Systematic Bias from Motion in Developmental Neuroimaging: Sources, Solutions, and Implications for Research Validity

Abstract

Motion artifacts introduce systematic bias into developmental neuroimaging data, threatening the validity of brain-behavior associations, particularly in pediatric and clinical populations. This article synthesizes current evidence to explore the origins and consequences of this bias, evaluates methodological approaches for artifact correction and data quality control, provides strategies for optimizing acquisition and processing pipelines, and introduces frameworks for validating findings against motion-related confounds. Targeted at researchers and drug development professionals, this resource aims to equip the field with practical knowledge to mitigate motion-induced bias, thereby enhancing the reliability of neurodevelopmental discoveries and their translation into clinical applications.

Unmasking the Problem: How Motion Creates Systematic Bias in Neurodevelopmental Data

In neurodevelopmental research, the prevailing assumption that larger sample sizes inherently mitigate noise represents a critical methodological pitfall. Motion artifacts in magnetic resonance imaging (MRI) do not constitute random noise but instead introduce systematic bias that correlates powerfully with key variables of interest such as age, clinical status, and cognitive ability. This technical analysis demonstrates how motion artifacts persist and even amplify in large-scale studies, threatening the validity of developmental findings. Through examination of artifact mechanisms, empirical evidence from major cohorts, and analysis of mitigation strategies, we establish that motion-related variance behaves as structured noise that conventional averaging cannot eliminate. The implications demand a fundamental shift in approach for researchers, scientists, and drug development professionals utilizing neuroimaging biomarkers.

Motion artifacts have emerged as a preeminent challenge for developmental neuroimaging, fundamentally distinct from random noise due to their structured pattern and non-random distribution across populations [1]. The critical insight revolutionizing the field is that in-scanner motion frequently correlates with central variables of interest—including age, clinical status, cognitive ability, and symptom severity—thereby introducing systematic bias rather than random error [1] [2]. This confounding relationship creates a methodological perfect storm where motion artifacts masquerade as neural effects, potentially invalidating conclusions from even the largest-scale studies.

The problem is particularly acute in developmental neuroscience and psychiatric drug development, where participant populations (children, older adults, clinical cohorts) systematically exhibit greater motion than healthy young adult controls [2]. As sample sizes expand to thousands of participants in initiatives like the Adolescent Brain Cognitive Development (ABCD) Study, the assumption that motion artifacts would "average out" has proven dangerously incorrect [3]. Contrary to this expectation, motion introduces directional biases that persist and potentially amplify in large datasets, creating spurious but systematic correlations in functional connectivity MRI networks [1] [3].

Empirical Evidence: Motion Artifacts in Large-Scale Studies

Evidence from Major Neuroimaging Cohorts

Recent findings from the ABCD Study challenge the foundational assumption that larger sample sizes counteract noisy images. A manual quality assessment of 10,295 structural MRI scans revealed that 55% were of suboptimal quality, with 2% deemed unusable [3]. Crucially, incorporating these scans introduced systematic bias rather than random error: lower-quality scans consistently underestimated cortical thickness and overestimated cortical surface area [3].

In one analysis, when only the 4,600 highest-quality scans were included, significant group differences in cortical volume for children with aggressive behaviors appeared in three brain regions [3]. When moderate-quality scans were added, this number jumped to 21 regions, and when all scans were pooled, 43 regions showed significant differences [3]. This inflation of effect sizes with decreasing data quality demonstrates how motion artifacts create spurious findings that larger samples amplify rather than mitigate.

Motion as a Confound in Developmental Trajectories

The relationship between motion and age creates particularly problematic confounding in developmental research. Motion follows a U-shaped trajectory across the lifespan, with high motion in young children decreasing to low values in late teens through the 30s, followed by a gradual rise in later decades [2]. This pattern directly conflates with developmental changes in brain structure and function.

Table 1: Age-Related Motion Patterns in Neurodevelopment

Age Group	Mean Framewise Displacement (FD)	Developmental Period	Impact on Connectivity
Middle Childhood	~0.50 mm [4]	Rapid synaptic pruning	Inflated short-distance connections [5]
Adolescence	~0.09 mm [4]	Network specialization	Diminished long-distance connections [5]
Adulthood	~0.05 mm [4]	Network stability	More accurate connectivity patterns [5]

Longitudinal studies confirm that head motion decreases significantly as children age, with one study of children ages 9-14 showing a significant age effect on framewise displacement during both diffusion (p < .001) and resting-state functional MRI (p < .001) [4]. This motion-age relationship systematically biases estimates of connectivity change during development, particularly inflating distance-dependent effects [5].

Mechanisms: How Motion Artifacts Create Systematic Bias

Physics of Motion Artifacts in MRI

The manifestation of motion artifacts in MRI stems from fundamental physical principles of image acquisition. Unlike photography, MRI data collection occurs in Fourier space (k-space), where each sample contains global information about the image [6]. Motion during acquisition creates inconsistencies between different portions of k-space data, violating the core assumption of stationary objects during reconstruction [6] [7].

The specific appearance of motion artifacts depends on both the nature of movement and the k-space sampling strategy:

• Cartesian sampling: Produces ghosting artifacts in the phase-encoding direction [6] [7]
• Radial sampling: Results primarily in image blurring [6] [7]
• Periodic motion: Creates coherent ghosting with replicas corresponding to motion frequency [6]
• Random motion: Generates incoherent ghosting appearing as stripes [6]

These artifacts produce spatially structured noise that correlates with subject factors rather than distributing randomly across populations.

Impact on Functional Connectivity Measures

In functional connectivity MRI (fc-MRI), motion artifacts introduce distance-dependent biases that systematically alter network properties. Higher motion causes spurious decreases in long-distance correlations and increases in short-distance connectivity [1] [4]. This pattern directly mimics—and potentially creates—the appearance of developmental changes, where brain maturation is characterized by strengthening long-range connections and weakening short-range connections [5].

The temporal properties of motion artifacts further complicate their removal. Motion produces both immediate signal drops following movement events and longer-duration artifacts (up to 8-10 seconds) that may result from motion-related changes in CO₂ from yawning or deep breathing [1]. These effects introduce nonlinear relationships that rigid-body correction models cannot fully capture [1].

Diagram Title: Motion Artifact Propagation in MRI

Population-Specific Vulnerabilities and Biases

Clinical Populations and Motion as Behavioral Phenotype

The non-random distribution of motion across clinical populations creates systematic exclusion biases that threaten the generalizability of neuroimaging findings. Patients with psychotic disorders exhibit significantly more head movement than healthy controls due to factors including psychomotor agitation, anxiety, paranoia, medication side effects, and difficulty following instructions [8]. This elevated motion may represent a behavioral phenotype rather than mere data quality issue, as patients who struggle to remain still may constitute a distinct neurobiological subtype with more severe symptoms [8].

In ADHD populations, children consistently display greater framewise displacement than controls across ages 9-14 [4]. Crucially, even children in remission from ADHD showed continued elevation in head motion compared to controls, suggesting that motion may represent a persistent trait rather than state marker [4]. This pattern indicates that motion itself may be part of the ADHD phenotype, with important implications for interpreting neuroimaging findings in this population.

Table 2: Motion Patterns Across Clinical Populations

Population	Motion Level	Primary Contributors	Impact on Data Generalizability
Psychosis	Significantly elevated [8]	Psychomotor agitation, disorganization, anxiety [8]	Exclusion of most severe cases creates spectrum bias [8]
ADHD	Consistently elevated [4]	Innate hyperactivity, impulsivity [4]	Motion as intrinsic trait rather than artifact [4]
Neurodevelopmental Disorders	Elevated [2]	Developmental immaturity, symptom-related restlessness [2]	Altered developmental trajectories [2]
Childhood Samples	Age-dependent [2] [4]	Developmental capacity for stillness [2]	Confounding of age and motion effects [5]

The Missing Not at Random (MNAR) Problem

The practice of excluding high-motion scans introduces a fundamental statistical problem: Missing Not at Random (MNAR) data [8]. When participants with the most severe clinical presentations produce unusable scans due to motion, the resulting dataset systematically underrepresents the most severe end of the illness spectrum. This exclusion biases effect size estimates and limits generalizability to the original target population [8].

For example, if patients with severe, disorganized schizophrenia cannot tolerate the scanner environment, their exclusion will bias hippocampal volume estimates toward larger values, potentially underestimating the true effect size of the disorder on brain structure [8]. This MNAR problem violates the assumptions of most inferential statistical approaches, potentially yielding biased parameter estimates and invalid inferences [8].

Mitigation Strategies and Methodological Recommendations

Acquisition Protocols and Experimental Design

Effective motion mitigation begins during study design and data acquisition. Strategic approaches include:

• Split-session designs: Distributing fMRI acquisition across multiple same-day sessions reduces head motion in children, while inside-scanner breaks benefit adults [9]
• Mock scanner training: Familiarization sessions improve participant comfort and compliance [4] [9]
• Participant preparation: Clear instructions, comfortable positioning, and reinforcement strategies minimize initial motion [8]
• Sequence optimization: Radial k-space sampling produces more tolerable blurring artifacts compared to Cartesian ghosting [6] [7]

These approaches address motion prevention rather than correction, reducing the problem at its source.

Processing and Analysis Techniques

When motion occurs despite prevention efforts, multiple processing strategies can mitigate its impact:

• Robust confound regression: Expanded nuisance regressors including motion parameters and their temporal derivatives [1] [5]
• Motion scrubbing: Identifying and removing high-motion volumes using framewise displacement (FD > 0.5mm) or DVARS thresholds [1] [8]
• ICA-based denoising: Algorithms like ICA-AROMA automatically identify and remove motion-related components [8]
• Volume-based correction: Realigning each volume to a reference using tools like FSL's MCFLIRT or AFNI's 3dvolreg [8]

Critically, each method has limitations. Scrubbing can introduce MNAR biases, ICA may remove neural signals along with noise, and regression assumes linear relationships that may not fully capture motion effects [8].

Diagram Title: Motion Mitigation Pipeline

Table 3: Research Reagent Solutions for Motion Management

Tool/Category	Specific Examples	Function	Limitations
Motion Quantification	Framewise Displacement (FD), DVARS [1] [8]	Quantifies volume-to-volume movement for scrubbing and QC	Varies across implementations; TR-dependent [1]
Realignment Tools	FSL's MCFLIRT, AFNI's 3dvolreg [8]	Corrects for between-volume motion via rigid registration	Cannot correct intra-volume motion [8]
ICA Denoising	ICA-AROMA, FSL's FIX [8]	Identifies and removes motion-related components	May remove neural signals; computational intensity [8]
Quality Metrics	Surface Hole Number (SHN) [3]	Automated quality assessment approximating manual rating	Does not eliminate error as effectively as manual rating [3]
Prospective Correction	PACE, Volumetric Navigators [8]	Real-time motion tracking and slice position updating	Hardware limitations; not widely available [8]

Motion artifacts in MRI represent a fundamental challenge that larger samples amplify rather than mitigate. The systematic nature of motion-induced bias, its correlation with key demographic and clinical variables, and its structured impact on connectivity measures collectively undermine the assumption that motion averages out in large datasets. For developmental researchers and drug development professionals, this necessitates a paradigm shift from considering motion as a nuisance variable to recognizing it as a potentially catastrophic confound that threatens inference validity.

Future directions must include: (1) development of universal motion metrics standardized across acquisition parameters; (2) integration of prospective correction techniques into clinical scanners; (3) adoption of standardized reporting of motion exclusion criteria and quality control procedures; and (4) implementation of sensitivity analyses demonstrating result robustness across quality thresholds. Most critically, the field must abandon the dangerous illusion that larger samples automatically solve the motion problem and instead confront motion artifacts as structured variance that demands sophisticated methodological attention throughout the research pipeline.

Structural magnetic resonance imaging (sMRI) has become a cornerstone of clinical neuroscience research, offering unparalleled insights into brain morphometry. However, the fidelity of this powerful tool is critically threatened by a pervasive source of systematic bias: in-scanner motion. Even minor, visually imperceptible movement during acquisition can introduce structured noise that systematically distorts key morphometric measures. This technical guide details the specific nature of this bias—a consistent underestimation of cortical thickness and overestimation of cortical surface area—within the critical context of developmental neuroimaging research. Understanding and mitigating this bias is paramount for researchers, scientists, and drug development professionals aiming to identify true neurobiological markers in developmental disorders and treatment effects.

The challenge is particularly acute in studies of children, adolescents, and individuals with neurodevelopmental or movement disorders, who tend to move more during scanning [10] [11]. As large-scale, population-level studies like the Adolescent Brain Cognitive Development (ABCD) Study become more common, the field is confronting a sobering realization: larger sample sizes alone do not overcome this systematic bias; they can, in fact, amplify it, leading to both false-positive and false-negative findings [3] [12]. This guide synthesizes recent evidence quantifying this impact, outlines methodologies for its detection and correction, and provides a practical toolkit for enhancing the rigor of structural neuroimaging.

Quantifying the Systematic Bias

The impact of motion on sMRI is not random error but a directionally consistent bias that mimics specific neuroanatomical patterns. Evidence from large-scale studies demonstrates that this bias systematically alters measurements in ways that can be misinterpreted as genuine neurodevelopmental effects.

Evidence from Major Studies

Table 1: Key Studies Quantifying Motion-Related Bias in sMRI

Study	Sample	Key Finding on Cortical Thickness	Key Finding on Cortical Surface Area
ABCD Study (Roffman et al.) [3]	>10,000 scans; children aged 9-10	Lower-quality scans consistently underestimate cortical thickness.	Lower-quality scans consistently overestimate cortical surface area.
ABCD Preprint (Roffman et al.) [12]	11,263 T1 scans from ABCD Study	Linear association between poorer quality and reduced thickness across much of the cortex.	Increased surface area in lateral/superior regions; mixed effects elsewhere.
Healthy Brain Network [11]	388 participants; ages 5-21	Image quality significantly impacts cortical thickness in ~23.4% of brain areas investigated.	Image quality significantly impacts cortical surface area in ~23.4% of brain areas investigated.
PMC Study [13]	127 children, adolescents, and young adults	Trend-level decrease in cortical thickness with greater motion.	N/A (Focused on cortical volume and curvature)

Incorporating lower-quality scans dramatically inflates effect sizes in group analyses. One analysis of the ABCD data found that when comparing cortical volume in children with versus without aggressive behaviors, the number of significant brain regions jumped from 3 to 43 as lower-quality scans were added to the analysis [3]. This demonstrates how motion bias can create the illusion of widespread, statistically significant findings.

Effect Sizes and Anatomical Specificity

The bias introduced by motion is not uniform across the brain and can be substantial in magnitude.

Table 2: Effect Sizes of Motion on Structural Measures

Metric	Direction of Bias	Effect Size (Cohen's d)	Anatomical Notes
Cortical Thickness	Systematic Underestimation	0.14 – 2.84 [12]	Effects are anatomically heterogeneous [13].
Cortical Surface Area	Systematic Overestimation	0.14 – 2.84 [12]	Overestimation is prominent in lateral/superior regions [12].
Cortical Gray Matter Volume	Decrease	Significant relationships found [13]	A product of thickness and area; bias direction can vary.
Cortical Curvature	Increase	Significant relationships found [13]	Increased mean curvature with greater motion.

The biomechanical reasons for this specific directional bias are rooted in how motion corrupts the image data. Motion causes blurring at the tissue boundaries, which complicates the accurate identification of the gray matter/white matter border (impacting thickness) and the pial surface (impacting surface area) [13] [11]. The result is an output that can appear biologically plausible, making the bias particularly insidious.

Experimental Protocols for Assessing Motion Bias

Rigorous assessment of motion bias relies on a combination of experimental design and quality control methodologies. The following protocols are considered best practice in the field.

Protocol 1: Using fMRI as a Proxy for sMRI Motion

Objective: To obtain a continuous, quantitative estimate of subject motion during a scanning session to correlate with structural MRI measures, even in the absence of visible sMRI artifacts [13].

