotations) describing how much a given volume must be moved to align with the reference [1]. These realignment parameters can be summarized as the frame displacement (FD), which provides a concise index of volume-to-volume motion [1]. Different methods exist for calculating FD, with work by Yan et al. showing that the matrix root mean squared formulation derived by Jenkinson et al. aligns best with voxel-specific measures of displacement [1].

Table 2: Motion Quantification Metrics in fMRI Research

Metric	Calculation Method	Typical Threshold	Limitations
Frame Displacement (FD)	Composite of 6 realignment parameters	0.2-0.5 mm	Difficult to compare across sequences
Voxel-Specific FD	Computed directly from image header	Varies by region	Computationally intensive
RMS Displacement per Minute	Standardized FD measure	Study-dependent	Enables cross-study comparison
Motion Artifact Score (MAS)	Automated segmentation in CCTA	0.6 (in CCTA)	Developed for coronary CT

It is important to recognize that all such volume-based measures are limited in temporal resolution, which is equivalent to the repetition time of the image, and thus do not effectively capture within-volume motion [1]. Furthermore, realignment estimates may be inaccurate in images substantially corrupted by motion-related artifacts [1]. With the advent of multiband imaging producing shorter repetition times, FD measures become more difficult to compare across studies, suggesting the utility of converting FD into a standardized measure such as millimeters of RMS displacement per minute [1].

Advanced Quantification Approaches

Beyond traditional motion parameters, researchers have developed specialized algorithms for quantifying motion artifact severity. The Motion Artifact Quantification algorithm developed for coronary CT angiography includes steps to identify regions of interest, segment vessel and shading artifacts, and calculate a motion artifact score (MAS) metric [6]. This approach demonstrated Dice coefficients of 0.84 for segmented vessel regions and 0.75 for segmented shading artifact regions, with MAS calculations within 10% of values obtained using ground-truth segmentations [6].

In unobtrusive health monitoring applications, researchers have introduced the shape-based signal-to-noise ratio (SNRS) to quantify the effect of motion on different sensing modalities [7]. Such approaches, combined with motion capture systems that provide sub-millimeter accuracy, allow precise quantification of motion artifacts relative to actual movement measurements [7].

Experimental Protocols for Motion Characterization

Controlled Motion Studies

Several methodological approaches have been employed to systematically characterize motion artifacts. These include:

• Motion Protocol Recordings: Subjects are instructed to perform specific movement patterns while simultaneous motion tracking and fMRI data are acquired. This approach allows direct correlation between motion parameters and signal changes [7]. For example, researchers have implemented protocols where subjects move with different amplitudes while reference sensors record cardiorespiratory activity, enabling comparison between unobtrusively acquired signals and gold standard measurements [7].

• Post-Hoc Analysis of Motion-Contaminated Volumes: By identifying volumes with elevated motion and examining their impact on connectivity metrics, researchers can characterize the spatial and temporal properties of motion artifacts [3]. This approach has revealed that volumes acquired during and up to ~10 seconds after movement systematically impact RSFC correlations in a distance-dependent manner [3].

Benchmarking Motion Correction Strategies

Comparative studies have been essential for evaluating the efficacy of different motion correction approaches. These typically involve:

• Processing the same dataset with multiple pipelines including no correction, regression-based methods, censoring approaches, and novel computational techniques [5] [3].

• Quantifying residual motion-connectivity relationships after applying each correction method, typically measured as the correlation between subject-level motion summary statistics and functional connectivity measures [3].

• Assessing network identifiability and test-retest reliability across correction approaches to ensure that motion reduction does not come at the cost of biological signal removal [5].

Diagram 1: Motion artifact characterization workflow illustrating the sequence from data acquisition through validation of correction methods.

Mitigation Strategies and Correction Protocols

Preprocessing and Denoising Techniques

Multiple preprocessing strategies have been developed to mitigate motion artifacts in fMRI data:

• Regression-Based Approaches: These include incorporating realignment parameters, their temporal derivatives, and quadratic terms as nuisance regressors. More advanced approaches include global signal regression (GSR), which effectively reduces motion-related variance but remains controversial due to potential removal of neurobiologically meaningful signal [1] [3].

• Censoring (Volume Removal): This approach identifies and removes motion-contaminated volumes based on frame displacement thresholds (typically FD > 0.2-0.5 mm) [3]. Censoring can reduce motion-related group differences to chance levels, particularly when combined with interpolation [3]. However, it creates discontinuities in the time series and may result in substantial data loss.

• Structured Matrix Completion: Advanced computational approaches formulate artifact reduction as a matrix completion problem, exploiting the structured low-rank properties of fMRI time series to recover missing entries after censoring [8]. This method enforces a low rank prior on a large structured matrix formed from the samples of the time series to recover missing entries while simultaneously performing slice-time correction [8].

Prospective Motion Correction

Beyond retrospective correction, prospective motion correction techniques modify pulse sequences in real time during scanning to counter the effects of motion [4]. This promising approach involves tracking head position and updating the imaging sequence accordingly, though it is not yet widely used in the field [4]. These methods can be combined with retrospective approaches for improved correction, though implementing them for routine practice requires further development [2] [4].

Experimental Design Considerations

Careful experimental design can minimize motion artifacts:

• Subject Training and Preparation: Proper instruction, practice sessions, and comfortable positioning can reduce motion [2]. Customized head molds have been shown to effectively reduce motion during resting state fMRI scans [4].

• Task Design Considerations: For task-based fMRI, minimizing motion correlated with task conditions reduces false positives. This may include careful timing of stimulus presentation and response collection to decouple task-related motion from neural activation patterns [4].

Table 3: Performance Comparison of Motion Correction Methods

Correction Method	Residual Distance-Dependence	Data Retention	Computational Demand	Key Limitations
Realignment Only	High	100%	Low	Ineffective for spin history effects
Motion Regression	Moderate	100%	Low	Limited efficacy for large motions
Global Signal Regression	Low	100%	Low	Removes neural signal; controversial
Volume Censoring	Low	Variable (50-90%)	Low	Data loss; discontinuities
ICA-Based Cleaning	Moderate-High	100%	High	Component misclassification
Structured Matrix Completion	Very Low	High (~95%)	Very High	Computational complexity

The Scientist's Toolkit: Research Reagent Solutions

Table 4: Essential Resources for Motion Artifact Research

Resource Category	Specific Tools/Solutions	Primary Function	Application Context
Motion Quantification Software	FSL (Jenkinson FD), AFNI, SPM	Calculate motion parameters from fMRI data	Standard preprocessing
Quality Control Metrics	Frame Displacement, DVARS	Identify motion-contaminated volumes	Data quality assessment
Denoising Algorithms	ICA-AROMA, COMPCOR, GSR	Remove motion-related variance	Data cleaning
Structured Matrix Completion	Custom MATLAB/Python implementations	Recover censored volumes	Advanced artifact correction
Prospective Correction	MRI-compatible cameras, MoTrack	Real-time motion compensation	Data acquisition
Multimodal Sensor Systems	Motion capture systems, cECG, PPG	Quantify motion and physiological signals	Experimental validation

Implications for Resting-State vs. Task fMRI Research

The differential impact of motion artifacts on resting-state versus task fMRI has important implications for research design and interpretation:

In resting-state fMRI, the primary concern is the introduction of distance-dependent bias in functional connectivity measures [3] [4]. This bias can create spurious group differences in studies comparing populations with different motion characteristics (e.g., children vs. adults, patients vs. controls) [1] [2]. The absence of known timing for neural events in resting-state designs makes it particularly challenging to distinguish motion artifacts from true neural signals based on temporal characteristics alone.

In task-based fMRI, motion artifacts introduce additional complexities related to the temporal coupling between motion and task design [4]. When motion is uncorrelated with task timing, it primarily increases noise and reduces statistical power [4]. However, when motion correlates with task conditions (e.g., due to button presses, visual stimulation, or cognitive state), it can produce false positives that mimic task-related activation [4]. This is particularly problematic for tasks that inherently elicit head movement, such as those involving speech, motor responses, or emotional stimuli.

These differences necessitate distinct mitigation strategies for each paradigm. Resting-state studies may prioritize rigorous global nuisance regression and censoring approaches, while task-based designs may focus on careful task timing, response method selection, and model inclusion of motion parameters.

The spatial distribution of motion artifacts—characterized by systematic distance-dependent effects on connectivity measures—represents a fundamental challenge for fMRI research. The spurious inflation of short-distance correlations coupled with suppression of long-distance connections can profoundly impact findings in both basic neuroscience and clinical applications. This is particularly critical for drug development professionals using fMRI as a biomarker for treatment response, where motion-related confounds could mimic or obscure therapeutic effects.

Robust characterization and mitigation of motion artifacts requires a multifaceted approach combining careful experimental design, optimized processing pipelines, and comprehensive reporting of motion-related quality metrics. While no single method completely eliminates motion artifacts, approaches such as structured matrix completion show promise for effectively addressing distance-dependent biases while maximizing data retention. Future methodological advances, particularly in prospective correction and multimodal monitoring, offer potential for further improving the fidelity of connectivity measurements in both resting-state and task-based fMRI paradigms.

Temporal Properties and Signal Dynamics in Resting-State vs. Task Contexts

Functional magnetic resonance imaging (fMRI) research primarily utilizes two experimental paradigms: task-based fMRI (tfMRI), which measures brain activity in response to specific cognitive stimuli, and resting-state fMRI (rs-fMRI), which records spontaneous brain fluctuations in the absence of directed tasks. These paradigms exhibit fundamentally different temporal properties and signal dynamics, with profound implications for data interpretation, especially concerning motion artifacts. Rs-fMRI signals are easily contaminated by artifacts arising from head motion, with complex spatiotemporal properties that can lead to substantial errors in functional connectivity estimates [9]. Even motions on the millimeter scale can be problematic, as the small amplitude of BOLD signals—typically a few percent or less—ensures that subtle movements may introduce significant confounds [9]. In task-based fMRI, head motion can be temporally correlated with task performance, and under many circumstances, the resulting "motion artifacts" cannot be distinguished from brain activity, compromising data interpretation [9]. Understanding these fundamental differences is crucial for researchers, scientists, and drug development professionals utilizing fMRI in experimental and clinical contexts.

Fundamental Differences in Signal Composition

Intrinsic Signal Patterns and Neural Correlates

Research has demonstrated intrinsic, fundamental differences in signal composition patterns that effectively characterize and differentiate these two types of fMRI signals. A novel two-stage sparse representation framework applied to Human Connectome Project (HCP) data revealed distinctive and descriptive atoms in cross-subject common dictionaries that can characterize and differentiate tfMRI and rsfMRI signals with 100% classification accuracy [10]. This indicates fundamental differences in how neural information is structured across these paradigms.

Task-based fMRI is widely adopted to identify brain regions functionally involved in specific task performance, while resting-state fMRI explores intrinsic functional segregation or specialization of brain regions/networks [10]. These functional differences manifest in distinct signal dynamics. Cognitive tasks can amplify differences across individuals in connectivity relevant for explaining behavioral outcomes, whereas resting-state data may be suboptimal for building connectome-based predictive models (CPMs) [11]. The enhanced predictive power of task-based fMRI likely stems from its ability to directly engage neural circuits associated with specific cognitive functions, thereby capturing more behaviorally relevant information [11].

Temporal Signal Properties and Artifact Vulnerability

The temporal characteristics of rs-fMRI and tfMRI signals differ significantly in their vulnerability to motion artifacts. Rs-fMRI signals do not display band-limited frequency content, making frequency filtering commonly applied in rs-fMRI (∼0.01–0.1 Hz) potentially ineffective for motion correction [9]. Low-frequency, autocorrelated trends are readily apparent in rs-fMRI data due to motion, requiring specialized methods beyond frequency filters to remove these artifacts while retaining true fMRI signal content [9].

In task-based fMRI, the structured timing of stimulus presentations provides temporal anchors that can help distinguish neural activity from motion-related artifacts, whereas resting-state lacks these reference points, making it more challenging to differentiate true neural signals from noise [9]. Furthermore, the absence of prescribed behavioral tasks in rs-fMRI may even exacerbate head motion when imaging individuals at rest [9].

Table 1: Comparative Signal Properties of Resting-State and Task fMRI

Property	Resting-State fMRI	Task-Based fMRI
Signal Origin	Spontaneous fluctuations	Task-evoked responses
Temporal Structure	Unconstrained, low-frequency oscillations	Structured around task timing
Motion Artifact Impact	Inflates short-range, reduces long-range connectivity	May correlate with task performance
Frequency Content	Non-band-limited motion artifacts	Task-locked responses help isolate signals
Predictive Power	Lower for behavioral outcomes	Higher for specific neuropsychological measures [11]
Optimal Application	Identifying intrinsic networks	Linking brain activity to specific functions

Quantitative Predictive Performance and Cost Efficiency

Predictive Power Across Experimental Paradigms

Substantial evidence indicates that task-based fMRI generally outperforms resting-state fMRI in predictive modeling of behavioral and clinical outcomes. When applied to a clinically heterogeneous transdiagnostic cohort, research has identified shared and distinct functional fingerprints of neuropsychological outcomes across seven fMRI conditions [11]. The robustness of these findings is supported by novel network science-driven Bayesian methods that substantially improve precision and robustness in imaging biomarker detection [11].

Different types of task fMRI show varying strengths of predictive power for different behaviors. For example, the emotional N-back memory task was found to be less optimal for negative emotion outcomes, while the gradual-onset continuous performance task demonstrated stronger links with sensitivity and sociability outcomes than with cognitive control outcomes [11]. This differential predictive power means there is untapped potential to maximize scientific return on investment by tailoring the fMRI task during scanning to achieve optimal predictive power for addressing specific research questions.

Cost Efficiency Considerations in Study Design

The cost efficiency of fMRI studies—due to specialized equipment, trained personnel, and substantial scanning time—is a critical consideration in research design. Evidence suggests that well-tailored fMRI tasks can significantly improve cost efficiency compared to resting-state protocols [11]. Unique optimal pairings of task-based fMRI conditions and neuropsychological outcomes should not be ignored when designing well-powered neuroimaging studies [11].

Resting-state acquisition protocols are often selected because they are easier to replicate and compare across studies, but this apparent advantage may be offset by reduced predictive power for many outcomes of interest. Literature shows that resting-state data may be suboptimal for building connectome-based predictive models (CPMs) compared to task-based approaches [11].

Table 2: Predictive Performance of fMRI Paradigms for Different Behavioral Domains

Behavioral Domain	Optimal fMRI Paradigm	Key Brain Regions	Performance Notes
Cognitive Control	Gradual-onset continuous performance task	Frontoparietal network	Stronger detection than resting-state [11]
Negative Emotion	Emotional tasks (excluding N-back)	Limbic system, prefrontal cortex	Emotional N-back less optimal [11]
Sensitivity/Sociability	Gradual-onset continuous performance task	Social brain network	Superior to resting-state [11]
General Cognitive Ability	N-back tasks	Prefrontal, parietal regions	Surpasses resting-state functional connectivity [11]

Motion Artifact Characteristics and Correction Methodologies

Motion Artifact Manifestations Across Paradigms

Head motion presents distinct challenges in resting-state versus task-based fMRI, with characteristic patterns of artifact manifestation in each paradigm. In rs-fMRI, motion artifacts exhibit distinctive "distance" and "orientation" dependencies, with decreased long-distance connectivity and increased local connectivity [9]. Increased lateral connectivity at the expense of connectivity in the inferior–superior and anterior–posterior directions has also been observed [9].

The temporal patterns of movement and associated artifacts do not display band-limited frequency content in rs-fMRI [9]. As such, conventional frequency filtering may be ineffective for motion correction and can even smear motion contamination across the entire dataset if not applied carefully. Low-frequency, autocorrelated trends are readily apparent in rs-fMRI data due to motion, requiring specialized methods to remove these artifacts while retaining true fMRI signal content [9].

In task-based fMRI, head motion can be temporally correlated with task performance, creating unique challenges for distinguishing motion artifacts from true brain activation [9]. For example, button presses or speech production during tasks can induce head movements that coincide precisely with expected activation periods, creating potentially severe confounds.

Advanced Correction Algorithms and Processing Pipelines

Multiple sophisticated approaches have been developed to address motion-related artifacts in fMRI data. For resting-state fMRI, the aCompCor (anatomical Component Based Noise Correction) method utilizes principal components analysis to estimate nuisance signals from white matter and cerebrospinal fluid regions [12]. This approach effectively removes motion artifacts more effectively than tissue-mean signal regression and improves the specificity of functional connectivity estimates [12].

For task-based fMRI, novel processing pipelines like OGRE (One-step General Registration and Extraction) have been developed to reduce inter-individual variability. OGRE implements one-step interpolation that combines motion correction, field map distortion correction, and spatial normalization into a single transformation, minimizing the spatial blurring associated with multiple sequential transformations in conventional pipelines [13]. This approach has demonstrated significantly lower inter-subject variability compared to standard FSL preprocessing and stronger detection of task-related activation in primary motor cortex [13].

Diagram 1: Motion Artifact Pathways and Corrections

Experimental Protocols for Signal Optimization

Protocol Design for Resting-State fMRI

Resting-state fMRI acquisition requires careful protocol design to minimize motion artifacts while capturing meaningful spontaneous fluctuations. Effective protocols incorporate multiple mitigation strategies:

• Subject Preparation: Comprehensive instruction, training, and mild restraints can reduce motion, though these are usually insufficient alone [9].
• Acquisition Parameters: Optimized echo planar imaging (EPI) sequences with reduced echo spacing and parallel imaging techniques can minimize distortion and artifact susceptibility [14].
• Retrospective Correction: Implementation of multiple motion correction algorithms, including aCompCor, which uses principal components analysis from noise regions of interest to effectively attenuate motion artifacts without requiring external physiological monitoring [12].

Recent advances in acquisition technology include integrated reconstruction pipelines that simultaneously address 2D Nyquist and aliasing artifacts in EPI data, substantially improving image quality and functional sensitivity [14]. These approaches can correct aliasing artifacts induced by 2D Nyquist phase errors, in-plane and through-plane acceleration, as well as shot-to-shot motion-induced phase variations.

Protocol Design for Task-Based fMRI

Task-based fMRI protocols must balance experimental design with motion mitigation strategies:

• Task Selection: Choosing tasks with established predictive power for specific neuropsychological outcomes of interest [11].
• Paradigm Optimization: Block designs with adequate baseline periods facilitate separation of task-related signals from low-frequency drifts.
• Processing Pipelines: Implementation of one-step interpolation methods, as in the OGRE pipeline, which reduces inter-subject variability and enhances task-related signal detection compared to multi-step interpolation approaches [13].

The OGRE pipeline combines FreeSurfer brain extraction, FSL FNIRT registration, and one-step interpolation of preprocessing transformations, demonstrating significantly lower inter-subject variability than conventional FSL preprocessing or fMRIPrep [13]. This approach is particularly valuable for clinical research where minimizing variability is essential for detecting group differences or treatment effects.

Table 3: Experimental Protocols for Motion Mitigation

Protocol Component	Resting-State fMRI	Task-Based fMRI
Subject Preparation	Instruction, training, mild restraints	Task practice, response method optimization
Acquisition Parameters	Reduced FOV, multi-band acceleration	Task-locked timing, optimized TR
Retrospective Correction	aCompCor, scan scrubbing, tissue regression	One-step interpolation, physiological noise modeling
Quality Metrics	Framewise displacement, DVARS	Task-activation robustness, inter-subject variability
Software Tools	fMRIPrep, 1000 Functional Connectomes scripts	OGRE pipeline, FSL FEAT, HCP pipelines

The Scientist's Toolkit: Research Reagent Solutions

Table 4: Essential Research Tools for fMRI Motion Mitigation

Tool/Software	Primary Function	Application Context	Key Features
OGRE Pipeline	One-step interpolation preprocessing	Task-based fMRI	Reduces inter-subject variability, improves activation detection [13]
aCompCor	Physiological noise removal	Resting-state fMRI	PCA-based noise estimation from WM/CSF, superior to mean signal regression [12]
fMRIPrep	Integrated preprocessing	Both paradigms	Modular design, multiple software integration, standardized processing [13]
FSL FEAT	General linear model analysis	Task-based fMRI	Widely adopted volumetric analysis, over 15,000 citations [13]
HCP Pipelines	Surface-based analysis	Both paradigms	One-step interpolation, multimodal data integration [13]
BrainSurfCNN	Deep learning prediction	Cross-paradigm mapping	Surface-based CNN predicting task contrasts from resting-state data [15]

Emerging Frontiers and Future Directions

Cross-Paradigm Predictive Modeling

Advanced computational approaches are emerging that bridge resting-state and task-based fMRI paradigms. BrainSurfCNN, a surface-based fully-convolutional neural network model, can predict individual task-based contrast maps from resting-state fMRI scans with exceptional accuracy [15]. This model achieves predictive performance on par with the test-retest reliability of measured subject-level contrast maps, demonstrating the fundamental relationship between intrinsic connectivity and task-evoked activity [15].

Such approaches have significant implications for clinical research and drug development, potentially enabling the derivation of task-related information from resting-state scans alone. This could be particularly valuable for populations unable to perform complex cognitive tasks, such as individuals with severe neurological disorders or pediatric patients.

Integrated Analysis Frameworks

Future methodological developments are moving toward integrated analysis frameworks that combine the strengths of both resting-state and task-based approaches. Network science-driven Bayesian generative models like LatentSNA bypass power and replicability issues common in existing fMRI predictive models while incorporating network theories in model building [11]. These approaches demonstrate substantially improved precision and robustness in imaging biomarker detection, strengthening connectivity signals by allowing true connectivity and internalizing signals to mutually inform each other [11].

The growing availability of large neurodevelopmental datasets, such as the ABCD Study, enables well-powered analyses using large samples of longitudinal neuroimaging data collected from diverse populations [16]. These resources facilitate the development of more reliable brain-behavior associations through advanced acquisition protocols, multivariate analysis methods, and more robust consideration of phenotypic complexity [16].

Diagram 2: Integrated fMRI Analysis Workflow

The temporal properties and signal dynamics of resting-state and task-based fMRI demonstrate fundamental differences that significantly impact their application in research and clinical contexts. Task-based fMRI generally provides superior predictive power for specific neuropsychological outcomes, while resting-state offers practical advantages for certain populations and research questions. Motion artifacts manifest differently across paradigms, requiring tailored mitigation strategies such as aCompCor for resting-state and one-step interpolation pipelines for task-based fMRI. Emerging computational approaches that bridge these paradigms, including deep learning models that predict task activation from resting-state data, represent promising frontiers for future research. For drug development professionals and clinical researchers, selection of the appropriate fMRI paradigm must consider the specific behavioral domains of interest, target population characteristics, and the trade-offs between predictive power and practical implementation constraints.

In functional magnetic resonance imaging (fMRI), head motion presents a pervasive challenge to data integrity. However, not all motion artifacts are created equal. While resting-state fMRI contends with motion that introduces distance-dependent biases in functional connectivity, task-based fMRI faces a uniquely critical problem: motion that is temporally correlated with the experimental design. This specific type of artifact occurs when a participant's head movements systematically coincide with the presentation of stimuli or the performance of behavioral tasks. Such correlation transforms motion from a simple source of noise into a potent confounder that can produce spurious, false-positive activations that are indistinguishable from genuine neural responses [2] [17] [4]. The central challenge, as noted in foundational literature, is the risk of "throwing the baby out with the bathwater"—over-correction may remove true activations along with artifacts, while under-correction permits motion to masquerade as brain function [17].

This technical guide examines the unique nature of task-correlated motion, its impact on fMRI signal interpretation, and the advanced methodologies required to mitigate its effects. We frame this discussion within a broader thesis on motion artifact characteristics, contrasting the problems of temporal correlation in task-fMRI with the distance-dependent correlation artifacts predominant in resting-state studies.

Contrasting Motion Artifacts: Task fMRI vs. Resting-State fMRI

The fundamental difference in how motion artifacts manifest in task versus resting-state fMRI stems from their distinct experimental designs and analytical goals. The table below summarizes the key differential characteristics:

Table 1: Characteristics of Motion Artifacts in Task vs. Resting-State fMRI

Characteristic	Task fMRI	Resting-State fMRI
Primary Nature of Problem	Temporal confound: Motion correlates with task paradigm	Spatial confound: Motion correlates with tissue proximity
Effect on Signals	Mimics or obscures true hemodynamic response to tasks	Inflates short-distance correlations; reduces long-distance correlations
Statistical Consequence	Increased false positives for "activation"; biased effect size estimates	Distance-dependent bias in functional connectivity matrices
Typical Correction Approach	Models separating task-locked motion from true BOLD response	Global signal regression; censoring of high-motion volumes
Key References	[17] [4]	[2] [1] [3]

In resting-state fMRI, the primary concern is motion's spatially structured impact on functional connectivity metrics. Motion artifacts introduce a distance-dependent bias, spuriously increasing correlations between nearby brain regions while weakening correlations between distant regions [2] [1] [3]. This pattern arises because motion-induced signal changes are most similar in voxels close to one another, creating the appearance of organized functional networks that are actually artifactual in nature [3]. Studies have shown that even sub-millimeter motions can distort connectivity estimates from seed correlation analyses, independent component analysis (ICA), and graph theoretic approaches [2].

In task-based fMRI, the critical issue is the temporal structure of motion. When head movements coincide with task conditions (e.g., a participant moves slightly each time they press a button during a motor task), the motion becomes confounded with the experimental design [17] [4]. This correlation is particularly problematic because:

• It mimics true activation: Motion-induced signal changes time-locked to task events are indistinguishable from genuine hemodynamic responses in univariate analyses [17].
• It resists standard correction: Conventional motion correction (realignment and regression of motion parameters) may remove both artifact and true signal when they are correlated, potentially increasing false negatives—a phenomenon known as "over-correction" [17].
• It threatens validity: In clinical and cognitive studies, groups that differ in their tendency to move during specific tasks (e.g., patients vs. controls) may show apparent "activation differences" that reflect motion artifacts rather than neural differences [2] [4].

Mechanisms: How Motion Creates Spurious Signal Changes

Understanding how physical head movement translates into BOLD signal artifacts is crucial for developing effective correction strategies. Motion affects fMRI data through multiple physical mechanisms:

• Spin History Effects: When a head moves between radiofrequency pulses, the spin magnetization is disrupted from its steady-state recovery. This causes signal intensity changes that can persist for several seconds after motion ceases, as the spin system gradually returns to equilibrium [17] [4] [12].

• Through-Plane Motion: As the brain moves in and out of the imaging slice plane, the composition of tissues within each voxel changes. This alters the local magnetic properties and consequently the measured signal, particularly at tissue boundaries where partial volume effects are most pronounced [17] [4].

• Magnetic Field Interactions: Head motion within spatially varying magnetic fields (B0 inhomogeneity) causes geometric distortions and intensity changes in echo-planar imaging (EPI), the most common fMRI acquisition sequence. These effects are especially problematic because they introduce non-linear signal changes that cannot be fully corrected by rigid-body realignment [17] [4].

• Interpolation Artifacts: During image realignment, data must be interpolated to new voxel grids. This mathematical process inevitably introduces correlations between neighboring voxels and can create spurious signal changes, particularly when motion is large or abrupt [17].

These mechanisms collectively produce complex, variable signal waveforms that can be shared across nearly all brain voxels and may persist for more than 10 seconds after motion ceases [3]. The temporal profile of these artifacts often includes an initial large signal change followed by a prolonged recovery period, creating a time course that can closely resemble a true hemodynamic response when correlated with task timing [3].

Methodological Approaches for Detection and Correction

Retrospective Correction Methods

Table 2: Retrospective Motion Correction Methods for Task fMRI

Method	Key Principle	Advantages	Limitations
Extended Motion Regression	Includes temporal derivatives and quadratic terms of realignment parameters	Accounts for delayed and non-linear spin history effects; relatively simple implementation	Limited efficacy for strongly task-correlated motion; may remove true signal
Independent Component Analysis (ICA)	Decomposes data into spatially independent components; identifies and removes motion-related components	Separates motion from neural signals based on spatial patterns; does not require temporal orthogonality	Requires accurate component classification; risk of removing neural signals with similar spatial patterns
aCompCor	Uses principal components of noise from white matter and CSF as nuisance regressors	Effectively models spatially heterogeneous noise; does not assume global motion effects	May capture neural signal if noise ROIs are contaminated; component selection can be subjective
Censoring (Scrubbing)	Removes motion-corrupted volumes from analysis; may be combined with interpolation	Directly eliminates severely corrupted data points; can be combined with other methods	Creates discontinuities in time series; reduces statistical power; requires choice of threshold

Several specialized approaches have been developed specifically for addressing task-correlated motion:

ICA-Based Removal: This method involves decomposing the realigned fMRI time-series into spatially independent components, automatically classifying components representing task-related residual motion effects based on their associated task-related changes in signal intensity and variance, and reconstructing the data without these components [17]. The crucial advantage of ICA is its ability to separate neuronal responses from motion artifacts based on their spatial independence, bypassing the temporal correlation problem that plagues regression-based methods [17]. Validation studies have demonstrated that ICA-based removal more effectively reduces task-related motion effects compared to conventional voxel-wise regression, resulting in fewer false negatives while effectively controlling false positives [17].

aCompCor (Anatomical Component-Based Noise Correction): This method extracts noise regressors via principal component analysis (PCA) from regions of interest (ROIs) in white matter and cerebrospinal fluid (CSF), where BOLD signals are presumed to be dominated by noise [12]. Unlike mean-signal regression, aCompCor captures multiple spatially coherent noise patterns, making it more effective at modeling the spatially heterogeneous effects of motion. Research has shown that aCompCor removes motion artifacts more effectively than tissue-mean signal regression and improves the specificity of functional connectivity estimates [12].

Censoring (Scrubbing) with Structured Matrix Completion: For severely motion-corrupted volumes, censoring (removing problematic volumes) is an effective strategy. However, this creates discontinuities in time series. Advanced approaches now use structured low-rank matrix completion to recover missing entries after censoring by exploiting the linear recurrence relations in BOLD signals [8]. This method formulates the artifact-reduction problem as the recovery of a super-resolved matrix from unprocessed fMRI measurements, enforcing a low-rank prior on a large structured matrix formed from the samples of the time series [8]. This approach not only compensates for motion but also performs slice-time correction at a fine temporal resolution, significantly improving functional connectivity estimates and seed-based correlation analyses [8].

Prospective Motion Correction (PMC)

Beyond retrospective methods, prospective motion correction (PMC) uses real-time head tracking (typically with MR-compatible optical systems) to update the imaging sequence parameters during data acquisition, effectively stabilizing the brain's position relative to the scanner [18]. The tracking system determines head motion at a high sampling rate (e.g., 80 Hz), sending this information to the scanner to update radio frequency and gradient pulses before acquiring each fMRI slice [18].

Studies demonstrate that PMC significantly improves the temporal signal-to-noise ratio (tSNR) of fMRI data acquired during head movement, preserves the spatial definition of major resting-state networks (including the default mode and visual networks), and maintains correlation matrices comparable to those obtained during still conditions [18]. PMC is particularly valuable for populations prone to movement (e.g., children, patient groups) and for tasks that inherently induce motion.

Experimental Protocols and Validation

Validating Correction Methods with Controlled Motion

Rigorous validation of motion correction methods requires ground truth data where the presence and timing of motion are precisely known. One innovative approach uses a dynamic phantom ("Brain dancer") that can reliably generate signals with controllable onset delays mimicking fMRI responses in a block-design paradigm [19]. The phantom comprises two agarose compartments with different T2* values (mimicking "active" and "inactive" neural states), with a cylindrical design that rotates about its axis to create dynamic signal changes in specific voxels [19]. This setup enables researchers to quantify the effective temporal resolution (ETR)—the minimal time delay that can be discerned between evoked responses with high statistical confidence—under various scanning parameters, including echo time (TE), repetition time (TR), voxel size, and contrast-to-noise ratio (CNR) [19].

Protocol for ICA-Based Motion Correction

For researchers implementing ICA-based motion correction, the following protocol adapted from [17] provides a methodological roadmap:

• Data Preprocessing: Perform standard spatial realignment (motion correction) using rigid-body transformation to align all volumes to a reference image.
• ICA Decomposition: Apply spatial ICA to the realigned time-series data to decompose it into statistically independent components, each with an associated time course and spatial map.
• Component Classification: Automatically identify motion-related components using criteria based on:
  • Task-related signal change: Correlation between component time course and task design.
  • Temporal variance characteristics: Identification of heteroscedasticity (changing variance) aligned with task conditions.
• Data Reconstruction: Reconstruct the fMRI time series without the motion-classified components, effectively removing the artifact while preserving neural signals.

This protocol has been experimentally validated using fMRI studies where head motion was explicitly included as part of the task design, demonstrating superior performance compared to conventional voxel-wise regression methods [17].

Quantitative Data on Motion Effects and Correction Efficacy

Table 3: Quantitative Measures of Motion Effects and Correction Efficacy

Metric	Value/Relationship	Experimental Context	Reference
Effective Temporal Resolution (ETR)	151 ms (cortex-mimic); 248 ms (basal ganglia-mimic) at TR=600 ms	Multi-echo EPI with dynamic phantom at 7T	[19]
Framewise Displacement (FD) Threshold	FD > 0.5 mm: marked correlation changes; FD = 0.15–0.2 mm: significant changes begin	Empirical analysis of motion-contaminated resting-state data	[3]
tSNR Reduction from Intentional Motion	45% reduction without PMC; 20% reduction with PMC	Resting-state fMRI with instructed leg movements	[18]
Variance Attributable to Motion	30–90% of variance in MR signal	Literature review of motion effects in fMRI	[17]
Duration of Motion Artifacts	Persistence >10 seconds after motion cessation	Empirical characterization of motion-related signal changes	[3]

Table 4: Essential Tools for Investigating and Mitigating Task-Correlated Motion

Tool/Resource	Function/Purpose	Example Use Case
Dynamic Phantom	Generates "ground truth" signal with controllable onset delays	Validating temporal resolution and motion correction methods [19]
Optical Motion Tracking System	Provides real-time head pose data for prospective correction	Prospective Motion Correction (PMC) during task fMRI [18]
Independent Component Analysis	Blind source separation of signal and noise components	Removing task-correlated motion without temporal orthogonality [17]
Structured Low-Rank Matrix Completion	Recovers missing data after motion censoring	Interpolating censored volumes while maintaining data structure [8]
Framewise Displacement (FD)	Quantifies volume-to-volume head movement	Identifying motion-corrupted volumes for censoring [3]
aCompCor Algorithm	Derives noise regressors from WM and CSF compartments	Modeling spatially heterogeneous motion effects [12]

Task-correlated motion represents a distinct and critical challenge in task-based fMRI, differing fundamentally from the motion-related problems in resting-state studies. Its capacity to produce spurious, false-positive activations threatens the validity of findings across cognitive, clinical, and developmental neuroscience. While traditional correction methods struggle with the temporal confound inherent in task-correlated motion, advanced approaches including ICA-based removal, aCompCor, structured matrix completion, and prospective motion correction offer increasingly powerful solutions. As fMRI continues to advance with higher magnetic fields and accelerated acquisition, developing ever more sophisticated methods to address this persistent challenge will remain essential for generating meaningful and reliable insights into brain function.

Motion artifacts represent a significant challenge in functional magnetic resonance imaging (fMRI), potentially confounding results in both resting-state and task-based studies [20]. The quantification and correction of these artifacts are prerequisites for obtaining valid functional connectivity and activation maps. This technical guide provides an in-depth examination of two fundamental components in motion artifact management: realignment parameters (RPs), which describe head position over time, and framewise displacement (FD), a scalar summary derived from these parameters that quantifies volume-to-volume head movement [21]. Within the broader thesis on motion artifact characteristics, a critical distinction emerges: resting-state fMRI (rs-fMRI) analyses, particularly functional connectivity measures, demonstrate heightened sensitivity to motion-induced correlations compared to many task-based fMRI analyses focused on activation magnitudes [20] [3]. This document details the measurement, quantification, and mitigation protocols for these motion metrics, tailored for researchers, scientists, and drug development professionals.

Realignment Parameters (RPs): The Foundation of Motion Description

Definition and Acquisition

Realignment parameters are estimated for each volume in a functional time series via rigid-body registration to a reference volume (typically the first or middle volume). This process corrects for simple head displacement in space but does not address the associated intensity artifacts caused by the disruption of spin history [20]. The six parameters constitute a complete description of head position at each time point:

• Three Translational Parameters: Measured in millimeters (mm), representing displacement along the X (left-right), Y (anterior-posterior), and Z (superior-inferior) axes.
• Three Rotational Parameters: Measured in degrees (°) or radians, representing rotation around the X (pitch), Y (yaw), and Z (roll) axes.

Limitations of Realignment and RPs

Spatial realignment alone is an incomplete solution for motion artifact correction. While it compensates for geometric displacement, it does not rectify the signal intensity changes that occur when head movement disrupts the magnetic field establishment and readout of the BOLD signal [20]. Consequently, the RPs themselves become crucial regressors in subsequent processing steps to remove motion-related variance from the BOLD time series. However, a significant limitation of using summary statistics (e.g., mean displacement) derived from RPs is their inability to distinguish between qualitatively different types of motion, such as a single large movement versus frequent small movements, which can have disparate impacts on data quality [20].

Definition and Calculation

Framewise Displacement (FD) is a succinct metric that quantifies the total volume-to-volume head movement by summing the absolute derivatives of the six realignment parameters [21]. The formula for FD at timepoint i is:

FDᵢ = |Δxᵢ| + |Δyᵢ| + |Δzᵢ| + |Δαᵢ| + |Δβᵢ| + |Δγᵢ|

Where:

• Δxᵢ, Δyᵢ, Δzᵢ are the differences in translational parameters (in mm) between volume i-1 and i.
• Δαᵢ, Δβᵢ, Δγᵢ are the differences in rotational parameters.

Rotational displacements are converted from angular units (degrees) to spatial units (mm) by calculating the arc length on a sphere of a specified radius, typically 50 mm [21]. This conversion allows for the coherent summation of translational and rotational movements into a single, comprehensive index of head motion.

FD Calculation Workflow

The following diagram illustrates the standard workflow for computing Framewise Displacement from a raw fMRI time series.

Impact of Motion on fMRI Data and the Role of FD

Systematic Artifacts in Functional Connectivity

Subject motion introduces systematic but spurious correlation structures in resting-state functional connectivity MRI (rs-fcMRI) data. The artifact is distance-dependent: motion often spuriously increases short-distance correlations and decreases long-distance correlations [20] [3]. This occurs because motion-induced signal changes are most similar at nearby voxels, artificially inflating their correlation, while disrupting the coupling between distant brain regions.

Persistence of Motion Artifacts

Motion-related signal changes are complex and can persist for more than 10 seconds after the physical movement has ceased [3]. This prolonged effect means that even brief movements can contaminate a substantial number of subsequent volumes, making simple regression of concurrent motion parameters an insufficient correction strategy.

Inadequacy of Standard Corrections

Common functional connectivity processing steps, including spatial registration and regression of motion estimates from the data, do not fully remove motion-related artifacts [20]. While global signal regression (GSR) is highly effective at reducing motion-related variance, it introduces other interpretational complexities [3]. This underscores the necessity of identifying and dealing with motion-contaminated volumes before final analysis.

Experimental Protocols for Motion Mitigation

The "Scrubbing" Protocol

A widely adopted method for mitigating motion artifacts is "scrubbing," which involves the identification and removal of motion-contaminated volumes [20] [3]. The standard protocol is as follows:

• Compute FD: Calculate the Framewise Displacement time series from the six realignment parameters.
• Set FD Threshold: Identify volumes with FD values exceeding a predetermined threshold. A common and conservative threshold is FD > 0.2 mm [3]. Higher thresholds (e.g., 0.3 mm or 0.5 mm) are more lenient but may retain more artifact.
• Flag and Remove: Flag identified volumes as outliers. These volumes are subsequently excluded from functional connectivity calculations or other downstream analyses. Some protocols also remove the volume immediately preceding and following a high-motion volume to account for the temporal smearing of the hemodynamic response.
• Censored Interpolation: Advanced implementations may interpolate across censored volumes using data from neighboring, low-motion volumes to maintain a continuous time series, though this is not always necessary for correlation-based connectivity measures.

Threshold Sensitivity and Comparison of QC Measures

The performance of scrubbing is highly dependent on the chosen FD threshold. The table below summarizes quantitative findings related to FD thresholds and their impact on data quality and functional connectivity (FC).

Table 1: Quantitative Data on FD Thresholds and Motion Artifact Impact

FD Threshold (mm)	Impact on Data and Functional Connectivity
> 0.2	Significant changes in FC correlations begin to be detectable [3].
> 0.3	A common, moderately conservative threshold for flagging motion-contaminated frames [21].
> 0.5	Marked and substantial changes in FC correlations are observed [3].

Protocol for Group-Level Analysis and Matching

When comparing cohorts (e.g., patients vs. controls, children vs. adults), it is critical to account for between-group differences in motion, as these can create spurious findings.

• Quantify Motion per Subject: Calculate a summary statistic for each subject, such as mean FD or the number of volumes with FD > 0.2 mm.
• Match Groups: Statistically match groups on these summary measures. If matching is not perfect, include the mean FD as a nuisance covariate in group-level analyses.
• Interpret with Caution: Be aware that group-level regression of motion metrics may inadvertently remove true effects of interest if they are correlated with motion (e.g., if a clinical group moves more and also has genuine connectivity alterations) [3].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 2: Essential Tools and Software for Motion Measurement and Correction in fMRI

Tool/Solution	Function/Brief Explanation	Example Software Packages
Realignment Algorithm	Performs rigid-body registration of each volume to a reference, outputting the 6 realignment parameters.	FSL MCFLIRT, SPM Realign, AFNI 3dVolreg
Framewise Displacement Calculator	Computes the FD time series from the 6 realignment parameters, handling unit conversion for rotations.	fMRIscrub R package [21], BRAMILA Matlab tools [22], in-house scripts
Data Censoring ("Scrubbing") Tool	Identifies and flags volumes exceeding a specified FD threshold for removal from subsequent analysis.	Integrated in CONN, DPABI, and custom pipelines
Global Signal Regressor	A powerful nuisance regressor that reduces motion-related variance but must be used with caution due to potential interpretive issues [3].	Can be generated and applied in most major fMRI software (SPM, FSL, AFNI)
High-Quality Template	The reference image (often the first or middle volume of the time series) used for realignment.	Provided by the user's own data or standard space templates (MNI152)

Comparative Analysis: Resting-State vs. Task fMRI and Motion Sensitivity

The characteristics and impact of motion artifacts differ meaningfully between resting-state and task-based fMRI paradigms, informing the broader thesis on motion artifact characteristics.

Table 3: Motion Artifacts in Resting-State vs. Task fMRI

Aspect	Resting-State fMRI (rs-fMRI)	Task-Based fMRI
Primary Analysis	Functional connectivity (correlation between time series).	Activation (signal change linked to task blocks/events).
Key Motion Effect	Systematic, distance-dependent spurious correlations [20]. Can profoundly alter the apparent network structure.	Signal dropouts and intensity changes at the single-volume level, potentially mislocalizing activation.
Sensitivity to Motion	Extremely high. Correlation metrics are directly biased by motion-induced signal changes shared across voxels [5].	Variable, but generally lower direct bias. Motion can be less correlated with the task paradigm, making it a tractable confound in the GLM.
Role of FD/Scrubbing	Often critical. Scrubbing is a cornerstone of modern rs-fMRI pipelines to avoid spurious connectivity results [20] [3].	Useful, but sometimes less central. Motion regressors in the GLM are often the primary correction method.

A critical finding is that the method used to estimate functional connectivity in rs-fMRI influences sensitivity to motion. For instance, full correlation demonstrates a higher residual relationship with motion compared to partial correlation and information-theoretic measures, even after rigorous motion correction [5]. This indicates that the choice of connectivity metric is a key consideration in study design, trading off motion sensitivity against other properties like test-retest reliability.

Framewise Displacement, derived from realignment parameters, is an indispensable metric for quantifying head motion in fMRI. Its primary strength lies in identifying specific volumes contaminated by motion, enabling targeted mitigation strategies like scrubbing. The pervasive and structured artifacts introduced by motion, particularly in resting-state functional connectivity, necessitate a rigorous, multi-stage approach to motion correction. This includes robust subject-level scrubbing, consideration of different connectivity metrics, and careful group-level analysis to prevent spurious findings. As fMRI continues to play a vital role in basic neuroscience and drug development, the precise measurement and mitigation of motion artifacts via FD and RPs remain foundational to generating valid and interpretable results.

Functional magnetic resonance imaging (fMRI) has become an indispensable tool for studying brain function across diverse populations. However, head motion remains a significant confound in fMRI data acquisition and analysis, particularly in resting-state fMRI (rs-fMRI) studies where spontaneous fluctuations are measured without external task constraints. The characteristics of motion vary substantially across different population groups, including clinical, pediatric, and elderly participants, creating unique challenges for researchers and clinicians. Understanding these population-specific motion patterns is crucial for developing effective motion mitigation strategies and interpreting neuroimaging findings accurately, especially within the broader context of differentiating motion artifact characteristics in resting-state versus task-based fMRI research.

This technical guide examines the distinct motion profiles across population groups, provides quantitative comparisons of motion parameters, outlines experimental protocols for motion reduction, and presents visualization tools to aid researchers in addressing these challenges. The insights presented here are particularly relevant for drug development professionals who rely on fMRI biomarkers for clinical trials and therapeutic development, where motion artifacts can significantly impact data quality and interpretation.

Motion Characteristics Across Populations

Pediatric Populations

Pediatric head motion presents unique challenges due to developmental factors. Research has consistently demonstrated an inverse relationship between head motion and age [23], with younger children exhibiting significantly greater motion than adolescents. A comprehensive analysis of a transdiagnostic pediatric sample (N = 1,388, ages 5-21 years) revealed that head motion in children is not random but follows specific biomechanical patterns [23].

• Motion Patterns: High movers exhibit motion dominated by x-rotation (pitch), combined with z- and y-translation, creating a characteristic "nodding" movement [23]. This specific pattern provides a focused target for motion reduction strategies.
• Physiological Influences: Children's respiration rates (14-22 breaths/minute) are significantly higher than adults (12-18 breaths/minute), contributing to a baseline of high-frequency motion best addressed during data preprocessing [23].
• Developmental Factors: Anatomically, children have proportionally larger heads relative to their bodies, weaker neck muscles and ligaments, and more flexed head position when supine—all contributing to increased motion [23].

Table 1: Pediatric Motion Characteristics and Mitigation Strategies

Characteristic	Impact on fMRI	Recommended Mitigation
Nodding pattern (x-rotation with z-y translation)	Significant signal dropouts; correlation artifacts	Targeted head stabilization; real-time motion correction
Higher respiratory rates	High-frequency motion components	Physiological noise modeling; band-pass filtering
Large head-to-body ratio	Increased neck strain; reduced comfort	Specialized head support; padding optimization
Limited attention span	Increased motion over time	Shorter acquisition protocols; engaging stimuli

Elderly Populations

Elderly populations present distinct motion-related challenges, often characterized by age-related physiological factors and comorbidities. While elderly participants typically exhibit less gross head movement than children, they face issues related to discomfort, arthritis, and neurological conditions that can affect data quality.

• Structural-Functional Relationships: Research has shown that structural degradation in white matter pathways, particularly in the anterior corpus callosum, may lead to altered functional connectivity patterns in older adults [24]. This presents a complex interaction where motion-related artifacts must be distinguished from genuine neurobiological changes.
• Compensatory Mechanisms: Some studies suggest that structural degradation may initially lead to increased functional connectivity as a potential compensatory mechanism, though this interpretation requires careful validation against motion artifacts [24].

Clinical Populations

Clinical populations present unique motion challenges that vary by specific disorder or condition. For example, individuals with autism spectrum disorder (ASD) often exhibit increased motion during scanning, while those with Parkinson's disease may have tremor-related movements that require specialized acquisition protocols [25] [26].

• Disorder-Specific Patterns: Research on Freezing of Gait (FOG) in Parkinson's disease has identified specific neural networks affected by the condition, including the extramedullary system, visual networks, default mode network, salience network, and supplementary motor area [26]. Each of these networks may be differentially affected by motion artifacts.
• Medication Effects: Clinical populations often undergo pharmacological treatments that can either increase restlessness or induce sedation, both of which impact motion characteristics during scanning sessions.

Motion in Resting-State vs. Task fMRI

The context of fMRI acquisition—whether during rest or task performance—significantly influences motion patterns and their impact on data quality. Understanding these differences is essential for appropriate experimental design and data interpretation.

Resting-State fMRI Motion Considerations

Resting-state fMRI examines spontaneous brain activity in the absence of external tasks, providing valuable insights into intrinsic brain networks. However, the lack of engaging stimuli presents particular challenges for motion management:

• Mind Wandering and Restlessness: During rest, participants, particularly children, may experience increased mind wandering and subsequent restlessness. Children have different relationships with spontaneous thoughts and may not sustain mind wandering as immersively as adults, leading to increased motion over time [23].
• Network Identification Challenges: Motion artifacts can mimic or obscure genuine functional connectivity patterns, potentially leading to erroneous conclusions about network organization and integrity [27] [28].

Task-Based fMRI Motion Considerations

Task-based fMRI incorporates experimental manipulations to probe specific cognitive processes, which introduces different motion considerations:

• Task Engagement Benefits: Appropriate task engagement can significantly reduce head motion. Movie-watching, for example, has been shown to reduce mean motion and temporal drift (linear increases in motion over time), particularly in high-motion participants [23].
• Stimulus-Correlated Motion: In task-based designs, motion may become correlated with specific task conditions, creating confounds that are difficult to disentangle from true neural activity [23].
• Post-Task Effects: Research has shown that brain states following task performance can exhibit altered functional gradient stability, with clinical populations such as those with lifelong premature ejaculation showing delayed recovery to baseline resting-state connectivity patterns [29].

Table 2: Motion Characteristics in Resting-State vs. Task fMRI

Parameter	Resting-State fMRI	Task-Based fMRI
Overall motion levels	Generally higher, especially in pediatric populations	Often reduced due to engagement
Motion patterns	More random, less predictable	May correlate with task conditions
Temporal drift	Increases over time	Reduced with engaging stimuli
Mitigation approaches	Mock scanning, comfort optimization	Task engagement, clear instructions
Data quality impact	Affects intrinsic connectivity estimates	May confound task-related activation

Quantitative Motion Metrics and Data Analysis

Robust quantification of head motion is essential for developing effective mitigation strategies and ensuring data quality. Several metrics and analytical approaches have been developed to characterize motion across populations.

Motion Quantification

• Framewise Displacement (FD): FD is widely used to quantify head motion by calculating the absolute positional and rotational displacement between consecutive volumes [25] [23]. Studies often apply FD thresholds (e.g., 0.10 mm, 0.15 mm, 0.20 mm) to classify scans as high- or low-motion [25].
• Spectral Analysis: Examination of motion in the frequency domain can identify physiological sources of motion, such as respiration and cardiac cycles, which occur at characteristic frequencies [23].

Analytical Considerations

• Motion Covariation: Including motion parameters as covariates in general linear models helps reduce but does not eliminate motion-related artifacts, particularly in functional connectivity analyses [23].
• Data-Driven Approaches: Machine learning methods applied to rs-fMRI data have shown excellent accuracy in classifying clinical populations (e.g., Parkinson's disease with Freezing of Gait), but require careful handling of motion confounds to avoid spurious classification features [26].

Experimental Protocols for Motion Mitigation

Several evidence-based protocols have been developed to mitigate motion artifacts across different populations. These approaches can be implemented during scanning preparation, data acquisition, and data processing stages.

Pre-Scanning Preparation Protocols

• Mock Scanner Training: Comprehensive mock scanning protocols that expose participants to the scanning environment can significantly reduce motion. One effective protocol includes:
  • Gradual exposure to scanner sounds and environment
  • Practice remaining still with feedback
  • Motion monitoring with camera systems
  • Desensitization to scanner anxiety [25]
• Participant Preparation: Adequate preparation includes explaining the importance of staying still, ensuring comfort, and addressing any concerns before entering the scanner.

In-Scanner Motion Reduction Techniques

• Multimodal Engagement: Combining mock scanning with in-scan steps such as weighted blankets and incentive systems can achieve low-motion data even in pediatric populations undergoing 60-minute scan protocols [25].
• Stimulus Selection: Movie-watching significantly reduces motion compared to rest, particularly for high-motion individuals. Naturalistic stimuli like movies decrease temporal drift in motion and increase intersubject correlations of framewise displacement [23].
• Real-Time Monitoring: Framewise Integrated Real-time MRI Monitoring (FIRMM) software allows researchers to monitor head motion in real-time and continue scanning until sufficient low-motion data is acquired [25].

Diagram 1: Comprehensive Motion Mitigation Workflow

The Scientist's Toolkit: Research Reagents and Materials

Implementing effective motion mitigation requires specific tools and approaches tailored to different population needs. The following table outlines essential components of a comprehensive motion management protocol.

Table 3: Essential Research Materials for Motion Management

Tool/Resource	Function/Purpose	Population Specificity
Mock Scanner Facility	Participant acclimation to scanner environment; motion behavior training	Critical for pediatric and clinical populations with anxiety
Weighted Blankets	Provides proprioceptive input; reduces restlessness and anxiety	Particularly effective for pediatric and ASD populations
Incentive Systems	Motivational tools to reinforce stillness during scanning	Effective for children; can be adapted for elderly with appropriate rewards
Real-Time Motion Monitoring (FIRMM)	Tracks head motion during acquisition; informs data collection decisions	Beneficial for all populations but essential for high-motion groups
Motion-Robust Acquisition Sequences	Pulse sequences designed to minimize motion sensitivity	Multi-echo sequences valuable for all populations
Naturalistic Stimuli	Movie clips or engaging tasks to maintain attention	Reduces motion in pediatric populations during long acquisitions
Specialized Head Coils and Padding	Improves comfort and immobilization without increasing anxiety	Customized approaches needed for different head sizes (pediatric vs. adult)
Physiological Monitoring Equipment	Records cardiac and respiratory signals for noise modeling	Essential for distinguishing physiological motion from neural signals

Advanced Methodologies and Analytical Approaches

Multi-Contrast fMRI Techniques

Advanced acquisition methods such as multi-contrast laminar fMRI at ultra-high fields (7T) enable simultaneous measurement of multiple hemodynamic parameters, potentially providing more robust metrics less susceptible to motion artifacts [30]. This technique allows comprehensive characterization of neurovascular responses across cortical layers, which may help distinguish motion-related artifacts from genuine neural signals.

Machine Learning Applications

Machine learning approaches applied to rs-fMRI data have demonstrated excellent accuracy in classifying clinical conditions such as Parkinson's disease with Freezing of Gait (AUC > 90%) [26]. These methods can potentially identify motion-resistant features that maintain diagnostic utility despite motion-related variance in the data.

Diagram 2: Motion Artifact Pathways in Resting-State vs Task fMRI

Understanding population-specific motion patterns is essential for advancing fMRI research across clinical, pediatric, and elderly populations. The distinct motion characteristics observed in each group necessitate tailored approaches to experimental design, data acquisition, and analysis.

Future research should focus on developing standardized motion assessment protocols that can be applied across populations, enabling more direct comparison of motion effects and mitigation strategy efficacy. Additionally, advances in acquisiition techniques, such as multi-echo fMRI and multi-contrast approaches, show promise for inherently reducing motion sensitivity [30]. The integration of machine learning methods for automated motion detection and correction represents another promising direction for addressing these persistent challenges [26].

For drug development professionals utilizing fMRI biomarkers, rigorous motion management is particularly critical, as motion artifacts can mimic or obscure treatment effects, potentially leading to erroneous conclusions about therapeutic efficacy. By implementing the population-specific strategies outlined in this guide, researchers can significantly improve data quality and enhance the validity of their findings across diverse participant groups.

Motion Correction Methodologies: From Retrospective Processing to Real-Time Solutions

Head motion remains one of the most significant confounding factors in functional magnetic resonance imaging (fMRI), threatening the validity of both resting-state and task-based studies. The blood oxygen level-dependent (BOLD) signal measured with fMRI is susceptible to various noise sources, with motion artifacts systematically altering correlations in functional connectivity measures and reducing statistical power in task activation maps [3] [31]. In resting-state fMRI (rs-fMRI), motion introduces spurious correlations that can be mistaken for meaningful neuronal connectivity, particularly creating distance-dependent artifacts where correlations are most increased between nearby voxels [3]. In task-fMRI, motion that correlates with the experimental paradigm can create false activations or obscure true activation patterns, complicating interpretation of results [32]. These challenges are particularly acute in clinical populations, including children, elderly patients, and individuals with neurological conditions such as multiple sclerosis, who may exhibit greater movement during scanning [31] [32].

Retrospective correction pipelines—those applied after data acquisition—represent the most common approach for mitigating motion artifacts in fMRI research. These pipelines typically involve a sequence of operations including realignment (correcting for head movement across volumes), regression (statistically removing motion-related variance), and filtering (temporally filtering out noise frequencies) [33]. However, the modular nature of these preprocessing steps introduces complex interactions that, if not properly managed, can lead to the reintroduction of artifacts previously removed in earlier stages [34]. This technical guide examines the core principles, methodologies, and practical implementation of retrospective correction pipelines, with particular attention to their application across resting-state and task-based fMRI paradigms.

Motion Artifact Characteristics: Differential Impacts in Resting-State vs. Task fMRI

Fundamental Properties of Motion Artifacts

Motion-induced signal changes in fMRI data exhibit several characteristic properties that complicate their removal. These artifacts are often complex and variable waveforms that are shared across nearly all brain voxels and can persist for more than 10 seconds after the physical movement has ceased [3]. This prolonged impact means that even brief movements can affect multiple subsequent volumes in a time series. The spatial signature of motion artifacts is particularly problematic for functional connectivity measures, as it introduces distance-dependent effects where correlations are most strongly increased between nearby brain regions, potentially mimicking biologically plausible connectivity patterns [3].

Motion affects the BOLD signal through multiple mechanisms. Rigid body movement (translation and rotation) causes spin-history effects where the longitudinal magnetization of spins is altered as they experience different excitation histories, and intra-slice motion that occurs during the acquisition of a single slice [31]. Furthermore, head displacements alter magnetic field homogeneity and shimming, creating dynamic distortions that persist even after perfect image realignment [31]. These complex physical underpinnings explain why motion artifacts cannot be fully addressed through a single correction approach.