Workflow:

• Image Acquisition: Acquire a T1-weighted structural scan followed by two (or more) resting-state fMRI scans during the same session.
• Motion Quantification (fMRI): Process the fMRI data through a standard realignment algorithm (e.g., in SPM or FSL) to generate framewise displacement (FD) timeseries for each subject.
• Summary Metric: Calculate the mean FD across the entire fMRI scan. This serves as a reliable, continuous proxy for the subject's tendency to move during the scanning session, including during the prior structural scan [13] [11].
• Morphological Analysis: Process the T1-weighted structural scans using automated pipelines (e.g., FreeSurfer, CIVET) to extract cortical thickness, surface area, and volume.
• Statistical Analysis: Perform whole-brain or region-of-interest regression analyses, modeling the morphometric measures (e.g., cortical thickness) as a function of the mean FD metric, while controlling for covariates like age, sex, and site.

Protocol 2: Manual Quality Control (MQC) and Group Comparison

Objective: To categorize scans based on visual quality and quantify the morphometric differences between quality groups in a large dataset [3] [12].

Workflow:

• Visual Inspection: A trained rater, blinded to subject information, views each T1 volume and assigns a quality rating based on a predefined scale (e.g., 1=minimal edits needed, 2=moderate edits, 3=substantial edits, 4=unusable) [12].
• Group Formation: Form groups based on MQC ratings. For example, a "High-Quality" group (rating 1) and a "Lower-Quality" group (ratings 2-4).
• Automated Processing: Process all scans through an automated pipeline like FreeSurfer to obtain morphometric measures.
• Group Analysis: Compare morphometric outputs (cortical thickness, surface area) between the High-Quality and Lower-Quality groups to quantify the systematic bias introduced by motion.

Protocol 3: Retrospective Deep Learning-Based Motion Correction

Objective: To correct for motion artifacts in structural MRI images after acquisition using a convolutional neural network (CNN) [14].

Workflow:

• Model Training:
  • Gather a dataset of motion-free, high-quality T1-weighted images.
  • Artificially corrupt these clean images by simulating motion artifacts in the Fourier domain (K-space) to create paired training data (corrupted vs. clean).
  • Train a 3D CNN to learn the mapping from the motion-corrupted image to the clean image.
• Application and Validation:
  • Apply the trained CNN to real motion-affected scans to generate corrected images.
  • Validate the correction using image quality metrics like Peak Signal-to-Noise-Ratio (PSNR) and by demonstrating improved cortical surface reconstructions [14].
  • Statistically compare morphometric measures (e.g., cortical thickness) derived from original and corrected images to show a reduction in motion-related bias.

Figure 1: Experimental Workflow for Quantifying and Correcting Motion Bias. This diagram outlines the three primary methodological approaches for assessing the impact of motion on cortical measurements.

The Scientist's Toolkit: Key Research Reagents and Solutions

Successfully navigating the challenge of motion bias requires a suite of methodological tools and quality control metrics.

Table 3: Essential Materials and Tools for Motion Bias Research

Tool/Solution	Function	Relevance to Motion Bias
FreeSurfer [10]	Automated surface-based and volume-based stream processing of T1 MRI.	The primary source of morphometric measures (thickness, area). Its outputs are the target of the bias.
CIVET Pipeline [13]	Automated MRI analysis pipeline for cortical segmentation and thickness.	Provides an alternative processing platform to validate that motion biases are not pipeline-specific.
Surface Hole Number (SHN) [3] [12]	An automated QC metric estimating imperfections in cortical reconstruction.	A robust, automated proxy for manual QC. Effectively differentiates lower-quality scans to "stress-test" findings.
CAT12 Toolbox [11]	Computational Anatomy Toolbox for SPM; provides image quality metrics.	Offers an independent, automated aggregate "grade" for a structural scan, correlating with human rating and morphometry.
Framewise Displacement (FD) [13] [11]	A quantitative measure of head motion derived from fMRI realignment parameters.	Serves as a continuous assay for a subject's motion propensity during the scanning session.
Convolutional Neural Network (CNN) [14]	Deep learning model for retrospective motion artifact correction.	Learns to map motion-corrupted T1 images to their clean counterparts, improving downstream surface reconstruction.

The systematic underestimation of cortical thickness and overestimation of surface area due to in-scanner motion is a critical, empirically validated source of bias in developmental neuroimaging. This bias is systematic, not random, leading to effect size inflation and potentially spurious findings that can misdirect research and drug development efforts. Mitigating this threat requires a proactive, multi-pronged strategy: implementing rigorous manual or automated quality control (with metrics like Surface Hole Number), incorporating motion as a covariate in statistical models, and exploring advanced retrospective correction techniques. For the field to produce reliable and replicable results, especially in large-scale studies of developmental populations, acknowledging and correcting for this systematic bias must become a non-negotiable standard in the analysis pipeline.

The pursuit of objective biomarkers in developmental neuroscience is fundamentally challenged by systematic biases, with motion artifacts representing a particularly pervasive source of error. While motion affects neuroimaging data universally, its impact disproportionately affects already vulnerable populations, including children and psychiatric patients, thereby distorting our understanding of brain development and pathology. This disproportionality arises not from biological inevitability but from a complex interplay of technical limitations, methodological oversights, and socio-structural barriers that converge upon these groups. Within the context of a broader thesis on systematic bias from motion in developmental neuroimaging, this technical guide examines how pre-existing vulnerabilities are amplified by research methodologies, creating a feedback loop that further marginalizes these populations. The consequences extend beyond scientific inconvenience to affect the validity, reliability, and generalizability of neurodevelopmental findings, ultimately perpetuating inequities in both knowledge generation and clinical translation [15] [16].

This whitepaper provides an in-depth analysis of the mechanisms through which motion-related biases disproportionately affect vulnerable populations, supported by quantitative data and experimental evidence. It further presents detailed methodologies for bias mitigation and a forward-looking framework for developing more equitable neuroimaging research practices targeted at researchers, scientists, and drug development professionals working at the intersection of neurodevelopment and psychiatric disorders.

The Disproportionate Impact of Motion on Vulnerable Populations

Quantitative Evidence of Disparities

Table 1: Documented Disparities in Neuroimaging Study Populations

Domain of Disparity	Reported Finding	Primary Source
Demographic Reporting	Only 10% of neuroimaging studies report race; 4% report ethnicity [17].	Systematic Review of 408 MRI Studies (2010-2020)
Sample Representation	Predominantly White participants (86%) in a pre-clinical Alzheimer's trial of ~6,000 individuals [17].	Raman et al.
Workforce Diversity	Lack of diversity in neuroscience workforce leads to unacknowledged bias in scientific agendas [16].	Firat et al.
Device Exclusion	EEG, fNIRS, skin conductance, and eye-tracking tools systematically exclude participants based on phenotypic differences (e.g., hair structure, skin pigmentation) [15].	Webb et al.

Mechanistic Pathways of Disproportionate Impact

The disproportionate effect of motion on children and psychiatric patients operates through several interconnected pathways, creating a bias propagation pipeline [18].

Physiological and Developmental Factors

Children, particularly young children and those with neurodevelopmental or psychiatric conditions, exhibit age-typical and condition-related motor restlessness. The capacity for volitional motion suppression during extended scanning sessions is a developmental achievement that hinges on prefrontal cortex maturation, which continues into early adulthood. Psychiatric populations, including those with Attention-Deficit/Hyperactivity Disorder (ADHD), Tourette's Syndrome, or anxiety disorders, may experience involuntary movements, tics, or heightened psychomotor agitation that directly conflict with data acquisition requirements. Furthermore, the MRI environment itself—a loud, confined, and novel space—can induce anxiety and stress, particularly in children with a history of trauma or those with conditions like autism spectrum disorder. This stress can potentiate threat hypervigilance and increase motion, thereby confounding neural activity patterns related to the experimental task with those related to stress and anxiety [16].

Socio-Structural and Economic Factors

Vulnerable populations often face logistical and economic barriers that compound their physiological predisposition to motion-related artifacts. Longitudinal studies, crucial for developmental neuroscience, show higher attrition rates among marginalized groups due to time-intensive protocols, transportation difficulties, and greater family responsibilities [15]. For example, the Generation R Study in the Netherlands became less diverse in terms of ethnicity and educational level with each wave despite concerted retention efforts [15]. This attrition bias means that studies progressing over time may systematically lose the very participants who are most susceptible to motion, thereby producing increasingly non-representative data. Furthermore, individuals from lower socioeconomic backgrounds may have less access to financial and digital resources, which can prevent initial participation and familiarity with research settings, potentially increasing anxiety and motion during their first and only scan session [15] [19].

Technological and Methodological Exclusion

A critical yet often overlooked pathway is the inherent design bias in neuroscientific tools. Electrophysiological devices like EEG were often not designed to handle human phenotypic variability. Participants with coarse, curly, or thick hair types (e.g., Afro-textured hair) or protective styles (e.g., braids, dreadlocks) have been systematically excluded from EEG studies due to difficulties in achieving adequate electrode-scalp contact and insufficient signal quality [15] [16]. This exclusion is not merely a recruitment failure but a fundamental bias in technological design. Similarly, the physical design of MRI head coils can restrict natural hairstyles, and sew-in hair extensions with metal tracks can prevent entry into the scanner bore altogether [16]. This technological exclusion means that entire groups are erased from datasets from the outset, and those who do participate may do so under suboptimal or stressful conditions that increase motion.

Experimental Protocols for Investigating and Mitigating Bias

Community-Based Participatory Research (CBPR) Framework

The Community-Based Participatory Research (CBPR) framework actively involves the population of interest in the research process to counter biases.

• Objective: To establish equitable researcher-community partnerships that enhance recruitment, retention, and protocol appropriateness for underrepresented groups, thereby reducing situational factors that lead to increased motion.
• Procedure:
  • Establish a Community Advisory Board (CAB): Recruit a diverse group of community members, including parents and youth from the target population, to collaborate throughout the research process [16].
  • Conduct Positionality Mapping: Researchers explicitly document their own social positions (e.g., race, gender, institutional affiliation) and reflect on how these positions may influence research questions, hypothesis formation, and interactions with participants [16].
  • Co-Design Study Protocols: The CAB provides feedback on all aspects of the study, including the consent process, assessment tools, and the MRI experience. This can involve creating child-friendly scanner mock-up training sessions to reduce anxiety and practicing remaining still in a simulated environment [16].
  • Iterative Review and Dissemination: The CAB reviews findings and helps disseminate results back to the community in accessible formats.
• Rationale: This approach builds trust, improves cultural competence, and directly addresses logistical and emotional barriers that contribute to motion in scanner environments, leading to more robust and generalizable data collection [16].

Quantitative Bias Analysis (QBA) for Motion Artifact

Quantitative Bias Analysis (QBA) provides a set of methodological techniques to quantitatively estimate the potential magnitude and direction of systematic error, such as selection bias introduced by motion-related exclusion.

• Objective: To move beyond qualitative discussion of motion as a limitation and instead model how motion-induced selection bias might have influenced the observed study results.
• Procedure (Probabilistic Bias Analysis):
  • Define the Bias Structure: Using a Directed Acyclic Graph (DAG), depict the relationships between motion, participant vulnerability factors, inclusion in the final analysis, and the outcome of interest [20] [21].
  • Specify Bias Parameters: Estimate the probability of exclusion from the final analysis due to motion for different groups (e.g., children with ADHD vs. typically developing controls). These parameters can be informed by internal study data or external validation studies [22] [20].
  • Model the Uncertainty: Assign probability distributions (e.g., beta distributions) to the bias parameters to account for uncertainty in their values [20].
  • Perform Probabilistic Adjustment: Run a large number of Monte Carlo simulations (e.g., 10,000). In each simulation, draw a value for each bias parameter from its specified distribution and use these values to probabilistically correct the original data.
  • Summarize Results: The output is a distribution of bias-adjusted estimates. Report the median adjusted estimate and a 95% simulation interval, which quantifies the uncertainty in the results after accounting for the systematic bias [20].
• Rationale: QBA transforms the discussion of motion bias from a speculative limitation into a quantifiable uncertainty, providing a more realistic interpretation of study findings and highlighting the potential for biased inference [22] [20] [21].

Protocol for Mitigating Motion in Pediatric & Psychiatric Populations

A practical, multi-faceted protocol can proactively reduce motion artifacts during data acquisition.

• Objective: To minimize the occurrence and impact of head motion during MRI scans in challenging populations.
• Procedure:
  • Pre-Scan Preparation:
    • Scanner Mock-Up Training: Conduct extensive behavioral training using a mock scanner. Have participants practice the scan protocol while listening to recorded scanner noises. Provide real-time feedback on head motion [16].
    • Social Story Videos: Create visual guides featuring participants from diverse backgrounds successfully completing a scan to reduce anxiety and set expectations.
  • In-Scan Strategies:
    • Positive Reinforcement: Implement a reward system (e.g., points, small prizes) for successful still periods. Use a real-time motion-tracking system with a display that provides visual feedback to the participant (e.g., a "game" where staying still keeps a character on screen).
    • Comfort Optimization: Use comfortable, pediatric-sized padding and cushions to minimize involuntary motion. Ensure the room is at a comfortable temperature.
    • Shortened Paradigms: Break long scanning sessions into shorter, manageable blocks with breaks in between.
  • Post-Scan Analysis:
    • Proactive Data Exclusion: Apply stringent, pre-registered motion-correction algorithms (e.g., ICA-AROMA, SCRUBBING).
    • Motion as a Covariate: In statistical models, include quantitative motion estimates as a nuisance variable to control for its residual effects.
• Rationale: A comprehensive approach that addresses anxiety, provides motivation, and maximizes physical comfort can significantly increase the yield of usable data from vulnerable participants, reducing the need for exclusion and mitigating selection bias [16].

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Resources for Equitable and Rigorous Neuroimaging Research

Tool or Resource	Function/Purpose	Specific Examples & Notes
Community Advisory Board (CAB)	Advises on all research stages, improves cultural competence, builds trust, and enhances participant comfort to reduce motion-inducing anxiety.	Composed of community members, parents, and youth; used for protocol co-design and feedback on dissemination [16].
Quantitative Bias Analysis (QBA)	Quantifies the impact of systematic errors (e.g., selection bias from motion-related exclusion) on study results, moving beyond speculative discussion.	Methods range from simple to probabilistic analysis; requires specification of bias parameters [22] [20] [21].
Directed Acyclic Graph (DAG)	A visual tool to identify and communicate potential sources of confounding and selection bias in the study design.	Used to map relationships between vulnerability, motion, exclusion, and outcomes, clarifying the bias structure for QBA [20].
Scanner Mock-Up System	A simulated MRI scanner environment for behavioral training; acclimates participants to the scanning experience, reducing anxiety and motion.	Includes a mock scanner bore and playback of scanner sounds; allows for practice with feedback [16].
Real-Time Motion Tracking	Software that provides immediate feedback on participant head motion during the scan, allowing for correction and coaching.	Can be integrated with visual feedback games for children to incentivize staying still [16].
Openly Shared Bias Analysis Code	Pre-written scripts (e.g., in R, Python, or SAS) to facilitate the implementation of QBA methods, lowering the barrier to adoption.	Resources are available from organizations like the International Society for Environmental Epidemiology (ISEE) QBA SIG [21].

The disproportionate impact of motion artifacts on children and psychiatric patients in developmental neuroimaging is not a mere technical obstacle but a profound methodological and ethical challenge that threatens the validity of the field's findings. As this guide has detailed, the issue is rooted in a self-reinforcing cycle where physiological predispositions, structural barriers, and technologically exclusionary designs converge to systematically marginalize vulnerable populations. This ultimately results in biased datasets, ungeneralizable models, and clinical tools that may perform poorly for the very groups they are intended to serve.

Breaking this cycle requires a fundamental shift from passive observation to active intervention. The experimental protocols and tools outlined—ranging from Community-Based Participatory Research and rigorous Quantitative Bias Analysis to practical motion-mitigation strategies—provide a roadmap for this transition. The future of equitable developmental neuroscience depends on the widespread adoption of such practices. This includes diversifying research teams, mandating demographic reporting, and investing in the development of inclusive technologies that accommodate natural human variation [15] [17] [16]. By reconceptualizing motion not as a nuisance variable but as a manifestation of systemic bias, researchers can build a more rigorous, reproducible, and just science of brain development.