Paradigm-Specific Considerations

The impact and characteristics of motion artifacts differ substantially between resting-state and task-based fMRI paradigms, necessitating tailored correction strategies:

• Resting-State fMRI: In rs-fMRI, the primary concern is motion introducing systematic biases in functional connectivity measures. Even small movements can create spurious correlations that vary systematically with age, clinical status, or other factors of interest, potentially leading to false group differences [3]. The low-frequency nature of resting-state BOLD fluctuations (typically <0.1 Hz) makes them particularly vulnerable to contamination by slow motion-related drifts.

• Task-Based fMRI: In task paradigms, the central challenge is temporal correlation between motion and task design. When movement synchronizes with task blocks or events, it can produce false activations or reduce sensitivity to detect true activation patterns [32]. Task designs also typically involve higher frequency signal changes than resting-state, bringing them closer to the frequency characteristics of motion artifacts and complicating their separation through temporal filtering.

Table 1: Motion Artifact Characteristics in Resting-State vs. Task fMRI

Characteristic	Resting-State fMRI	Task-Based fMRI
Primary Impact	Alters functional connectivity measures	Creates false activations or reduces detection power
Temporal Pattern	Low-frequency fluctuations vulnerable to slow drifts	Event-related or block-designed patterns may correlate with motion
Spatial Pattern	Distance-dependent artifactual correlations	Focal artifacts in activated regions
Group Differences	Systematic biases between clinical populations	Confounded activation differences between groups
Correction Challenges	Separating neural correlations from motion-induced correlations	Disentangling task-locked BOLD responses from task-correlated motion

Core Methodologies in Retrospective Correction

Volume Realignment and Rigid Body Transformation

The foundational step in retrospective motion correction is volume realignment, which estimates and corrects for head movement across successive volumes. This approach treats the head as a rigid body with six degrees of freedom—three translational (displacement along X, Y, and Z axes) and three rotational (pitch, yaw, and roll) [33]. The standard implementation involves:

• Reference Selection: A single functional volume (typically the first or an average) is selected as the reference.
• Cost Function Optimization: Each volume is iteratively rotated and aligned to the reference using a cost function (typically mean-squared difference) minimized through nonlinear least-squares routines like Levenberg-Marquardt.
• Spatial Transformation: The calculated transformation parameters are applied to create a motion-corrected dataset through spatial interpolation [33].

While realignment effectively addresses gross head movement, it has important limitations. It cannot correct for spin-history effects or intra-slice motion, and it fails to account for dynamic changes in magnetic field homogeneity caused by head displacement [31]. Furthermore, interpolation during spatial transformation may introduce spatial smoothing and autocorrelations into the data [33].

Nuisance Regression Techniques

Following realignment, nuisance regression approaches statistically remove motion-related variance from the BOLD signal by including motion parameters as regressors of no interest in general linear models. The most common implementations include:

• 6-Parameter Model: Includes the three translation and three rotation parameters derived from realignment.
• 12-Parameter Model: Adds the temporal derivatives of the six motion parameters to account for lagged motion effects.
• 24-Parameter Model: Further expands to include quadratic terms of the original parameters and their derivatives, attempting to capture nonlinear motion effects [32].

Comparative studies suggest that the 6-parameter model often represents the best trade-off between motion correction and preservation of valuable neural signal, as more complex models may remove meaningful variance along with motion artifacts [32]. Beyond motion parameters, additional nuisance regressors may include:

• Physiological Signals: Cardiac and respiratory measures, when available.
• Tissue-Based Regressors: Mean signals from white matter and cerebrospinal fluid to remove physiological noise.
• Global Signal Regression: The global mean signal across all voxels, though this remains controversial as it can alter the interpretation of anticorrelated networks [3].

Temporal Filtering Approaches

Temporal filtering addresses frequency-specific noise in the BOLD signal. High-pass filtering removes low-frequency drifts (typically <0.01 Hz) that may arise from scanner instabilities or slow motion-related trends. Low-pass filtering can reduce high-frequency noise, though it is less commonly applied as it may remove neural signal of interest, particularly in task designs [33].

The order of filtering operations is critical, as filtering after regression can reintroduce motion-related variance previously removed. This occurs because filtering operates on the entire time series, including periods with motion artifacts, and can spread their influence to other time points [34]. As such, current best practices often recommend performing temporal filtering within the general linear model framework rather than as a separate preprocessing step.

Advanced Motion Correction Strategies

• Censoring (Scrubbing): Identifies and removes volumes affected by excessive motion, typically defined by framewise displacement (FD > 0.5 mm) or DVARS thresholds. Censoring can be implemented through complete removal or by including spike regressors in the GLM [3] [32].
• Volume Interpolation: Replaces motion-corrupted volumes with interpolated data from adjacent non-corrupted volumes, preserving the temporal structure of the data while reducing artifact influence [32].
• Component-Based Methods: Algorithms like ICA-AROMA use independent component analysis to automatically identify and remove motion-related components from the data [34].

Table 2: Performance Comparison of Motion Correction Methods

Method	Mechanism	Advantages	Limitations
Volume Realignment	Spatial alignment of volumes to reference	Standardized, widely implemented	Cannot correct spin-history or field inhomogeneity effects
6-Parameter Regression	Statistical removal of motion variance	Simple, preserves degrees of freedom	Limited capture of complex motion effects
24-Parameter Regression	Expanded regressors including derivatives and quadratics	Captures more complex motion variance	May remove neural signal, reduces power
Scrubbing	Removal of contaminated volumes	Effective for large motion events	Reduces temporal degrees of freedom
Volume Interpolation	Replaces corrupted volumes with estimated data	Preserves temporal structure	Potential smoothing of signal transitions
ICA-AROMA	Data-driven component classification and removal	Adaptive to specific motion patterns	Requires careful component classification

Implementation Protocols and Workflow Design

Integrated Preprocessing Pipeline

Implementing an effective retrospective correction pipeline requires careful consideration of the order and interaction between processing steps. The following workflow represents a current best-practice approach for integrating realignment, regression, and filtering operations:

Diagram 1: Integrated Preprocessing Workflow (77 characters)

This workflow emphasizes the integrated handling of motion correction where realignment, outlier detection, and regression are performed in a coordinated manner. Critical implementation considerations include:

• Simultaneous Nuisance Modeling: Rather than performing sequential regressions, include all nuisance regressors (motion parameters, spikes, tissue signals) in a single model to prevent artifact reintroduction [34].
• Filtering After Regression: Perform temporal filtering after nuisance regression to prevent the reintroduction of filtered motion variance.
• Covariate Orthogonalization: If sequential processing is necessary, orthogonalize later covariates with respect to earlier ones to maintain separation between removed components [34].

Experimental Protocol for Method Comparison

When evaluating motion correction efficacy in a specific experimental context, researchers can implement the following systematic comparison protocol:

• Data Acquisition: Acquire fMRI data with embedded motion conditions, including both task and resting-state scans to evaluate paradigm-specific effects.
• Pipeline Implementation: Process data through multiple correction approaches (e.g., 6-parameter, 24-parameter, scrubbing, volume interpolation).
• Quality Metric Calculation: Compute quantitative metrics including Framewise Displacement (FD), DVARS (standardized variance of spatial differences), and quality control measures (e.g., qualitative assessment ratings).
• Outcome Assessment: Evaluate correction efficacy through:
  • Residual correlation between motion and functional connectivity (for rs-fMRI)
  • Activation maps and false positive rates (for task-fMRI)
  • Between-group differences in motion-related metrics

This protocol was implemented in a recent multiple sclerosis study comparing correction methods, which found that models with 6 motion parameters combined with volume interpolation outperformed more complex approaches in reducing motion-related artifacts while preserving neural signal of interest [32].

The Scientist's Toolkit: Essential Research Reagents

Table 3: Essential Software Tools for Retrospective Motion Correction

Tool Name	Primary Function	Application Notes
fMRIPrep	Integrated preprocessing pipeline	Analysis-agnostic tool that automatically adapts workflows to dataset idiosyncrasies; combines tools from FSL, ANTs, FreeSurfer, AFNI [35] [36]
FSL	FMRIB Software Library (FEAT, MCFLIRT)	Provides motion correction (MCFLIRT), independent component analysis (MELODIC), and physiological noise modeling; widely used for volume realignment [36]
AFNI	Analysis of Functional NeuroImages (3dvolreg, 3dTshift)	Offers volume registration, slice timing correction, and nonlinear distortion correction; includes extensive time series analysis tools [36]
SPM	Statistical Parametric Mapping (spm_realign)	Provides realignment, coregistration, and spatial normalization; implements mass-univariate GLM approach for statistical analysis [36]
ICA-AROMA	Automatic Removal of Motion Artifacts	Data-driven component classification that identifies and removes motion-related independent components; particularly effective for high-motion datasets [34]
RABIES	Rodent Automated Bold Improvement of EPI Sequences	Standardized platform for rodent fMRI preprocessing with robust registration and quality control; adapted for cross-species translational research [37]

Quality Assessment and Validation Framework

Quantitative Quality Control Metrics

Robust validation of motion correction efficacy requires multiple quantitative measures:

• Framewise Displacement (FD): Quantifies the spatial extent of head movement between consecutive volumes, calculated as the sum of absolute translational displacements and rotational changes (converted to millimeters). Volumes with FD > 0.5 mm are typically flagged as motion outliers [32].
• DVARS: Measures the rate of change of BOLD signal across the entire brain at each time point, calculated as the root mean square of voxel-wise differentiated data. Peaks in DVARS indicate abrupt changes often associated with motion [32].
• Quality Indexes: Include measures such as temporal signal-to-noise ratio (tSNR), contrast-to-noise ratio, and spatial smoothness estimates that may be affected by motion correction procedures.

These metrics should be evaluated both before and after correction to assess the specific impact of each processing step. However, researchers should note that improvements in these metrics may sometimes represent cosmetic improvements rather than true correction of motion artifacts, particularly when filtering operations simply smooth out motion-related spikes without addressing underlying signal distortions [3].

Experimental Validation Approaches

Several experimental designs can specifically validate motion correction efficacy:

• Test-Retest Reliability: Acquiring repeated scans from the same individuals under different motion conditions (e.g., instructed movement vs. still periods) allows direct comparison of correction methods in recovering consistent neural signals [34].
• Phantom Studies: Using artificial phantoms with simulated motion provides ground truth comparisons for evaluating different correction approaches.
• Group Comparison Controls: Including motion-matched control groups in clinical studies helps distinguish true neural effects from residual motion artifacts.

Recent multi-site studies have demonstrated that standardized preprocessing pipelines like fMRIPrep can achieve high-quality results across diverse datasets, with success rates exceeding 99% for critical operations like cross-subject alignment and susceptibility distortion correction [36].

Retrospective correction pipelines incorporating realignment, regression, and filtering approaches represent essential tools for mitigating motion artifacts in fMRI research. The current evidence supports integrated pipeline designs that address the sequential nature of preprocessing steps to prevent artifact reintroduction. For most applications, volume realignment with 6 motion parameters combined with targeted approaches for motion outliers (such as volume interpolation or scrubbing) provides an effective balance between artifact removal and signal preservation.

The differential impact of motion in resting-state versus task-based fMRI necessitates paradigm-specific validation of correction approaches. In rs-fMRI, the focus should be on eliminating distance-dependent correlations without introducing global signal alterations, while in task-fMRI, the priority is preventing false activations from task-correlated motion.

Future developments in retrospective correction will likely include more sophisticated dynamic distortion correction methods, machine learning approaches for artifact identification, and improved integration with prospective motion correction systems. Furthermore, standardized quality control frameworks and reporting standards will enhance reproducibility across the field. As these technical advances mature, they will strengthen the validity of fMRI findings in both basic neuroscience and clinical drug development applications.

Head motion is one of the largest sources of noise in functional magnetic resonance imaging (fMRI), particularly in resting-state fMRI (rs-fMRI) where it can significantly distort estimates of functional connectivity [38] [3] [2]. Even sub-millimeter movements can introduce spurious but spatially structured patterns that bias results, especially in between-group studies involving pediatric, clinical, or elderly populations who typically exhibit higher motion levels [3] [2] [39]. The BOLD signal changes induced by motion are often complex and variable waveforms that can persist for more than 10 seconds after the physical movement ceases, creating prolonged artifacts that contaminate the data [3].

In task-based fMRI, motion can be temporally correlated with task performance, while in resting-state fMRI, the absence of a prescribed behavioral task does not eliminate the problem—in fact, motion may be exacerbated when individuals are at rest [2]. The artifacts manifest differently across fMRI paradigms: in resting-state analyses, motion spuriously increases short-range correlations while decreasing long-distance connectivity, creating a characteristic distance-dependent artifact pattern [3] [39]. This differential impact underscores why advanced motion correction techniques must be tailored to the specific research context and why a "one-size-fits-all" approach often proves insufficient.

Limitations of Traditional Motion Correction Approaches

The Linearity Assumption Problem

The most common traditional approach for motion correction involves regressing out the six rigid-body realignment parameters (3 translations and 3 rotations) obtained from volume registration, often with the addition of their temporal derivatives (creating a 12-parameter model) [38] [40] [41]. Some variants extend this further to 24 parameters by including time-shifted and squared versions of these motion parameters [38] [41]. However, these approaches rest on a fundamental limitation: they assume that motion-induced signal changes are linearly related to the estimated realignment parameters [38] [40].

This linearity assumption fails to account for the complex, nonlinear ways that motion actually affects the MR signal. At curved edges of image contrast, motion in one direction may cause a signal increase while the same motion in the opposite direction may not produce a commensurate decrease [38] [40]. Similarly, in regions with nonlinear intensity gradients, displacements in opposite directions produce asymmetric signal changes [40]. Motion can also cause voxels to sample different proportions of tissue classes, resulting in signal changes that may be positive, negative, or neutral depending on the specific proportions sampled [38] [40].

Residual Artifacts After Traditional Correction

Even after applying traditional nuisance regression with motion parameters, significant motion-related artifacts persist in the data. Research has demonstrated that multiple quality control measures continue to show residual dependence on head motion after standard processing, particularly influencing correlations between nearby brain regions [3] [39]. This residual variance occurs because rigid-body transformations cannot compensate for nonlinear effects such as field inhomogeneity changes, spin-excitation history artifacts, motion during slice acquisition, and interpolation errors [2] [33].

The consequence of these limitations is that traditional correction methods alone are often insufficient to completely remove motion-related biases, especially when studying populations with different motion characteristics or when investigating traits that correlate with motion levels [39]. This has led to the development of more sophisticated approaches that better model the complex relationship between head movement and signal changes.

Advanced Regression Techniques for Motion Correction

Expanded Motion Parameter Models

The expansion beyond the basic 6-parameter model represents an important evolution in motion regression techniques. The 12-parameter model incorporates the first temporal derivatives of the motion parameters, while the 24-parameter model (often referred to as the Friston model) includes the original parameters, their derivatives, their squared values, and the squared derivatives [38] [41]. Some implementations extend this further to 36 parameters by incorporating these regressors at previous time points [41].

Table 1: Standard Motion Regression Models

Model Name	Parameters Included	Key Characteristics
6-parameter	3 translations + 3 rotations	Basic linear motion correction
12-parameter	6 parameters + their temporal derivatives	Accounts for instantaneous velocity
24-parameter (Friston)	6 parameters + derivatives + squares + squared derivatives	Attempts to capture some nonlinear relationships
36-parameter	24 parameters + lagged versions	Includes temporal shifts to model persistent effects

Motion Simulation (MotSim) and PCA Approaches

The Motion Simulation (MotSim) approach represents a significant advancement in modeling motion-related signal changes by simulating the actual signal changes that the estimated motion would produce [38] [40]. This method involves:

• Selecting a single brain volume from the acquired data
• Creating a 4D dataset by rotating and translating this volume according to the inverse of the estimated motion parameters
• Motion-correcting this simulated dataset with rigid body volume registration
• Performing temporal principal components analysis (PCA) on the resultant time series

The MotSim method generates three distinct sets of nuisance regressors: the "forward" model (PCA of the MotSim dataset), the "backward" model (PCA of the registered MotSim dataset), and the "both" model (PCA of both datasets concatenated) [38]. Research has demonstrated that these MotSim regressors account for a significantly greater fraction of variance, result in higher temporal signal-to-noise ratio, and lead to functional connectivity estimates that are less correlated with motion compared to traditional parameter-based approaches [38] [40].

Convolutional Neural Network Approaches

Recent research has explored using convolutional neural networks (CNNs) to derive optimal motion regressors from the basic motion parameters [41]. These networks typically consist of temporal convolutional layers that can non-parametrically model the prolonged effects of head motion, addressing the limitation of traditional approaches that use fixed mathematical transformations of the motion parameters [41].

The CNN approach is optimized using time series from white matter and cerebrospinal fluid, which share similar motion-related artifacts with gray matter but contain minimal neural signals, thus avoiding the removal of neurally-related BOLD signals during the nuisance regression process [41]. Studies have demonstrated that CNN-derived regressors can more effectively reduce motion-related artifacts compared to the same number of traditional regressors [41].

Integrated Correction Frameworks

The most effective motion correction strategies typically combine multiple approaches in an integrated framework. A common effective combination includes:

• Volume realignment using rigid-body transformations
• Nuisance regression with expanded motion parameters or MotSim regressors
• Volume censoring ("scrubbing") of high-motion time points
• Incorporation of physiological regressors from white matter and CSF
• Optional global signal regression (though this remains controversial due to potential introduction of artificial anti-correlations) [3] [41]

The combination of scrubbing and motion regression has been shown to provide the greatest reduction in motion-related artifacts, though there is an important tradeoff between data quality and the number of remaining time points, particularly when studying populations with higher motion levels [3] [41].

Quantitative Comparison of Motion Correction Techniques

Table 2: Performance Comparison of Motion Correction Techniques

Technique	Variance Explained	tSNR Improvement	Residual Motion-FC Correlation	Computational Demand
12-parameter model	Baseline	Baseline	High	Low
24-parameter model	Moderate improvement	Moderate improvement	Moderate	Low
MotSim PCA models	Significantly greater than 12-parameter [38]	Higher than standard approaches [38]	Significantly reduced [38] [40]	Medium-High
CNN-derived regressors	Superior to same number of traditional regressors [41]	Not reported	Effectively reduces artifacts [41]	High (initial training)
Volume censoring (FD < 0.2 mm)	N/A	N/A	Reduces overestimation artifacts [39]	Low (but reduces data points)

Empirical studies have systematically compared these advanced techniques. For instance, MotSim approaches have been shown to account for a significantly greater fraction of variance compared to the standard 12-parameter model [38] [40]. In one study, the MotSim "Both" model (incorporating both forward and backward components) demonstrated particularly strong performance in reducing the correlation between motion and functional connectivity estimates [38].

The effectiveness of censoring approaches depends heavily on the selected threshold. Research indicates that framewise displacement (FD) thresholds in the range of 0.15-0.2 mm represent a reasonable balance between removing motion-contaminated volumes and retaining sufficient data for analysis, with FD > 0.5 mm indicating clearly problematic volumes that should typically be censored [3] [39].

Implementation Protocols and Methodologies

Motion Simulation (MotSim) Protocol

The implementation of the MotSim approach follows these key steps [38]:

• Volume Selection: Choose a representative EPI volume (typically after removing initial transient volumes and performing slice-timing correction)
• Motion Simulation: Apply the inverse of the estimated motion parameters to this base volume using interpolation (linear interpolation is often sufficient, with higher-order methods providing minimal additional benefit)
• Motion Correction: Perform rigid-body volume registration on the simulated motion dataset
• PCA Decomposition: Conduct temporal principal components analysis on voxels from the entire brain, including edge voxels dilated 2 voxels outward from the brain mask
• Regressor Selection: Retain the first N principal components (typically 12 to match the standard approach) as nuisance regressors
• Nuisance Regression: Include these regressors in the general linear model alongside other nuisance signals

Volume Censoring Protocol

Effective volume censoring involves these critical steps [3] [39]:

• Calculate Framewise Displacement: Compute FD for each volume based on the motion parameters
• Identify Contaminated Volumes: Flag volumes exceeding threshold (typically FD > 0.2-0.5 mm, depending on study requirements)
• Extended Censoring: Also censor volumes following high-motion volumes (research shows artifacts can persist >10 seconds after motion ceases)
• Spike Regression: Create regressors with single non-zero values at identified time points for inclusion in nuisance regression
• Quality Control: Ensure sufficient volumes remain after censoring (typically >50% of original data)

Advanced Regression Implementation

Table 3: Essential Research Tools for Advanced Motion Correction

Tool/Resource	Function/Purpose	Implementation Examples
Framewise Displacement (FD)	Quantifies volume-to-volume head motion	Compute from realignment parameters; used for censoring [3] [39]
MotSim Algorithm	Generates motion-related signal changes via simulation	Implement in AFNI, SPM, or FSL; requires base volume and motion parameters [38] [40]
Temporal PCA	Reduces dimensionality of motion-simulated data	Apply to MotSim datasets; select first 12 components [38]
CNN Motion Modeling	Non-parametrically models prolonged motion effects	Train on WM/CSF signals to avoid neural signal removal [41]
Volume Censoring (Scrubbing)	Removes motion-contaminated time points	Identify volumes with FD > 0.2-0.5 mm; create spike regressors [3] [39]
Global Signal Regression	Removes whole-brain signal fluctuations	Controversial: reduces motion artifacts but may introduce bias [3] [41]
Tissue-Based Regressors	Accounts for physiological noise	Extract mean signals from WM and CSF masks [41]

Advanced regression techniques that incorporate nonlinear and derivative motion parameters represent a significant improvement over traditional motion correction approaches in fMRI research. By better modeling the complex, nonlinear relationship between head movement and signal changes, methods like MotSim and CNN-derived regressors more effectively address the pervasive challenge of motion artifacts [38] [40] [41].

The integration of these advanced regression techniques with volume censoring and physiological noise modeling provides the most robust protection against motion-related biases, particularly important when studying populations with different motion characteristics or when investigating traits correlated with motion [3] [39]. Future developments will likely focus on real-time correction methods, improved dynamic modeling of motion effects, and standardized implementation across major processing platforms.

As the field moves toward increasingly large-scale datasets and more diverse participant populations, the rigorous application of these advanced motion correction techniques will be essential for ensuring the validity and reproducibility of fMRI research findings across both resting-state and task-based paradigms.

Functional magnetic resonance imaging (fMRI) data are notoriously susceptible to contamination from artifacts arising from a myriad of sources, with subject head motion representing one of the most significant challenges. These artifacts cause deviations from common distributional assumptions, introduce spatial and temporal outliers, and reduce the signal-to-noise ratio of the data—all of which can have negative consequences for the accuracy and power of downstream statistical analysis [42] [43]. Head motion is the largest source of artifact in structural and functional MRI signals, and its impact is particularly problematic in developmental or clinical populations where motion is often correlated with the independent variable of interest (e.g., age, diagnosis) [44] [39]. Even with highly compliant participants, involuntary sub-millimeter head movements systematically alter fMRI data, with non-linear characteristics of MRI physics making complete removal of motion artifact during post-processing exceptionally difficult [39].

The manifestation of motion artifacts differs considerably between resting-state and task-based fMRI paradigms. For resting-state functional connectivity (FC), which relies on correlational patterns of spontaneous brain activity, motion introduces systematic bias that persists even after extensive denoising [39]. This artifact manifests spatially as decreased long-distance connectivity and increased short-range connectivity, most notably in the default mode network [39]. In contrast, task fMRI measures are generally less sensitive to these effects, as they rely on aggregations of trial-related activity and subtractions between conditions rather than the inter-regional correlations that can be artifactually inflated by motion [45]. This fundamental difference in how motion contaminates these two types of fMRI data has profound implications for the selection and application of censoring strategies.

Core Methodological Approaches

Scrubbing Techniques

Motion scrubbing, one of the two primary censoring approaches, relies on subject head motion-derived measures, typically framewise displacement (FD), to identify and remove contaminated volumes [42] [39]. This method requires researchers to select a threshold (e.g., FD < 0.2 mm) beyond which volumes are flagged as outliers [39]. While this approach can effectively reduce motion-related artifacts, it suffers from several drawbacks: the need to choose an often-arbitrary threshold, lack of generalizability to multiband acquisitions, and high rates of censoring that can lead to exclusion of entire subjects from analysis [42].

Table 1: Comparison of Primary Scrubbing Techniques

Method	Basis	Advantages	Limitations
Motion Scrubbing	Head motion-derived measures (e.g., FD)	Direct measurement of physical head movement; Widely implemented	Arbitrary threshold selection; High data loss; Excludes entire subjects
DVARS	Temporal derivative of time courses & root mean square variance over voxels	Data-driven; Responsive to actual signal changes	May flag neural activations as artifacts; Limited spatial specificity
Projection Scrubbing	Statistical outlier detection & dimension reduction (ICA)	Statistically principled; Minimizes unnecessary censoring; Improved data retention	Computational complexity; Requires specialized implementation

Data-driven scrubbing methods like DVARS are based on observed noise in the processed fMRI timeseries and may avoid some issues inherent to motion scrubbing [42] [43]. DVARS calculates the temporal derivative of time courses and the root mean square variance over voxels, flagging volumes where signal changes exceed expected levels [43]. A more recent innovation, "projection scrubbing," represents a novel data-driven approach based on a statistical outlier detection framework and strategic dimension reduction, including independent component analysis (ICA), to isolate artifactual variation [42]. This method operates by identifying abnormal patterns in the data itself rather than relying solely on head motion parameters, resulting in more targeted removal of truly contaminated volumes [42].

Spike Regression

Spike regression represents a distinct technical approach to handling motion-contaminated volumes. Rather than completely removing flagged timepoints from the analysis, spike regression incorporates additional regressors into the general linear model to account for the signal spikes associated with motion artifacts [44]. This method effectively "de-weights" the influence of contaminated volumes while preserving the overall temporal structure of the data, thereby maintaining the continuity of the time series and conserving temporal degrees of freedom that would be lost with traditional scrubbing [44].

The implementation of spike regression typically involves identifying outlier volumes based on motion parameters (similar to motion scrubbing), then creating dummy regressors for each flagged timepoint that are included in the connectivity or activation model. This approach can be particularly valuable in cases where scrubbing would result in excessive data loss or when analyzing task-based fMRI where maintaining precise timing is crucial for detecting hemodynamic responses. However, spike regression may be less effective than scrubbing for severe motion artifacts and introduces additional covariates that must be accounted for in statistical modeling [44].

Quantitative Comparisons of Method Efficacy

The comparative efficacy of various censoring strategies has been systematically evaluated across multiple large-scale studies, providing empirical evidence to guide methodological decisions.

Table 2: Efficacy Metrics Across Censoring Methods

Method	Residual Motion-FC Relationship	Distance-Dependent Artifact	Network Identifiability	Data Retention
No Censoring	Severe [39]	Pronounced [44] [39]	Limited [44]	100%
Motion Scrubbing (FD < 0.2 mm)	Substantially reduced [39]	Mitigated [39]	Improved [42]	Low (High exclusion) [42]
Data-Driven Scrubbing	Moderate reduction [42]	Moderately mitigated [42]	Best performance [42]	High [42]
Spike Regression	Moderate reduction [44]	Moderately mitigated [44]	Improved [44]	High (preserves degrees of freedom) [44]

Benchmarking studies have revealed that stringent motion scrubbing (e.g., FD < 0.2 mm) effectively reduces motion-related artifacts but at the cost of significant data loss. In the Adolescent Brain Cognitive Development (ABCD) Study, this approach reduced significant motion overestimation from 42% (19/45) to just 2% (1/45) of traits [39]. However, it did not decrease the number of traits with significant motion underestimation scores, highlighting the complex relationship between censoring and trait-specific motion effects [39].

Data-driven methods demonstrate a superior balance between noise reduction and data retention. Projection scrubbing, in particular, has been shown to exclude only a fraction of the number of volumes or entire sessions compared to motion scrubbing while still improving typical benchmarks such as the validity, reliability, and identifiability of functional connectivity [42]. This ability to improve data retention without negatively impacting downstream analysis quality has major implications for sample sizes in population neuroscience research, potentially preserving statistical power while controlling for motion artifacts [42].

Differential Applications in Resting-State vs. Task fMRI

Resting-State fMRI Considerations

Resting-state functional connectivity is especially vulnerable to motion artifact because the timing of the underlying neural processes is unknown [39]. The distance-dependent effects of motion on resting-state FC are well-established, manifesting as decreased long-distance connectivity and increased short-range connectivity [39]. This systematic bias poses particular challenges for studies of populations with inherently higher motion levels, such as children, older adults, and individuals with psychiatric or neurological disorders [39].

The lingering impact of motion even after denoising was starkly demonstrated in analyses of the ABCD dataset, where after extensive denoising (ABCD-BIDS preprocessing), 23% of signal variance was still explained by head motion [39]. Furthermore, the motion-FC effect matrix showed a strong negative correlation (Spearman ρ = -0.58) with the average FC matrix, indicating that connection strength tended to be weaker in participants who moved more [39]. This residual relationship highlights the critical importance of effective censoring strategies for resting-state fMRI, particularly when studying traits that correlate with motion propensity.

Task fMRI Considerations

Task-based fMRI presents a different set of challenges and opportunities for motion censoring. Because task analyses typically rely on aggregations of trial-related activity and subtractions between conditions, they are generally less sensitive to motion effects than resting-state correlational approaches [45]. However, task-based functional connectivity analyses are becoming increasingly common and face similar motion vulnerability as resting-state FC [45].

Notably, task fMRI paradigms may capture more behaviorally relevant information than resting-state functional connectivity, with FC patterns derived from fMRI tasks outperforming resting-state FC at predicting individual differences in cognitive task performance [46]. This advantage appears to be driven largely by the FC patterns associated with the task design itself [46]. The optimal pairing of specific task paradigms with behavioral outcomes of interest suggests that tailored task selection can enhance sensitivity to individual differences while potentially mitigating some motion-related concerns [11].

Experimental Protocols and Implementation

Procedural Workflow for Censoring Implementation

The implementation of effective censoring strategies follows a structured workflow that integrates with broader fMRI preprocessing pipelines. The following diagram illustrates the key decision points in selecting and applying appropriate censoring methods:

Successful implementation of censoring strategies requires both computational tools and methodological considerations. The following table outlines key components of the researcher's toolkit for addressing motion artifacts:

Table 3: Essential Research Tools for Motion Censoring

Tool/Category	Specific Examples	Function/Purpose
Motion Quantification	Framewise Displacement (FD), DVARS	Quantify head motion at each timepoint to identify contaminated volumes
Data Processing Software	fMRIPrep, ABCD-BIDS, AFNI, FSL, SPM	Implement standardized preprocessing pipelines including motion correction and censoring
Censoring Algorithms	Projection Scrubbing, ICA-AROMA, CompCor	Data-driven approaches for identifying artifactual components and volumes
Quality Control Metrics	SHAMAN, Distance-Dependence	Evaluate residual motion effects after censoring and assess trait-specific motion impacts
Reference Datasets	ABCD Study, HCP, UK Biobank	Provide large-scale benchmarks for method validation and comparison

The SHAMAN framework (Split Half Analysis of Motion Associated Networks) represents a particularly advanced tool for assigning a motion impact score to specific trait-FC relationships, distinguishing between motion causing overestimation or underestimation of trait-FC effects [39]. This method capitalizes on the observation that traits are stable over the timescale of an MRI scan whereas motion is a state that varies from second to second, enabling more precise quantification of motion's impact on specific research questions [39].

Impact on Brain-Behavior Associations and Recommendations

Motion censoring strategies have profound implications for the validity and reproducibility of brain-behavior associations. In large-scale brain-wide association studies (BWAS), failure to adequately address motion artifacts can lead to both false positive and false negative findings, particularly for traits that correlate with motion propensity [39]. The natural tension in censoring lies between the need to remove 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 [39].

Evidence-based recommendations for selecting censoring strategies must consider several factors:

• For resting-state fMRI with populations prone to motion (e.g., children, clinical groups), data-driven scrubbing methods like projection scrubbing provide an optimal balance of artifact removal and data retention [42].

• For task-based fMRI where maintaining trial timing is crucial, spike regression may be preferable to preserve temporal structure while modeling motion effects [44].

• When studying motion-correlated traits, implement trait-specific motion impact assessments like SHAMAN to quantify and account for residual motion effects [39].

• Always report the specific censoring thresholds and methods used, along with the amount of data retained, to enable evaluation of potential biases and facilitate replication [42] [39].

The integration of robust censoring strategies with appropriate experimental design, including task selection tailored to target behaviors, represents the most promising path forward for maximizing the validity and reproducibility of fMRI studies of brain-behavior relationships [46] [11].

Functional magnetic resonance imaging (fMRI) has become an indispensable tool for studying brain function in both healthy and pathological states. However, the fMRI signal is easily contaminated by artifacts arising from head movement during data collection, with motion being the largest source of artifact in structural and functional MRI signals [39]. These motion artifacts persist even after common corrective processing procedures have been applied and exhibit complex spatiotemporal properties that can lead to substantial errors in functional connectivity estimates [2]. The small amplitude of blood oxygenation level-dependent (BOLD) signals—typically a few percent or less—ensures that millimeter-scale head motions may be problematic, introducing systematic biases that affect downstream analyses [2] [39].

The challenge of motion artifacts manifests differently across experimental paradigms. In resting-state fMRI (rs-fMRI), where participants are not performing any specific task, head motion can be exacerbated when individuals are at rest [2]. Rs-fMRI is especially vulnerable to motion artifact because the timing of the underlying neural processes is unknown [39]. Characteristic distance-dependent and orientation-dependent errors have been reported in correlation-based estimates, with decreased long-distance connectivity and increased local connectivity [2]. For task-based fMRI, the problems are equally severe but manifest differently. Task-related motion can be temporally correlated with task performance, making it particularly difficult to distinguish motion artifacts from true brain activity [17]. In experiments where conditions might cause slight head movements (such as motor tasks or speech), these movements are often highly correlated with the experimental design, creating confounding effects that are challenging to address with conventional correction methods [17].

This whitepaper explores two advanced computational approaches for mitigating motion artifacts: structured low-rank matrix completion and Independent Component Analysis (ICA)-based methods. These approaches represent significant advancements over traditional correction strategies, offering more sophisticated ways to distinguish true functional networks from motion-related noise in both resting-state and task-based fMRI research.

Theoretical Foundations of Motion Artifacts

Characteristics and Mechanisms of Motion Artifacts

Head motion during fMRI scanning results in misalignment of one volume to the next, introducing measurement inaccuracies as imaging voxels do not represent identical brain regions over time [47]. While primary effects of participant head motion are often corrected by realigning volumes using linear alignment algorithms, secondary effects related to partial voluming, interpolation effects, magnetic field inhomogeneities, intra-volume motion, and spin-history effects persist despite realignment [47]. Through-plane head motion during 2D planar image acquisition is especially problematic as it can cause regions of adjacent slices to be subsequently excited before they have fully recovered—a phenomenon known as spin history artifact [12].

Motion artifacts display distinct spatial and temporal characteristics. Spatially, they cause a characteristic pattern of decreased long-distance connectivity and increased short-range connectivity, most notably in the default mode network [39]. Temporally, motion artifacts do not display band-limited frequency content, making frequency filtering (typically applied in rs-fMRI to isolate the ~0.01-0.1 Hz range) ineffective for motion correction and potentially problematic as it can smear motion contamination across the entire dataset [2].

Differential Impact on Resting-State vs. Task-Based fMRI

The impact of motion artifacts differs significantly between resting-state and task-based fMRI, necessitating specialized correction approaches for each paradigm:

Table 1: Motion Artifact Characteristics in Resting-State vs. Task-Based fMRI

Characteristic	Resting-State fMRI	Task-Based fMRI
Temporal Correlation	Uncorrelated with underlying neural processes	Potentially correlated with task design and condition transitions
Primary Manifestation	Distance-dependent biases in functional connectivity; inflated short-range connections, weakened long-range connections	Confounding of task-related activation estimates; both false positives and false negatives
Correction Challenges	Unknown timing of neural processes makes separation difficult	Risk of removing true neural signal correlated with task conditions
Population Concerns	Problematic in developmental, aging, and clinical populations with higher motion	Particularly challenging for motor, speech, or other tasks that induce movement

In resting-state fMRI, motion introduces systematic bias that is not completely removed by standard denoising algorithms [39]. Even after extensive processing, residual motion artifacts can lead to spurious brain-behavior associations, particularly concerning for studies of traits associated with motion (e.g., psychiatric disorders) [39]. For task-based fMRI, the central challenge is that stimulus-related motion cannot be separated from stimulus-related signal variations of interest in regression models, potentially decreasing sensitivity to functional activation [47].

Structured Low-Rank Matrix Completion Approaches

Theoretical Framework and Algorithmic Principles

Structured low-rank matrix completion represents a principled approach to reduce motion artifacts from rs-fMRI data by exploiting the inherent structure of neural time series [8]. This method addresses the common practice of "censoring" or "scrubbing," where volumes with high frame-by-frame motion are excised from processed fMRI data prior to functional connectivity analysis. While censoring mitigates spurious effects in correlation, it creates discontinuities in the time series and may lead to significant data loss, resulting in unreliable correlation estimates [8].

The fundamental innovation of structured low-rank methods lies in modeling the temporal signal at every voxel location as governed by a linear recurrence relation (LRR), which expresses the voxel intensity at the current time point as a linear combination of its intensities from the past [8]. This compact representation leads to the construction of a low-rank Hankel matrix, whose entries correspond to the signal samples. The Hankel matrices from different voxels are stacked vertically to form a large structured matrix, which also exhibits low-rank structure that can be exploited for recovery [8].

The forward model for motion-compensated recovery can be formulated as:

Yi = Mi(Si(X)) + ηi, i = 1,2,…,nv

Where Yi denotes the ith unprocessed volume, X corresponds to the reconstructed fMRI time series with higher temporal resolution, Si is a linear sampling operator, Mi is a motion operator encoding motion parameters, and ηi represents the modeling error [8]. Successful recovery of X from the Yis ensures both motion compensation and slice time correction, addressing multiple sources of artifact simultaneously.

Implementation and Computational Considerations

Implementing the structured low-rank matrix completion algorithm involves solving a large-scale system of equations and forming a large structured matrix, which presents challenges in terms of memory demand and computational complexity [8]. To address these issues, a variable splitting strategy has been developed that decouples the original problem into two simpler sub-problems [8]. This approach eliminates the need to evaluate the large structured matrix directly and enables significant speed-up through efficient parallel computations in each sub-problem.

The algorithm proceeds through several key stages:

• Motion-Corrupted Volume Identification: Detection of volumes with elevated motion using framewise displacement (FD) or similar metrics
• Matrix Formation: Construction of the structured low-rank matrix from the remaining unprocessed fMRI data
• Matrix Completion: Recovery of missing entries through optimization techniques that enforce the low-rank prior
• Signal Reconstruction: Generation of the motion-compensated time series with slice-time correction

Experimental Validation and Performance

The structured low-rank matrix completion approach has been validated through simulations, data acquired under different motion conditions, and datasets from the Adolescent Brain Cognitive Development (ABCD) study [8]. Functional connectivity analysis demonstrated that the proposed reconstruction resulted in connectivity matrices with lower errors in pair-wise correlation than both non-censored and censored time series based on a standard processing pipeline [8]. Additionally, seed-based correlation analyses showed improved delineation of the default mode network, confirming the method's effectiveness in reducing the adverse effects of motion in fMRI analysis [8].

ICA-Based Motion Correction Methods

Fundamental ICA Framework and Methodological Variations

Independent Component Analysis (ICA) has emerged as a powerful data-driven approach for identifying and removing motion artifacts from fMRI data. Applied to fMRI data, ICA decomposes the data into a set of spatially independent component maps (ICs) and associated time courses [47]. The resulting components represent brain activity, structured noise (e.g., motion-related, physiological, or scanner-induced noise), or other artifacts. Components identified as representing noise can be regressed out from the data, effectively denoising the fMRI signals [47].

Several ICA-based approaches have been developed, each with distinctive characteristics:

• ICA-AROMA (ICA-based Automatic Removal Of Motion Artifacts): Utilizes a small (n=4) but robust set of theoretically motivated temporal and spatial features to automatically identify motion components without requiring classifier re-training [47]. This method preserves the data's autocorrelation structure and largely maintains temporal degrees of freedom.

• CICADA (Comprehensive Independent Component Analysis Denoising Assistant): A novel denoising method that uses manual classification guidelines to automatically, comprehensively, and accurately capture most common sources of fMRI noise [48]. CICADA performs nearly as well as manual IC classification—the current gold-standard—while automating the process.

• Traditional ICA with Manual Classification: Requires expert reviewers to label components as signal or noise, which is time-consuming and requires extensive training but can achieve high accuracy [48].

Component Classification and Feature Extraction

The critical step in ICA-based denoising is the accurate classification of components as neural signal or noise. ICA-AROMA employs four specific features to identify motion-related components: (1) high-frequency content, (2) correlation with motion parameters, (3) edge fraction, and (4) cerebrospinal fluid fraction [47]. This feature set enables robust identification without needing dataset-specific classifier training.

CICADA leverages a more comprehensive set of criteria derived from manual classification guidelines, achieving 97.9% mean overall accuracy in IC classification—significantly more accurate than FIX (92.9%) and ICA-AROMA (83.8%) across multiple datasets [48]. The method also greatly eases implementation of manual ICA denoising by decreasing the number of ICs a user must inspect by an average of 75% [48].

For task-based fMRI, specialized criteria have been developed that utilize task-related changes in signal intensity and variance (heteroscedasticity) to identify motion components that are temporally correlated with the experimental design [17]. This approach is particularly valuable for separating motion-induced signal changes from true neural activation in paradigms where motion is condition-dependent.

Performance Across Experimental Paradigms

ICA-based methods have demonstrated substantial efficacy in both resting-state and task-based fMRI. In resting-state data, ICA-AROMA removed motion-related spurious noise more effectively than regression using 24 motion parameters or spike regression [47]. For task-based fMRI, ICA-AROMA increased sensitivity to group-level activation, demonstrating its value for experimental paradigms where motion may confound activation detection [47].

Comparative evaluations across multiple denoising pipelines reveal that pipelines combining ICA-FIX and global signal regression (GSR) demonstrate a reasonable trade-off between motion reduction and behavioral prediction performance [49]. However, inter-pipeline variations in predictive performance are modest, suggesting that optimal pipeline selection may depend on specific research goals and data characteristics.

Comparative Analysis and Integration

Methodological Trade-offs and Complementary Strengths

Structured low-rank matrix completion and ICA-based approaches offer distinct advantages and limitations for motion artifact correction. The table below summarizes their key characteristics:

Table 2: Comparison of Motion Correction Methodologies

Feature	Structured Low-Rank Matrix Completion	ICA-Based Methods
Theoretical Basis	Linear recurrence relations in time series; low-rank matrix structure	Spatial independence of neural signals and motion artifacts
Data Requirements	Requires identification of motion-corrupted volumes	Works on complete time series without volume removal
Computational Demand	High (large-scale matrix operations) but optimized via variable splitting	Moderate (matrix decomposition and classification)
Primary Output	Motion-compensated time series with slice-time correction	Denoised time series with motion components removed
TDOF Preservation	Complete preservation	Mostly preserved (only noise components removed)
Applicability	Resting-state fMRI [8]	Both resting-state and task-based fMRI [47]
Automation Level	Fully automatic once motion-corrupted volumes identified	Varies (manual, automated, or hybrid classification)

Table 3: Research Reagent Solutions for Motion Artifact Correction

Resource	Type	Function	Implementation Considerations
ABCD-BIDS Pipeline	Software Pipeline	Default denoising for large-scale studies; includes global signal regression, respiratory filtering, spectral filtering, despiking, and motion timeseries regression [39]	Achieves 69% relative reduction in motion-related variance compared to minimal processing alone [39]
ICA-AROMA	Algorithm/Software	Automatic removal of motion artifacts using ICA without classifier re-training [47]	Preserves temporal degrees of freedom; applicable to both resting-state and task-based fMRI
CICADA	Algorithm/Software	Automated ICA denoising using manual classification guidelines [48]	97.9% mean accuracy in component classification; reduces manual inspection workload by 75%
SHAMAN	Analytical Framework	Quantifies motion impact on specific trait-FC relationships using split-half analysis [39]	Distinguishes between motion causing overestimation or underestimation of trait-FC effects
aCompCor	Methodological Approach	Estimates nuisance signals from white matter and CSF using principal components rather than mean signals [12]	More effectively attenuates motion artifacts than mean signal regression; enhances connectivity specificity
Structured Low-Rank Completion	Algorithmic Framework	Recovers missing entries from censoring using structured matrix priors [8]	Simultaneously provides motion compensation and slice-time correction

Integrated Approaches and Future Directions

The most effective motion correction strategies often combine multiple approaches. For instance, the ABCD-BIDS pipeline integrates global signal regression, respiratory filtering, motion timeseries regression, and despiking/interpolation of high-motion frames [39]. Similarly, comprehensive evaluation studies suggest that pipelines combining ICA-FIX and GSR demonstrate a reasonable trade-off between motion reduction and behavioral prediction performance [49].

Emerging methods like SHAMAN (Split Half Analysis of Motion Associated Networks) address the critical need for quantifying residual motion effects on specific trait-FC relationships, helping researchers determine whether their findings are impacted by motion to avoid reporting false positive results [39]. This is particularly important for studies of populations with motion-correlated traits (e.g., psychiatric disorders).

Future developments will likely focus on real-time correction methods that can be combined with retrospective approaches, potentially offering better correction and increased fMRI signal sensitivity [2]. Additionally, methods that explicitly account for the differential impact of motion on resting-state versus task-based fMRI will enhance the validity of findings across experimental paradigms.

Experimental Protocols and Validation Metrics

Validation Frameworks and Benchmarking

Robust validation of motion correction methods requires multiple approaches assessing different aspects of performance. The following protocols represent current best practices:

• Motion-FC Effect Quantification: Compute the correlation between each participant's framewise displacement (FD) and their functional connectivity to generate a motion-FC effect matrix with units of change in FC per mm FD [39]. The motion-FC effect matrix typically shows a strong negative correlation with the average FC matrix (Spearman ρ ≈ -0.58), indicating that connection strength tends to be weaker in participants who move more [39].

• Split-Half Analysis (SHAMAN): Capitalize on the relative stability of traits over time by measuring differences in correlation structure between split high- and low-motion halves of each participant's fMRI timeseries [39]. A significant difference indicates that state-dependent differences in motion impact the trait's connectivity.

• Distance-Dependent Correlation Analysis: Examine the relationship between motion and connectivity as a function of the distance between brain regions, as motion typically inflates short-range connectivity while weakening long-range connectivity [2] [39].

• Network Specificity Assessment: Evaluate the specificity of functional connectivity metrics using the known anatomy of established networks like the default mode and motor control networks [12].

Performance Benchmarks Across Methods

Recent large-scale evaluations provide compelling evidence of the relative performance of different denoising strategies:

• After minimal processing (motion-correction by frame realignment only), 73% of signal variance is explained by head motion [39]
• After denoising using ABCD-BIDS (including respiratory filtering, motion timeseries regression, and despiking), 23% of signal variance is explained by head motion, representing a 69% relative reduction compared to minimal processing alone [39]
• Censoring at framewise displacement (FD) < 0.2 mm reduces significant motion overestimation from 42% to 2% of traits but does not decrease the number of traits with significant motion underestimation scores [39]
• CICADA achieves 97.9% mean overall accuracy in IC classification, significantly outperforming FIX (92.9%) and ICA-AROMA (83.8%) across multiple datasets [48]

These benchmarks highlight that while modern denoising methods substantially reduce motion artifacts, residual effects persist and may still impact findings, particularly for motion-correlated traits. Comprehensive motion correction therefore requires multiple complementary approaches tailored to specific research questions and experimental designs.

Functional magnetic resonance imaging (fMRI) is an indispensable tool for studying brain function, but its signals are contaminated by physiological noise from cardiac pulsations and respiratory cycles. These fluctuations can mimic or mask true neural activity, compromising data interpretation [50] [9]. Physiological noise is particularly problematic as the field moves to higher magnetic field strengths, where it becomes a dominant confound limiting fMRI sensitivity [50]. In resting-state fMRI (rs-fMRI), even sub-millimeter head motions introduce complex artifacts that distort functional connectivity estimates, reducing long-range connections while strengthening short-range ones [51] [9]. Effective physiological noise correction is therefore essential for accurate interpretation of both task-based and resting-state fMRI data.

Within this context, anatomical Component Based Noise Correction (aCompCor) has emerged as a powerful data-driven approach that eliminates the need for external physiological monitoring [50] [51]. This technical guide details the implementation, optimization, and application of aCompCor and related tissue-based regressor methods, with particular emphasis on their efficacy for addressing motion-related artifacts in resting-state versus task fMRI research designs.

Theoretical Foundations of aCompCor

Core Algorithm and Principles

The aCompCor method is grounded in a fundamental assumption: signals from specific noise regions-of-interest (ROIs) can accurately model physiological fluctuations throughout the brain. These noise ROIs comprise areas such as white matter (WM) and cerebrospinal fluid (CSF) where time series data are unlikely to contain neurally-modulated BOLD signals [50]. Instead, fluctuations in these regions primarily reflect physiological noise of non-neural origin.

The algorithm employs principal component analysis (PCA) to compactly characterize time-series data from noise ROIs. Significant principal components are then incorporated as nuisance parameters within general linear models (GLMs) for BOLD and perfusion-based fMRI data [50]. Mathematically, for BOLD data, the GLM with aCompCor takes the form:

b = Xh + Sd + Pc + n

Where b represents measured BOLD data, Xh models the stimulus response, Sd represents nuisance parameters, and Pc represents physiological noise terms with P containing the noise regressors from aCompCor and c representing their weights [50].

Noise Region of Interest Selection Methods

aCompCor implementations utilize two primary approaches for defining noise ROIs:

• Anatomically-defined ROIs: This method uses high-resolution anatomical data to identify voxels consisting primarily of white matter or cerebrospinal fluid [50] [51]. This approach benefits from clear biological justification but requires quality anatomical scans and accurate tissue segmentation.

• Temporal standard deviation (tSTD) ROIs: This data-driven approach identifies voxels with the highest temporal standard deviation in the time-series data, based on the finding that areas of high tSTD often correspond to ventricles, edge regions, and vessels affected by physiological noise [50]. This method offers the advantage of not requiring a separate anatomical scan.

Implementation and Workflows

Core aCompCor Procedure

The standard aCompCor implementation follows a systematic workflow to estimate and remove physiological noise components from fMRI data.

Figure 1: aCompCor Computational Workflow

For functional studies with evoked responses, an additional processing step is recommended to minimize inclusion of stimulus-related fluctuations in noise components. This involves performing a preliminary GLM analysis using the appropriate design matrix to exclude voxels from the noise ROI that show significant task-related activation (typically using a liberal threshold of p < 0.2) [50].

Practical Implementation in Software Pipelines

Widely-used preprocessing tools like fMRIPrep have incorporated aCompCor into their standardized outputs. The toolbox generates multiple noise components derived from different anatomical masks:

• WM-based components: Principal components from white matter mask
• CSF-based components: Principal components from cerebrospinal fluid mask
• Combined components: Principal components from merged WM and CSF masks [52]

Modern implementations output a large number of components (potentially equaling the number of TRs), allowing researchers flexibility in selecting the optimal number for their specific analysis rather than being limited to an arbitrary fixed number [53].

Experimental Validation and Protocol Details

Key Experimental Evidence

The efficacy of aCompCor has been demonstrated across multiple validation studies employing varied experimental protocols and comparison metrics.

Table 1: Key Experimental Studies Validating aCompCor Performance

Study	Subjects & Design	Comparison Methods	Primary Outcomes
Behzadi et al. (2007) [50]	10 healthy adults; Periodic and block visual stimuli; Resting-state scans	No correction; RETROICOR	Significant reduction in temporal standard deviation; Increased activated voxels in functional data; Superior to RETROICOR for BOLD
Muschelli et al. (2014) [51]	130 typically developing children; Resting-state fMRI	Mean tissue signal regression; Scrubbing	Better motion artifact reduction; Improved connectivity specificity; Eliminated need for scrubbing when used
fMRIPrep Implementation [52]	Standardized pipeline; Various datasets	Multiple denoising strategies	Flexible component output; Integration with anatomical processing; Automated confound generation

Detailed Imaging Parameters

To ensure reproducibility, documented validation studies have employed specific acquisition protocols:

Behzadi et al. protocol: Data were collected on a GE Signa Excite 3 Tesla system with a PICORE QUIPPS II arterial spin labeling sequence for perfusion imaging (TR=2s, TI1/TI2=600/1500ms) and dual-echo spiral readout (TE1/TE2=9.1/30ms, FOV=24cm, 64×64 matrix) [50].

Muschelli et al. protocol: Imaging data were collected using a Philips Achieva 3T scanner with a 2D-SENSE EPI sequence (TR/TE=2500/30ms, flip angle=70°, 47 contiguous 3-mm slices, in-plane resolution 3.05×3.15mm) [51].

Performance Comparison with Alternative Methods

Relative Effectiveness for Motion Correction

aCompCor demonstrates distinct advantages over alternative physiological noise correction approaches, particularly for mitigating motion-related artifacts.

Table 2: Method Comparison for Physiological Noise Correction

Method	Requires External Monitoring	Motion Reduction Efficacy	Implementation Complexity	Key Limitations
aCompCor	No	High (superior to mean signal) [51]	Moderate	Component selection criteria vary
RETROICOR	Yes (ECG, respiratory belt) [54]	Moderate to High	High	Technical challenges with external recordings [55]
Mean Signal Regression	No	Low to Moderate	Low	Insufficient for motion artifacts [51]
HRAN	No	High for fast fMRI [54]	High	Optimized for TR < 500ms
SCRUBBING	No	Moderate (when used alone) [51]	Low	Data loss from volume removal

When compared specifically to mean tissue-based signal regression, aCompCor more effectively removes motion artifacts because it can capture multiple, spatially distinct nuisance signals that might cancel each other out when averaged across voxels [51]. A particular strength is its ability to account for delayed and non-linear motion effects without requiring explicit modeling of these complex relationships.

Impact on Functional Connectivity Specificity

Research has demonstrated that aCompCor preserves neurobiologically plausible connectivity patterns better than alternative methods. In assessments using the known anatomy of functional networks, aCompCor maintained stronger specificity for established network boundaries in both the default mode and motor control networks compared to mean signal regression approaches [51].

Motion Artifacts in Resting-State vs. Task fMRI

Differential Characteristics and Challenges

Head motion presents distinct challenges in resting-state versus task fMRI designs, influencing how aCompCor should be applied and interpreted within these contexts.

In resting-state fMRI, motion artifacts exhibit characteristic distance and orientation dependencies, typically reducing long-range connectivity while increasing short-range correlations [9]. This pattern is particularly problematic because it systematically biases connectivity measures in ways that can confound group comparisons, especially when studying populations with different inherent motion characteristics (e.g., children vs. adults, patients vs. controls) [9] [56].

For task-based fMRI, the primary concern is the potential for temporal correlations between motion and task paradigms. Even small motions can introduce false activations if they coincide systematically with task conditions [9]. The block design commonly used in task fMRI may produce distinctive motion artifacts that differ from the more stochastic motion patterns in resting-state scans.

Population-Specific Considerations

Recent large-scale studies have identified specific factors that influence head motion across different populations. Analysis of 40,969 subjects from the UK Biobank revealed that Body Mass Index (BMI) and ethnicity were the strongest predictors of head motion, with a ten-point increase in BMI corresponding to a 51% motion increase [56]. Contrary to expectations, psychiatric disorders alone were not significant motion indicators, while hypertension was associated with significantly increased motion [56].

Optimization Guidelines

Component Selection Strategies

A critical implementation decision involves determining how many principal components to include as nuisance regressors. While early implementations used a fixed number of components (typically 6), modern approaches favor data-driven criteria:

• Kaiser-Guttman criterion: Retaining components with eigenvalues greater than the mean eigenvalue
• Broken stick model: Comparing explained variance to random distribution expectations
• Cross-validation: Testing different component numbers on subset of data

In practice, 6-10 components often capture the majority of structured noise, but optimal numbers may vary with data quality, acquisition parameters, and motion severity [53].

Integration with Complementary Methods

For comprehensive noise correction, aCompCor is most effective when combined with other strategies:

• Realignment parameters: Always include the 6 rigid-body motion parameters and their temporal derivatives
• Motion scrubbing: For severely contaminated datasets, consider combining with selective volume removal [51]
• Temporal filtering: Maintain appropriate high-pass filtering to remove slow drifts

Notably, research indicates that when aCompCor is properly implemented, additional scrubbing provides no significant benefit for motion reduction or connectivity specificity [51].

The Scientist's Toolkit

Table 3: Essential Research Reagents and Computational Tools

Tool/Resource	Function	Implementation Considerations
fMRIPrep [52]	Automated preprocessing pipeline	Standardized aCompCor implementation; outputs multiple component types; compatible with BIDS format
PhysIO Toolbox [55]	Model-based noise correction	RETROICOR implementation; requires physiological recordings; integration with SPM
FSL MCFLIRT [56]	Head motion estimation	Quantifies framewise displacement; used for motion censoring
White Matter & CSF Masks [50] [51]	Define noise ROIs	Require accurate tissue segmentation; can be derived from T1-weighted images
PCA Algorithms [50] [53]	Component extraction	Memory efficient implementation for high-dimensional data; standardized scaling procedures

aCompCor represents a robust, data-driven approach for physiological noise modeling that effectively addresses motion-related artifacts in both resting-state and task fMRI. Its ability to capture complex spatiotemporal noise patterns without external monitoring makes it particularly valuable for large-scale studies and clinical populations. When implementing aCompCor, researchers should carefully consider noise ROI definition, component selection criteria, and appropriate integration with other preprocessing steps based on their specific experimental design and research questions. As fMRI continues to advance with higher field strengths and faster acquisition sequences, data-driven correction methods like aCompCor will remain essential tools for maximizing signal fidelity and biological validity in neuroimaging research.

Optimizing fMRI Protocols: Mitigation Strategies for Challenging Populations and Designs

In functional magnetic resonance imaging (fMRI), head motion artifacts represent a significant methodological challenge, particularly because in-scanner motion is frequently correlated with variables of clinical and developmental interest. This correlation creates systematic bias that can profoundly impact study conclusions in both resting-state (rs-fMRI) and task-based fMRI research [1] [2]. When studying groups that inherently differ in motion characteristics—such as children vs. adults, healthy controls vs. patients with neurological disorders, or individuals with cognitive impairments vs. those without—motion artifacts can create spurious group differences or mask genuine effects [1]. In functional connectivity (fc-MRI) studies, these artifacts have been shown to distort correlation-based measures, producing characteristic distance and orientation dependencies with decreased long-distance connectivity and increased local connectivity [2]. The problem is particularly acute in clinical and developmental neuroimaging because the groups with the greatest brain impairment often exhibit the highest levels of in-scanner motion [2], creating a systematic confound that threatens the validity of findings across numerous research domains.