In clinical research, particularly in developmental neuroimaging, head motion during data acquisition has emerged as a critical source of systematic bias rather than mere random noise. This motion systematically excludes specific patient phenotypes from study populations, creating a Missing Not at Random (MNAR) problem that fundamentally compromises the validity and generalizability of research findings. In neuroimaging studies of psychiatric disorders such as schizophrenia and bipolar disorder, patients exhibit significantly more head movement during scanning compared to healthy controls [8]. This increased motion is not random but is intrinsically linked to core symptoms of these conditions, including psychomotor agitation, disorganized behavior, anxiety, or medication side effects such as akathisia [8]. When researchers exclude data from participants with excessive motion—a standard quality control practice—they systematically remove the most severely affected individuals from their samples. This introduces substantial bias by shifting the study population toward the less severe end of the clinical spectrum, potentially obscuring crucial brain-behavior relationships and producing misleading conclusions about the neurobiological underpinnings of psychiatric illness.

The MNAR Problem: Theoretical Framework and Clinical Implications

Defining Missing Not at Random (MNAR) in Clinical Contexts

In statistical terms, MNAR refers to situations where the probability of data being missing is directly related to the actual values of the missing data themselves. In the context of clinical neuroimaging, this occurs because patients with more severe symptoms are more likely to produce unusable scans due to motion artifacts [8]. This creates a fundamental violation of the assumptions underlying most standard statistical approaches, including t-tests and ANOVA, which assume that any missing data are ignorable (missing due to random reasons) [8]. When data are MNAR, these analyses yield biased parameter estimates and invalid inferences in hypothesis testing, potentially leading to false conclusions about disease mechanisms and treatment effects.

Evidence from electronic monitoring studies in bipolar disorder further confirms the MNAR phenomenon in clinical research. One study found that missing data were lowest for participants in depressive episodes, intermediate for those with subsyndromal symptoms, and highest for euthymic participants [23]. Furthermore, when participants' clinical status changed during the study (e.g., transitioning from euthymia to depression), missing data for self-rating scales increased significantly [23]. This pattern demonstrates that missingness is directly tied to clinical state rather than occurring randomly.

Motion as a Behavioral Phenotype in Psychosis Spectrum Disorders

Rather than being mere noise that corrupts data, head movements during scanning may carry important information about the study population. Patients who struggle to remain still for an MRI could represent a distinct behavioral or neurobiological subtype of psychosis [8]. For instance, marked restlessness or inability to comply with scanner instructions may be a proxy for high levels of psychomotor agitation, anxiety, or disorganization [8]. Similarly, severe paranoia might make it challenging for patients to tolerate the confined, noisy scanner environment, leading to increased movements. These symptoms typically indicate a more severe or acute presentation, meaning that excluding scans from these subjects necessarily biases the study population toward milder cases and limits the generalizability of findings to the full clinical population [8].

Table 1: Clinical Correlates of In-Scanner Motion in Psychosis Spectrum Disorders

Clinical Feature	Relationship to Motion	Impact on Data Exclusion
Psychomotor Agitation	Directly increases movement	High exclusion risk for agitated patients
Disorganized Behavior	Reduces ability to follow instructions	Systematic exclusion of disorganized subtype
Paranoia	Increases discomfort in scanner environment	Exclusion of patients with reality distortion
Medication Side Effects	Akathisia can cause restlessness	Exclusion of patients with treatment complications
Poor Insight	Reduces compliance with instructions	Exclusion of patients lacking illness awareness

Quantitative Evidence of Motion-Induced Bias in Large-Scale Studies

Systematic Bias in Major Neuroimaging Datasets

The assumption that larger sample sizes automatically counteract the effects of noisy data has been challenged by recent findings from the Adolescent Brain Cognitive Development (ABCD) Study, which revealed that poor image quality introduces systematic bias into large-scale neuroimaging analyses [3]. When researchers manually assessed the quality of over 10,000 sMRI scans from the ABCD Study, they found more than half (55%) were of suboptimal quality, even after standard automated quality-control steps [3]. Critically, incorporating these lower-quality scans introduced systematic bias rather than random noise.

The study demonstrated that lower-quality scans consistently underestimate cortical thickness and overestimate cortical surface area, with these errors growing as scan quality decreases [3]. In one analysis, when examining cortical volume differences between children with or without aggressive and rule-breaking behaviors, the number of significant brain regions inflated dramatically as lower-quality scans were added: from 3 regions in the highest-quality scans (n=4,600) to 21 regions when moderate-quality scans were included, and to 43 regions when all scans were pooled [3]. This demonstrates how motion artifacts can create false positive findings or inflate effect sizes in large datasets.

Table 2: Impact of Scan Quality on Statistical Results in the ABCD Study

Sample Composition	Number of Significant Brain Regions	Effect Size Changes
Highest-quality scans only (n=4,600)	3 regions	Baseline effect sizes
Including moderate-quality scans	21 regions	Effect sizes more than doubled in some regions
All scans pooled	43 regions	Further inflation of effect sizes

Statistical Consequences of Motion Exclusion

The exclusion of high-motion scans has direct statistical consequences for research findings. In a hypothetical example described in the literature, if the full schizophrenia population has smaller hippocampal volumes than healthy controls by a specific value, excluding high-motion scans (which disproportionately come from patients with more severe illness) would bias the estimated average hippocampal volume toward a larger value [8]. This leads to an underestimation of the true effect in the study and potentially masks genuine neurobiological differences associated with the disorder.

The problem is particularly pronounced in studies comparing clinical populations to healthy controls, as the exclusion rates are typically asymmetrical between groups. One analysis found that controlling for motion artifacts reduced the number of significant findings in brain-wide association studies by over 50%, with the largest reductions occurring in brain regions previously associated with motion [3].

Methodological Approaches: From Motion Correction to Phenotyping

Motion Mitigation Strategies in MRI Research

Several strategies have been developed to manage head motion during MRI scans, ranging from preventive approaches to technical corrections:

• Preventive Methods: Providing clear instructions during patient preparation, using physical restraints (foam padding), practice mock scan sessions, reward incentives, and displaying media content during scan breaks [8].
• Prospective Motion Correction: Real-time tracking and correction during scanning by updating slice acquisition coordinates based on detected movement [8].
• Real-time Monitoring: Systems that monitor head motion frame-by-frame and can pause scanning or extend acquisition until sufficient low-motion data are collected [8].
• Retrospective Correction: Algorithms applied after data acquisition, including volume realignment tools like FSL's MCFLIRT and AFNI's 3dvolreg [8].

Despite these innovations, real-time motion correction tools are not yet in widespread use due to complexity and hardware limitations, meaning most studies still rely heavily on retrospective correction methods [8].

Analytical Approaches for Motion-Affected Data

For data already affected by motion, researchers have developed multiple analytical strategies to mitigate bias:

• Motion Scrubbing: Removing volumes with motion beyond a specific threshold (e.g., framewise displacement >0.5mm) from fMRI time series [8]. When too many volumes are scrubbed (typically >20%), the entire participant may be excluded.
• Covariate Adjustment: Including motion parameters (mean framewise displacement, number of removed frames) as covariates in group-level statistical models [8].
• Denoising Techniques: Using Independent Component Analysis (ICA) to identify noise components associated with motion and regress them out of the data without discarding entire volumes [8]. Tools include FSL's ICA-Based X-noiseifier (FIX) and ICA-AROMA.
• Data Imputation: Experimental machine learning-based predictive models to impute missing timepoints due to large motion in fMRI data [8].

A systematic comparison of 14 retrospective motion correction pipelines found that those combining various strategies of signal regression and volume scrubbing reduced the fraction of connectivity edges contaminated by motion to <1%, compared to significant residual bias when using only simple rigid body motion correction [8].

Advanced Motion Phenotyping Frameworks

Beyond treating motion as a confounder, researchers are developing frameworks to quantify motion as a behavioral phenotype itself. The Motion Sensing Superpixels (MOSES) computational framework represents one such approach, measuring and characterizing biological motion with a superpixel "mesh" formulation [24]. This method enables systematic quantification of complex motion phenotypes in time-lapse imaging data, capturing both single-cell and collective migration patterns without requiring precise cell segmentation [24].

MOSES differs fundamentally from traditional Particle Image Velocimetry (PIV) approaches by enabling continuous tracking of cellular motion and extraction of rich feature sets that can be used to create distinctive motion "signatures" for different biological conditions [24]. This approach has been applied to study boundary formation dynamics between different epithelial cell types, revealing how complex cellular dynamics relate to pathological processes [24].

Diagram 1: MNAR Data Exclusion Pathway

The Scientist's Toolkit: Essential Materials and Methods

Table 3: Research Reagent Solutions for Motion Management and Phenotyping

Tool/Category	Specific Examples	Function/Application
Motion Correction Software	FSL's MCFLIRT, AFNI's 3dvolreg, ICA-AROMA	Retrospective motion correction for fMRI data
Motion Metrics	Framewise Displacement (FD), DVARS	Quantifying head motion for scrubbing thresholds
Motion Phenotyping Tools	MOSES (Motion Sensing Superpixels)	Quantitative analysis of cellular motion patterns
Quality Control Metrics	Surface Hole Number (SHN)	Automated quality assessment approximating manual ratings
Experimental Assays	Co-culture boundary formation assays	Study cell population dynamics and motion phenotypes

Experimental Protocols for Motion Phenotyping

The MOSES framework employs a specific methodology for motion phenotyping that can be adapted to various research contexts:

• Cell Culture and Preparation: Three epithelial cell lines (EPC2, CP-A, OE33) are cultured in pairwise combinations to model tissue boundaries [24].
• Fluorescent Labeling: Different cell populations are labeled with lipophilic membrane dyes (e.g., red and green fluorescent markers) to enable tracking [24].
• Co-culture Setup: Cells are separated by a removable divider (500µm width) in a 24-well plate, which is removed after 12 hours to allow migration [24].
• Time-Lapse Imaging: Live cell imaging over extended periods (up to 6 days) captures motion dynamics [24].
• MOSES Analysis: Application of the superpixel mesh formulation to extract motion features without requiring precise cell segmentation [24].
• Phenotype Classification: Unsupervised analysis of motion signatures to identify distinct motion phenotypes across experimental conditions [24].

Diagram 2: Motion as Phenotype vs. Exclusion

Addressing the MNAR problem in clinical research requires a fundamental shift in how we conceptualize and handle motion-related data. Rather than treating motion solely as a confounder to be eliminated, researchers should recognize it as a meaningful behavioral phenotype that provides insights into clinical status and symptom severity. Moving forward, the field must adopt more sophisticated approaches that maximize data retention through advanced correction methods while simultaneously analyzing motion patterns as clinically relevant variables. This dual approach—combining technical improvements in motion correction with analytical frameworks that incorporate motion as a behavioral measure—will enhance the validity, reproducibility, and clinical relevance of neuroimaging research across developmental and psychiatric disorders.

Large-scale neuroimaging datasets, such as the Adolescent Brain Cognitive Development (ABCD) Study, have been transformative for developmental neuroscience, offering unprecedented statistical power to detect subtle brain-behavior relationships. However, this power is predicated on data quality. A critical and pervasive challenge is in-scanner head motion, a systematic source of artifact that is not random noise. In developmental populations and individuals with certain behavioral traits, motion is more prevalent, creating a systematic bias that can inflate effect sizes and produce spurious findings. This case study examines how motion artifact led to inflated effect sizes within the ABCD Study, the methodologies used to uncover this bias, and the subsequent framework developed to quantify and mitigate its impact, a crucial consideration for both basic research and clinical drug development.

Quantitative Evidence of Motion-Induced Bias

Analyses of the ABCD dataset have provided concrete, quantitative evidence demonstrating how motion artifacts systematically bias structural and functional imaging measures.

Structural MRI (sMRI) Findings

A manual quality assessment of over 10,000 sMRI scans from the ABCD Study revealed that more than half (55%) were of suboptimal quality, even after passing standard automated quality control. This systematic data quality issue led to predictable biases in cortical measurements [3].

Table 1: Impact of Scan Quality on Cortical Measurements and Group Differences in ABCD sMRI Data

Scan Quality Inclusion	Impact on Cortical Measurements	Number of Significant Brain Regions Showing Group Differences (Aggressive Behavior)	Key Finding
High-Quality Scans Only (n=4,600)	Reference standard	3	Baseline effect
Including Moderate-Quality Scans	Consistent underestimation of cortical thickness; overestimation of cortical surface area [3]	21	Effect size in some regions more than doubled [3]
Including All Scans (n>10,000)	Introduction of systematic error growing as quality decreases [3]	43	Catapulted number of significant findings, indicating spurious associations [3]

Functional MRI (fMRI) Findings

In resting-state fMRI, head motion introduces a spatially systematic signature, decreasing long-distance connectivity and increasing short-range connectivity [25]. Even after rigorous denoising, residual motion artifact significantly confounds trait-FC relationships.

Table 2: Trait-Specific Motion Impact on Functional Connectivity (FC) in ABCD fMRI Data

Analysis Condition	Percentage of Traits with Significant Motion Overestimation	Percentage of Traits with Significant Motion Underestimation	Key Implication
After Standard Denoising (ABCD-BIDS)	42% (19/45 traits) [25]	38% (17/45 traits) [25]	Motion artifact remains a major confound even after standard processing
After Motion Censoring (FD < 0.2 mm)	Reduced to 2% (1/45 traits) [25]	No reduction (17/45 traits remained significant) [25]	Censoring mitigates overestimation but is ineffective for/ may worsen underestimation

Experimental Protocols for Quantifying Motion Artifact

Manual Quality Control of sMRI

• Objective: To evaluate the assumption that large sample sizes inherently overcome the noise introduced by poor-quality scans [3].
• Method: A team manually rated 10,295 sMRI scans from 9- and 10-year-olds in the ABCD Study using a 4-point scale (1 = minimal correction needed, 4 = unusable). This manual rating served as the gold standard against which automated metrics were compared [3].
• Analysis: Researchers examined how the inclusion of scans with different quality ratings (1-4) affected standard sMRI measures (cortical thickness, surface area) and the effect sizes in brain-behavior analyses (e.g., comparing children with/without aggressive behaviors) [3].

The SHAMAN Framework for fMRI

• Objective: To devise a trait-specific "motion impact score" to determine if specific brain-behavior relationships are confounded by residual motion [25].
• Method: Split Half Analysis of Motion Associated Networks (SHAMAN) capitalizes on the stability of traits over time. For each participant, the fMRI timeseries is split into high-motion and low-motion halves. SHAMAN measures the difference in the correlation structure between these halves [25].
• Analysis:
  • A significant difference between the halves indicates that motion impacts the trait-FC effect.
  • A motion impact score aligned with the trait-FC effect direction indicates overestimation.
  • A score opposite to the trait-FC effect indicates underestimation.
  • Permutation testing yields a p-value for the motion impact score, distinguishing significant from non-significant motion confounding [25].

Evaluating Automated Quality Metrics

• Objective: To identify a practical, automated alternative to labor-intensive manual quality control for large datasets [3].
• Method: The performance of automated quality-control metrics was compared against manual ratings. The metric known as Surface Hole Number (SHN), which estimates imperfections in cortical reconstruction, was found to best approximate manual quality ratings [3].
• Application: While not as effective as manual control, using SHN as a covariate or to "stress-test" results by analyzing how effect sizes change as low-quality scans are added/removed was proposed as a feasible best practice [3].

Motion Artifact Workflow and Impact

The following diagram illustrates the procedural pathway through which head motion introduces systematic bias into neuroimaging data analysis, leading to inflated effect sizes and spurious findings.

The Scientist's Toolkit: Key Research Reagents and Solutions

Implementing rigorous motion correction requires a suite of methodological "reagents." The table below details essential tools and approaches for mitigating motion bias, as identified in research on the ABCD Study.