Motion Artifact Characteristics in Resting-State vs. Task fMRI

Fundamental Differences in Vulnerability

Head motion affects both resting-state and task-based fMRI, but the nature of the bias differs substantially between these paradigms. In task-based fMRI, motion often becomes temporally correlated with task performance through cognitive engagement [2]. For example, participants may move more when responding to stimuli or during cognitively demanding blocks. This creates a direct confound where motion-induced signal changes are indistinguishable from true BOLD responses to the task [2]. In resting-state fMRI, while no explicit task is administered, motion artifacts remain problematic and may even be exacerbated when individuals are imaged while at rest [2]. The absence of a structured task means motion patterns are less predictable, but systematic differences between groups in their tendency to move can still produce profound biases in functional connectivity measures [1].

Spatial and Temporal Characteristics of Motion Artifacts

The spatial distribution of motion artifacts follows characteristic patterns that reflect biomechanical constraints of head movement. Motion is minimal near the atlas vertebrae (where the skull attaches to the neck) and increases with distance from this anchor point [1]. Frontal cortex shows particularly high motion susceptibility, likely due to the preponderance of y-axis rotation (nodding movement) [1]. The artifacts manifest differently across brain regions: signal intensity drops occur across brain parenchyma, while areas at the brain's edge demonstrate large signal increases due to partial volume effects [1].

Temporally, motion produces both immediate, circumscribed signal changes and longer-duration artifacts. Large movements generate substantial signal drops immediately following the movement event, scaling with motion magnitude [1]. Additionally, sporadic longer-duration artifacts (persisting up to 8-10 seconds) may occur, potentially due to motion-related changes in CO₂ from yawning or deep breathing [1]. These complex spatiotemporal properties make motion artifacts particularly challenging to address with standard denoising approaches.

Table 1: Comparative Characteristics of Motion Artifacts in Resting-State vs. Task fMRI

Characteristic	Resting-State fMRI	Task-Based fMRI
Temporal Correlation	Less predictable patterns	Often correlated with task structure
Systematic Bias Risk	High between clinical groups	High within subjects across conditions
Primary Manifestation	Altered functional connectivity	False activation/deactivation
Spatial Pattern	Distance-dependent connectivity changes	Task-specific activation patterns
Frequency Content	Not band-limited	Often overlaps with task frequency

Quantitative Assessment of Motion Artifacts

Motion Quantification Metrics

In-scanner motion is typically estimated from the functional time series itself using rigid body realignment, producing six realignment parameters (RPs: 3 translations, 3 rotations) that describe how each volume must be moved to align with a reference [1]. These parameters are commonly summarized as frame displacement (FD), which provides a concise index of volume-to-volume motion [1]. Different implementations of FD exist, with the formulation from Jenkinson et al. (implemented in FSL) aligning best with voxel-specific measures of displacement [1]. It is critical to note that these measures are limited by the temporal resolution of the acquisition (repetition time) and cannot effectively capture within-volume motion [1]. Furthermore, with the advent of multiband imaging producing shorter repetition times, FD values become more difficult to compare across studies, suggesting a need for standardized measures such as millimeters of RMS displacement per minute [1].

Performance of Motion Correction Methods

Recent methodological advances have produced increasingly sophisticated approaches for motion correction. The modified slice-oriented motion correction (mSLOMOCO) pipeline represents one such advancement, demonstrating superior performance in removing residual motion artifacts compared to traditional approaches. In tests using gold-standard simulated motion data with various motion patterns, mSLOMOCO with 12 volume/slice-wise motion parameters and partial volume regressors reduced residual signal by 29-45% compared to standard volume-based correction (VOLMOCO) and by 28-31% compared to the original SLOMOCO pipeline [57]. These quantitative improvements highlight the importance of addressing both intervolume and intravolume motion components.

Table 2: Performance Comparison of Motion Correction Methods on SIMPACE Data [57]

Correction Method	Components	Residual Signal Reduction (1× motion)	Residual Signal Reduction (2× motion)
VOLMOCO	6 Vol-mopa + PV regressors	Baseline	Baseline
oSLOMOCO	14 voxel-wise regressors	Not reported	Not reported
mSLOMOCO	12 Vol-/Sli-mopa + PV regressors	29% improvement over VOLMOCO	45% improvement over VOLMOCO
mSLOMOCO vs oSLOMOCO	12 Vol-/Sli-mopa + PV regressors	28% improvement over oSLOMOCO	31% improvement over oSLOMOCO

Experimental Protocols for Motion Correction

Retrospective Correction Methods

Retrospective correction methods applied after data collection represent the most common approach to addressing motion artifacts. These include:

• Volume-based realignment: Each volume in the time series is rigidly realigned to a reference volume, producing the six realignment parameters that form the basis for many subsequent corrections [1].
• Nuisance regression: Motion parameters and their expansions (temporal derivatives, squared terms) are regressed out from the signal [57]. The partial volume (PV) regressor has been shown to effectively reduce residual motion artifact as a motion nuisance regressor after both VOLMOCO and mSLOMOCO correction [57].
• Volume censoring ("scrubbing"): Volumes with excessive motion (typically defined by FD thresholds) are removed from analysis [1].
• Slice-level correction: Methods like SLOMOCO address intravolume motion by applying corrections at the slice level, accounting for motion that occurs during the acquisition of a single volume [57].

The mSLOMOCO pipeline exemplifies an advanced retrospective approach, combining volume-wise rigid intervolume motion parameters (Vol-mopa), slice-wise rigid intravolume motion parameters (Sli-mopa), and the proposed PV motion nuisance regressor to achieve superior artifact reduction [57].

Prospective Correction and Experimental Design

Prospective methods aim to prevent motion artifacts during data acquisition:

• Real-time motion correction: Using optical tracking, navigator echoes, or real-time image registration to adjust imaging planes during acquisition [2].
• Subject preparation and stabilization: Proper instruction, training, and mild physical restraints can reduce motion, though these are usually insufficient alone [2].
• Task design considerations: For task-based fMRI, designing tasks that minimize systematic motion differences between conditions can reduce bias.

A critical consideration in study design is ensuring that motion does not systematically differ between groups of interest. When such systematic differences are unavoidable (as in many clinical comparisons), including motion as a covariate in group-level analyses becomes essential [2].

Special Considerations for Ultra-High Field fMRI

At ultra-high field strengths (7T and above), additional considerations emerge for laminar fMRI studies. Gradient echo (GE) sequences commonly used in laminar fMRI exhibit superficial bias, where larger signals are observed in superficial layers relative to deep layers [58]. This bias arises from both the presence of intra-cortical ascending draining veins and variations in baseline physiological parameters across cortical depths [58]. Correction methods for this specific bias include:

• Z-scoring timecourses [58]
• L2 normalization of beta estimates across conditions [58]
• Ratio of activations in two experimental conditions [58]
• Deming regression - a method that offers robust correction for superficial bias without requiring conditions to differ in regional-mean offset [58]

Simulation studies demonstrate that accurate correction of superficial bias is crucial to avoid drawing erroneous conclusions from laminar analyses of GE fMRI data [58].

Table 3: Research Reagent Solutions for Motion Correction Research

Tool/Resource	Function/Purpose	Availability
SIMPACE Sequence	Generates motion-corrupted MR data by altering imaging plane coordinates; creates gold-standard simulated motion data for validation	Custom sequence [57]
SLOMOCO Pipeline	Implements slice-oriented motion correction addressing both intervolume and intravolume motion	GitHub: https://github.com/wanyongshinccf/SLOMOCO [57]
Ex Vivo Brain Phantom	Provides motion-free ground truth data for validating correction methods; formalin-fixed brain in Fomblin [57]
Frame Displacement (FD) Metrics	Quantifies volume-to-volume motion from realignment parameters; multiple implementations available	FSL, AFNI, SPM [1]
Color Contrast Analyser (CCA)	Checks color contrast ratios in visualization tools to ensure accessibility	Third-party tool [59]
Deming Regression	Corrects superficial bias in laminar fMRI without requiring condition differences in regional-mean offset	Custom implementation [58]

Addressing systematic bias when motion correlates with clinical status or age requires a multifaceted approach combining rigorous experimental design, comprehensive data processing, and appropriate statistical correction. Based on current evidence, the following practices are recommended:

• Report motion metrics transparently: Provide detailed motion-related quality control metrics (mean FD, maximum FD, number of censored volumes) to enable comparison between studies and groups [2].
• Implement multi-stage correction: Combine prospective and retrospective methods, considering advanced approaches like mSLOMOCO that address both intervolume and intravolume motion [57].
• Include motion covariates: In group-level analyses, include motion as a covariate when studying groups with systematic motion differences [2].
• Validate with realistic simulations: Use gold-standard simulated motion data like SIMPACE to validate correction pipelines for specific research contexts [57].
• Account for superficial bias in high-field studies: When conducting laminar fMRI at ultra-high field, implement appropriate correction methods for superficial bias such as Deming regression or ROI ratio methods [58].

The continued development and validation of motion correction methods remains essential for ensuring the validity of fMRI findings, particularly in clinical populations where motion may systematically correlate with variables of interest.

Handling Task-Correlated Motion in Motor, Speech, and Cognitive Paradigms

Task-correlated motion (TCM) presents a unique confound in functional magnetic resonance imaging (fMRI) studies, particularly in paradigms involving motor execution, speech production, or cognitive tasks that elicit unintentional movement. Unlike resting-state fMRI where motion is relatively random, TCM occurs synchronously with task performance, creating systematic artifacts that can be mistaken for genuine blood oxygenation level-dependent (BOLD) activation [60] [61]. This systematic correlation makes traditional motion correction methods less effective, as removing task-correlated signals risks eliminating the neuronal activity of interest [2]. The problem is especially pronounced in clinical populations such as stroke patients, Parkinson's disease patients, children, and elderly individuals who may exhibit amplified movements during task performance [60] [2]. This technical guide examines the characteristics of TCM across different paradigms, evaluates correction methodologies, and provides practical solutions for researchers conducting task-based fMRI studies in both basic and clinical research settings.

Motion Artifact Characteristics: Resting-State vs. Task fMRI

Fundamental Differences in Motion Properties

Head motion artifacts manifest differently in resting-state versus task-based fMRI, necessitating distinct correction approaches. The table below summarizes the key differences:

Table 1: Characteristics of motion artifacts in resting-state vs. task-based fMRI

Characteristic	Resting-State fMRI	Task-Based fMRI
Temporal Pattern	Random, sporadic movements	Systematic, time-locked to task stimuli
Primary Correction Challenge	Removing noise without altering intrinsic connectivity	Separating motion artifact from true BOLD signal
Spatial Manifestation	Distance-dependent effects: increased short-range, decreased long-range connectivity [62] [63]	False activation at brain edges, high-contrast regions, and motor/speech areas [61]
Group-Level Impact	Can inflate/deflate correlations similarly across subjects [63]	May produce consistent false activation patterns across subjects [61]
Effective Correction Strategies	Global signal regression, censoring, aCompCor [51] [63]	Multi-echo ICA, selective detrending, targeted regression [60] [61]

Physiological Basis of Motion Artifacts

Motion artifacts in fMRI originate from multiple physical mechanisms that affect the MR signal. Through-plane motion modulates the steady-state magnetization established by repeated RF excitation pulses, creating spin-history artifacts [63]. Motion also causes time-dependent modulations in the background magnetic field, resulting in image distortions [61]. During speech production, local nonrigid movements of the pharynx, tongue, and jaw alter magnetic field distribution within the brain, particularly affecting inferior frontal and temporal regions where signal changes can exceed 100% [61]. These artifacts persist despite image registration because they alter the fundamental relationship between tissue properties and measured signal intensity.

Paradigm-Specific Challenges and Solutions

Motor Paradigms

Motor tasks such as hand grasping present unique challenges because head motion naturally correlates with task execution. Clinical populations like stroke patients exhibit approximately twice the head motion during hand grasp and ankle flexion tasks compared to healthy individuals [60]. These motions can introduce false positive activation in motor regions or mask true activation patterns.

Solution: Multi-Echo ICA (ME-ICA) ME-ICA leverages multi-echo fMRI acquisitions to separate BOLD signals from motion-related artifacts based on their distinct physical properties. The method calculates two parameters—kappa (κ) and rho (ρ)—that quantify T2* and S0 weighting of signal components respectively [60]. BOLD components exhibit high κ and low ρ, while motion artifacts show the opposite pattern.

Table 2: Performance comparison of motion correction methods for motor-task fMRI

Method	Principle	Effectiveness for High Motion	Limitations
Single-Echo (SE) + Motion Regression	Regresses out motion parameters from realignment	Limited for task-correlated motion [60]	Can remove neural signals of interest
Multi-Echo Optimally Combined (ME-OC)	Combines echoes weighted by T2* contrast	Moderate improvement in motor region t-statistics [60]	Less effective at dissociating BOLD from motion
Multi-Echo ICA (ME-ICA)	ICA decomposition with echo-time information	High; better dissociates motion from BOLD, reduces noise [60]	Requires multi-echo sequences, computational complexity

Experimental Protocol for ME-ICA:

• Acquire multi-echo fMRI data with echo times optimized for BOLD sensitivity (e.g., TE = 12-35ms)
• Perform slice-time correction and volume realignment
• Decompose data using ICA while leveraging multi-echo information
• Calculate kappa and rho parameters for each component
• Classify components as BOLD or non-BOLD based on their TE-dependence
• Reconstruct denoised data using only BOLD components
• Perform general linear model analysis on denoised data

Speech Paradigms

Overt speech production generates particularly complex artifacts due to concurrent head movement, magnetic susceptibility changes from air volume variations in vocal tract, and modulation of breathing patterns [61]. These artifacts are most pronounced in inferior frontal and temporal regions, potentially obscuring language-related activation.

Solution: Selective Detrending Selective detrending identifies and removes TCM artifacts based on their temporal and spatial characteristics while preserving legitimate BOLD signals [61]. The method leverages differences in the evolution of speech-related TCM artifacts and hemodynamic responses.

Experimental Protocol for Selective Detrending:

• Identify artifact-corrupted voxels (typically at brain edges or near vocal structures)
• Extract reference time courses from these voxels
• Compute principal components of artifact time courses
• Regress out components that are temporally correlated with task but distinct from expected hemodynamic response
• Verify preservation of legitimate BOLD signals in regions of interest

Table 3: Comparison of speech artifact correction methods

Method	Principle	Effectiveness	Limitations
Motion Parameter Regression (MPR)	Regresses out rigid-body motion parameters	Limited for non-rigid speech motions [61]	Does not address local magnetic field changes
Image Ignoring/Screening	Removes images acquired during speech	Moderate when HDR and TCM are temporally separated [61]	Compromises statistical power, ineffective with HDR overlap
Nonselective Detrending	Removes global artifact patterns from all voxels	High artifact reduction but removes true signal [61]	Redensitivity in language areas
Selective Detrending	Targeted removal based on temporal and spatial features	High artifact reduction with signal preservation [61]	Requires careful parameter optimization

Cognitive Paradigms

While cognitive paradigms typically involve less overt movement, they still suffer from subtle task-correlated motion such as micro-movements associated with button presses, shifts in attention, or emotional responses. These subtle motions can introduce systematic confounds that affect between-group comparisons, particularly when comparing populations with different motion characteristics (e.g., children vs. adults, patients vs. controls).

Solution: Anatomical Component-Based Correction (aCompCor) aCompCor addresses motion artifacts by extracting noise components from regions without BOLD signal, such as white matter and cerebrospinal fluid [51]. Unlike mean signal regression, aCompCor uses principal component analysis to capture multiple spatially coherent noise patterns in these regions, better accounting for the complex spatial manifestations of motion artifacts.

Experimental Protocol for aCompCor:

• Segment structural T1-weighted image to identify white matter and CSF regions
• Project these masks into functional space
• Extract multiple principal components from each region
• Include these components as nuisance regressors in general linear model
• Optionally combine with "scrubbing" (removing high-motion volumes)

Advanced Correction Methodologies

Multi-Echo Acquisition and Analysis

Multi-echo fMRI represents a powerful approach for addressing TCM by enabling separation of BOLD and non-BOLD signals based on their distinct TE-dependence. The ME-ICA pipeline has demonstrated particular effectiveness for motor-task fMRI with high motion levels [60]. The method exploits the fact that BOLD signals scale with TE (T2* weighting), while many motion-related artifacts exhibit different TE dependence.

Prospective Motion Correction

Prospective motion correction techniques use external tracking systems or navigator echoes to update imaging planes in real-time during data acquisition [57]. The SIMPACE (Simulated Prospective Acquisition Correction) sequence can emulate various motion patterns to optimize these methods, potentially reducing spin-history effects and other acquisition-specific artifacts [57].

Functional Connectivity Measures Robust to Motion

Different functional connectivity measures exhibit varying sensitivity to motion artifacts. Recent research has systematically evaluated multiple connectivity metrics:

Table 4: Sensitivity of functional connectivity measures to motion artifacts

Connectivity Measure	Residual Motion Sensitivity	Test-Retest Reliability	Recommended Use
Full Correlation	High distance-dependent motion effects [5]	High	Studies prioritizing reliability over motion robustness
Partial Correlation	Low motion sensitivity [5]	Low	Motion-sensitive studies with adequate sample size
Coherence	Moderate motion sensitivity [5]	Intermediate	Frequency-specific connectivity questions
Mutual Information	Low motion sensitivity [5]	Intermediate	Non-linear connectivity applications

The Scientist's Toolkit: Essential Research Reagents

Table 5: Key analytical tools and resources for tackling task-correlated motion

Tool/Resource	Function	Application Context
ME-ICA Pipeline	Denoising of multi-echo fMRI data	Motor-task fMRI with high motion [60]
aCompCor	Noise component extraction from WM/CSF	Resting-state and cognitive task fMRI [51]
SLOMOCO	Slice-oriented motion correction	Addressing intravolume motion effects [57]
Homer2 fNIRS Processing	Motion correction for fNIRS data	Pediatric populations with high motion [64]
fMRIPrep	Automated preprocessing pipeline	Standardized handling of motion parameters
ART (Artifact Detection Tools)	Identification of motion-corrupted volumes	Creating scrubbing regressors for denoising

Experimental Design Considerations

Paradigm Optimization

Careful experimental design can minimize TCM without compromising task validity. For motor paradigms, consider using isometric contractions rather than overt movements when possible. For speech production tasks, use block designs with sufficient rest periods between utterances to allow motion artifacts to dissipate. Event-related designs with jittered inter-stimulus intervals can help disentangle motion artifacts from hemodynamic responses [61].

Acquisition Parameters

Multi-echo acquisitions with optimized echo times (typically 3+ echoes spanning 12-35ms) significantly enhance the ability to distinguish BOLD from non-BOLD signals [60]. Smaller voxel sizes can reduce intra-voxel dephasing artifacts, though this must be balanced against signal-to-noise considerations. Prospective motion correction sequences should be employed when available.

Visualization of Key Methodological Approaches

ME-ICA Processing Workflow

Motion Artifact Characterization Framework

Task-correlated motion remains a significant challenge in task-based fMRI, particularly for clinical populations who may exhibit amplified movements during task performance. The most effective approach involves a combination of optimized acquisition parameters, careful experimental design, and advanced processing techniques tailored to specific paradigm requirements. Multi-echo methods like ME-ICA show particular promise for motor paradigms, while selective detrending approaches offer advantages for speech production tasks. Future developments in real-time motion correction, deep learning-based artifact detection, and optimized acquisition sequences will further enhance our ability to dissociate motion artifacts from true neural signals. As these methods evolve, researchers should prioritize transparency in reporting motion correction procedures and motion metrics to enable proper evaluation and comparison across studies.

Head motion remains a significant challenge in functional magnetic resonance imaging (fMRI), systematically altering signal intensity and compromising data integrity in both resting-state and task-based paradigms. The characteristics of motion artifacts differ meaningfully between these experimental approaches, necessitating tailored optimization strategies. In resting-state fMRI (rs-fMRI), motion systematically alters correlations in functional connectivity, producing spurious but spatially structured patterns that can be difficult to distinguish from true neural signals [65]. These motion-induced signal changes often persist for more than 10 seconds after visible motion ceases and increase observed RSFC correlations in a distance-dependent manner [65]. For task-based fMRI, motion can become temporally correlated with task performance, creating artifacts that cannot be distinguished from true brain activity related to the experimental paradigm [2]. This focused review examines protocol optimization strategies for mitigating motion artifacts, with particular emphasis on subject preparation, physical padding techniques, and real-time monitoring technologies, contextualized within the distinct motion artifact characteristics of resting-state versus task-based fMRI research.

Motion Artifact Characteristics: Resting-State vs. Task fMRI

Understanding the fundamental differences in how motion artifacts manifest across experimental paradigms is crucial for developing effective optimization strategies. The table below summarizes key distinctions in motion artifact characteristics between resting-state and task-based fMRI.

Table 1: Motion Artifact Characteristics in Resting-State vs. Task-Based fMRI

Characteristic	Resting-State fMRI	Task-Based fMRI
Primary Motion Concern	Distance-dependent functional connectivity alterations [65]	Temporal correlation with task performance [2]
Typical Motion Pattern	Variable waveforms, often shared across voxels [65]	Often coupled with task demands or responses [66]
Signal Impact	Increases short-range correlations, decreases long-range connections [65] [39]	Introduces signal changes confounded with activation patterns [2]
Artifact Persistence	Can persist >10 seconds after motion ceases [65]	Typically limited to movement periods
Spatial Pattern	Creates spatially structured noise resembling default mode network [67]	Localized to regions involved in task performance
Vulnerable Populations	Clinical, pediatric, elderly populations [65]	Populations with movement-related task demands

The physical principles underlying motion artifacts stem from the fact that even millimeter-scale head motions create complex spatiotemporal properties that can lead to substantial errors in functional connectivity estimates [2]. In rs-fMRI, motion introduces spurious variance that is most similar at nearby voxels, spuriously increasing correlations between BOLD timeseries across all voxels, with the greatest increases between nearby voxels [65]. This creates a characteristic distance-dependent artifact that can persist despite various post-processing correction approaches.

Subject Preparation Strategies

Effective subject preparation represents the first line of defense against motion artifacts in fMRI acquisition. These strategies aim to minimize motion at its source through behavioral interventions and clear communication.

Pre-Scan Training and Instruction

Comprehensive pre-scan preparation should include:

• Detailed motion minimization instructions: Provide clear, explicit guidance on the importance of remaining still, with specific examples of what movements to avoid [66].
• Behavioral training: For task-based fMRI, practice sessions outside the scanner can familiarize subjects with task requirements, potentially reducing movement associated with task uncertainty [66].
• Habituation procedures: For anxious populations or pediatric participants, mock scanner sessions using simulated scanner sounds can reduce anxiety-driven motion [68].

Strategic Instruction Framing

The framing of motion instructions significantly impacts compliance and effectiveness:

• Positive feedback framing: Instead of emphasizing punishment for movement, frame instructions around the ability to maintain stillness, as evidenced by studies showing that subjects can learn to minimize motion when provided with appropriate feedback [66].
• Clear feedback interpretation: Instruct subjects on how to interpret real-time feedback, noting that colored cross displays (white for <0.2 mm, yellow for 0.2-0.3 mm, red for ≥0.3 mm) help subjects understand their motion status without causing excessive frustration [66].

Table 2: Subject Preparation Protocols for Different Populations

Population	Preparation Strategy	Evidence of Efficacy
Adults	Combination of verbal instructions and real-time feedback	Significant reduction in average FD from 0.347 to 0.282 mm [66]
Infants/Toddlers	Scanning during natural sleep, respiratory motion filtering	Improved data retention and quality with age-specific filters [69]
Clinical Populations	Customized habituation, anesthesia protocols	Sevoflurane associated with reduced RSN robustness (OR = 0.2) [70]
Pediatric	Mock scanner training, positive reinforcement	FIRMM increased usable data (FD ≤ 0.2 mm) in infant cohorts [68]

Padding and Physical Stabilization

Physical restraint systems complement behavioral strategies by providing mechanical limitation of head movement. Optimal padding strategies must balance motion restriction with subject comfort.

Head Stabilization Techniques

• Customized head molds: Expensive but highly effective for reducing motion, particularly in pediatric populations or high-resolution studies [66].
• Foam padding systems: Standard foam padding placed snugly between the head and head coil provides sufficient stabilization for most research applications [66] [2].
• Bite bars: While effective for minimizing motion, these are often poorly tolerated and can introduce discomfort that itself leads to movement.
• Vacuum-based immobilization systems: Emerging technologies that provide customized fit and superior motion restriction.

Whole-Body Positioning

Comprehensive positioning addresses motion sources beyond the head:

• Comfortable body support: Ensure subjects are comfortably positioned with proper support under knees and neutral spine alignment to reduce whole-body repositioning movements.
• Restraint systems: For pediatric populations, swaddling or specialized blankets can provide security and reduce startle reflexes [68].
• Interface management: Ensure all monitoring equipment and cables are securely positioned to prevent discomfort or distraction.

Real-Time Motion Monitoring

Real-time monitoring technologies represent a significant advancement in motion mitigation, allowing for immediate correction and intervention during scanning sessions.

Implementation Protocols

The Framewise Integrated Real-Time MRI Monitoring (FIRMM) software exemplifies the real-time monitoring approach:

• Real-time motion tracking: Software uses rapid image reconstruction and rigid-body alignment to estimate frame-by-frame movement during acquisition [66] [68].
• Visual feedback display: Participants view a colored cross that changes based on motion levels (white for <0.2 mm, yellow for 0.2-0.3 mm, red for ≥0.3 mm) [66].
• Between-run feedback: After each scanning sequence, participants receive a Head Motion Report with a percentage score (0-100%) and graphical representation of their motion over time, with encouragement to improve their score on subsequent runs [66].

Efficacy Across Populations

Real-time monitoring demonstrates significant efficacy across diverse populations:

• Adult populations: Implementation during a word repetition task resulted in a statistically significant reduction in average framewise displacement (FD) from 0.347 mm to 0.282 mm, with a small-to-moderate effect size [66].
• Infant populations: Scanning with FIRMM (n = 407) versus without FIRMM (n = 295) significantly increased the amount of usable fMRI data (FD ≤ 0.2 mm) acquired per infant [68].
• Task-based applications: Effective even during cognitively demanding tasks, demonstrating that participants can process motion feedback while attending to task requirements [66].

The following diagram illustrates the integrated workflow for implementing real-time motion monitoring in fMRI studies:

Table 3: Research Reagent Solutions for Motion Mitigation

Tool/Resource	Primary Function	Application Context
FIRMM Software	Real-time motion tracking and feedback	Resting-state and task-based fMRI [66] [68]
Custom Head Molds	Physical motion restraint	High-resolution studies, pediatric populations [66]
Age-Specific Notch Filters	Respiratory artifact removal	Infant and toddler populations [69]
Visual Feedback Display	Real-time motion awareness	Task-based fMRI with participant engagement [66]
Respiratory Monitoring	Physiological noise identification	Populations with high respiratory rates [69]
Siemens Prisma Scanner	High-quality image acquisition	Research-grade fMRI with multiband sequences [66] [69]

Integrated Optimization Framework

Successful motion mitigation requires an integrated approach combining preparation, physical stabilization, and monitoring technologies. The most effective protocols implement strategies from all three domains tailored to specific population needs and research questions.

Population-Specific Optimization

• Pediatric populations: Combine natural sleep scanning, age-specific respiratory filters (20-30 bpm for infants vs. 12-18 bpm for adults), and real-time monitoring for technologist intervention [68] [69].
• Clinical populations: Consider anesthesia protocols when necessary, noting that sevoflurane administration reduces the odds of having robust resting-state networks (OR = 0.2) [70].
• Elderly populations: Focus on comfort optimization and clear communication to reduce adjustment-related movement.

Data Quality Assessment

Implementation of these protocols should be validated through rigorous quality assessment:

• Framewise displacement thresholds: Apply censoring at FD < 0.2 mm to reduce significant motion overestimation to 2% of traits (from 42% without censoring) [39].
• Data retention metrics: Monitor the amount of usable data (FD ≤ 0.2 mm) as a key performance indicator [68].
• Motion-impact scores: Utilize methods like SHAMAN to assign trait-specific motion impact scores for identifying overestimation or underestimation of trait-FC effects [39].

Effective mitigation of motion artifacts in fMRI research requires a comprehensive, multi-layered approach addressing both the behavioral and physical sources of movement while leveraging advanced monitoring technologies. The distinct characteristics of motion artifacts in resting-state versus task-based fMRI necessitate tailored optimization strategies that account for their different artifact patterns and confounding mechanisms. By implementing robust subject preparation protocols, optimizing physical stabilization techniques, and integrating real-time monitoring systems, researchers can significantly improve fMRI data quality and reliability. These optimization approaches are particularly crucial for studies involving motion-vulnerable populations, including pediatric, clinical, and elderly participants, where motion-related artifacts may systematically covary with factors of scientific interest. As fMRI continues to evolve as a fundamental tool in neuroscience research and drug development, rigorous attention to motion mitigation protocols remains essential for generating valid, interpretable results.