Table 3: Essential Materials and Methods for Motion Mitigation in Neuroimaging

Tool/Solution Category	Specific Example	Function & Rationale
Automated QC Metrics	Surface Hole Number (SHN) [3]	An automated proxy for image quality that estimates imperfections in cortical surface reconstruction; used to flag potentially problematic scans without manual inspection.
Motion Censoring	Framewise Displacement (FD) thresholding (e.g., FD < 0.2 mm) [25]	Post-hoc removal of individual fMRI volumes with excessive motion. Effective at reducing motion overestimation but can introduce bias by disproportionately excluding data from certain populations [25].
Trait-Specific Motion Quantification	SHAMAN Framework [25]	A statistical method that assigns a specific "motion impact score" to a given brain-behavior association, distinguishing between overestimation and underestimation.
Inclusive Analysis Methods	Motion-Ordering & Bagging [26]	Statistical techniques that retain high-motion participants in analyses, improving sample representation and reproducibility of effect sizes, particularly for minoritized youth [26].
Denoising Algorithms	ICA-AROMA, FIX [8]	ICA-based algorithms that automatically identify and remove motion-related components from fMRI data without censoring entire volumes.
Effect Size Benchmarking	BrainEffeX Web App [27]	A resource providing "typical" effect sizes from large datasets, allowing researchers to compare their findings against benchmarks to identify potentially inflated results.

The investigation into motion-inflated effect sizes within the ABCD Study yields critical lessons for developmental neuroimaging and the application of large datasets in clinical neuroscience. Key takeaways include:

• Motion is a Systematic Bias, Not Random Noise: It introduces predictable, spatially structured artifacts that can inflate or deflate effect sizes, leading to both false positives and false negatives [3] [25].
• Quality Trumps Quantity: Simply increasing sample size without rigorous quality control can compound errors and produce specious, overly optimistic findings [3].
• Standard Pipelines Are Insufficient: The default preprocessing and QC pipelines of large, publicly available datasets may not adequately remove motion artifact, requiring additional vigilance from researchers [3] [25].
• Equity Implications: Standard exclusion practices for high-motion data can disproportionately affect clinically severe or minoritized populations, biasing samples and limiting generalizability [26] [8].

To ensure robust and reproducible results, researchers should adopt a multi-pronged strategy: employ trait-specific motion impact analyses like SHAMAN; use automated metrics like SHN to stress-test the robustness of findings; consider inclusive methods like motion-ordering to maintain representative samples; and benchmark observed effect sizes against realistic expectations from resources like BrainEffeX. For drug development professionals relying on neuroimaging biomarkers, a critical appraisal of motion correction methodologies is essential to de-risk decisions based on potentially confounded data.

Correcting the Signal: A Methodological Toolkit for Motion Artifact Mitigation

In-scanner head motion represents a fundamental confound in developmental neuroimaging, introducing systematic bias that impedes our understanding of neurodevelopmental mechanisms [28]. This technical challenge is particularly acute in pediatric populations, where increased head motion can lead to severe noise and artifacts in magnetic resonance imaging (MRI) studies, inflating correlations between adjacent brain areas and decreasing correlations between spatially distant territories [28]. The ramifications extend beyond technical inconvenience to potentially skew scientific findings, as motion artifacts have been shown to create spurious brain-behavior associations that can masquerade as neural effects [2] [25]. This whitepaper examines three core acquisition-based strategies—mock scanner training, real-time motion correction, and pediatric-friendly protocols—that collectively address this challenge at the data collection stage, before systematic biases become embedded in research datasets.

The problem is especially pronounced in large-scale neuroimaging initiatives. Recent analyses of major datasets like the Adolescent Brain Cognitive Development (ABCD) Study reveal that more than half of structural MRI scans may be of suboptimal quality, even after standard automated quality-control steps [3]. Incorporating these scans introduces systematic bias, consistently underestimating cortical thickness and overestimating cortical surface area in analyses [3]. For functional MRI (fMRI), the situation is equally concerning, as head motion introduces spatially systematic artifacts that decrease long-distance connectivity while increasing short-range connectivity, particularly affecting default mode network measurements [25]. Given that motion tendencies follow a U-shaped trajectory across development—with high motion in young children decreasing through adolescence and rising again in later adulthood—failure to address these acquisition challenges disproportionately affects studies of developmental populations [2].

Mock Scanner Training: Principles and Implementation

Mock scanner training involves placing participants in an environment designed to mimic the actual MRI scanning environment, with the dual purpose of desensitizing them to the unusual surroundings and training them to limit movement. This approach capitalizes on behavioral preparation and systematic desensitization to reduce anxiety and increase compliance, particularly crucial for children who may find the scanning environment intimidating. By familiarizing participants with scanner noises, confined spaces, and the requirement to remain still, mock training addresses both the psychological and physiological aspects of motion control.

Efficacy and Empirical Support

Recent research demonstrates that even brief mock scanner sessions yield substantial benefits for data quality. A growth curve study with 123 Chinese children and adolescents found that a single 5.5-minute training session in an MRI mock scanner effectively suppressed head motion during subsequent actual scanning [28] [29]. The study revealed that younger children (aged 6-9 years) derived the greatest benefit from such training, suggesting that mock scanning should be particularly prioritized for early childhood studies [28]. Another investigation examining longer scanning protocols found that mock scanner training, when combined with complementary in-scanner methods like weighted blankets and an incentive system, enabled the acquisition of low-motion fMRI data from pediatric participants (age 7-17) undergoing a 60-minute scan protocol [30]. This finding is significant because shortened scan protocols—a common approach to minimizing motion—reduce the reliability of functional connectivity measures, creating a tension between data quantity and quality [30].

The quantitative benefits of mock scanner training are substantial across multiple metrics, as summarized in Table 1.

Table 1: Efficacy Metrics of Mock Scanner Training

Metric	Without Mock Training	With Mock Training	Improvement	Study
Scans with mean FFD >0.10 mm	71.4%	32.3%	54.8% reduction	[30]
Scans with mean FFD >0.15 mm	50.0%	9.38%	81.2% reduction	[30]
Scans with mean FFD >0.20 mm	33.9%	4.17%	87.7% reduction	[30]
Optimal training duration	-	5.5 minutes	-	[28]
Maximum benefit age group	-	6-9 years	-	[28]

FFD = Frame-to-frame displacement

Implementation Protocol

Successful implementation of a mock scanner protocol involves multiple components that collectively prepare the child for the actual scanning environment:

• Environment Replication: The mock scanner should closely mimic the actual MRI environment, including bore size, lighting, and acoustic properties. Playing audio recordings of MRI sequence sounds throughout the session helps desensitize participants to the unusual noises they will encounter [31].

• Behavioral Training: Participants practice lying still with verbal feedback provided about head position. This often incorporates a visual feedback system where children can see real-time metrics of their head movement and learn to control it [30].

• Habituation Sessions: Multiple brief sessions may be more effective than a single extended session, particularly for children with anxiety or neurodevelopmental disorders.

• Positive Reinforcement: An incentive system that rewards successful stillness helps motivate participation and compliance. This can include token economies or small rewards for meeting motion thresholds [30].

The following workflow diagram illustrates a comprehensive mock scanning protocol that integrates these elements:

Mock Scanner Implementation Workflow

Real-Time Motion Monitoring and Correction

Real-time motion monitoring represents a technological approach to the motion challenge, providing immediate feedback to researchers and technicians during data acquisition. Unlike mock scanning, which is preventive, real-time monitoring is an active acquisition strategy that enables adaptive scanning protocols based on participant performance.

FIRMM and Real-Time Monitoring Efficacy

Framewise Integrated Real-Time MRI Monitoring (FIRMM) software exemplifies this approach by calculating and displaying head motion metrics (specifically framewise displacement, FD) to the MRI technician in real time during fMRI scans [31]. This enables technicians to extend scanning periods when participants are exhibiting low motion and potentially conclude scanning once sufficient high-quality data has been acquired, optimizing scanning efficiency. The software has demonstrated particular value in infant neuroimaging, where motion is especially prevalent and challenging to control.

In a comparative study of infant scanning with (n = 407) and without (n = 295) FIRMM, researchers found that adding real-time motion monitoring to state-of-the-art infant scanning protocols significantly increased the amount of usable fMRI data (defined as FD ≤ 0.2 mm) acquired per infant [31]. This advantage persisted across diverse infant populations, including both preterm and term-born infants, indicating its robustness across developmental stages. The real-time feedback enables a more dynamic approach to data acquisition than fixed-duration protocols, potentially reducing the need for repeated scanning sessions or excessive data collection to compensate for motion-contaminated frames.

Implementation Considerations

Successful implementation of real-time motion monitoring requires both technical infrastructure and procedural adaptations:

• Software Integration: FIRMM or comparable systems must be integrated with the MRI scanner's data output, typically requiring specific software installations and compatibility checks.

• Technician Training: MRI technicians must be trained to interpret real-time motion metrics and make informed decisions about scan continuation based on both motion data and protocol requirements.

• Protocol Adaptation: Scanning protocols may need adjustment to accommodate the flexible scanning durations enabled by real-time monitoring, particularly for task-based fMRI where complete paradigm administration remains important.

• Complementary Techniques: Real-time monitoring is most effective when combined with other motion mitigation strategies. For example, one study demonstrated successful implementation of FIRMM alongside natural sleep protocols in infants, using feed-and-swaddle techniques and vacuum-based immobilizers [31].

Pediatric-Friendly Scanning Protocols

Pediatric-friendly scanning protocols encompass a range of adaptations to the MRI environment and procedures that acknowledge the unique needs and characteristics of developing populations. These strategies focus on creating a supportive, minimally stressful environment that naturally facilitates reduced motion.

Core Protocol Components

Several evidence-based components comprise effective pediatric-friendly scanning protocols:

• Natural Sleep Protocols: For infants and young children, scanning during natural sleep represents one of the most effective motion reduction strategies. The "feed and swaddle" approach involves modifying feeding schedules to ensure feeding 30-45 minutes before scanning, followed by snug swaddling in pre-warmed sheets [31]. This approach is often complemented by vacuum immobilizers (e.g., MedVac Bag) that gently secure the infant's position when air is evacuated [31].

• Acoustic Adaptations: Playing audio recordings of MRI sequence sounds throughout the scan session can help infants stay asleep by minimizing disruptive changes in ambient noise [31]. Additionally, appropriate ear protection that is comfortable for extended wear is essential.

• Environmental Comfort: Pre-warming blankets, minimizing transitions between environments, and creating a calm, dimly lit scanning suite can reduce anxiety and promote stillness.

• Age-Appropriate Engagement: For older children who remain awake during scanning, age-appropriate explanations, engaging visual stimuli, and breaks when needed can improve compliance. Some protocols incorporate movie viewing during scans to maintain engagement and reduce motion [30].

Specialized Populations

Pediatric-friendly protocols require particular adaptation for special populations, including children with neurodevelopmental disorders such as autism spectrum disorder (ASD). Research indicates that while age is the strongest determinant of head motion across all pediatric populations, children with neurodevelopmental disorders may not display the typical pattern of decreasing motion with age seen in neurotypical children [2]. This highlights the need for persistent motion mitigation strategies throughout childhood for these populations. Studies have demonstrated that with comprehensive protocols incorporating mock scanning and in-scanner adaptations, even children with ASD can successfully complete extended scanning protocols with low motion [30].

Successful implementation of acquisition-based motion mitigation strategies requires specific materials and resources. The following table catalogues essential components of an effective motion mitigation toolkit for developmental neuroimaging research.

Table 2: Research Reagent Solutions for Motion Mitigation

Tool Category	Specific Examples	Function & Application	Evidence
Mock Scanner Systems	Mock scanner with audio recording of sequence sounds	Familiarizes participants with MRI environment; enables practice with stillness	[28] [29] [30]
Real-Time Monitoring Software	FIRMM (Framewise Integrated Real-Time MRI Monitoring)	Provides real-time head motion metrics to guide acquisition length	[31]
Immobilization Devices	MedVac Vacuum Splint Infant Immobilizer; weighted blankets	Gently secures position without distress; provides proprioceptive input	[31] [30]
Visual Feedback Systems	Real-time head motion display for participants	Enables children to visualize and control their head movement	[30]
Acoustic Adaptation Tools	MRI sequence sound recordings; appropriate ear protection	Maintains sleep state by minimizing disruptive noise changes	[31]
Environmental Comfort Items	Pre-warmed blankets; dimmable lighting; calming visuals	Reduces anxiety and promotes relaxation during scanning	[31] [30]

Integration and Best Practices

The most effective approach to mitigating motion artifacts in developmental neuroimaging involves integrating multiple strategies rather than relying on a single solution. The following diagram illustrates how these strategies can be combined throughout the research timeline:

Integrated Motion Mitigation Timeline

Synthesized Best Practices

Based on current evidence, the following integrated practices represent the state of the art in acquisition-based motion mitigation:

• Implement Brief Mock Scanner Training: A single 5.5-minute mock scanner session provides substantial motion reduction, particularly for children aged 6-9 years [28]. This should be standard practice in pediatric neuroimaging studies.

• Leverage Real-Time Monitoring for Scanning Efficiency: FIRMM software should be incorporated to guide acquisition length based on motion metrics, particularly for resting-state fMRI [31]. This approach reduces the need for oversampling while ensuring adequate high-quality data.

• Adapt Protocols to Developmental Stage: Infant protocols should prioritize natural sleep with feed-and-swaddle approaches, while protocols for older children should incorporate engagement strategies and clear, age-appropriate instructions [31] [30].

• Employ Complementary Immobilization: Weighted blankets and vacuum immobilizers provide gentle physical reminders to remain still without causing distress [30].

• Assess Motion Impact Post-Hoc: For studies examining traits associated with motion (e.g., psychiatric conditions), methods like SHAMAN should be employed to calculate motion impact scores for specific trait-FC relationships, distinguishing between overestimation and underestimation effects [25].

Acquisition-based strategies for mitigating head motion represent a crucial frontier in developmental neuroimaging methodology. As evidence mounts regarding the systematic biases introduced by motion artifacts—particularly in large-scale datasets—the implementation of robust, multi-modal approaches becomes increasingly imperative. Mock scanner training, real-time motion monitoring, and pediatric-friendly protocols each contribute distinct advantages to this effort, but their integration yields the most powerful protection against motion-related confounds.

The field is moving toward standardized incorporation of these methods in major neuroimaging initiatives, as evidenced by their adoption in studies like the HEALthy Brain and Child Development (HBCD) Study [32]. This trajectory acknowledges that rigorous science requires not only sophisticated analytical approaches but also meticulous attention to data quality at the acquisition stage. By implementing the strategies outlined in this whitepaper, developmental neuroscientists can substantially enhance the validity and reproducibility of their findings, ultimately accelerating our understanding of typical and atypical neurodevelopment.

In-scanner head motion represents a profound methodological confound in functional magnetic resonance imaging (fMRI), particularly for resting-state functional connectivity (RSFC) studies investigating neurodevelopmental trajectories. This technical challenge is especially acute when studying populations with naturally higher movement levels, such as children, or individuals with neurodevelopmental disorders like attention-deficit/hyperactivity disorder (ADHD) [33]. Motion introduces systematic biases that can create spurious group differences or mask genuine neurobiological relationships, potentially leading to erroneous conclusions about brain development and pathology [34] [25].

The fundamental problem stems from how motion affects the blood-oxygen-level-dependent (BOLD) signal. Head movement causes spatially varying signal changes that rigid-body realignment alone cannot fully correct, as it fails to address associated intensity variations and spin-history effects [34]. These motion artifacts manifest in RSFC as a characteristic pattern: decreased long-distance connectivity coupled with increased short-range connectivity [25]. Critically, developmental research faces the confounding reality that motion itself follows a developmental trajectory—children exhibit higher motion that decreases with age [33], creating the risk that observed age-related connectivity changes might reflect motion artifacts rather than neural maturation.

This whitepaper provides a comprehensive technical guide to the post-processing arsenal for mitigating motion-related bias, with particular emphasis on developmental neuroimaging applications where motion represents both a technical and interpretative challenge.

Fundamental Concepts: Quantifying and Characterizing Motion Artifacts

Framewise Displacement: The Standard Metric

Framewise displacement (FD) quantifies head motion between consecutive volumetric acquisitions (frames). It is calculated as the sum of the absolute values of the derivatives of the six realignment parameters (three translations and three rotations) [35]. Rotational displacements are converted from degrees to millimeters by calculating displacement on the surface of a sphere of radius 50 mm [35]. FD provides a single scalar value for each time point that represents the total extent of head movement, serving as the primary metric for identifying motion-corrupted volumes in censoring pipelines.