Head motion represents the most significant source of artifact in functional magnetic resonance imaging (fMRI), systematically altering signals and potentially leading to both false positive and false negative findings in brain-behavior associations [39] [71]. Even sub-millimeter movements can introduce spurious variance to the blood-oxygenation-level-dependent (BOLD) signal, complicating the interpretation of functional connectivity (FC) in both resting-state and task-based paradigms [49] [72]. The challenge is particularly pronounced in resting-state fMRI (rs-fMRI), where the timing of underlying neural processes is unknown, making it especially vulnerable to motion artifacts that can mimic or obscure genuine neural correlations [39] [73].

Establishing robust motion inclusion/exclusion criteria requires balancing competing priorities: eliminating motion-contaminated data to reduce spurious findings while avoiding the systematic exclusion of participants who may exhibit important variance in traits of interest [39] [74]. This technical guide examines current methodologies for quantifying motion effects, evaluates exclusion criteria across experimental contexts, and provides evidence-based recommendations for establishing quality control pipelines that maintain both data quality and sample representativeness.

Motion Artifact Characteristics: Resting-State vs. Task fMRI

Fundamental Differences in Vulnerability

The impact of motion manifests differently across resting-state and task fMRI paradigms, necessitating tailored quality control approaches. Resting-state functional connectivity relies on low-frequency temporal correlations in spontaneous BOLD fluctuations, making it particularly susceptible to motion-induced signal alterations that occur at similar timescales [39] [73]. Task-based fMRI generally demonstrates greater tolerance to motion artifacts due to the known timing of stimulus presentation and evoked responses, though certain tasks may elicit movements that correlate with experimental conditions [66].

Table 1: Motion Characteristics Across fMRI Paradigms

Characteristic	Resting-State fMRI	Task-Based fMRI
Primary Motion Concern	Global spurious correlations; distance-dependent connectivity changes [39]	Condition-specific motion; task-correlated movement [66]
Typical Motion Profile	Sustained lower-amplitude drift; spontaneous movements [71]	Phasic, event-related movements; response-contingent motion [66]
Spatial Impact Pattern	Decreased long-distance, increased short-range connectivity [39]	Focal activation artifacts; distorted hemodynamic response functions
Temporal Structure	Low-frequency fluctuations (<0.1 Hz) vulnerable to motion artifacts [73]	Higher-frequency evoked responses less overlapping with motion
Optimal Censoring Approach	Aggressive frame censoring (FD < 0.2 mm) [39]	Moderate censoring thresholds; condition-specific models [66]

Quantifying Motion Effects

Framewise displacement (FD) has emerged as the standard metric for quantifying head motion, representing the sum of absolute derivatives of the three translation and three rotation parameters [66] [39]. FD provides a comprehensive measure of volume-to-volume movement, with values ≥0.2 mm typically indicating motion-contaminated volumes requiring censoring [39]. However, different populations and research questions may warrant adjusted thresholds, particularly when studying clinical groups that typically exhibit higher motion [74].

Recent evidence demonstrates that motion artifacts persist despite sophisticated denoising approaches. In the large Adolescent Brain Cognitive Development (ABCD) sample, even after implementing the ABCD-BIDS denoising pipeline (including global signal regression, respiratory filtering, and motion parameter regression), 23% of signal variance remained explainable by head motion [39]. This represents a substantial 69% reduction from the 73% of variance explained by motion following minimal processing, yet underscores the persistent influence of motion on downstream analyses.

Establishing Inclusion/Exclusion Criteria: Evidence-Based Thresholds

Framewise Displacement Censoring Thresholds

The selection of FD censoring thresholds represents a critical decision point in fMRI quality control, with significant implications for both data quality and sample composition. Research indicates that stringent censoring at FD < 0.2 mm effectively reduces motion-induced overestimation of brain-behavior relationships, decreasing significant motion overestimation from 42% to just 2% of traits in the ABCD sample [39]. However, this approach comes with tradeoffs, as excluding high-motion participants may systematically bias samples by removing individuals with specific demographic, clinical, or cognitive characteristics [74].

Table 2: Motion Censoring Thresholds and Their Implications

Censoring Threshold	Benefits	Limitations	Recommended Context
FD < 0.2 mm (stringent)	Maximally reduces motion overestimation (to 2% of traits) [39]	Excludes more data; may bias sample composition [74]	Studies where motion correlates with trait of interest; case-control designs
FD < 0.4 mm (moderate)	Balances data quality and retention; appropriate for many research contexts	Residual motion effects may persist in motion-correlated traits [39]	Typical adult populations; studies with limited statistical power
FD < 1.0 mm (lenient)	Maximizes data retention and sample representativeness [75]	Substantial risk of motion-related artifacts and spurious findings [39]	Preliminary analyses; special populations where data is scarce

Participant-Level Exclusion Criteria

Beyond frame-wise censoring, establishing participant-level exclusion criteria is necessary when data quality is sufficiently compromised. While standards vary across laboratories, common exclusion thresholds include:

• Excessive motion: Participants with >10% of volumes censored at FD < 0.2 mm are often excluded, though more liberal thresholds of 15-25% have been used in pediatric populations [71].
• Minimum data requirements: Retaining participants with at least 5 minutes of usable data after censoring ensures sufficient signal-to-noise ratio for reliable connectivity estimates [71].
• Motion correlation with traits: For traits strongly correlated with motion (e.g., ADHD symptoms), more stringent exclusion criteria may be warranted to avoid spurious associations [39].

Recent evidence highlights the consequential nature of these exclusion decisions. Analyses of the ABCD study demonstrate that exclusion based on motion criteria systematically relates to "a broad spectrum of behavioral, demographic, and health-related variables," potentially biasing research findings [74]. This underscores the importance of both transparent reporting of exclusion criteria and analytical approaches that account for missing data patterns.

Quality Control Pipelines and Methodologies

Integrated Quality Control Procedures

Effective quality control requires both quantitative metrics and qualitative visual inspection at multiple processing stages [71] [75]. Quantitative measures alone may miss subtle artifacts, processing errors, or incidental findings that warrant exclusion [71]. A comprehensive QC protocol should include:

• Real-time monitoring: Using software such as FIRMM (Real-time fMRI Motion Monitoring) to provide feedback during scanning sessions, reducing average FD by approximately 19% in task-based fMRI [66].
• Visual inspection: Examining structural and functional images for alignment errors, artifacts, and incidental findings not captured by quantitative metrics [71].
• Multi-stage assessment: Evaluating data quality after initial preprocessing, registration, and time series analysis to identify errors at each processing step [75].

The following workflow diagram illustrates a comprehensive quality control procedure integrating both quantitative and qualitative assessments:

Motion Impact Detection Methods

Novel methodologies have emerged to detect trait-specific motion impacts, addressing the limitation that standard motion quantification approaches are agnostic to the hypothesis under study [39]. The Split Half Analysis of Motion Associated Networks (SHAMAN) method quantifies motion impact by comparing correlation structures between high- and low-motion halves of each participant's fMRI timeseries [39]. This approach distinguishes between motion causing overestimation versus underestimation of trait-FC effects, providing a more nuanced understanding of motion's influence on specific research questions.

The following diagram illustrates the SHAMAN methodology for calculating motion impact scores:

Table 3: Essential Tools and Resources for fMRI Motion Quality Control

Tool/Resource	Function	Application Context
FIRMM (Real-time fMRI Motion Monitoring)	Provides real-time feedback to participants during scanning; reduces motion by 19% in task-based fMRI [66]	All fMRI studies, particularly with motion-prone populations
Framewise Displacement (FD)	Quantifies volume-to-volume head movement; primary metric for censoring decisions [66] [39]	Standardized motion quantification across studies
SHAMAN (Split Half Analysis of Motion Associated Networks)	Computes trait-specific motion impact scores; distinguishes over/underestimation [39]	Brain-behavior association studies; motion-correlated traits
ABCD-BIDS Pipeline	Implements comprehensive denoising (global signal regression, respiratory filtering, motion regression) [39]	Large-scale studies; standardized processing across sites
Visual Inspection Protocols	Identifies artifacts, processing errors, and incidental findings missed by quantitative metrics [71] [75]	Essential complement to all automated QC procedures

Establishing evidence-based motion inclusion/exclusion criteria requires careful consideration of research context, sample characteristics, and analytical goals. While stringent framewise displacement thresholds (FD < 0.2 mm) effectively mitigate motion-related artifacts, researchers must remain cognizant of potential selection biases introduced by excluding high-motion participants [39] [74]. Integrating multiple quality control approaches—including real-time monitoring, quantitative metrics, visual inspection, and trait-specific motion impact assessments—provides the most comprehensive protection against spurious findings while maintaining methodological transparency.

Future methodological developments should focus on denoising approaches that preserve neural signal in high-motion data, thereby reducing the need for exclusion and enhancing the representativeness of neuroimaging samples. Until such methods are perfected, clear reporting of motion QC procedures and their impact on sample composition remains essential for interpreting and replicating fMRI research findings.

Motion artifacts represent a fundamental challenge in functional magnetic resonance imaging (fMRI), introducing systematic biases that can compromise data integrity and lead to spurious findings. The characteristics and impacts of these artifacts differ significantly between resting-state fMRI (rs-fMRI) and task-based fMRI (task-fMRI), necessitating distinct processing strategies. In rs-fMRI, head motion has been shown to systematically affect resting-state functional connectivity measures, potentially creating false patterns that could be mistaken for neuronal effects [34]. For task-fMRI, motion can interact with task timing and introduce confounds that vary across experimental conditions [76] [77]. This technical guide examines the core principles for selecting processing pipelines that effectively balance artifact correction with preservation of biological signal across these two fundamental fMRI paradigms.

The modular nature of most preprocessing pipelines introduces specific vulnerabilities. As later preprocessing steps can reintroduce artifacts previously removed in earlier stages, the order of operations becomes critically important [34]. Furthermore, recent large-scale evaluations have revealed that inappropriate pipeline choices can produce results that are not only misleading but systematically so, with the majority of pipelines failing at least one key validation criterion [78]. This underscores the importance of informed pipeline selection based on empirical evidence rather than convention alone.

Core Principles of Pipeline Selection

Foundational Concepts for Pipeline Optimization

Effective pipeline selection requires understanding several foundational concepts that govern the interaction between processing steps and their ultimate impact on data quality:

• Projection and Reintroduction Artifacts: Each regression step in a preprocessing pipeline constitutes a geometric projection of data onto a subspace. Performing sequences of these projections can move data into subspaces no longer orthogonal to previously removed nuisance covariates, effectively reintroducing artifact-related signal [34]. This mathematical reality explains why linear filtering operations are not commutative and why operation order profoundly impacts outcomes.

• The Reliability-Validity Trade-off: Aggressive denoising might improve reliability metrics while simultaneously reducing validity by removing biologically meaningful signal. A suitable pipeline must minimize spurious (motion-induced) differences while remaining sensitive to meaningful individual differences and experimental effects [78]. This balance is particularly crucial for brain-wide association studies (BWAS), where overly aggressive motion correction might systematically exclude individuals with high motion who may exhibit important variance in traits of interest [39].

• Trait-Specific Motion Contamination: Motion artifacts do not affect all research questions equally. Recent methods like SHAMAN (Split Half Analysis of Motion Associated Networks) can assign motion impact scores to specific trait-FC relationships, distinguishing between motion causing overestimation or underestimation of effects [39]. This recognizes that motion correction cannot be "one-size-fits-all" but must account for how specific traits correlate with motion propensity.

Differential Challenges: Resting-State vs. Task fMRI

Table 1: Fundamental Differences in Motion Artifact Challenges Between Resting-State and Task fMRI

Characteristic	Resting-State fMRI	Task-Based fMRI
Primary Motion Effect	Alters intrinsic correlation structure	Creates task-condition-specific confounds
Temporal Characteristics	Continuous artifact throughout scan	Often correlated with task timing/transitions
Key Vulnerability	Distance-dependent connectivity bias	Spurious activation/deactivation patterns
Connectivity Impact	Systemic decrease in long-distance connectivity [39]	Altered task-modulated functional connectivity [76]
Critical Processing Step	Global signal regression controversy [78]	Physiological interaction methods [76]

Pipeline Selection Frameworks by fMRI Modality

Resting-State fMRI Pipeline Optimization

Rs-fMRI presents unique challenges due to the absence of known timing for underlying neural processes, making it especially vulnerable to motion artifacts [39]. Comprehensive evaluations of rs-fMRI processing pipelines have identified several critical considerations:

• Pipeline Performance Variability: Systematic evaluation of 768 data-processing pipelines for network reconstruction from rs-fMRI revealed vast variability in suitability for functional connectomics. While most pipelines failed at least one criterion, optimal pipelines consistently satisfied all criteria across different datasets, spanning minutes, weeks, and months [78]. These optimal pipelines shared common characteristics including appropriate parcellation schemes and filtering approaches.

• Denoising Efficacy vs. Biological Preservation: Different denoising pipelines show varying effectiveness in simultaneously achieving two key objectives: mitigating motion-related artifacts and augmenting brain-behavior associations. No single pipeline universally excels at both objectives across different cohorts, though pipelines combining ICA-FIX and global signal regression (GSR) demonstrate a reasonable trade-off [49].

• The Global Signal Regression Controversy: GSR remains one of the most contentious issues in rs-fMRI preprocessing. While it effectively reduces motion artifacts, it may also remove biologically relevant signal. The decision to include GSR must be informed by research questions and potential biases introduced [78].

Table 2: Resting-State fMRI Pipeline Recommendations Based on Recent Evidence

Processing Step	Recommendation	Rationale	Key Evidence
Global Signal Regression	Context-dependent use	Balances motion reduction with signal preservation	Li et al., 2019 [49]
Motion Censoring	Framewise displacement < 0.2mm	Reduces spurious motion-connectivity relationships	Power et al., 2014; [39]
Temporal Filtering	Combine with motion regression in single step	Prevents artifact reintroduction	PMC6865661 [34]
Network Construction	Portrait divergence evaluation	Assesses whole-network topology preservation	Nature Communications, 2024 [78]
Multi-Site Harmonization	MP-PCA denoising	Improves SNR while compensating for site differences	Kochunov et al., 2018 [79]

Task-fMRI Pipeline Optimization

Task-fMRI introduces different challenges, as motion artifacts often correlate with task conditions, creating spurious activation patterns. The enormous analytical flexibility in task-fMRI—with one study identifying nearly as many unique analysis pipelines as there were studies—creates significant reproducibility challenges [77]:

• Pipeline Community Patterns: Community detection algorithms applied to task-fMRI pipelines have identified that pipelines sharing specific parameters (e.g., number of motion regressors, software packages) tend to produce similar results. This suggests that rather than evaluating all possible pipelines, researchers can focus on representative pipelines from different "communities" that cover the analytical space [77].

• Software-Dependent Variability: The choice of software package (AFNI, FSL, SPM) significantly impacts final group-level results, with variation largely attributable to a handful of individual analysis stages. The most influential stages differ across datasets, with first-level signal modeling, noise modeling, and group-level modeling being particularly variable [80].

• Task-Modulated Functional Connectivity (TMFC): For estimating context-dependent reconfiguration of intrinsic network organization during tasks, methods performance varies significantly by design type. Simulation studies controlling ground-truth TMFC reveal that the most sensitive methods for rapid event-related designs and block designs are sPPI and gPPI with deconvolution procedures, while for other designs, beta-series correlations using least-squares separate (BSC-LSS) perform best [76].

Table 3: Task-fMRI Pipeline Recommendations Based on Experimental Evidence

Processing Step	Recommendation	Rationale	Key Evidence
Software Selection	Consistent use within study	Reduces introduced variability	Bowring et al., 2022 [80]
Motion Regressors	24-parameter model	Comprehensive motion capture	Germani et al., 2023 [77]
TMFC Method (Event-related)	BSC-LSS	Optimal sensitivity for most designs	Communications Biology, 2024 [76]
TMFC Method (Block)	sPPI/gPPI with deconvolution	Best sensitivity for block designs	Communications Biology, 2024 [76]
HRF Modeling	Account for variability	Significant impact on TMFC sensitivity	Communications Biology, 2024 [76]

Quantitative Comparison of Pipeline Performance

Recent large-scale evaluations provide empirical data to guide pipeline selection based on specific research objectives and data characteristics:

Table 4: Quantitative Performance Metrics Across Pipeline Variants

Pipeline Characteristic	Performance Metric	Impact on Results	Context
GSR vs. No-GSR	Differential topology reliability	Affects motion susceptibility & biological sensitivity [78]	rs-fMRI
Parcellation Resolution	Test-retest reliability	200-node parcellations show optimal reliability [78]	rs-fMRI
Motion Censoring Threshold	Trait-FC effect accuracy	FD < 0.2mm reduces overestimation from 42% to 2% of traits [39]	rs-fMRI
Software Package	Group-level statistic similarity	AFNI, FSL, SPM show significant divergence [80]	task-fMRI
HRF Derivative Inclusion	Activation pattern similarity	Affects community assignment of pipelines [77]	task-fMRI
Deconvolution in PPI	TMFC sensitivity	Prominently increases sensitivity [76]	task-fMRI

Experimental Protocols for Pipeline Validation

Protocol 1: Motion Impact Assessment with SHAMAN

The Split Half Analysis of Motion Associated Networks (SHAMAN) provides a robust method for quantifying trait-specific motion impacts:

• Data Requirements: One or more rs-fMRI scans per participant with framewise displacement (FD) calculations and trait measures of interest.

• Procedure:

  • Split each participant's timeseries into high-motion and low-motion halves based on FD median.
  • Compute correlation structure for each half separately.
  • Measure differences in trait-FC effects between halves.
  • Permutation testing with non-parametric combining across connections generates motion impact scores.

• Interpretation: Positive scores aligned with trait-FC effect direction indicate motion overestimation; opposite directions indicate underestimation [39].

Protocol 2: Community Detection for Pipeline Stability

This protocol evaluates pipeline stability across different subject groups and experimental contexts:

• Data Preparation: Process identical dataset with multiple pipeline variants (varying software, smoothing, motion parameters, HRF modeling).

• Graph Construction:

  • Compute similarity matrices between pipelines using Pearson's correlation between statistical maps.
  • Build weighted graphs where nodes represent pipelines and edges represent similarity.

• Community Detection: Apply Louvain algorithm to identify pipelines producing similar results.

• Stability Assessment: Calculate how frequently pipeline pairs cluster together across different subject samples [77].

Protocol 3: Joint Denoising and Artifact Correction

For structural MRI with simultaneous noise and motion artifacts:

• Framework Implementation: Employ Joint image Denoising and Artifact Correction (JDAC) with iterative learning.

• Model Architecture:

  • Adaptive denoising model with noise level estimation using gradient map variance.
  • Anti-artifact model with U-Net architecture and gradient-based loss preservation.

• Iteration Process: Alternating application of denoising and anti-artifact models with early stopping based on noise level estimation [81].

Implementation Tools and Computational Solutions

The Scientist's Toolkit: Essential Research Reagents

Table 5: Key Computational Tools for Pipeline Implementation and Evaluation

Tool/Resource	Function	Application Context
fMRIPrep	Standardized preprocessing	Automated, reproducible pipeline generation [49]
SHAMAN	Motion impact quantification	Trait-specific artifact assessment [39]
Portrait Divergence (PDiv)	Network topology comparison	Whole-network similarity assessment [78]
ENIGMA rsfMRI Pipeline	Multi-site harmonization	Cross-site genetic analyses of connectivity [79]
JDAC Framework	Joint denoising/artifact correction	Structural MRI quality enhancement [81]
Community Detection	Pipeline similarity analysis	Identifying equivalent analytical paths [77]

Visual Guide to Pipeline Selection Logic

Effective fMRI pipeline selection requires careful consideration of the fundamental differences between resting-state and task-based paradigms, with explicit strategies to balance correction efficacy against signal preservation. The empirical evidence demonstrates that optimal pipelines do exist—those that consistently satisfy multiple validation criteria across diverse datasets and time intervals [78]. However, these optimal pipelines must be selected with recognition that no single approach universally excels across all research contexts [49].

Future directions in pipeline optimization should emphasize trait-specific motion impact assessment [39], community-based pipeline evaluation [77], and joint processing frameworks that explicitly model interactions between different artifact types [81]. By adopting these evidence-based frameworks, researchers can navigate the complex analytical space of fMRI processing with greater confidence, producing more reproducible and biologically valid findings that advance our understanding of brain function in health and disease.

Validating Correction Efficacy: Benchmarking Methods and Comparative Performance

Functional magnetic resonance imaging (fMRI) has become a cornerstone technique for investigating human brain function in both research and clinical contexts. However, the blood oxygen level-dependent (BOLD) signal measured by fMRI is vulnerable to confounding artifacts from multiple sources, with head motion representing the most significant technical challenge. Motion artifacts systematically alter fMRI time series data, potentially inducing spurious correlations that can be misinterpreted as neurobiologically meaningful functional connectivity. These artifacts present distinct challenges across different experimental paradigms, particularly when comparing resting-state fMRI with task-based fMRI designs. In resting-state fMRI, where the timing of underlying neural processes is unknown, the data is especially vulnerable to motion artifacts that can mimic or obscure intrinsic brain networks. Understanding, quantifying, and mitigating these motion-related artifacts requires a suite of specialized evaluation metrics, principally Framewise Displacement (FD) and DVARS, along with careful interpretation of functional connectivity measures.

The impact of motion on fMRI data is not merely a technical nuisance but constitutes a substantial threat to validity across numerous research domains. Motion artifacts have been shown to cause systematic biases in functional connectivity patterns, typically decreasing long-distance correlations while increasing short-distance correlations [20]. This distance-dependent effect has led to spurious findings in studies of developmental populations, clinical groups, and aging cohorts where motion may correlate with the variable of interest. For researchers in both academic and pharmaceutical development settings, ensuring that brain-behavior associations reflect genuine neurobiological phenomena rather than motion-induced artifacts is therefore paramount for basic research validity and drug target identification.

Core Motion Quantification Metrics

Framewise Displacement (FD)

Conceptual Foundation and Calculation Framewise Displacement quantifies the total head movement between consecutive brain volumes by aggregating translational and rotational displacement parameters. Derived from the rigid-body realignment parameters estimated during image preprocessing (three translational directions: X, Y, Z; three rotational directions: pitch, yaw, roll), FD represents the spatial magnitude of head movement from one volume to the next.

The standard calculation for FD approximates the cumulative displacement of all brain voxels by summing the absolute values of translational displacements with the rotational displacements converted to linear distance at a given radius from the brain center [20]. A common approach calculates FD using the following formula:

FDₜ = |Δxₜ| + |Δyₜ| + |Δzₜ| + |Δαₜ| × r + |Δβₜ| × r + |Δγₜ| × r

Where Δxₜ, Δyₜ, Δzₜ represent translational changes, and Δαₜ, Δβₜ, Δγₜ represent rotational changes in radians between timepoints t-1 and t. The radius r is typically set at 50 mm, roughly approximating the average distance from the cerebral cortex to the center of the head.

Experimental Application and Thresholding In practical application, FD timecourses are computed for each subject, and volumes exceeding predetermined thresholds are flagged as potentially contaminated. The selection of FD thresholds represents a critical balance between artifact removal and data retention. Common thresholds found in the literature include:

• FD < 0.2 mm: Stringent threshold that significantly reduces motion-related artifacts but may exclude substantial data, particularly in high-motion populations [39]
• FD < 0.3 mm: Moderate threshold shown to be effective in pediatric populations while retaining a majority of participants (83% in one study of first-grade children) [82]
• FD < 0.5 mm: More liberal threshold sometimes used in adult populations with low motion

It is important to note that FD provides a summary statistic of movement between volumes but does not distinguish between qualitatively different movement types (e.g., sudden large movements versus frequent small movements) that may have different artifact profiles [20].

DVARS

Conceptual Foundation and Calculation DVARS (D referring to temporal derivative of timecourses, VARS referring to RMS variance over voxels) is a data-driven metric that quantifies the rate of change of the BOLD signal across the entire brain at each timepoint. Unlike FD, which measures estimated head movement, DVARS captures the actual signal changes in the data, reflecting both motion-induced artifacts and true neural signal fluctuations.

The calculation involves computing the spatial standard deviation of the temporal derivative of the BOLD signal:

DVARSₜ = √[ mean ( (Iₜ - Iₜ₋₁)² ) ]

Where Iₜ represents the brain volume at time t, and the mean is taken across all brain voxels. The resulting metric is typically expressed in units of percentage BOLD signal change.

Interpretation and Thresholding Elevated DVARS values indicate volumes with substantial overall signal change, which may result from head motion, scanner artifacts, or physiological noise. DVARS is often standardized using the median and median absolute deviation across timepoints to facilitate outlier detection. Common thresholding approaches include:

• Dispersion-based thresholds: Typically 2-3 standard deviations above the mean
• Percentile-based thresholds: Often the top 10-20% of values
• Absolute signal change thresholds: Varies by study but often around 0.5% BOLD signal change

DVARS provides complementary information to FD, as it captures signal changes regardless of their origin and can detect artifacts not fully captured by motion parameters [42].

Table 1: Comparative Analysis of Primary Motion Metrics

Metric	Data Source	Primary Unit	Measures	Strengths	Limitations
Framewise Displacement (FD)	Rigid-body realignment parameters	Millimeters (mm)	Estimated head movement between volumes	Direct physical interpretation; standardized calculation	May not capture all spin history effects; threshold selection arbitrary
DVARS	BOLD signal timecourses	% BOLD signal change	Rate of whole-brain signal change	Data-driven; captures actual signal abnormalities	Cannot distinguish motion from neural signal; sensitive to global signal fluctuations
Projection Scrubbing	Multivariate projections (e.g., ICA components)	Statistical outlier measure	Abnormal patterns in dimensionality-reduced data	Statistically principled; combines multiple artifact sources	Computationally intensive; requires specialized implementation

Functional Connectivity Measures and Motion Sensitivity

Traditional Connectivity Metrics

Functional connectivity (FC) is conventionally defined as the "temporal coincidence of spatially distant neurophysiological events" [83], typically operationalized through various statistical dependence measures between regional BOLD timecourses. The most ubiquitous FC measure is Pearson's correlation, which captures linear temporal dependencies between brain regions. However, motion artifacts systematically influence Pearson correlation in distinctive patterns, generally decreasing long-distance correlations while increasing short-distance correlations [20] [39]. This distance-dependent effect arises because motion often causes transient, widespread signal changes that affect distant regions differently than proximal ones.

Beyond Pearson correlation, numerous alternative FC measures exhibit varying sensitivity to motion artifacts:

• Cross-correlation: Measures linear dependence at different temporal lags
• Coherence: Quantifies frequency-based synchronization
• Wavelet coherence: Captures time-frequency dependencies
• Mutual information: Detects non-linear statistical dependencies
• Distance metrics (Euclidean, cityblock): Quantify signal dissimilarity
• Time-warping measures (dynamic time warping): Capture similar temporal patterns with phase differences

Research indicates that no single FC measure consistently outperforms others across all contexts, and each may be differentially sensitive to motion artifacts [84]. A composite approach utilizing multiple complementary measures may provide more robust characterization of true functional connections.

Motion Impact on Connectivity Interpretation

The systematic relationship between motion and spurious connectivity presents particular challenges for studies comparing populations with differential motion characteristics (e.g., children vs. adults, clinical vs. control groups). Even after standard denoising procedures, residual motion artifacts can significantly impact brain-behavior associations. Recent research using the ABCD dataset demonstrated that after standard denoising with ABCD-BIDS processing, 23% of signal variance was still explained by head motion [39]. Furthermore, the motion-FC effect matrix showed a strong negative correlation (Spearman ρ = -0.58) with the average FC matrix, indicating that participants who moved more had systematically weaker connections across most brain connections.

The Split Half Analysis of Motion Associated Networks (SHAMAN) framework has been developed to quantify trait-specific motion impact, distinguishing between motion causing overestimation or underestimation of trait-FC relationships [39]. Applying this method to 45 traits in the ABCD study revealed that after standard denoising without motion censoring, 42% of traits had significant motion overestimation scores and 38% had significant underestimation scores. Stringent censoring at FD < 0.2 mm reduced significant overestimation to just 2% of traits but did not decrease the number with significant underestimation scores, indicating complex, trait-specific motion effects.

Table 2: Motion Artifact Characteristics Across fMRI Paradigms

Characteristic	Resting-State fMRI	Task-Based fMRI
Primary Motion Vulnerability	Inter-regional correlations throughout scan	Trial-related activity and condition subtractions
Typical Motion Correction Approach	Global signal regression, motion censoring, ICA-based denoising	Regression of motion parameters, inclusion as trial regressors
Connectivity Analysis Method	Seed-based, ICA, graph theory	Psychophysiological interactions, beta-series correlation, task-based connectivity
Impact on Brain-Behavior Associations	High vulnerability to spurious correlations	Less vulnerable but task-connectivity analyses still affected
Optimal Scrubbing Strategy	Data-driven scrubbing (projection scrubbing, DVARS)	Trial removal based on movement during trial epochs

Differential Motion Effects in Resting-State vs Task fMRI

Resting-State fMRI Vulnerabilities

Resting-state functional connectivity MRI (rs-fcMRI) exhibits particular vulnerability to motion artifacts due to its reliance on spontaneous low-frequency fluctuations in the BOLD signal. The absence of structured task timing means that motion-related signal changes can be easily misattributed to intrinsic brain activity. Power et al. (2011) demonstrated that subject motion produces substantial changes in rs-fcMRI timecourses despite spatial registration and regression of motion estimates, causing "systematic but spurious correlation structures" [20].