Table 1: Common Framewise Displacement Thresholds in Developmental Neuroimaging

FD Threshold	Typical Application Context	Trade-offs
0.2 mm	Conservative censoring for high-quality data [34] [25]	Maximizes artifact removal but may discard excessive data
0.3-0.4 mm	Moderate censoring for typical developmental studies	Balance between data retention and artifact removal
0.5 mm	Liberal censoring when using PACE [34] or with limited data	Preserves more data but may retain residual artifacts

The Developmental Motion Context

Understanding motion in developmental research requires acknowledging its systematic relationship with age and clinical status. Longitudinal research demonstrates that head motion significantly decreases as age increases during both childhood and adolescence [33]. Furthermore, children with ADHD display consistently greater FD than controls across development, and critically, this elevation persists even in children who experience symptomatic remission of ADHD [33]. These observations confirm that motion is not merely a technical nuisance but potentially reflects meaningful behavioral traits, complicating its removal without introducing sample bias.

The Motion Correction Pipeline: From Acquisition to Denoising

Effective motion mitigation requires an integrated approach spanning acquisition, prospective correction, and retrospective denoising strategies. The following workflow illustrates a comprehensive pipeline for addressing motion artifacts in developmental neuroimaging studies:

Prospective Motion Correction

Prospective methods correct for motion during data acquisition, potentially offering superior artifact reduction compared to purely retrospective approaches.

Prospective Acquisition Correction (PACE): This image-based online motion detection and correction sequence tracks head position to maintain fixed orientation relative to the scanner coordinate system, thereby reducing spin-history effects [34]. PACE provides two principal advantages: (1) effective elimination of significant negative motion-BOLD relationships associated with signal dropouts, and (2) enabling less aggressive censoring thresholds (0.5 mm versus 0.2 mm with conventional EPI) while achieving equivalent motion reduction [34]. Implementation requires no external devices or subject markers, making it suitable for high-throughput developmental studies.

Volumetric Navigators (vNavs): These prospective correction systems have demonstrated significant reduction in motion-induced bias and variance in brain morphometry [36], which is particularly valuable for developmental studies tracking structural brain changes across time.

Retrospective Denoising Algorithms

Retrospective methods operate on acquired data and represent the core of the post-processing arsenal.

Nuisance Regression Techniques

Nuisance regression removes unwanted signal components by modeling potential confounds.

Table 2: Nuisance Regression Methods for Motion Denoising

Method	Procedure	Advantages	Limitations
Global Signal Regression (GSR)	Regresses out average signal across all brain voxels [37]	Effectively removes widespread motion artifacts [34]	Introduces artificial negative correlations; may remove neural signal [34] [37]
WM-CSF Regression	Regresses out signals from white matter and cerebrospinal fluid masks [37]	Targets non-gray matter signals	Assumes WM/CSF contain only noise; may retain motion artifacts [37]
Motion Parameter Regression	Includes 6-36 motion parameters as regressors [34]	Directly models movement effects	Incomplete motion removal; may overfit data [34]
aCompCor	Uses principal components from WM/CSF masks as regressors [37]	Accounts for spatially varying noise	May remove neural signal along with noise
tCompCor	Uses principal components from high-variance voxels [37]	Data-driven noise component identification	Risk of removing neural signal from active regions

Volume Censoring (Scrubbing)

Censoring identifies and removes motion-corrupted time points exceeding specific FD thresholds. In conventional EPI data, conservative thresholds (FD < 0.2 mm) are often necessary, but with PACE-correction, higher thresholds (FD < 0.5 mm) can provide qualitatively equivalent artifact reduction [34]. For multiband acquisitions with sub-second TRs, adaptations like low-pass filtering motion parameters prior to FD calculation (LPF-FD) may improve performance [38].

ICA-Based Denoising

ICA-AROMA (Automatic Removal of Motion Artifacts) applies independent component analysis to decompose fMRI data into spatially independent components, then automatically identifies and removes motion-related components based on their spatial and temporal characteristics [37]. Compared to other ICA approaches, ICA-AROMA requires no classifier training, making it suitable for diverse developmental datasets.

Integrated Preprocessing Pipelines

Several integrated pipelines combine multiple denoising approaches:

fMRIPrep: A robust, standardized preprocessing pipeline that performs minimal preprocessing including motion correction, field unwarping, normalization, and brain extraction [39]. Its "glass box" philosophy emphasizes transparency and quality assessment through visual reports.

FuNP (Fusion of Neuroimaging Preprocessing): A wrapper software that combines components from AFNI, FSL, FreeSurfer, and Workbench to incorporate recent preprocessing developments into a single package [40]. FuNP provides both volume- and surface-based processing streams.

Experimental Protocols and Methodological Considerations

Implementing Framewise Displacement Censoring

A standardized protocol for FD-based censoring includes:

• Calculate FD from the 6 motion parameters (3 translations, 3 rotations) using the formula: FD(t) = |Δx| + |Δy| + |Δz| + |Δα| + |Δβ| + |Δγ|, with rotations converted to mm [35].
• Identify corrupted volumes where FD exceeds threshold (typically 0.2-0.5 mm depending on data characteristics and correction methods) [34] [25].
• Remove identified volumes along with one preceding and two following volumes to account for spin-history effects [34].
• Document data loss and ensure balanced retention across experimental groups to avoid introducing bias.

For multiband data, implement low-pass filtering of motion parameters (cutoff: 0.2 Hz) prior to FD calculation to reduce physiological contamination [38].

Evaluating Denoising Efficacy

Assessing denoising performance is crucial for methodological rigor:

QC-FC Correlations: Calculate correlations between subject-level motion (mean FD) and functional connectivity measures; effective denoising should minimize these relationships [38].

Distance-Dependent Artifacts: Examine whether connectivity artifacts (increased short-distance and decreased long-distance connectivity) persist after processing [25].

SHAMAN (Split Half Analysis of Motion Associated Networks): A novel method that assigns motion impact scores to specific trait-FC relationships, distinguishing between motion causing overestimation or underestimation of effects [25].

Table 3: Denoising Algorithm Efficacy Across Metrics (Based on [37])

Method	Physiological Noise Removal	Low-Frequency Signal Retention	Age-Related fcMRI Differences
ICA-AROMA	High	Low	Lower
GSR	High	Low	Lower
aCompCor	Moderate (better for high-frequency)	High	Higher
tCompCor	Moderate (better for high-frequency)	High	Higher
WM-CSF Regression	Variable	Moderate	Variable

Table 4: Essential Tools for Motion Denoising in Developmental Neuroimaging

Tool/Resource	Function	Application Context
Framewise Displacement Scripts	Quantifies head motion between volumes	Essential quality metric for all fMRI studies [35]
fMRIPrep	Standardized automated preprocessing pipeline	Reproducible preprocessing across large datasets [39]
ICA-AROMA	Automatic removal of motion components via ICA	Data-driven denoising without requiring physiological recording [37]
FuNP	Fusion pipeline combining multiple software tools	Integrated volume and surface-based processing [40]
SHAMAN	Quantifies motion impact on specific trait-FC relationships	Testing for residual motion effects in brain-behavior associations [25]
ABCD-BIDS Pipeline	Integrated denoising with respiratory filtering, GSR	Large-scale consortium studies (e.g., ABCD) [25]

Discussion and Future Directions

Despite extensive methodological advances, motion remains a persistent challenge in developmental neuroimaging. Recent evidence indicates that even after comprehensive denoising (e.g., with ABCD-BIDS pipeline), motion can still explain 23% of signal variance [25], and censoring at FD < 0.2 mm reduces but does not eliminate motion-related biases in trait-FC associations [25].

The field is moving toward approaches that explicitly quantify and account for residual motion effects rather than assuming complete elimination. Methods like SHAMAN [25] represent promising directions for estimating motion impact on specific research questions, particularly important when studying traits correlated with motion (e.g., ADHD symptoms). Future work should prioritize developing standardized reporting guidelines for motion mitigation procedures and thresholds specific to developmental populations, enabling better cross-study comparisons and replication.

For developmental researchers, the optimal post-processing arsenal combines prospective correction when feasible, appropriate volume censoring thresholds balanced against data retention needs, and denoising methods matched to specific research questions—acknowledging that motion-related bias can never be fully eliminated but must be thoughtfully managed and quantified.

::: {.author-info} This technical whitepaper is framed within the context of a broader thesis on systematic bias from motion in developmental neuroimaging research. :::

Motion artifacts represent a significant source of systematic bias in neuroimaging, particularly in developmental populations where subject movement is often unavoidable. This whitepaper provides an in-depth comparative analysis of four prominent motion artifact correction techniques—Principal Component Analysis (PCA), Spline Interpolation, Wavelet Analysis, and Kalman Filtering. Based on empirical evaluations using functional Near-Infrared Spectroscopy (fNIRS) and Magnetic Resonance Imaging (MRI) data, each method demonstrates distinct strengths and weaknesses in mitigating motion-induced noise. The evidence indicates that Wavelet-based analysis and Spline Interpolation often yield superior performance in recovering the hemodynamic response, significantly reducing mean-squared error (MSE) and enhancing contrast-to-noise ratio (CNR). For researchers and drug development professionals, the selection of an appropriate correction algorithm is paramount to ensuring the validity and reliability of neuroimaging biomarkers, especially in clinical trials and longitudinal developmental studies where motion can confound results.

In developmental neuroimaging research, systematic bias introduced by subject motion is a fundamental methodological challenge. Motion artifacts can severely compromise data quality, leading to inaccurate interpretations of brain function and structure. This is especially critical when studying pediatric, geriatric, or clinical populations, where the ability to remain still is often limited. In functional near-infrared spectroscopy (fNIRS), motion causes rapid shifts in the optical coupling between fibers and the scalp, producing high-frequency noise and baseline shifts that can obscure the underlying hemodynamic response function (HRF) [41]. Similarly, in magnetic resonance imaging (MRI), motion introduces ghosting, blurring, and non-linear signal perturbations that complicate anatomical and functional analyses [42] [43]. Left uncorrected, these artifacts introduce systematic error, threatening the internal validity of research findings and the accurate assessment of therapeutic interventions in drug development. This paper systematically evaluates four post-processing techniques designed to mitigate this bias at the algorithmic level.

Core Algorithmic Methodologies

This section details the fundamental principles and experimental protocols underlying each motion correction technique.

Principal Component Analysis (PCA)

• Core Principle: PCA is an orthogonal linear transformation that converts a set of observations of possibly correlated variables into a set of values of linearly uncorrelated variables called principal components. It operates on the premise that motion artifacts are a major source of variance in the data and appear coherently across multiple channels. The technique decomposes the multi-channel NIRS data (arranged as a matrix of time points by channels) and projects out the first few principal components that are assumed to represent the motion artifact [41] [44].
• Experimental Protocol (Standard PCA):
  • Input Data: A matrix of NIRS measurements (optical density or hemoglobin concentrations) with dimensions m (time points) × n (channels).
  • Decomposition: Compute the covariance matrix of the data and perform eigenvalue decomposition to identify the principal components, ranked by the percentage of variance they explain.
  • Component Removal: Remove a predetermined number of the first (largest variance) components.
  • Reconstruction: Reconstruct the data using the remaining components [41].
• Advanced Iteration (Targeted PCA - tPCA): A modified approach applies PCA not to the entire dataset, but only to epochs identified as containing motion artifacts. This targeted application seeks to overcome the issue of filtering desired physiological signals that can occur when standard PCA is applied globally [45].
  • Automatically identify motion-contaminated epochs on any channel.
  • Combine these epochs from all channels into a new data matrix.
  • Apply PCA to this motion-only matrix and remove components explaining up to 97% of its variance.
  • Stitch the corrected epochs back into the original time series, aligning the mean values of adjacent segments. This process is typically iterated 2-3 times for optimal results [45].

Spline Interpolation

• Core Principle: This method models the periods of motion artifacts using cubic spline interpolation. The interpolated spline, which represents the estimated motion artifact, is then subtracted from the original signal [41] [44].
• Experimental Protocol:
  • Motion Detection: Automatically identify the start and end points of motion artifacts on a channel-by-channel basis using an algorithm (e.g., hmrMotionArtifact in the HOMER2 package). Detection is typically based on thresholds for changes in signal amplitude (AMPthresh) and standard deviation (STDthresh) within a moving window (tMotion) [41] [45].
  • Spline Modeling: For each identified motion segment, fit a cubic spline curve to the data points. The parameter p (e.g., 0.99) controls the smoothness of the spline [45].
  • Artifact Subtraction: Subtract the fitted spline from the original data segment.
  • Baseline Adjustment: Shift the resulting segment by the difference between the mean value of the segment and the mean value of the preceding, motion-free segment to ensure continuity [45].

Wavelet Analysis

• Core Principle: Wavelet analysis uses a multi-resolution approach to decompose a signal into different frequency components. Motion artifacts, characterized by abrupt changes, appear as outliers in the wavelet coefficient distribution. These outlier coefficients are then thresholded and removed before signal reconstruction [41] [44].
• Experimental Protocol:
  • Decomposition: Perform a discrete wavelet transform (DWT) on the NIRS time-series, decomposing it into approximation and detail coefficients across multiple levels.
  • Statistical Thresholding: Assume the wavelet coefficients have a Gaussian distribution, with physiological signals centered around zero and motion artifacts as outliers. A tuning parameter α (e.g., 0.1) determines the threshold for identifying motion-corrupted coefficients [45]. Adaptive thresholding methods may also incorporate a linear prediction factor to minimize the MSE between noisy and original signals [46].
  • Coefficient Correction: Apply a thresholding function (e.g., soft or hard thresholding) to the detail coefficients to attenuate or zero out those attributed to motion.
  • Signal Reconstruction: Perform an inverse discrete wavelet transform using the corrected coefficients to reconstruct the denoised NIRS signal [45].

Kalman Filtering

• Core Principle: The Kalman filter is a recursive, state-space algorithm that estimates the state of a dynamic system from a series of noisy measurements. For motion correction, it models the physiological NIRS signal as the hidden state to be estimated, treating motion as part of the observation noise. It is particularly effective when the data contains periods of motion-free activity that can inform the model [41] [44].
• Experimental Protocol:
  • Model Definition: Define a state-space model where the system state represents the true, motion-free hemodynamic signal.
  • Prediction and Update: The filter operates in a two-step recursive process:
    • Prediction Step: Predict the current state and its uncertainty based on the previous state.
    • Update Step: Update the prediction using the new measurement, calculating a weighted average between the predicted state and the measured value, with more weight given to estimates with higher certainty.
  • Iterative Processing: As each new data point is acquired, the filter updates its estimate of the true underlying physiological signal, effectively filtering out the high-frequency, high-amplitude noise characteristic of motion artifacts [41].

Quantitative Performance Comparison

Empirical studies directly comparing these algorithms on real and simulated NIRS data provide critical performance metrics to guide method selection. The tables below summarize key quantitative findings.

Table 1: Performance Comparison in Recovering the Hemodynamic Response Function (HRF) from fNIRS Data [41].

Method	Average Reduction in MSE	Average Increase in CNR	Key Strengths
Spline Interpolation	55%	Not Specified	Most accurate recovery of simulated HRF
Wavelet Analysis	Not Specified	39%	Highest contrast-to-noise ratio improvement
PCA	Significant	Significant	Effective when motion is primary variance source
Kalman Filtering	Significant	Significant	Recursive processing; no additional inputs needed

Table 2: Performance of Advanced Methods (tPCA) vs. Established Leaders [45].

Method	Mean-Squared Error (MSE)	Pearson's Correlation (R²) to True HRF	Contextual Notes
Targeted PCA (tPCA)	Lowest	Highest	Statistically superior to wavelet and spline for Velcro probes
Wavelet Filtering	Higher than tPCA	Lower than tPCA	Effective denoising capability
Spline Interpolation	Higher than tPCA	Lower than tPCA	Performance depends on accurate motion detection

Recent research consolidates these findings, identifying Wavelet filtering and Temporal Derivative Distribution Repair (TDDR) as the most effective methods for preserving functional connectivity and network topology in fNIRS, a critical consideration for developmental brain research [47].

The Scientist's Toolkit: Essential Research Reagents and Materials

The following table outlines key software and analytical tools utilized in the empirical studies cited in this whitepaper.

Table 3: Key Research Tools and Software for Motion Correction Analysis.

Tool Name	Function / Application	Relevance to Motion Correction
HOMER2 Package	A comprehensive NIRS processing environment written in MATLAB.	Provides standardized functions (e.g., hmrMotionArtifact) for motion artifact detection and serves as a platform for implementing and comparing correction algorithms like spline interpolation and tPCA [41] [45].
WaveLab Toolbox	A MATLAB toolbox for wavelet analysis.	Used to implement the discrete wavelet transform and associated thresholding techniques required for wavelet-based motion artifact correction [45].
TechEn CW6 System	A continuous-wave NIRS acquisition system.	The specific hardware platform used in multiple foundational studies comparing motion correction techniques, ensuring consistency in data acquisition [41] [45].
Collodion-Fixed Fiber Probe	A probe secured with a clinical adhesive (Collodion).	Not an algorithm, but a hardware solution that minimizes motion artifact at the source. Serves as a gold-standard control in performance comparisons, often outperifying software-corrected data from standard Velcro probes [45].

Integrated Workflow and Decision Framework

The following diagram illustrates a logical workflow for selecting and applying motion correction strategies within a neuroimaging data processing pipeline, integrating the insights from the comparative analysis.

Systematic bias from motion is an unavoidable obstacle in developmental neuroimaging that demands robust, algorithmic solutions. This analysis demonstrates that while all four major correction techniques—PCA, Spline Interpolation, Wavelet Analysis, and Kalman Filtering—provide significant improvement over simple trial rejection or no correction, they are not interchangeable. Wavelet-based analysis consistently emerges as a top-performing and versatile method, excelling in CNR improvement, handling task-correlated motion, and preserving functional connectivity patterns [41] [47] [44]. Spline interpolation is highly effective for distinct, large-amplitude artifacts, offering the best fidelity in HRF recovery when motion segments are accurately identified [41]. The evolution of PCA into targeted PCA (tPCA) shows promise, particularly for data where motion is the dominant source of variance across channels [45].

For the research and pharmaceutical development community, these findings highlight that the choice of motion correction strategy is non-trivial and can directly impact study conclusions. The recommended path forward is a critical, data-informed approach rather than a one-size-fits-all application. Future work will focus on the development of fully automated, hybrid pipelines that intelligently select and combine the strengths of these algorithms, and on the deeper integration of prospective motion correction (e.g., via optical tracking [42]) with retrospective post-processing to further mitigate the pervasive influence of motion in vulnerable populations.

In developmental and psychiatric neuroimaging research, head motion introduces systematic bias that disproportionately affects vulnerable populations, including children and individuals with severe mental illnesses. This bias is not random noise; it constitutes a methodological confound that can spuriously alter measurements of cortical thickness, surface area, and functional connectivity [8] [25] [3]. Patients with psychotic disorders, for instance, exhibit significantly more head movement during scanning than healthy controls, potentially due to symptoms like psychomotor agitation, disorganization, or medication side effects [8]. When researchers exclude high-motion scans from analysis—a common practice—they introduce selection bias by systematically removing the most severely affected individuals, thus compromising the generalizability of findings to the full patient population [8]. This creates an urgent need for computational solutions that can salvage data quality from motion-corrupted acquisitions rather than discarding them.

Deep learning approaches for motion correction have emerged across imaging modalities. In brain PET imaging, supervised deep learning models have been developed to predict rigid motion parameters without external tracking devices [48]. Similarly, in ultrasound microvessel imaging, deep learning-based motion correction has improved diagnostic accuracy for thyroid nodule classification by preserving critical vascular structures [49]. However, these approaches typically address isolated artifacts. The Joint image Denoising and motion Artifact Correction (JDAC) framework represents a paradigm shift by simultaneously addressing two interrelated sources of image degradation—noise and motion artifacts—through an iterative learning approach specifically designed for 3D brain MRI [50] [51] [52]. This integrated methodology is particularly valuable for large-scale neuroimaging studies where data quality variability can introduce systematic errors that undermine the validity of brain-behavior associations.

Technical Foundation of the JDAC Framework

The JDAC framework addresses a critical limitation of conventional approaches: treating denoising and motion artifact correction as separate tasks. When severe noise and motion artifacts occur simultaneously in low-quality images, processing them independently leads to suboptimal results [50] [53] [52]. JDAC formulates this as a joint optimization problem, leveraging the Alternating Direction Method of Multipliers (ADMM) to decompose the problem into simpler subproblems that are solved iteratively [50] [53].

The framework consists of two core deep learning models working in tandem: an adaptive denoising model and an anti-artifact model [50] [51]. These models are employed iteratively within a unified framework that progressively refines image quality by implicitly exploring the underlying relationships between noise and motion artifacts [50]. A key innovation is the incorporation of a novel noise level estimation strategy that enables conditional processing and determines early stopping criteria for the iterative process [50] [52].

Table: Core Components of the JDAC Framework

Component	Architecture	Key Innovation	Function
Adaptive Denoising Model	U-Net with conditional normalization	Noise level estimation via gradient map variance	Progressively reduces noise while preserving anatomical details
Anti-Artifact Model	U-Net with gradient-based loss	Gradient L1 loss between corrected and ground truth images	Eliminates motion artifacts while maintaining structural integrity
Iterative Learning Framework	ADMM optimization	Early stopping based on noise level estimation	Coordinates both models to jointly improve image quality

The Iterative Learning Process

The JDAC framework employs an iterative refinement process where the denoising and anti-artifact models are applied sequentially through multiple cycles. The ADMM optimization decomposes the original problem into simpler subproblems [53]:

The problem is formulated as:

This leads to the augmented Lagrangian function:

where u is the Lagrange multiplier and ρ is a penalty parameter [53].

Within this optimization framework, the denoising subproblem is handled by the adaptive denoising model, while the anti-artifact subproblem is addressed by the motion correction model [53]. The iterative process continues until the estimated noise level falls below a predefined threshold Δ, implementing an early stopping strategy that prevents oversmoothing and accelerates computation [50] [53].

Core Methodological Components

Adaptive Denoising with Noise Level Estimation

The adaptive denoising model incorporates a novel noise estimation strategy that uses the variance of gradient maps to quantitatively estimate noise levels [50] [53] [52]. Specifically, the noise level σe is estimated as:

where ▽x represents the image gradient map [53]. This estimation enables conditioned denoising through a U-Net backbone with feature normalization based on the estimated noise variance [50] [51].

The denoising model is trained on 9,544 T1-weighted MRIs from the ADNI dataset with manually added Gaussian noise as supervision [50] [52]. The model is trained to predict the noise component of the input image, with the denoised result obtained by subtracting this predicted noise from the input [53]. The training utilizes a standard L1 loss function between the predicted and true noise [53].

Motion Artifact Correction with Anatomical Preservation

The anti-artifact model employs another U-Net architecture but incorporates a specialized gradient-based loss function designed specifically to maintain the integrity of brain anatomy during the motion correction process [50] [51] [52]. The loss function is defined as:

where m̂ is the motion-corrected MRI estimate and m is the ground truth motion-free MRI [53].

This model was trained on 552 T1-weighted MRIs with motion artifacts and paired motion-free images, incorporating both the standard L1 loss and the innovative gradient-based loss to ensure that critical anatomical details are preserved throughout the correction procedure [50] [52]. The gradient-based loss specifically helps prevent distortion of original brain structures in 3D MR images [50].

Research Reagent Solutions

Table: Essential Research Components for JDAC Implementation

Component	Specification	Function in Framework
Training Data (Denoising)	9,544 T1-weighted MRIs from ADNI	Provides supervised training data with synthetic Gaussian noise for denoising model
Training Data (Artifact Correction)	552 T1-weighted MRIs with paired motion-free scans	Enables training of anti-artifact model with real motion artifacts
U-Net Architecture	3D convolutional neural network	Backbone for both denoising and anti-artifact models; preserves spatial information
Noise Estimation Module	Gradient variance calculation	Quantifies noise levels for conditional denoising and early stopping criteria
ADMM Optimizer	Alternating Direction Method of Multipliers	Enables iterative decomposition of the joint optimization problem

Experimental Validation and Performance Analysis

Experimental Design and Datasets

The JDAC framework was rigorously validated on multiple datasets, including two public datasets and a real motion-affected MRI dataset from a clinical study [50] [52]. All MRIs underwent minimal preprocessing including skull stripping and intensity normalization to the range [0,1] [50]. The evaluation compared JDAC against several state-of-the-art methods including BM4D (a 3D extension of the classical BM3D denoising algorithm) and DDPM (denoising diffusion probabilistic models) [50] [52].

Table: Dataset Composition for JDAC Validation

Dataset	Sample Size	Purpose	Preprocessing
ADNI	9,544 T1-weighted MRI scans	Pretraining and validating denoising models	Skull stripping, intensity normalization to [0,1]
MR-ART	552 T1-weighted MRI scans	Training anti-artifact model	Skull stripping, intensity normalization to [0,1]
NBOLD	Clinical motion-affected MRIs	Real-world validation	Skull stripping, intensity normalization to [0,1]

Quantitative Results and Comparative Performance

Experimental results demonstrated JDAC's effectiveness in both denoising and motion artifact correction tasks compared to state-of-the-art methods [50] [51]. The framework's iterative learning strategy proved particularly beneficial for motion-affected MRIs with severe noise [50].

A critical ablation study examined the importance of noise level estimation by comparing JDAC with a degraded variant (JDACw/oE) that omitted this component [50]. Results demonstrated that JDAC outperformed JDACw/oE across all evaluation metrics for both denoising and anti-artifact tasks, confirming that explicit noise level estimation enhances performance [50].

The gradient-based loss function in the anti-artifact model successfully maintained brain anatomy integrity throughout the motion correction process, addressing a crucial requirement for neuroimaging research where preserving structural relationships is essential for valid morphometric analyses [50] [52].

Implications for Developmental Neuroimaging Research

The JDAC framework offers significant potential for addressing systematic biases in developmental neuroimaging research. By effectively correcting motion artifacts in 3D brain MRI, it enables researchers to retain data from participants who would otherwise be excluded due to excessive movement [50] [8]. This is particularly crucial for studies involving children, adolescents, and clinical populations with neurological or psychiatric conditions, as these groups typically exhibit higher motion levels that, if not properly addressed, can introduce spurious findings [8] [25] [3].

Research has shown that incorporating lower-quality motion-affected images in analyses can substantially alter effect sizes and lead to false positives [3]. In the ABCD Study, for instance, including moderate-quality scans more than doubled effect sizes in some brain regions when examining cortical volume differences related to behavioral problems [3]. The JDAC approach mitigates this issue by salvaging data quality rather than discarding valuable data points, thus enhancing both statistical power and result validity.

For large-scale brain-wide association studies (BWAS), the ability to process 3D volumetric data without slice-by-slice processing represents a significant advantage, as it preserves important 3D anatomical information that is lost when using 2D methods [50] [52]. This ensures continuity across imaging planes and maintains the integrity of brain structures in all dimensions, providing more reliable data for investigating brain-behavior relationships in developmental populations.

The JDAC framework represents a significant advancement in addressing the dual challenges of noise and motion artifacts in brain MRI through an iterative learning approach that jointly optimizes both tasks. By integrating adaptive denoising with motion artifact correction in a 3D processing framework, it overcomes limitations of conventional methods that process these artifacts separately or rely on 2D slice-by-slice approaches [50] [51] [52].

Future research directions include extending the framework to other imaging modalities such as functional MRI and diffusion tensor imaging, which face similar challenges with motion artifacts [8] [25]. Additionally, adapting the approach for real-time correction during image acquisition could further enhance its clinical utility [48]. As neuroimaging studies continue to grow in scale and scope, methodologies like JDAC will play an increasingly vital role in ensuring data quality and minimizing systematic biases that undermine the validity of research findings, particularly in developmental populations and clinical groups prone to movement during scanning.

Systematic bias from participant motion is a fundamental challenge in developmental neuroimaging research, particularly in studies involving children, adolescents, and clinical populations. This bias does not constitute random noise but rather introduces systematic artifacts that can invalidate research findings and lead to false scientific conclusions [3]. In functional MRI (fMRI), head motion induces spurious correlations and reduces long-distance connectivity, while in structural MRI (sMRI), it can systematically underestimate cortical thickness and overestimate cortical surface area [3] [25]. These effects are particularly pronounced in large-scale studies like the Adolescent Brain Cognitive Development (ABCD) Study, where over half of scans may be of suboptimal quality, potentially introducing confounds that correlate with traits of clinical interest [3] [25].

Reference-based techniques, including Real-Time Location Systems (RLAS) and external motion tracking, offer promising approaches to quantify, mitigate, and correct for these motion-induced artifacts. By providing independent measures of head movement, these technologies enable researchers to distinguish true neural signals from motion-related artifacts, thereby increasing the validity and reproducibility of developmental neuroimaging findings.

Understanding Motion Artifacts in Neuroimaging

Characterization and Impact of Motion Artifacts

Motion artifacts manifest differently across neuroimaging modalities but consistently introduce systematic bias rather than random noise. The table below summarizes key motion artifact characteristics across major imaging modalities:

Table 1: Characteristics of Motion Artifacts in Different Neuroimaging Modalities

Modality	Primary Motion Effects	Systematic Biases Introduced	Vulnerable Populations
Resting-state fMRI	Decreased long-distance connectivity; Increased short-range connectivity [25]	Spurious brain-behavior associations; Altered default mode network connectivity [25]	Children with ADHD, autism, psychiatric conditions [25]
Structural MRI	Underestimated cortical thickness; Overestimated surface area [3]	Inflated effect sizes in group comparisons; False positive findings [3]	Young children, elderly, neuropsychiatric populations [3]
Task-based fMRI	Signal distortions; Misalignment of brain volumes	Altered activation maps; Reduced statistical power	Populations with limited compliance

In fMRI, motion produces spatially systematic artifacts because its effects follow the physics of MRI signal acquisition [25]. The systematic nature of these artifacts means they do not average out with larger sample sizes and can create the illusion of meaningful brain-behavior relationships where none exist [3]. For example, in the ABCD Study, standard denoising pipelines leave approximately 23% of signal variance explained by head motion, which can still produce significant spurious associations [25].

Quantitative Assessment of Motion Impact

Recent methodological advances enable precise quantification of motion's impact on specific research findings. The Split Half Analysis of Motion Associated Networks (SHAMAN) method assigns a motion impact score to specific trait-functional connectivity relationships, distinguishing between motion causing overestimation or underestimation of effects [25].

Application of SHAMAN to ABCD Study data revealed that after standard denoising, 42% (19/45) of traits had significant motion overestimation scores, while 38% (17/45) had significant underestimation scores [25]. Motion censoring at framewise displacement (FD) < 0.2 mm reduced significant overestimation to 2% (1/45) of traits but did not decrease the number of traits with significant motion underestimation scores, indicating complex motion-effect relationships that require sophisticated correction approaches [25].

Reference-Based Motion Tracking Technologies

Real-Time Location Systems (RLAS)

Real-Time Location Systems provide continuous, precise tracking of head position within the MRI scanner. While traditional camera-based systems face limitations including line-of-sight requirements and limited tracking accuracy, emerging technologies offer enhanced capabilities:

Table 2: Comparison of Motion Tracking Technologies

Technology	Key Features	Accuracy/Precision	Limitations
Camera-based Optical	Market leaders: Hexagon, FARO, Nikon [54]	Sub-millimeter to millimeter	Requires line-of-sight; Limited to controlled environments [55]
Magnetic Induction	Wireless, wearable, real-time; No drift [55]	High precision for relative position	Emerging technology; Limited commercial availability
Inertial Measurement Units	Wearable sensors	High temporal resolution	Position drift over time [55]
Laser Tracking	High-precision metrology [54]	Sub-millimeter accuracy	Expensive; Complex setup

Emerging Tracking Technologies

Recent innovations in motion tracking focus on overcoming limitations of traditional approaches. Magnetic induction-based systems represent a particularly promising development:

Magnetic Induction Tracking: Researchers at USC's ACME Lab have developed the first wireless, wearable, real-time motion-tracking network based on magnetic induction (MI) [55]. This system uses a network of lightweight sensors with custom-designed MI transceiver chips that measure changes in magnetic coupling between sensor pairs to track position and orientation [55]. Unlike inertial measurement systems that suffer from drift, this approach provides absolute positioning similar to GPS, enabling drift-free motion tracking without requiring calibration [55].