These motion-induced correlations display distinctive spatial patterns: reduced long-distance functional connections (e.g., between prefrontal cortex and posterior regions) and enhanced short-distance connections (e.g., within sensory cortices). This pattern notably resembles certain neurodevelopmental trajectories and psychopathological alterations, creating potential for spurious findings in developmental and clinical research. For example, early studies interpreting reduced long-distance connectivity in autism as a neurobiological characteristic failed to adequately account for increased head motion in the autistic participants [39].

Task-Based fMRI Considerations

Task-based fMRI analyses are generally less sensitive to motion artifacts than resting-state approaches, primarily because they typically rely on aggregations of trial-related activity and subtractions between conditions rather than inter-regional correlations [45]. However, specific task-based connectivity methods—including psychophysiological interactions (PPI) and beta-series correlation—remain vulnerable to motion confounds.

Interestingly, task fMRI can provide unique opportunities to investigate motion-resistant aspects of brain organization. The similarity in functional connectivity between rest and task states (interpreted as reconfiguration efficiency) has been shown to relate to cognitive performance and clinical symptoms [85]. This similarity metric may be less susceptible to motion artifacts than absolute connectivity strength, as motion tends to affect rest and task states similarly within individuals.

Advanced Motion Mitigation Methodologies

Motion Regression Techniques

Nuisance regression represents the most common approach for motion correction in fMRI, utilizing various parameterizations of head motion estimates as regressors:

• 6-parameter model: Basic approach using three translational and three rotational parameters
• 12-parameter model: Adds temporal derivatives of motion parameters
• 24-parameter model: Includes squares of parameters and their derivatives (Friston et al., 1996)
• 36-parameter model: Further expands with delayed and averaged motion parameters

Recent advances have introduced more sophisticated regression approaches, including Convolutional Neural Network (CNN) models that non-parametrically learn the relationship between motion parameters and BOLD signal changes [41]. These methods demonstrate superior motion artifact reduction compared to traditional expansions while avoiding excessive removal of neural signals.

Data-Driven Scrubbing Methods

Volume censoring (scrubbing) removes motion-contaminated volumes rather than attempting to statistically model their effects. While traditional motion scrubbing uses FD and/or DVARS thresholds, emerging data-driven approaches offer advantages:

Projection Scrubbing This statistically principled method identifies artifactual volumes by detecting outliers in multivariate projections of the data, such as independent component analysis (ICA) components [42]. Projection scrubbing improves the validity, reliability, and identifiability of functional connectivity compared to motion scrubbing while dramatically increasing data retention and sample sizes by avoiding unnecessary censoring.

ICA-Based Denoising Independent component analysis automatically separates BOLD signals into spatially independent components, allowing for identification and removal of motion-related components [82]. When combined with volume censoring, ICA denoising effectively removes residual motion artifact, making it particularly valuable for high-motion pediatric cohorts.

Integrated Processing Pipelines

Comprehensive motion mitigation typically combines multiple approaches. The ABCD-BIDS pipeline exemplifies this integrated strategy, incorporating global signal regression, respiratory filtering, spectral filtering, despiking, and motion parameter regression [39]. This combination achieves a 69% relative reduction in motion-related signal variance compared to minimal processing alone, though significant motion artifacts persist, necessitating additional censoring for sensitive brain-behavior analyses.

Experimental Protocols for Motion Management

Protocol for High-Motion Pediatric Populations

Studying developmental populations requires specialized approaches to manage characteristically high motion levels:

• Real-time motion monitoring: Visual feedback or automated tracking during scanning [82]
• Behavioral training: Practice sessions in mock scanners to acclimatize children
• Targeted censoring threshold: FD < 0.3 mm balances artifact removal and data retention [82]
• ICA-based denoising: Removes residual motion artifact after censoring
• Concatenation approach: Processes data in segments around high-motion volumes

This protocol enabled retention of 83% of participants in a first-grade cohort despite extreme motion, demonstrating feasibility of obtaining usable resting-state data in challenging populations [82].

Protocol for Motion Impact Assessment

The SHAMAN framework provides a method for quantifying trait-specific motion impact:

• Split half analysis: Divide each participant's data into high-motion and low-motion halves based on median FD
• Trait-FC effect calculation: Compute correlation between trait and FC separately for each half
• Motion impact score: Difference between trait-FC effects in high-motion vs low-motion halves
• Directional interpretation: Positive scores indicate motion overestimation; negative scores indicate underestimation of trait-FC effects
• Permutation testing: Non-parametric combining across connections yields significance values [39]

This protocol enables researchers to identify which specific trait-FC relationships in their analysis are most vulnerable to residual motion confounds.

Research Reagent Solutions

Table 3: Essential Resources for Motion Management in fMRI Research

Resource Category	Specific Tools/Methods	Primary Function	Implementation Considerations
Motion Quantification Software	FSL (FSLMOTION), AFNI (@1dtool.py), SPM (realign)	Calculate FD, DVARS, and other motion metrics	Pipeline integration; threshold selection
Data Scrubbing Tools	Artifact Detection Tools (ART), Nipype (DropTRs)	Identify and remove contaminated volumes	Data retention tradeoffs; interpolation methods
Denoising Algorithms	ICA-AROMA, CompCor, GSR	Remove motion-related variance from BOLD signal	Risk of neural signal removal; validation requirements
Statistical Frameworks	SHAMAN, PRONTO	Quantify motion impact on specific research questions	Trait-specific sensitivity; computational demands
Experimental Controls	Mock scanner training, real-time feedback	Reduce in-scanner motion during acquisition	Participant engagement; feasibility in clinical populations

Evaluation metrics for motion artifacts—particularly Framewise Displacement and DVARS—provide essential tools for protecting the validity of functional connectivity findings in both resting-state and task-based fMRI. The characteristically different vulnerabilities of these experimental paradigms necessitate tailored approaches to motion management. While resting-state fMRI requires aggressive motion correction to prevent spurious correlations, task-based designs benefit from focused attention to trial-specific motion and connectivity analyses.

Emerging methodologies, including data-driven scrubbing and trait-specific motion impact assessment, offer promising avenues for improving the motion robustness of fMRI research. Particularly for brain-behavior association studies and clinical trials measurement, comprehensive motion management must extend beyond standard preprocessing to include rigorous quantification of residual motion effects. By implementing the metrics, methods, and experimental protocols outlined in this technical guide, researchers and drug development professionals can significantly enhance the validity and reproducibility of their functional connectivity findings.

Comparative Performance of Major Correction Pipelines (e.g., aCompCor vs. Global Signal Regression)

Motion artifacts represent a significant source of non-neuronal noise in functional magnetic resonance imaging (fMRI) data, threatening the validity of both resting-state and task-based studies. These artifacts arise from head movements during scanning, which introduce spin history effects and signal changes that can mimic or obscure true neural connectivity and activation patterns [51]. The challenge is particularly pronounced when comparing populations with different movement characteristics, such as children versus adults or clinical populations versus healthy controls [51]. The persistence of motion-related artifacts even after standard realignment procedures has necessitated the development of sophisticated preprocessing pipelines, with global signal regression (GSR) and anatomical CompCor (aCompCor) emerging as two prominent but philosophically distinct approaches [86] [51]. This technical guide provides a comprehensive comparison of these pipelines, framed within the context of motion artifact characteristics across resting-state and task fMRI paradigms.

Pipeline Fundamentals: Philosophical and Methodological Divergence

Global Signal Regression (GSR)

GSR is a preprocessing technique that removes the global signal—the average time course across all voxels in the brain—from each voxel's signal using general linear model (GLM) regression [86] [87].

• Primary Rationale: Effectively removes global artifacts driven by motion, respiration, and cardiac rhythms that affect the entire brain [87].
• Key Controversy: Mathematically introduces negative correlations (anti-correlations) by shifting the correlation distribution, making functional connectivity interpretation challenging [86] [88].
• Neuronal Information Debate: The global signal may contain neurally relevant information related to vigilance and arousal, raising concerns about removing biologically meaningful variance [87].

Anatomical CompCor (aCompCor)

aCompCor is a component-based noise correction method that addresses physiological noise without removing the global signal [86] [51].

• Core Methodology: Estimates nuisance signals from noise regions-of-interest (ROIs)—specifically, eroded white matter (WM) and cerebral spinal fluid (CSF) masks—where BOLD signals are unlikely to be neurally derived. Principal components analysis (PCA) is applied to these regions, and the top components are regressed out from the BOLD time series [86].
• Advantage Over Mean Signal: PCA components better account for voxel-specific phase differences in physiological noise compared to simple averaging of WM and CSF signals [86].
• Motion Correction Efficacy: Effectively mitigates motion-related artifacts by capturing spatially disparate nuisance signals that might cancel each other out when averaged [51].

Table 1: Fundamental Characteristics of GSR and aCompCor Pipelines

Feature	Global Signal Regression (GSR)	Anatomical CompCor (aCompCor)
Core Approach	Regression of whole-brain average signal	Regression of principal components from noise ROIs (WM/CSF)
Primary Noise Target	Global artifacts (motion, respiration)	Physiological noise (cardiac, respiratory) and motion
Signal Domain	Entire brain or gray matter	Anatomically defined noise regions
Effect on Correlation Distribution	Shifts distribution toward negative values; introduces anti-correlations [86]	Preserves native correlation distribution; does not force anti-correlations [86]
Key Strengths	Highly effective at removing global artifacts; strengthens brain-behavior associations in some cases [87]	Avoids mathematical introduction of anti-correlations; improves connectivity specificity [86] [51]
Principal Limitations	Removes potential neural signal; complicates interpretation of negative correlations [87]	May be less effective for global, spatially uniform artifacts [86]

Comparative Experimental Evidence and Performance Metrics

Anticorrelations: Biological Validty vs. Mathematical Artifact

A fundamental debate in resting-state fMRI concerns whether anti-correlations between networks like the default mode network (DMN) and task-positive network (TPN) reflect genuine biological organization or preprocessing artifacts.

• Evidence from GSR Studies: GSR robustly reveals anti-correlations between the DMN and TPN, but these have been questioned as potentially artificial due to the mathematical properties of GSR [86] [88].
• Evidence from aCompCor Studies: Research using aCompCor (without GSR) has also demonstrated robust anticorrelations between a medial prefrontal cortex (MPFC) seed in the DMN and TPN regions. This suggests that anticorrelations are not solely an artifact of GSR and may have a biological basis [86] [88].
• Specificity Findings: The specificity of these anticorrelations was similar between GSR and aCompCor methods. However, aCompCor demonstrated higher specificity and sensitivity for positive correlations compared to GSR [86].

Motion Artifact Reduction and Connectivity Specificity

Comparative studies have directly evaluated the efficacy of these pipelines in mitigating motion artifacts and improving the specificity of functional connectivity estimates.

• Superior Motion Mitigation: aCompCor has been shown to remove motion artifacts more effectively than mean tissue-based signal regression (a common alternative to GSR). The inclusion of more components from noise ROIs better mitigates motion-related artifacts [51].
• Connectivity Specificity: aCompCor improves the specificity of functional connectivity estimates. This has been validated using the known anatomy of the default mode and motor control networks as a ground truth [51].
• Interaction with "Scrubbing": The data-scrubbing technique (removing volumes with excessive motion) provides no additional benefit for motion reduction or connectivity specificity when using aCompCor, whereas it offers improvement when used with mean signal regression [51].

Impact on Brain-Behavior Associations

The utilitarian value of a preprocessing pipeline can be measured by its ability to strengthen associations between functional connectivity and behavior.

• GSR Strengthens Associations: In studies of young healthy adults, applying GSR increased the behavioral variance explained by whole-brain resting-state functional connectivity (RSFC) by an average of 47% across 23 measures in one dataset and 40% across 58 measures in another [87].
• Task fMRI Context: While not directly comparing pipelines, evidence suggests that task-based functional connectivity can capture more behaviorally relevant information than resting-state connectivity, as task paradigms elicit brain states more directly related to the cognitive constructs being measured [46].

Table 2: Quantitative Performance Comparison of GSR and aCompCor

Performance Metric	GSR Findings	aCompCor Findings
Anticorrelation Specificity	Robustly identifies DMN-TPN anticorrelations, but concerns about artificial introduction [86]	Identifies biologically plausible DMN-TPN anticorrelations without forced negative distribution [86] [88]
Positive Correlation Sensitivity	Reduced specificity for positive correlations compared to aCompCor [86]	Higher sensitivity and specificity for positive correlations [86]
Motion Artifact Reduction	Effective at reducing global motion artifacts [87]	More effective than mean signal regression; captures spatially disparate motion effects [51]
Behavioral Variance Explained	Increases RSFC-behavior associations (e.g., +47% variance explained) [87]	Specific quantitative comparisons with GSR on behavior are less commonly reported
Data Scrubbing Benefit	Scrubbing provides additional motion reduction benefit [51]	Scrubbing provides no additional benefit after aCompCor application [51]

Motion Artifact Characteristics in Resting-State vs. Task fMRI

The impact and characteristics of motion artifacts differ meaningfully between resting-state and task fMRI paradigms, influencing pipeline performance.

• Resting-State fMRI: Lacks external timing references, making it highly vulnerable to motion-induced correlations that can mimic intrinsic connectivity. Motion typically inflates short-range connectivity while weakening long-range connectivity [51]. The absence of task structure makes it difficult to distinguish noise from neural signal based on temporal patterns alone.
• Task-Based fMRI: The presence of known task timing (e.g., block or event-related designs) provides a temporal template for expected neural responses, offering some protection against motion confounds. However, task designs can introduce systematic correlations between motion and task conditions (e.g., greater movement during more demanding tasks), creating spurious activation effects [46].

These differences necessitate careful consideration when choosing a preprocessing pipeline. The consensus in the field is moving toward the view that there is no single "right" way to preprocess data, and different approaches may reveal complementary insights about brain function [89].

Experimental Protocols for Pipeline Comparison

To ensure rigorous comparison between correction pipelines, researchers should adopt standardized experimental and processing protocols. The following methodology is synthesized from key studies cited in this guide [86] [51].

Data Acquisition Parameters

• Participants: 15 healthy adults (mean age 37.3 ± 2.4, 9 males) were used in the foundational study [86]. Larger samples (~130 children) provide greater power for motion effects analysis [51].
• Scanner Sequence: Acquire data on a 3T scanner using a T2*-weighted EPI sequence. Example parameters: TR/TE = 2500/30 ms, flip angle = 70°, 47 contiguous 3-mm slices, in-plane resolution 3.05×3.15-mm [51].
• Scan Duration: 10-minute resting-state scan with instructions to "keep your eyes open and think of nothing in particular" [86].

Preprocessing Workflow

• Standard Spatial Preprocessing: Perform slice-time correction, realignment, normalization to standard space (e.g., MNI), and spatial smoothing (e.g., 4-mm kernel) [86].
• Nuisance Regression: Apply either:
  • GSR Pipeline: Regress out the global mean signal, average signals from white matter and CSF ROIs, and 6-24 motion parameters (with derivatives and squares) [86] [90].
  • aCompCor Pipeline: Segment T1-weighted image into WM, CSF, and gray matter. Erode WM and CSF masks by one voxel to minimize partial voluming. Extract principal components from these noise ROIs and regress out the top components (e.g., 5 components each) along with motion parameters [86] [51].
• Temporal Filtering: Apply a band-pass filter (e.g., 0.009-0.08 Hz) to retain low-frequency fluctuations of interest [86].
• Optional Scrubbing: Identify and censor volumes with framewise displacement (FD) exceeding a threshold (e.g., 0.9 mm) [51] [90].

Outcome Measures and Validation

• Framewise Displacement (FD): Quantifies head motion between volumes.
• DVARS: Measures the rate of change of BOLD signal across the entire brain.
• Functional Connectivity Specificity: Assess using known network anatomy (e.g., DMN and motor networks) as a ground truth [51].
• Behavioral Associations: Correlate connectivity measures with behavioral performance or clinical scores [87].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Tools for fMRI Preprocessing and Analysis

Tool/Resource	Function/Purpose	Example Software/Package
fMRI Data Preprocessing	Spatial realignment, normalization, coregistration	SPM, FSL, AFNI
Noise Regression Toolbox	Implementation of GSR, aCompCor, and other denoising methods	CONN toolbox, fMRIPrep, NITRC scripts
Head Motion Estimation	Calculation of framewise displacement (FD) and DVARS	SPM, FSL, custom scripts
Physiological Noise Modeling	Estimation of noise components from WM and CSF	aCompCor implementation (Behzadi et al.)
Connectivity Analysis	Seed-based correlation, independent component analysis (ICA)	CONN, FSL MELODIC
Statistical Analysis	Group-level inference, brain-behavior correlations	SPM, FSL, R, Python (nilearn)

The choice between GSR and aCompCor represents a fundamental trade-off in fMRI preprocessing. GSR effectively removes global artifacts and can strengthen brain-behavior associations but introduces interpretive challenges regarding anti-correlations. aCompCor avoids forced negative correlations and provides superior motion mitigation without removing global neural information, but may be less effective for purely global artifacts. The emerging consensus suggests that no single pipeline reveals the "true" nature of the brain, and different approaches offer complementary insights [89]. The optimal pipeline may depend on the specific research question, participant population, and data acquisition parameters. Future methodological developments will likely focus on adaptive pipelines that selectively leverage the strengths of each approach based on data quality characteristics and specific analytical goals.

Motion artifacts represent a significant methodological challenge in functional magnetic resonance imaging (fMRI), particularly in resting-state functional connectivity (rs-fcMRI) research. Unlike task-based fMRI, where motion can be temporally correlated with stimulus presentation, resting-state fMRI is especially vulnerable to motion artifacts because the timing of underlying neural processes is unknown [2] [39]. These artifacts persist despite extensive denoising efforts and can systematically bias connectivity estimates, particularly affecting the distinction between long-distance and short-distance connections [1] [12]. This technical guide examines the characteristics of residual motion artifacts, their differential impact on connectivity measures, and provides detailed methodologies for their detection and mitigation in the context of neuroimaging research and clinical drug development.

The fundamental challenge lies in the spatial and temporal properties of motion artifacts. Even after applying standard denoising procedures, residual artifacts exhibit a characteristic distance-dependent bias, artificially decreasing long-range connectivity while inflating short-range connectivity [12]. This pattern is particularly problematic for studies investigating populations prone to increased head motion, such as children, older adults, and individuals with neurological or psychiatric disorders, where motion-related artifacts can be misinterpreted as genuine neurobiological phenomena [2] [39].

Motion Artifact Characteristics in Resting-State vs. Task fMRI

Fundamental Differences in Vulnerability

Resting-state fMRI possesses unique vulnerabilities to motion artifacts compared to task-based paradigms. Without the temporal structure provided by task stimuli, rs-fcMRI lacks a straightforward method to distinguish motion-induced signal changes from spontaneous neural fluctuations [2]. This fundamental difference necessitates specialized approaches for residual artifact detection in resting-state analyses.

Table 1: Comparative Characteristics of Motion Artifacts in Resting-State vs. Task fMRI

Characteristic	Resting-State fMRI	Task-Based fMRI
Temporal Correlation with Confounds	Uncorrelated with designed experimental paradigm	Often correlated with task performance, creating confusion with true activation [2]
Primary Manifestation	Systematic bias in functional connectivity networks [1]	Distortion of activation maps and statistical inference [2]
Spatial Pattern	Decreased long-distance connectivity, increased short-distance connectivity [12]	Region-specific artifacts dependent on motion type and direction
Frequency Content	Not band-limited; affects standard frequency bands (0.01-0.1 Hz) [2]	Can be more readily separated from stimulus-frequency responses
Denoising Challenges	Lack of reference model for neural activity; greater reliance on data-driven approaches [39]	Task reference model available but can be contaminated by motion-correlated artifacts

Spatial Distribution of Motion Artifacts

The spatial distribution of motion artifacts follows consistent patterns rooted in biomechanical constraints. Research demonstrates that motion is minimal near the atlas vertebrae (where the skull attaches to the neck) and increases with distance from this pivot point [1]. Furthermore, frontal cortex regions typically experience higher motion magnitudes, likely due to the predominance of y-axis rotation (nodding movement) [1]. This spatial heterogeneity results in region-specific vulnerabilities to artifact contamination.

Motion causes characteristic signal changes across the brain: a substantial drop in signal intensity across the entire brain parenchyma immediately following movement events, while areas at the brain edge demonstrate large signal increases due to partial volume effects [1]. These global effects lead to increased image smoothness, which must be accounted for in post-processing strategies [1].

Quantitative Effects on Connectivity Measures

Distance-Dependent Bias Patterns

The most consistently reported effect of motion artifacts in functional connectivity is a distance-dependent bias. This systematic artifact manifests as spurious inflation of short-range connections and attenuation of long-range connections [12]. The strength of this effect varies across brain networks, with default mode and retrosplenial temporal sub-networks showing particularly high vulnerability [5].

Table 2: Quantitative Effects of Motion on Functional Connectivity Measures

Connectivity Measure	Sensitivity to Motion	Residual Motion Correlation After Denoising	Test-Retest Reliability	Key Characteristics
Full Correlation	High	Relatively high residual distance-dependent relationship with motion [5]	High [5]	Most commonly used but highly motion-sensitive
Partial Correlation	Low	Low sensitivity to motion artifact [5]	Low to intermediate [5]	Reduces spurious connections but decreases reliability
Coherence	Intermediate	Moderate sensitivity to motion [5]	Intermediate	Frequency-domain approach
Information Theory Measures	Low	Low sensitivity to motion artifact [5]	Varies by specific measure	Non-linear approach
aCompCor	Low	Effectively attenuates motion artifacts [12]	Intermediate	PCA-based approach outperforms mean signal regression

Magnitude of Motion Effects in Large-Scale Studies

Recent large-scale studies have quantified the substantial impact of residual motion artifacts even after extensive denoising. Analysis of the Adolescent Brain Cognitive Development (ABCD) Study dataset (n = 9,652) revealed that after minimal processing (motion correction by frame realignment only), 73% of signal variance was explained by head motion [39]. After comprehensive denoising with the ABCD-BIDS pipeline (including global signal regression, respiratory filtering, motion timeseries regression, and despiking), 23% of signal variance remained explained by head motion, representing a 69% relative reduction but still substantial residual contamination [39].

The motion-FC effect matrix shows a strong negative correlation (Spearman ρ = -0.58) with the average FC matrix, indicating that connection strength is systematically weaker in participants who move more [39]. This effect persists even after aggressive motion censoring at framewise displacement (FD) < 0.2 mm (Spearman ρ = -0.51) [39]. Critically, the decrease in FC due to head motion often exceeds trait-related FC effect sizes, potentially obscuring or mimicking genuine neurobiological relationships [39].

Experimental Protocols for Residual Artifact Detection

Motion Impact Assessment Using SHAMAN Framework

The Split Half Analysis of Motion Associated Networks (SHAMAN) framework provides a robust method for quantifying trait-specific motion impact on functional connectivity [39]. This approach capitalizes on the stability of traits over time while motion varies from second to second.

Protocol Steps:

• Data Preparation: Process resting-state fMRI data through standard denoising pipelines (e.g., ABCD-BIDS including global signal regression, respiratory filtering, spectral filtering, despiking, and motion parameter regression) [39].
• Framewise Displacement Calculation: Compute Framewise Displacement (FD) using the Jenkinson formulation, which aligns best with voxel-specific measures of displacement [1].
• Split-Half Partitioning: For each participant, split the fMRI timeseries into high-motion and low-motion halves based on median FD.
• Connectivity Calculation: Compute separate functional connectivity matrices for high-motion and low-motion halves.
• Trait-FC Effect Estimation: Calculate the correlation between a trait of interest and FC for each half.
• Motion Impact Score Computation:
  • Compute difference in trait-FC effects between high-motion and low-motion halves
  • Aligned direction with trait-FC effect indicates motion overestimation score
  • Opposite direction indicates motion underestimation score
• Statistical Testing: Use permutation testing (e.g., 1,000 permutations) and non-parametric combining across connections to establish significance [39].

aCompCor-Based Artifact Detection Protocol

The anatomical Component-Based Noise Correction (aCompCor) method provides an effective alternative to mean signal regression for mitigating motion artifacts [12].

Protocol Steps:

• Tissue Mask Definition: Create masks of white matter (WM) and cerebrospinal fluid (CSF) using anatomical segmentation.
• Noise Component Extraction:
  • Extract multiple principal components (typically 5) from WM and CSF regions separately
  • Avoid global signal regression to prevent potential introduction of negative correlations [12]
• Nuisance Regression: Include realignment parameters, their derivatives, quadratic terms, and aCompCor components in a multiple regression model to clean BOLD signals.
• Connectivity Specificity Assessment: Evaluate the specificity of functional connectivity using known network templates (e.g., default mode network, motor control network).
• Performance Validation: Assess the framewise relationship between head motion (FD) and signal change (DVARS) to quantify residual artifacts [12].

Distance-Dependent Correlation Analysis

This method systematically evaluates the relationship between motion and connectivity as a function of the distance between brain regions.

Protocol Steps:

• Parcellation: Apply a standardized brain atlas to define regions of interest.
• Distance Calculation: Compute the Euclidean distance between centroid coordinates of each region pair.
• Motion-FC Correlation: Calculate the correlation between subject-specific motion (mean FD) and functional connectivity strength for each connection.
• Bin Analysis: Group connections into bins based on distance and compute the average motion-FC correlation within each bin.
• Significance Testing: Use permutation tests to evaluate whether the distance-dependent relationship exceeds chance levels.
• Pipeline Comparison: Compare different processing pipelines based on their ability to minimize the distance-dependent relationship [12].

Visualization Frameworks

Figure 1: Motion Artifact Effects on Connectivity. This diagram illustrates the pathway from head motion through various artifact mechanisms to the characteristic pattern of artificially inflated short-range connectivity and attenuated long-range connectivity.

Figure 2: Residual Artifact Detection Workflow. This workflow illustrates the comprehensive pipeline from raw data preprocessing through specialized residual artifact detection using the SHAMAN framework.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Research Toolkit for Residual Artifact Detection

Tool/Reagent	Function/Application	Specifications/Protocol Details
Framewise Displacement (FD)	Quantifies volume-to-volume head motion [1]	Use Jenkinson formulation (FSL) for voxel-specific accuracy [1]
aCompCor	Principal components noise correction from WM/CSF [12]	Extract 5 PCA components each from WM and CSF masks; more effective than mean signal regression [12]
SHAMAN Framework	Quantifies trait-specific motion impact [39]	Split-half analysis comparing high/low motion segments; provides over/under-estimation scores [39]
Distance-Dependent Analysis	Measures motion-connectivity relationship vs. distance [12]	Bin connections by Euclidean distance; compute motion-FC correlation per bin [12]
Global Signal Regression	Removes global motion-related fluctuations [1]	Controversial: can introduce negative correlations but effective for motion reduction [1]
Motion Censoring (Scrubbing)	Removes high-motion volumes from analysis [39]	Threshold recommendation: FD < 0.2 mm; balances artifact reduction vs. data retention [39]
Non-Linear Motion Parameters	Accounts for delayed and spin-history effects [1]	Include derivatives and quadratic terms of 6 realignment parameters [1]
Partial Correlation	Alternative connectivity measure less sensitive to motion [5]	Reduces spurious connections; shows lower motion sensitivity than full correlation [5]

Implications for Clinical Research and Drug Development

The impact of residual motion artifacts extends beyond methodological considerations to directly affect clinical trial design and neuropharmacological outcome measures. In longitudinal treatment studies, motion-related connectivity changes can be misinterpreted as drug effects, particularly when medications influence movement patterns (e.g., sedating agents, stimulants, or treatments for motor symptoms) [39]. The SHAMAN framework enables researchers to quantify and account for these motion-related confounds, protecting against false positive findings in brain-behavior association studies [39].

Recent evidence indicates that even with state-of-the-art denoising approaches, residual motion artifacts continue to impact a substantial proportion of trait-FC relationships. Analysis of 45 traits in the ABCD study revealed that after standard denoising without motion censoring, 42% (19/45) of traits had significant motion overestimation scores and 38% (17/45) had significant underestimation scores [39]. Aggressive censoring (FD < 0.2 mm) reduced significant overestimation to 2% (1/45) but did not decrease the number of traits with significant motion underestimation scores, highlighting the complex relationship between motion correction and trait-specific effects [39].

For drug development professionals, incorporating rigorous residual artifact detection protocols is particularly crucial when studying populations with inherent motion differences, such as neurodevelopmental disorders or neurodegenerative conditions. Implementing the experimental frameworks outlined in this guide can strengthen the validity of functional connectivity as a biomarker for treatment response and target engagement.

Motion artifacts represent one of the most significant methodological challenges in functional magnetic resonance imaging (fMRI), potentially introducing systematic biases that can compromise the validity of research findings and clinical applications. The persistence of motion-related confounds necessitates robust validation frameworks encompassing simulated data, test-retest reliability assessments, and biological plausibility checks. This technical guide examines these interconnected validation pillars within the context of motion artifact characteristics across resting-state and task-based fMRI paradigms. As motion exhibits differential impacts across these modalities—with resting-state fMRI potentially more susceptible to certain motion-related confounds due to the absence of task structure—understanding these distinctions becomes paramount for researchers, scientists, and drug development professionals relying on neuroimaging biomarkers. The validation approaches discussed herein provide essential tools for ensuring that observed effects reflect genuine neurobiological phenomena rather than motion-induced artifacts, thereby enhancing the rigor and reproducibility of neuroimaging research.