The system features compact, low-power, low-cost sensors that can function in any environment without line-of-sight restrictions, making them particularly suitable for tracking naturalistic movements in clinical populations and children [55]. The technology has been experimentally validated on both joint models and human volunteers, demonstrating potential for neuroimaging applications [55].

Experimental Protocols and Methodologies

Implementation Framework for Reference-Based Motion Correction

The following diagram illustrates a comprehensive workflow for implementing reference-based motion correction in developmental neuroimaging studies:

Motion Correction Workflow - A systematic approach from study design to analysis validation.

SHAMAN Analytical Protocol

The Split Half Analysis of Motion Associated Networks (SHAMAN) provides a robust method for quantifying trait-specific motion effects. The methodology capitalizes on the observation that traits (e.g., cognitive measures) remain stable during scanning while motion varies second-to-second [25].

Experimental Protocol:

• Data Requirements: One or more resting-state fMRI scans per participant with framewise displacement (FD) values calculated for each volume [25].
• Data Splitting: Divide each participant's timeseries into high-motion and low-motion halves based on FD values [25].
• Trait-FC Calculation: Compute trait-functional connectivity effects separately for high-motion and low-motion halves [25].
• Impact Score Calculation:
  • Calculate difference in correlation structure between halves
  • Aligned direction with trait-FC effect = motion overestimation score
  • Opposite direction = motion underestimation score [25]
• Statistical Validation: Use permutation testing and non-parametric combining across connections to derive significance values [25].

The following diagram illustrates the SHAMAN analytical process:

SHAMAN Analysis Flow - A method to quantify motion's impact on brain-behavior relationships.

Quality Control and Validation Protocols

Robust quality control is essential for mitigating motion-related bias. For structural MRI, the following protocol is recommended:

• Image Quality Assessment:
  • Manual quality rating using a standardized scale (1-4) [3]
  • Automated metrics like Surface Hole Number (SHN) as proxy for manual ratings [3]
• Stratified Analysis:
  • Conduct primary analysis on highest-quality scans only
  • Systematically add lower-quality scans while monitoring effect sizes
  • Identify regions where effects remain stable versus those that inflate [3]
• Validation Sampling: Apply methods like imputation or reweighting using validation samples with more comprehensive confounder data [56].

Implementation of these protocols in the ABCD Study revealed that including scans with moderate quality (rating 2) more than doubled effect sizes in some brain regions, indicating substantial bias introduction [3].

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Tools for Motion Tracking and Correction in Neuroimaging

Tool/Category	Specific Examples	Function/Purpose	Implementation Considerations
Motion Tracking Systems	Hexagon laser trackers [54]; Magnetic induction sensors [55]	Provide independent measure of head position	Compatibility with MRI environment; Accuracy vs. cost trade-offs
Quality Control Metrics	Framewise Displacement (FD); Surface Hole Number (SHN) [3]	Quantify scan quality and motion contamination	SHN approximates manual quality ratings for sMRI [3]
Analytical Tools	SHAMAN implementation [25]; BrainEffeX web app [27]	Calculate motion impact scores; Explore effect sizes	SHAMAN requires timeseries data; BrainEffeX provides reference effect sizes
Denoising Algorithms	ABCD-BIDS pipeline [25]; Global signal regression	Remove motion artifacts from fMRI data	ABCD-BIDS includes respiratory filtering, motion regression, despiking [25]
Validation Frameworks	Consensus clustering [57]; Classifier-based corroboration	Verify clustering results against motion confounds	Protects against false cluster identification from motion artifacts [57]

Implementation Guidelines and Best Practices

Technology Selection Framework

Choosing appropriate motion tracking technology requires consideration of multiple factors:

• Population Characteristics: Studies of children or clinical populations may prioritize wearable, unobtrusive systems that minimize participant burden [55].
• Research Questions: Investigations requiring precise anatomical measurements (e.g., cortical thickness) may necessitate higher-precision tracking systems [3].
• Environmental Constraints: Research conducted in naturalistic settings may benefit from wireless systems without line-of-sight requirements [55].
• Budget and Expertise: Commercial laser tracking systems represent established solutions but require significant investment, while emerging technologies may offer cost-effective alternatives [54].

Analytical Implementation Guidelines

Successful implementation of reference-based techniques requires careful analytical planning:

• A Priori Thresholds: Define motion exclusion thresholds before data analysis based on published standards (e.g., FD < 0.2mm for fMRI) and study population characteristics [25].
• Validation Analyses: Conduct sensitivity analyses excluding participants with high motion and compare effect sizes across quality strata [3].
• Motion Impact Assessment: Apply methods like SHAMAN to quantify potential bias in key findings, particularly for traits correlated with motion [25].
• Effect Size Interpretation: Use resources like BrainEffeX to reference typical effect sizes in neuroimaging and contextualize findings [27].

Reference-based techniques for motion tracking and correction represent essential methodologies for ensuring the validity of developmental neuroimaging research. By providing independent measures of head movement, technologies like RLAS and emerging magnetic induction systems enable researchers to distinguish true neural signals from motion-induced artifacts. When combined with robust analytical frameworks like SHAMAN and comprehensive quality control protocols, these approaches significantly reduce the systematic bias that has plagued many neuroimaging findings. As the field moves toward larger-scale studies and more diverse populations, implementing these reference-based techniques will be crucial for producing reproducible, valid insights into brain development and function.

Optimizing Your Pipeline: Practical Strategies for Robust Developmental Imaging

In large-scale developmental neuroimaging research, the presumption that large sample sizes inherently mitigate the effects of noisy data is challenged by a critical source of systematic bias: motion artifact. This technical guide delineates protocols for implementing rigorous quality control (QC) frameworks that integrate manual visual rating with automated metrics, specifically Surface Hole Number (SHN), to identify and correct for motion-related bias in structural MRI (sMRI). Evidence from the Adolescent Brain Cognitive Development (ABCD) Study demonstrates that inadequate QC can undermine sample size advantages, leading to biased effect sizes and spurious clinical associations. This whitepaper provides researchers and drug development professionals with detailed methodologies and tools to enhance the validity of neurodevelopmental findings.

Large, population-based MRI studies promise transformative insights into neurodevelopment and mental illness risk. However, MRI studies of youth are especially susceptible to motion artifacts, which are more frequent in children and individuals with neuropsychiatric conditions [12] [3]. Contrary to the assumption that large sample sizes average out this noise, recent research establishes that motion artifact introduces systematic bias rather than random error.

Analyses incorporating lower-quality sMRI scans consistently underestimate cortical thickness and overestimate cortical surface area [12] [3]. This bias is non-random; it can alter apparent effect sizes, increasing risks for both false-positive and false-negative findings in clinical association analyses [12]. In the ABCD Study, for example, including scans of moderate quality more than doubled the number of significant brain regions in an analysis of aggressive behaviors, indicating inflation of effect sizes rather than stabilization [3]. This evidence underscores that rigorous QC is not merely a data-cleaning step but a fundamental requirement for valid inference in developmental neuroimaging.

Manual Quality Control (MQC): The Gold Standard

Manual QC involves the visual inspection of sMRI scans by trained raters to assess overall image quality and identify artifacts. It remains the benchmark against which automated metrics are evaluated.

Experimental Protocol and Rating Scale

The following methodology, derived from the ABCD Study's analysis of over 10,000 scans, provides a replicable protocol for manual QC [12].

• Preprocessing: Process T1-weighted volumes through a standardized segmentation pipeline (e.g., FreeSurfer).
• Rater Training and Calibration: Employ a single rater or a team of raters with demonstrated high inter-rater reliability. Continuous calibration is essential to prevent "rater drift."
• Blinding: Raters must be blinded to all participant-level information (e.g., demographics, clinical status) to prevent confirmation bias.
• Rating Procedure: View each MRI volume individually and assign a rating based on a pre-defined scale, as detailed in the table below.

Table 1: Manual Quality Control (MQC) Rating Scale

MQC Rating	Description	Approx. Proportion of ABCD Baseline Scans	Implication for Analysis
1 (Excellent)	Minimal or no artifacts; requires minimal to no manual edits.	45.0% (n=4,630)	Ideal for analysis; reference standard.
2 (Good)	Moderate artifacts; requires moderate manual edits.	39.5% (n=4,063)	Introduces measurable bias; inclusion requires caution.
3 (Poor)	Substantial artifacts; requires substantial manual edits.	13.4% (n=1,383)	Introduces significant bias; strong recommendation for exclusion.
4 (Unusable)	Severe artifacts; segmentation failures.	2.1% (n=219)	Must be excluded from analysis.

Impact of Manual QC on Cortical Measures

Quantitative analysis reveals a linear relationship between MQC rating and cortical measurements. Compared to MQC=1 scans, lower-quality scans are associated with:

• Decreased cortical thickness across much of the cortical mantle.
• Increased cortical surface area in lateral/superior regions.
• Decreased cortical surface area in medial/inferior regions [12].

The following workflow diagram summarizes the manual QC process and its impact on the analytical pipeline:

Automated Metrics: Surface Hole Number (SHN)

While manual QC is the gold standard, it is prohibitively time-consuming for massive datasets. Automated QC metrics are essential for scalability, with Surface Hole Number (SHN) showing particular promise.

What is Surface Hole Number?

SHN is an automated index of topological complexity generated during cortical surface reconstruction (e.g., in FreeSurfer). It estimates the number of holes or imperfections in the reconstructed surface model. These "holes" often occur due to motion artifacts, signal dropout, or segmentation errors, reflecting the brain's surface's failure to be reconstructed as a topologically correct, continuous sheet.

Protocol for Validating SHN Against MQC

To implement SHN as a proxy for scan quality, researchers must first validate its performance against manual ratings on a subset of their data.

• Data Sampling: Select a representative sample of scans (e.g., n=500-1000) from your study that have undergone manual QC.
• SHN Extraction: Extract the SHN value for each scan from the surface reconstruction output.
• Statistical Analysis: Perform a Receiver Operating Characteristic (ROC) analysis to determine SHN's ability to discriminate between different MQC rating tiers (e.g., MQC=1 vs. MQC≥2).

Table 2: Performance of Surface Hole Number (SHN) as an Automated QC Metric

Performance Metric	Value at Baseline	Value at Year 2 Follow-up	Interpretation
Specificity	0.81 – 0.93	0.88 – 1.00	SHN excels at correctly identifying high-quality scans.
Sensitivity	(Information not specified in search results)	(Information not specified in search results)	Requires validation on your dataset.
Key Strength	Differentiates lower-quality scans with good specificity.	Performance maintained or improved in a longitudinal sample.	Effective as a screening tool to flag potentially problematic scans.

Integrating SHN into the QC Workflow

SHN should not blindly replace manual QC but rather be used to stratify data and stress-test results. The following diagram illustrates a hybrid QC workflow:

The Scientist's Toolkit: Essential Research Reagents and Materials

This table details key resources and tools required for implementing the QC framework described in this guide.

Table 3: Research Reagent Solutions for Neuroimaging QC

Item Name	Function/Description	Example/Note
Structural MRI Data	Raw T1-weighted image volumes are the primary data for analysis.	Acquired from large-scale studies (e.g., ABCD) or local cohorts.
Processing Software	Software suite for cortical reconstruction and volumetric segmentation.	FreeSurfer [12]; provides automated metrics like SHN.
Visualization Platform	Tool for displaying and manually inspecting 3D MRI volumes.	FreeSurfer's freeview, FSLeyes, or similar DICOM viewers.
Surface Hole Number (SHN)	Automated metric of topological defects in surface reconstruction.	Extracted from FreeSurfer output (e.g., ?h.orig.nofix surface files).
Statistical Software	Environment for data analysis, ROC curves, and covariate control.	R, Python with pandas/scikit-learn, SPSS, or MATLAB.
Quality Control Registry	A structured database (e.g., CSV, SQL) to store MQC ratings and SHN values.	Critical for tracking quality and stratifying datasets for analysis.

Systematic bias from motion artifact presents a formidable challenge to the integrity of developmental neuroimaging research. This guide demonstrates that a rigorous, multi-layered quality control approach is non-negotiable. By integrating the gold standard of manual rating with scalable, validated automated metrics like Surface Hole Number, researchers can mitigate bias, stress-test their findings, and ensure that the transformational insights promised by large-scale studies are built upon a foundation of reliable data.

In-scanner head motion represents a fundamental challenge in developmental neuroimaging, introducing systematic biases that compromise the integrity of brain-behavior associations. This technical guide examines the censoring dilemma—the critical tension between removing motion-contaminated data to reduce false positives and preserving sample representativeness to avoid selection bias. Drawing on recent large-scale studies including the Adolescent Brain Cognitive Development (ABCD) Study and Human Connectome Project, we synthesize evidence demonstrating that conventional denoising approaches leave substantial residual motion artifact while common exclusion practices systematically bias samples against vulnerable populations. We present quantitative frameworks for evaluating trait-specific motion impacts and provide methodological recommendations for optimizing this balance in developmental neuroimaging research and clinical drug development applications.

Head motion constitutes the largest source of artifact in functional MRI data, introducing systematic biases that disproportionately affect developmental neuroimaging and clinical populations [25]. The physical characteristics of MRI physics create non-linear artifacts that resist complete removal through standard denoising algorithms, particularly in resting-state functional connectivity (FC) where the timing of underlying neural processes is unknown [25]. This vulnerability is especially problematic when studying traits associated with motion, such as psychiatric disorders, attention-deficit/hyperactivity disorder, and autism spectrum conditions, where participants inherently demonstrate higher in-scanner movement [25].

The censoring dilemma emerges from two competing statistical problems: false positive inflation from motion-induced spurious correlations versus selection bias from systematic exclusion of high-motion participants. Even after comprehensive denoising, residual motion artifact continues to impact trait-FC relationships, with recent studies finding 42% of traits examined showed significant motion overestimation and 38% showed significant underestimation [25]. Simultaneously, exclusion based on motion thresholds creates samples that no longer represent the target populations, particularly for developmental studies where motion correlates with important demographic and clinical variables [58].

Quantitative Evidence: The Impact of Censoring Decisions

Efficacy of Denoising and Residual Motion Effects

Recent evidence demonstrates that even sophisticated denoising pipelines leave substantial motion-related variance. After minimal processing (motion correction only), 73% of fMRI signal variance is explained by head motion. Following comprehensive denoising using the ABCD-BIDS pipeline (including global signal regression, respiratory filtering, motion timeseries regression, and despiking), 23% of signal variance remains explained by motion—a 69% relative reduction but substantively significant residual contamination [25].

The residual motion effects demonstrate systematic spatial patterns, with motion-FC effect matrices showing strong negative correlation (Spearman ρ = -0.58) with average FC matrices, indicating that connection strength is systematically weaker in participants who move more [25]. This motion-related decrease in FC often exceeds the magnitude of trait-related FC effects, potentially overwhelming genuine neurobiological signals.

Table 1: Motion Artifact Reduction Through Processing Pipelines

Processing Stage	Variance Explained by Motion	Relative Reduction	Key Components
Minimal processing	73%	Baseline	Motion correction only
ABCD-BIDS denoising	23%	69%	Global signal regression, respiratory filtering, motion parameter regression, despiking
FD < 0.2 mm censoring	Further reduction	Varies by trait	Exclusion of high-motion frames

Threshold Effects in Motion Censoring

The stringency of motion censoring thresholds creates a fundamental tradeoff between data quality and sample retention. Analysis of the ABCD dataset demonstrates that censoring at framewise displacement (FD) < 0.2 mm reduces significant motion overestimation from 42% to just 2% of traits [25]. However, this same threshold does not reduce the number of traits with significant motion underestimation scores, indicating complex and trait-specific impacts of censoring decisions.

Critically, the relationship between exclusion thresholds and sample characteristics follows a predictable pattern, with more stringent thresholds resulting in greater systematic bias. Analysis of ABCD study data reveals that the odds of participant exclusion relate to a broad spectrum of behavioral, demographic, and health-related variables [58]. This creates a self-reinforcing cycle where the populations most critical for understanding developmental psychopathology are systematically excluded from neuroimaging samples.