Core Validation Frameworks

Simulated Data for Ground Truth Validation

Simulated datasets provide critical opportunities for method validation by enabling researchers to test analytical approaches against known ground truth. A recent multisite effort generated simulated longitudinal neurodevelopmental datasets to address fundamental challenges in brain-behavior research [91]. This collaborative project involved five independent research groups creating simulated datasets embedding their assumptions about brain development, cognition, and behavior, producing 15 datasets totaling 150,000 participants with seven longitudinal waves spanning ages 7-20 years.

The simulation methodology incorporated realistic developmental trajectories derived from established normative growth charts and large-scale datasets like the ABCD Study [91]. Key brain measures included intracranial volume, total gray and white matter volume, hippocampal and amygdala volumes, and frontal lobe cortical thickness, alongside cognitive and behavioral measures. This approach allows researchers to:

• Test statistical models against known underlying patterns
• Identify false positives and false negatives in analytical methods
• Understand how analytical choices affect results when ground truth is known
• Examine biases and assumptions in brain-behavior relationship modeling

Table 1: Simulated Dataset Specifications from Multisite Collaboration

Characteristic	Specification
Number of datasets	15
Total participants	150,000
Timepoints per subject	7 longitudinal waves
Age range	7-20 years
Brain measures	Intracranial volume, gray/white matter volumes, hippocampal/amygdala volumes, frontal cortical thickness
Behavioral measures	IQ, internalizing/externalizing symptoms, attention problems
Key features	Known ground truth, realistic developmental trajectories, missing data structures

Test-Retest Reliability Assessment

Test-retest reliability quantifies the consistency of measurements across multiple testing sessions, providing critical information about the temporal stability of biomarkers and experimental measures. Poor reliability places an upper limit on the ability to detect true effects in experimental and clinical studies.

Recent investigations have revealed concerning reliability patterns in computationally-derived measures. In a study examining the test-retest reliability of a computational assay probing advice-taking under volatility, researchers found largely poor reliability (intra-class correlation coefficient or ICC < 0.5) for both behavioral and computational measures [92]. Similarly, research on affective bias tasks found reliabilities ranging from poor to moderate (ICC: 0.18-0.49) for standard summary statistics, with computational model parameters often demonstrating even lower reliability than summary statistics [93].

Key factors influencing test-retest reliability include:

• Intrinsic measurement noise: Parameter recovery analyses indicate substantial impact on reliability [92]
• Practice effects: Performance changes between sessions due to familiarity with tasks
• State-like fluctuations: Unexplained within-subject variance across testing sessions
• Model over-parameterization: Excessively complex models may fit noise rather than signal

Table 2: Test-Retest Reliability Findings Across Methodologies

Methodology	Reliability Range	Key Findings
Computational measures of advice-taking	ICC < 0.5 (Poor)	Parameter recovery limitations substantially impact reliability
Affective bias task summary statistics	ICC: 0.18-0.49 (Poor-Moderate)	Standard behavioral measures show limited reliability
Affective bias computational parameters	Below summary statistics	Model parameters often less reliable than behavioral measures
Embedding session covariance in generative models	Improved reliability	Accounting for between-session variance enhances stability estimates

Biological Plausibility and Face Validity

Biological plausibility establishes whether findings align with established neurobiological principles, while face validity assesses whether a measurement or simulation appears to measure what it intends to measure based on superficial characteristics. For virtual reality (VR) simulations, face validity represents whether the simulation "looks and feels realistic" from the user's perspective [94].

The concept of plausibility illusion in VR research complements biological plausibility, referring to the illusion that the depicted scenario is actually occurring [94]. When successful, this illusion prompts users to behave in VR as they would in similar real-world circumstances, which is crucial for both experimentation and training applications.

Validation of biological plausibility often involves:

• Establishing correspondence between key elements of real and virtual tasks
• Ensuring realistic kinematic and depth information rather than superficial visual realism
• Demonstrating that simulations elicit realistic emotional, physiological, and behavioral responses
• Verifying that neural activation patterns align with established neurobiological principles

Motion Artifact Characteristics in Resting-State vs Task fMRI

Fundamental Differences in Motion Artifact Impact

Head motion systematically alters correlations in functional connectivity fMRI, with particularly pronounced effects in resting-state studies [1] [3]. The characteristics and impacts of motion artifacts differ substantially between resting-state and task fMRI paradigms:

Resting-State fMRI Vulnerabilities:

• Absence of task structure may increase motion propensity [2]
• Motion introduces distance-dependent correlations, spuriously increasing short-distance connectivity while decreasing long-distance connectivity [1] [3]
• Even submillimeter motions can distort functional connectivity estimates [2]
• Motion artifacts exhibit complex spatial distributions, with anterior brain regions typically more affected due to biomechanical constraints [1]

Task fMRI Considerations:

• Motion can be temporally correlated with task performance, creating confounds that are difficult to distinguish from true activation [2]
• Task engagement may suppress overall motion but introduce motion tied to specific task components
• Blocked designs may allow for censoring of motion-contaminated periods without complete data loss

Motion Characteristics and Measurement

Head motion is typically parameterized using six degrees of freedom (three translations and three rotations), with Framewise Displacement (FD) serving as a common summary metric [1]. Key motion characteristics include:

• Spatial Distribution: Motion is minimal near the atlas vertebrae (where the skull attaches to the neck) and increases with distance from this anchor point [1]. Anterior frontal and orbitofrontal areas show greater displacement than posterior regions.
• Temporal Properties: Motion produces characteristic signal changes including sharp intensity drops immediately following movement events and longer-duration artifacts persisting up to 8-10 seconds [1] [3].
• Frequency Content: Motion artifacts do not display band-limited frequency content, making frequency filtering ineffective for complete correction [2].

Systematic Bias in Group Comparisons

Motion artifacts introduce particularly problematic systematic biases in studies comparing groups with different motion characteristics. Multiple studies have demonstrated that patient populations, older adults, and pediatric subjects exhibit larger motions compared to young healthy adults [2] [1]. This differential motion creates spurious group differences that can be misinterpreted as neurobiological effects, potentially invalidating findings in developmental, clinical, and aging research.

Methodological Approaches for Motion Artifact Mitigation

Preprocessing and Correction Strategies

Numerous strategies have been developed to mitigate motion artifacts in fMRI data, each with distinct strengths and limitations:

Prospective Motion Correction:

• Subject instruction, training, and mild restraints during scanning [2]
• Real-time motion correction systems that adjust scanning parameters during acquisition

Retrospective Correction Methods:

• Nuisance regression: Including motion parameters as regressors of no interest
• Global signal regression: Effectively reduces motion-related variance but may introduce other biases [3]
• Volume censoring (scrubbing): Removing motion-contaminated volumes from analysis [3]
• ICA-based denoising: Identifying and removing motion-related components

Table 3: Motion Correction Methods and Efficacy

Method	Mechanism	Advantages	Limitations
Nuisance regression	Models motion as linear confound	Simple implementation, standard in pipelines	Limited efficacy for nonlinear effects
Global signal regression	Removes widespread signal changes	Highly effective for motion reduction	May remove neural signal, introduces negative correlations
Volume censoring	Removes contaminated volumes	Preserves uncontaminated data	Reduces data, requires careful threshold selection
ICA-based denoising	Identifies motion components	Data-driven, can capture complex artifacts	Component classification challenges
Real-time correction	Adjusts acquisition during scan	Prevents artifacts at source	Limited availability, technical complexity

Validation of Correction Efficacy

Establishing the effectiveness of motion correction strategies requires multiple validation approaches:

• QC metric improvement: Demonstrating reduction in motion-related quality control metrics
• Distance-dependent correlation reduction: Showing decreased spurious short-distance connections
• Group difference normalization: Reduction of motion-related group differences to chance levels [3]
• Comparison to ground truth: Using simulated data with known motion artifacts to test correction methods

Research indicates that combining multiple correction strategies typically yields the best results. For instance, integrating volume censoring with global signal regression and interpolation has been shown to effectively reduce motion-related group differences to chance levels [3].

Integrated Validation Framework

Comprehensive Validation Workflow

A robust validation framework integrates simulated data, test-retest reliability, and biological plausibility assessments within a comprehensive workflow. This integrated approach enables researchers to address motion artifact challenges at multiple levels:

Table 4: Essential Research Tools for Validation Studies

Tool/Resource	Function	Application in Validation
Simulated datasets with ground truth	Method testing and benchmarking	Evaluating analytical performance when truth is known [91]
Framewise Displacement (FD) metrics	Motion quantification	Standardized motion measurement across studies [1]
Computational modeling frameworks	Parameter estimation from behavior	Extracting latent cognitive variables from task performance [92] [93]
Multi-echo fMRI sequences	Enhanced artifact removal	Improving BOLD signal isolation from motion-related artifacts
Physiological recording equipment	Monitoring cardiac/respiratory signals	Modeling physiological confounds alongside motion [2]
Quality control pipelines	Data quality assessment	Standardized quality evaluation across datasets and sites
Intraclass correlation coefficient (ICC)	Reliability quantification	Measuring test-retest reliability of measures [92] [93]

Robust validation frameworks encompassing simulated data, test-retest reliability assessment, and biological plausibility checks are essential for advancing neuroimaging research amid the persistent challenges posed by motion artifacts. The differential impact of motion across resting-state and task fMRI paradigms necessitates tailored approaches for each modality. By integrating these validation pillars within a comprehensive framework and utilizing the methodological tools outlined in this guide, researchers can enhance the rigor, reproducibility, and interpretability of neuroimaging findings, ultimately strengthening the foundation for both basic cognitive neuroscience and clinical applications in drug development.

Functional magnetic resonance imaging (fMRI) has become an indispensable tool for probing brain function in health and disease. However, a persistent challenge complicates the interpretation of fMRI data: head motion. Even sub-millimeter motions introduce complex artifacts that can compromise data quality and lead to spurious findings [2] [9]. This challenge is particularly acute when studying behavioral phenotypes, where motion is often systematically correlated with the very demographic, clinical, or behavioral variables of interest [1] [4]. This technical guide examines the core problem of motion artifact in fMRI research, comparing its effects across resting-state and task paradigms. It further explores the critical challenge of disentangling true neural signals from motion-induced confounds, particularly when motion itself is a predictable biomarker of population characteristics.

Motion Artifact Characteristics: A Comparison of Resting-State and Task fMRI

Head motion during fMRI scanning introduces artifacts through multiple physical mechanisms, including spin-history effects, partial volume effects, and disruptions of the steady-state assumption in rapid EPI acquisitions [4]. While a pervasive problem, the impact and manifestation of motion can differ between resting-state and task-based fMRI.

Table 1: Characteristics of Motion Artifact in Resting-State vs. Task fMRI

Feature	Resting-State (rs-fMRI)	Task-Based fMRI
Primary Impact	Introduces distance-dependent spurious correlations in functional connectivity [2] [9].	Can mimic or obscure true task-related BOLD activation, leading to false positives or negatives [4].
Spatial Pattern	Increased short-range connectivity; decreased long-range connectivity [2] [9].	Artifacts localized to brain regions with greatest displacement (e.g., frontal regions) [2].
Temporal Correlation	Motion artifacts are often temporally structured but not tied to an external paradigm [95].	Critically, motion can be correlated with the task timing (e.g., moving in response to a stimulus) [4].
Typical Correction	Global signal regression, volume censoring ("scrubbing"), ICA-based denoising [2] [95].	Inclusion of motion parameters as regressors in general linear model [4].

The following diagram illustrates the pathways through which motion confounds the relationship between demographic variables and fMRI-derived behavioral phenotypes.

Quantitative Evidence: Demographic Predictors of In-Scanner Motion

Understanding which subject factors predict motion is essential for designing studies and interpreting results. A large-scale study of the UK Biobank cohort (n=40,969) provides robust quantitative data on the association between various indicators and fMRI head motion [96].

Table 2: Association of Demographic and Clinical Factors with fMRI Head Motion (Adapted from [96])

Motion Indicator	Association with Motion (Standardized Beta)	Statistical Significance (p-value)	Key Finding
Body Mass Index (BMI)	β = .050	p < .001	Strongest indicator; a 10-point BMI increase linked to 51% more motion [96].
Ethnicity	β = .068	p < .001	A significant indicator, though cohort was 97.6% White-British [96].
Hypertension	β = N/R	p = 0.048	Significantly increased motion in this subgroup [96].
Psychiatric Disorders	β = N/S	Not Significant	No significant motion increase in this broad category [96].
Cognitive Task Performance	t = 110.83	p < .001	Associated with increased head motion [96].
Prior Scan Experience	t = 7.16	p < .001	Associated with increased head motion [96].

Conversely, some assumed motion indicators, such as a history of smoking, caffeine consumption, and age, were found to be relatively less influential in this large cohort [96]. It is critical to note that these relationships can vary by population. For instance, a pediatric study found an interaction between age and neurodevelopmental diagnoses, where typical age-related decreases in head motion were attenuated in children with such disorders [97].

Experimental Protocols for Motion Detection and Correction

A multi-stage, rigorous methodology is required to detect, characterize, and mitigate motion artifacts. The following workflow details a comprehensive protocol derived from current literature [96] [1] [95].

Quantification and Detection Protocol

• Step 1: Realignment Parameter Estimation. For each subject's fMRI time series, rigid-body realignment (e.g., using FSL's MCFLIRT or SPM) estimates six head position parameters (3 translations, 3 rotations) for each volume relative to a reference volume [1] [4].
• Step 2: Calculate Framewise Displacement (FD). Compute a scalar summary metric of volume-to-volume head motion. The Jenkinson formulation of FD (derived from the RMS of the six realignment parameters) is recommended for its alignment with voxel-specific displacement [1].
• Step 3: Identify Compromised Volumes. Set a threshold for "excessive" motion, typically at FD > 0.2 mm - 0.5 mm. Volumes exceeding this threshold, as well as the subsequent volume (to account for spin-history effects), are flagged for censoring [95].

Correction and Mitigation Protocol

• Step 4: Retrospective Correction via Nuisance Regression. Incorporate the 6 realignment parameters, their temporal derivatives, and their squared terms (24 regressors total) into a general linear model to regress out motion-related variance from the BOLD signal [1] [95].
• Step 5: Global Signal Regression (GSR). While controversial as it removes globally shared neural signal, GSR is highly effective at reducing motion-related artifacts, particularly the distance-dependent bias in resting-state connectivity [1] [95].
• Step 6: Volume Censoring ("Scrubbing"). Remove (censor) the volumes flagged in Step 3 from subsequent connectivity or activation analyses. This is a powerful method for eliminating the influence of high-motion time points [95].
• Step 7: Prospective Correction. Emerging techniques use external tracking systems (e.g., Moiré Phase Tracking, camera-based systems) to update the scanner's field of view or slice position in real-time during acquisition, effectively "chasing" the moving head [2] [4].

The diagram below synthesizes this multi-layered protocol into a standard experimental workflow.

The Scientist's Toolkit: Essential Materials and Reagents

Table 3: Key Research Reagent Solutions for Motion-Corrected fMRI Studies

Tool / Resource	Function / Purpose	Example Software / Implementation
Realignment Algorithm	Estimates 3 translational and 3 rotational head motion parameters for each volume in the fMRI time series.	FSL MCFLIRT, SPM Realign, AFNI 3dVolreg [1].
Framewise Displacement (FD) Calculator	Computes a scalar summary metric of volume-to-volume head motion from realignment parameters for quality control and censoring.	FSL (fsl_motion_outliers), in-house scripts based on Jenkinson et al. formulation [1].
Nuisance Regression Tool	Removes variance associated with motion parameters and other confounds (e.g., white matter, CSF signals) from the BOLD data.	FSL FEAT, SPM, CONN toolbox, AFNI 3dTproject [1] [95].
Volume Censoring ("Scrubbing") Tool	Identifies and removes motion-corrupted volumes from analysis to prevent artifacts from influencing results.	FSL (fsl_motion_outliers), Power et al. (2012) method [95].
Prospective Motion Correction System	Tracks head position in real-time and updates the scanner's imaging plane to compensate for motion during data acquisition.	Moiré Phase Tracking, camera-based systems (e.g., Metria Innovation), cloverleaf navigators [2] [4].

Implications for Drug Development and Clinical Research

The challenge of motion artifact is particularly salient in clinical trials and pharmacological fMRI (phMRI), where fMRI is used as a potential biomarker for drug efficacy and target engagement [98]. Systematic differences in motion between patient and control groups, or changes in motion pre- and post-treatment, can create spurious drug effects or mask real ones [1]. This is a critical concern in emerging fields like psychedelic research, where the acute administration of a substance may directly increase head motion, creating a confound that is inextricably linked to the intervention itself if not properly controlled [99]. Regulatory agencies like the FDA and EMA require high standards of evidence for biomarker qualification, which includes demonstrating that measured effects are not driven by technical confounds like motion [98]. Therefore, the implementation of rigorous motion correction protocols, as outlined in this guide, is not merely a methodological preference but a fundamental requirement for generating interpretable and translatable fMRI results in drug development.

Head motion remains one of the most significant methodological challenges in fMRI research. Its effects are complex, varying between resting-state and task-based paradigms, and are often collinear with demographic and clinical variables of interest. As large-scale studies have shown, factors like BMI can be more predictive of motion than clinical diagnoses. The path forward requires a multi-pronged approach: a thorough understanding of motion's biomechanics and spatiotemporal characteristics, the rigorous application of combined retrospective correction methods like nuisance regression and censoring, and the future adoption of prospective real-time correction technologies. For researchers investigating behavioral phenotypes, acknowledging and robustly addressing the "Predictive Value Challenge" is essential to ensure that findings reflect the true neurobiology of the brain, rather than the confounding movement of the head.

Conclusion

Motion artifacts present distinct yet equally critical challenges in both resting-state and task-based fMRI, with the fundamental difference lying in the potential for temporal correlation with experimental design in task paradigms. Effective mitigation requires a multi-pronged approach combining prospective subject preparation with sophisticated retrospective processing pipelines. While methods like realignment parameter regression, censoring, aCompCor, and novel computational approaches show significant efficacy, no single technique completely eliminates motion-related bias, particularly in studies involving populations with different motion characteristics. Future directions should focus on developing standardized reporting metrics for motion-related quality control, optimizing real-time correction technologies, and establishing consensus pipelines for specific experimental contexts. For drug development and clinical research, rigorous motion correction is essential for ensuring the validity of functional connectivity and activation findings, particularly in longitudinal trials and studies involving patient populations prone to movement.

References

• 1. Motion artifact in studies of functional connectivity [https://pmc.ncbi.nlm.nih.gov/articles/PMC5930165/]
• 2. Resting State fMRI: Going Through the Motions [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2019.00825/full]
• 3. Methods to detect, characterize, and remove motion artifact in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3849338/]
• 4. Functional MRI [https://www.sciencedirect.com/science/article/abs/pii/B9780128244609000212]
• 5. Evaluating the sensitivity of functional connectivity ... [https://www.sciencedirect.com/science/article/pii/S1053811921006832]
• 6. Automated quantification and evaluation of motion artifact ... [https://pubmed.ncbi.nlm.nih.gov/30339290/]
• 7. Motion Artifact Quantification and Sensor Fusion for ... [https://www.mdpi.com/1424-8220/18/1/38]
• 8. Reducing the effects of motion artifacts in fMRI: A structured ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8934405/]
• 9. Resting State fMRI: Going Through the Motions - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6700228/]
• 10. Characterizing and Differentiating Task-based and Resting ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4559495/]
• 11. Cost Efficiency of fMRI Studies Using Resting‐State Vs. Task ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12182254/]
• 12. Reduction of motion-related artifacts in resting state fMRI ... [https://www.sciencedirect.com/science/article/abs/pii/S105381191400175X]
• 13. Reducing Inter-Individual Differences in Task fMRI ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12484288/]
• 14. Correction of aliasing artifact in accelerated echo-planar ... [https://www.sciencedirect.com/science/article/abs/pii/S0730725X25002401]
• 15. Predicting individual task contrasts from resting‐state ... [https://www.sciencedirect.com/science/article/pii/S1053811921011204]
• 16. Task and resting state fMRI modelling of brain-behavior ... [https://www.sciencedirect.com/science/article/pii/S0006322325014878]
• 17. Removing the effects of task-related motion using ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811904007839]
• 18. Prospective Motion Correction of fMRI: Improving the quality of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7238750/]
• 19. Investigating the effective temporal resolution in a task- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12290576/]
• 20. Spurious but systematic correlations in functional connectivity ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3254728/]
• 21. FD • fMRIscrub - Framewise Displacement [https://neuroconductor.org/help/fMRIscrub/reference/FD.html]
• 22. Beyond motion correction: Framewise displacement [https://crimia.wordpress.com/2016/02/07/beyond-motion-correction-framewise-displacement/]
• 23. Getting the nod: Pediatric head motion in a transdiagnostic sample... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9009630/]
• 24. Increased frontal functional connectivity correlates with ... [https://www.nature.com/articles/s41598-025-19143-y]
• 25. Low- motion data can be obtained in fMRI participants... pediatric [https://www.nature.com/articles/s41598-020-78885-z?error=cookies_not_supported&code=7e395e8d-27a5-46d6-8523-85a32b07577e]
• 26. Resting-state functional MRI metrics to detect freezing of ... [https://www.sciencedirect.com/science/article/pii/S0010482525005955]
• 27. a meta-analysis of resting-state fMRI [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2025.1576932/full]
• 28. Brain resting state functional connectivity changes with ... [https://www.sciencedirect.com/science/article/pii/S2666337625000800]
• 29. Task Modulation of Resting-State Functional Gradient ... [https://www.eneuro.org/content/12/6/ENEURO.0066-25.2025]
• 30. Laminar multi-contrast fMRI at 7T allows differentiation of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12254000/]
• 31. Prospective motion correction in functional MRI [https://www.sciencedirect.com/science/article/abs/pii/S1053811916306218]
• 32. Task-based functional MRI challenges in clinical ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2022.1017211/full]
• 33. What are the steps for preprocessing fMRI data? [https://mriquestions.com/data-pre-processing.html]
• 34. Modular preprocessing pipelines can reintroduce artifacts ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6865661/]
• 35. fMRIPrep - Tools [https://www.ebrains.eu/tools/fmriprep]
• 36. FMRIPrep: a robust preprocessing pipeline for functional MRI [https://pmc.ncbi.nlm.nih.gov/articles/PMC6319393/]
• 37. A standardized image processing and data quality platform ... [https://www.nature.com/articles/s41467-024-50826-8]
• 38. An Improved Model of Motion-Related Signal Changes in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5533292/]
• 39. Motion impact score for detecting spurious brain-behavior ... [https://www.nature.com/articles/s41467-025-63661-2]
• 40. An improved model of motion-related signal changes in fMRI [https://www.sciencedirect.com/science/article/abs/pii/S1053811916304360]
• 41. Robust Motion Regression of Resting-State Data Using a ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2019.00169/full]
• 42. Less is more: balancing noise reduction and data retention ... [https://www.sciencedirect.com/science/article/pii/S1053811923001180]
• 43. Less is more: balancing noise reduction and data retention in ... [https://pubmed.ncbi.nlm.nih.gov/36842522/]
• 44. Benchmarking of participant-level confound regression ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5483393/]
• 45. Task and resting state fMRI modelling of brain-behavior ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12593242/]
• 46. Task fMRI paradigms may capture more behaviorally ... [https://www.sciencedirect.com/science/article/pii/S1053811923000939]
• 47. A robust ICA-based strategy for removing motion artifacts ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811915001822]
• 48. An automated and flexible tool for comprehensive fMRI ... [https://pubmed.ncbi.nlm.nih.gov/40851911/]
• 49. The efficacy of resting-state fMRI denoising pipelines for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12336063/]
• 50. A Component Based Noise Correction Method (CompCor ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC2214855/]
• 51. Reduction of motion-related artifacts in resting state fMRI ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4043948/]
• 52. Outputs of fMRIPrep [https://fmriprep.org/en/23.2.0a1/outputs.html]
• 53. fMRIPrep outputs very high number of aCompCors (up to ... [https://neurostars.org/t/fmriprep-outputs-very-high-number-of-acompcors-up-to-1000/5451]
• 54. Model-based physiological noise removal in fast fMRI - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6911832/]
• 55. The PhysIO Toolbox for Modeling Physiological Noise in ... [https://www.sciencedirect.com/science/article/pii/S016502701630259X]
• 56. Comparison of MRI head motion indicators in ... [https://www.nature.com/articles/s41598-024-79827-9]
• 57. Estimation and Removal of Residual Motion Artifact in Retrospectively Motion-Corrected fMRI Data: A Comparison of Intervolume and Intravolume Motion Using Gold Standard Simulated Motion Data | Published in Aperture Neuro [https://apertureneuro.org/article/123369-estimation-and-removal-of-residual-motion-artifact-in-retrospectively-motion-corrected-fmri-data-a-comparison-of-intervolume-and-intravolume-motion-u]
• 58. Correcting for Superficial Bias in 7T Gradient Echo fMRI [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2021.715549/full]
• 59. How to check and document design accessibility in your ... [https://stephaniewalter.design/blog/how-to-check-and-document-design-accessibility-in-your-figma-mockups/]
• 60. Denoising task-correlated head motion from motor-task fMRI data with multi-echo ICA - PMC [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10370165/]
• 61. Selective detrending method for reducing task‐correlated ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3010868/]
• 62. Typicality of functional connectivity robustly captures motion ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7670643/]
• 63. Methods to detect, characterize, and remove motion artifact ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811913009117]
• 64. Comparison of motion correction techniques applied to ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9900395/]
• 65. Methods to detect, characterize, and remove motion artifact in resting state fMRI - PMC [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3849338/]
• 66. Real-time feedback reduces participant motion during task ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9882243/]
• 67. Are Movement Artifacts in Magnetic Resonance Imaging a Real Problem?—A Narrative Review - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC5447676/]
• 68. Real-time motion monitoring improves functional MRI data ... [https://www.sciencedirect.com/science/article/pii/S1878929322000597]
• 69. Filtering respiratory motion artifact from resting state fMRI data in infant and toddler populations - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC8803544/]
• 70. Imaging and Anesthesia Protocol Optimization in Sedated ... [https://pubmed.ncbi.nlm.nih.gov/39134370/]
• 71. Quality control in resting-state fMRI: the benefits of visual ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10192849/]
• 72. Measuring the effects of motion corruption in fetal fMRI - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11755121/]
• 73. The diagnostic potential of resting state functional MRI [https://www.sciencedirect.com/science/article/pii/S1053811925003374]
• 74. Balancing Data Quality and Bias: Investigating Functional ... [https://pubmed.ncbi.nlm.nih.gov/39788921/]
• 75. Quality control procedures and metrics for resting-state ... [https://www.frontiersin.org/journals/neuroimaging/articles/10.3389/fnimg.2023.1072927/full]
• 76. Comparison of whole-brain task-modulated functional ... [https://www.nature.com/articles/s42003-024-07088-3]
• 77. Uncovering communities of pipelines in the task-fMRI ... [https://arxiv.org/html/2312.06231v1]
• 78. Systematic evaluation of fMRI data-processing pipelines for ... [https://www.nature.com/articles/s41467-024-48781-5]
• 79. Comparison of heritability estimates on resting state fMRI ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6218292/]
• 80. Isolating the sources of pipeline-variability in group-level task ... [https://pubmed.ncbi.nlm.nih.gov/34773436/]
• 81. Iterative Learning for Joint Image Denoising and Motion ... [https://www.sciencedirect.com/science/article/pii/S1361841525004189?dgcid=rss_sd_all]
• 82. Can this data be saved? Techniques for high motion in ... [https://www.sciencedirect.com/science/article/pii/S1878929322001219]
• 83. Functional Connectivity - an overview [https://www.sciencedirect.com/topics/medicine-and-dentistry/functional-connectivity]
• 84. Rethinking Measures of Functional Connectivity via ... [https://www.nature.com/articles/s41598-020-57915-w]
• 85. Combining fMRI during resting state and an attention bias ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6911838/]
• 86. Anticorrelations in resting state networks without global ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3230748/]
• 87. Global Signal Regression Strengthens Association ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6585462/]
• 88. Anticorrelations in resting state networks without global ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811911009657]
• 89. Towards a consensus regarding global signal regression ... [https://www.sciencedirect.com/science/article/pii/S1053811916306711]
• 90. Motion correction and scrubbing for task-based fMRI [https://neurostars.org/t/motion-correction-and-scrubbing-for-task-based-fmri/28079]
• 91. A multisite effort to generate and share simulated datasets [https://www.nature.com/articles/s41597-025-04740-3]
• 92. Test-retest reliability of behavioral and computational ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0312255]
• 93. Test-Retest Reliability of Two Computationally ... [https://cpsyjournal.org/articles/10.5334/cpsy.92]
• 94. A Framework for the Testing and Validation of Simulated ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7136518/]
• 95. Methods to detect, characterize, and remove motion artifact ... [https://www.sciencedirect.com/science/article/pii/S1053811913009117]
• 96. Comparison of MRI head motion indicators in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11603305/]
• 97. Differences in Head Motion During Functional Magnetic ... [https://www.sciencedirect.com/science/article/pii/S2667174324001599]
• 98. The role of fMRI in drug development - PMC - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC5931333/]
• 99. Functional imaging studies of acute administration ... [https://www.sciencedirect.com/science/article/abs/pii/S0149763423003901]