Table 2: Participant Characteristics Associated with Exclusion in Developmental Samples

Characteristic Domain	Specific Variables	Impact on Exclusion Odds
Socioeconomic factors	Area Deprivation Index, Child Opportunity Index, parental education, family income	Significantly increased
Clinical characteristics	ADHD symptoms, autism features, impulsivity	Significantly increased
Cognitive performance	Executive function, inhibitory control	Significantly increased
Trauma exposure	Single and multiple traumatic events	Significantly increased
Physical characteristics	Body mass index	Significantly increased

Methodological Frameworks for Quantifying Motion Impacts

The SHAMAN Framework for Trait-Specific Motion Impact Scoring

The Split Half Analysis of Motion Associated Networks (SHAMAN) approach provides a novel method for computing trait-specific motion impact scores that can distinguish between motion causing overestimation or underestimation of trait-FC effects [25]. This framework capitalizes on the observation that traits (e.g., cognitive abilities, clinical symptoms) remain stable over the timescale of an MRI scan, while motion represents a state that varies second-to-second.

Experimental Protocol:

• Split each participant's fMRI timeseries into high-motion and low-motion halves based on framewise displacement
• Measure differences in correlation structure between split halves
• Compute motion impact score by comparing trait-FC effects between halves
• Aligned direction (positive) indicates motion overestimation of trait-FC effect
• Opposite direction (negative) indicates motion underestimation of trait-FC effect
• Permutation testing and non-parametric combining across connections yields significance values

This methodological innovation enables researchers to move beyond universal motion thresholds to trait-specific evaluations of motion impact, acknowledging that different brain-behavior relationships demonstrate varying vulnerability to motion artifacts.

Normative Modeling and Sample Size Considerations

Normative modeling approaches for neuroimaging markers provide critical insights into the sample size requirements for robust developmental neuroscience. Simulation studies demonstrate that precise estimation of outlying percentiles (1st, 5th, 10th)—the most clinically relevant metrics—requires surprisingly large samples (N ≫ 1000) [59]. Performance evaluation of these models must assess both bias and variance, with uncertainty dramatically increasing at the ends of age ranges where fewer data points exist.

Experimental Protocol for Normative Model Evaluation:

• Create plausible ground truth distributions of neuroimaging markers (e.g., hippocampal volumes across age)
• Repeatedly simulate samples for sizes ranging from 50 to 50,000 data points
• Fit range of normative models to each simulated sample
• Compare fitted models and variability across repetitions to ground truth
• Focus evaluation on outer percentiles (1st, 5th, 10th) as most clinically relevant
• Quantify both bias and variance across the age range

These simulations reveal that flexible models perform better across sample sizes, especially for non-linear ground truth, and highlight the substantial uncertainty in model performance even with what would typically be considered large samples in neuroimaging contexts [59].

Alternative Methodological Approaches

Pairwise Interaction Statistics for Functional Connectivity Mapping

Beyond conventional Pearson's correlation, researchers can select from numerous pairwise interaction statistics that demonstrate varying sensitivity to motion artifacts and different network properties. A comprehensive benchmarking study evaluated 239 pairwise statistics from 6 families of measures, finding substantial quantitative and qualitative variation across FC methods [60].

Measures such as covariance, precision, and distance display multiple desirable properties, including improved correspondence with structural connectivity and enhanced capacity to differentiate individuals and predict behavioral differences [60]. Precision-based statistics consistently demonstrate strong alignment with multiple biological similarity networks and may be particularly robust to certain motion effects due to their accounting for shared network influences.

Experimental Protocol for FC Method Benchmarking:

• Extract regional time series from resting-state fMRI data
• Calculate multiple pairwise interaction statistics (covariance, precision, spectral, distance, information-theoretic)
• Evaluate topological features (hub distribution, weight-distance relationships)
• Quantify structure-function coupling with diffusion MRI
• Assess individual fingerprinting capacity
• Test brain-behavior prediction performance

Artifact Suppression and Evaluation Techniques

Advanced signal processing techniques can potentially salvage physiological data from artifact-contaminated recordings. In magnetoencephalography (MEG) with deep brain stimulation implants, methods like temporal signal space separation (tSSS), independent component analysis, and null beamforming have demonstrated efficacy in suppressing magnetic artifacts while preserving neural signals [61].

Evaluation of these techniques using machine learning classification of spatiotemporal patterns during visual perception tasks (unaffected by stimulation) confirms that neural data can be recovered from artifact-contaminated recordings, with comparable classification performance between artifact-suppressed and clean conditions [61].

Table 3: Research Reagent Solutions for Motion Management

Tool Category	Specific Methods	Function/Purpose	Implementation Considerations
Motion quantification	Framewise displacement, DVARS	Quantify degree of head motion	Standardized metrics enable cross-study comparison
Denoising pipelines	ABCD-BIDS, global signal regression, respiratory filtering	Reduce motion-related variance	Multiple approaches often needed; order of operations matters
Censoring methods	Scrubbing, spike regression	Remove high-motion frames	Threshold selection critical for bias-quality balance
Trait-specific evaluation	SHAMAN framework	Quantify motion impact on specific brain-behavior relationships	Requires sufficient data for split-half analysis
Artifact suppression	tSSS, ICA, adaptive filtering	Remove artifacts while preserving signal	Validation needed for specific populations
Normative modeling	GAMLSS, flexible nonlinear models	Establish reference distributions for individual assessment	Large samples (N>1000) required for precision
Missing data handling	Multiple imputation, full information maximum likelihood	Address bias from non-random missingness	Requires understanding of missing data mechanisms

Recommended Best Practices and Future Directions

Based on current evidence, we recommend a tiered approach to addressing the censoring dilemma in developmental neuroimaging:

• Implement trait-specific motion impact assessment using methods like SHAMAN to evaluate whether specific brain-behavior relationships of interest are vulnerable to motion artifacts [25].

• Formally account for missing data using multiple imputation or related approaches rather than listwise deletion, particularly when studying motion-correlated traits [58].

• Evaluate multiple pairwise statistics for functional connectivity mapping, selecting measures optimized for specific research questions and with known robustness properties [60].

• Prioritize transparent reporting of quality control procedures, exclusion rates, and characteristics associated with missingness to enable evaluation of potential selection biases [58].

• Develop cohort-specific optimization of processing pipelines rather than universal thresholds, acknowledging that motion artifacts and denoising efficacy may vary across populations.

Future methodological development should focus on techniques that minimize the impact of motion during data acquisition, improve artifact suppression without signal loss, and establish normative reference distributions for functional connectivity across development. The integration of hardware improvements with advanced analytical approaches represents the most promising path forward for resolving the censoring dilemma in developmental neuroimaging.

In-scanner head motion represents one of the most significant methodological challenges in developmental neuroimaging research, particularly in large-scale studies involving children, adolescents, and clinical populations. Motion artifacts systematically alter functional connectivity (FC) measurements, potentially leading to both false-positive and false-negative findings in brain-behavior associations [25]. This technical burden is especially pronounced in psychiatric research, where patient populations such as those with psychotic disorders exhibit significantly more head movement than healthy controls, creating a systematic bias that can distort our understanding of brain-based disorders [8]. The Adolescent Brain Cognitive Development (ABCD) Study, with its extensive neuroimaging and behavioral data from over 11,000 children, has revealed that poor image quality affects more than half of structural MRI scans and introduces systematic bias that undermines the advantages of large sample sizes [3] [62]. In this context of non-random noise, methods for quantifying and correcting motion-related artifacts have become essential for validating trait-functional connectivity relationships.

The Motion Artifact Problem in Functional Connectivity

Systematic Nature of Motion Artifacts

Head motion during fMRI acquisition introduces spatially systematic biases in functional connectivity metrics. Analyses of the ABCD dataset demonstrate that motion primarily decreases long-distance connectivity while increasing short-range connectivity, most notably within the default mode network [25]. This pattern creates a distinctive signature where the motion-FC effect matrix shows a strong negative correlation (Spearman ρ = -0.58) with the average FC matrix, indicating that participants who move more consistently show weaker connection strengths across the brain [25]. Critically, the effect sizes of motion on FC can be larger than the trait-FC effects of interest, potentially obscuring or mimicking true neurobiological relationships [25].

The problem is particularly acute in developmental and psychiatric neuroimaging because motion is not randomly distributed across populations. Individuals with attention-deficit/hyperactivity disorder, autism spectrum disorder, and psychotic disorders typically exhibit higher in-scanner head motion than neurotypical participants [8] [25]. This creates a systematic confound wherein motion-correlated traits are especially vulnerable to spurious brain-behavior associations. Even with standard denoising procedures like the ABCD-BIDS pipeline, which includes global signal regression, respiratory filtering, and motion parameter regression, 23% of signal variance remains explained by head motion [25].

Limitations of Current Denoising Approaches

Common motion mitigation strategies include:

• Volume censoring (scrubbing): Removing high-motion frames exceeding framewise displacement thresholds
• Motion parameter regression: Including realignment parameters as nuisance regressors
• Global signal regression: Removing global signal variations potentially associated with motion
• ICA-based approaches: Identifying and removing motion-related components (e.g., ICA-AROMA)

However, these approaches have significant limitations. Volume censoring creates a natural tension between removing motion-contaminated data and maintaining statistical power and sample representativeness [25] [8]. This is particularly problematic for clinical populations, as excluding high-motion participants systematically removes the most severely affected individuals, creating a "missing not at random" (MNAR) problem that limits generalizability and biases effect size estimates [8]. Furthermore, Siegel et al. note that standard denoising methods cannot distinguish whether residual motion artifact causes overestimation or underestimation of specific trait-FC effects [25].

The SHAMAN Framework: Principles and Methodology

Conceptual Foundation

The Split Half Analysis of Motion Associated Networks (SHAMAN) framework was developed to address critical gaps in existing motion correction methodologies. SHAMAN capitalizes on a fundamental observation about the temporal characteristics of traits versus motion: traits (e.g., cognitive abilities, psychiatric symptoms) are stable over the timescale of an MRI scan, while motion is a state that varies from second to second [25]. This conceptual foundation enables researchers to determine whether specific trait-FC relationships are genuinely neural in origin or substantially impacted by residual motion artifact.

The method assigns a motion impact score to specific trait-FC relationships, with directional information indicating whether motion causes overestimation or underestimation of effects [25] [63]. This is particularly valuable for brain-wide association studies (BWAS) involving thousands of participants, where subtle effects might be disproportionately influenced by motion artifacts despite sophisticated denoising pipelines.

Analytical Workflow

The SHAMAN methodology operates through a structured workflow:

• Data Preparation: Resting-state fMRI data is processed through standard denoising pipelines (e.g., ABCD-BIDS) without motion censoring to preserve temporal structure.

• Split-Half Partitioning: For each participant, the fMRI timeseries is divided into high-motion and low-motion halves based on framewise displacement (FD) metrics.

• Connectivity Calculation: Functional connectivity matrices are computed separately for high-motion and low-motion halves.

• Trait-FC Effect Estimation: The relationship between traits and FC is quantified within each motion half.

• Motion Impact Scoring: Differences in trait-FC effect sizes between high-motion and low-motion halves are computed, with permutation testing to establish statistical significance.

• Directional Classification: Significant differences aligned with the direction of the trait-FC effect are classified as "motion overestimation scores," while opposite effects are classified as "motion underestimation scores" [25].

The following diagram illustrates the core logical workflow of the SHAMAN framework:

Quantitative Evidence from Large-Scale Applications

Impact on Behavioral Traits in the ABCD Study

Application of SHAMAN to 45 traits from n = 7,270 participants in the ABCD Study revealed substantial motion-related impacts on trait-FC relationships:

Table 1: Motion Impact on Traits in ABCD Study After Standard Denoising

Motion Impact Category	Percentage of Traits	Number of Traits	Key Characteristics
Significant Overestimation	42%	19/45	Effect sizes inflated by motion artifact
Significant Underestimation	38%	17/45	True effects masked by motion artifact
Minimal Motion Impact	20%	9/45	Robust trait-FC relationships

After standard denoising with ABCD-BIDS without motion censoring, the vast majority of traits (80%) showed significant motion impact scores [25]. This demonstrates that even with comprehensive denoising, residual motion artifact substantially influences most trait-FC relationships.

Effectiveness of Motion Censoring Thresholds

The SHAMAN framework was used to evaluate the effectiveness of different motion censoring strategies:

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

Censoring Threshold	Overestimation Impact	Underestimation Impact	Residual Motion Effects
No Censoring	42% of traits	38% of traits	Strong negative correlation between motion-FC and average FC (ρ = -0.58)
FD < 0.2 mm	2% of traits	38% of traits	Reduced but persistent negative correlation (ρ = -0.51)
Optimal Threshold	Dramatically reduces overestimation	Does not address underestimation	Balanced approach needed for specific research questions

Censoring at FD < 0.2 mm effectively reduced significant overestimation from 42% to 2% of traits, demonstrating its utility for controlling false positives [25]. However, this stringent threshold did not decrease the number of traits with significant motion underestimation scores, highlighting how aggressive motion removal can perpetuate false negatives [25].

Implementation Protocols for Developmental Neuroimaging

Experimental Design Considerations

Implementing motion impact validation requires strategic experimental design:

• Scanning Protocol Optimization:

  • Acquire longer resting-state scans (≥15 minutes) to enable robust split-half analysis
  • Implement real-time motion monitoring (e.g., FIRMM software) to identify problematic sessions while the participant is still in the scanner [25]
  • Include structured breaks to reduce participant fatigue and cumulative motion

• Participant Preparation:

  • Develop age-appropriate behavioral training protocols, especially for pediatric and clinical populations
  • Use mock scanners to acclimatize participants to the scanning environment
  • Implement motivational frameworks to encourage compliance during scanning

• Data Acquisition Parameters:

  • Consider multi-echo sequences for improved motion robustness [25]
  • Ensure sufficient spatial and temporal resolution to balance sensitivity and motion resilience
  • Collect physiological monitoring data (cardiac, respiratory) for advanced denoising

Analytical Implementation

The analytical implementation of motion impact scoring involves both standard and specialized processing steps:

• Preprocessing Pipeline:

  • Implement standard preprocessing (motion correction, normalization, smoothing)
  • Apply denoising strategies (global signal regression, CompCor, ICA-AROMA)
  • Generate framewise displacement (FD) and DVARS metrics for quality assessment

• SHAMAN-Specific Processing:

  • Divide preprocessed timeseries into high-motion and low-motion halves based on FD median split
  • Compute functional connectivity matrices for each half using predefined atlas parcellations
  • Calculate trait-FC correlations for each motion half separately
  • Compute motion impact score as the difference in trait-FC effects between halves

• Statistical Validation:

  • Perform permutation testing (typically 1,000-10,000 iterations) to establish significance
  • Apply false discovery rate (FDR) correction for multiple comparisons across connections
  • Calculate confidence intervals for motion impact scores using bootstrapping methods

Research Reagent Solutions Toolkit

Table 3: Essential Tools for Motion Impact Analysis

Research Tool	Function	Implementation Considerations
Framewise Displacement (FD)	Quantifies head movement between successive fMRI volumes	Standard metric, but threshold selection requires balancing data quality and retention
SHAMAN Algorithm	Computes motion impact scores for specific trait-FC relationships	Requires sufficient temporal data for split-half analysis; adaptable to different denoising pipelines
Surface Hole Number (SHN)	Automated quality metric for structural MRI	Identifies topological errors in cortical reconstruction; good specificity for scan quality [3]
ICA-AROMA	ICA-based automatic removal of motion artifacts	Effectively removes motion-related components; may remove neural signal with high motion [8]
FIRMM Software	Real-time motion monitoring during scanning	Enables prospective motion correction; reduces costs by identifying problematic sessions early [25]
ABCD-BIDS Pipeline	Standardized denoising for ABCD Study data	Includes respiratory filtering, motion regression, despiking; reduces motion-related variance by 69% [25]

Integration with Quality Control Frameworks

Multidimensional Quality Assessment

Effective motion impact validation requires integration with comprehensive quality control frameworks:

• Structural Data Quality:

  • Implement manual quality ratings using standardized scales (1-4 point quality ratings) [3]
  • Utilize automated metrics like Surface Hole Number (SHN) to identify topological errors
  • Assess bias patterns (e.g., systematic underestimation of cortical thickness in low-quality scans)

• Functional Data Quality:

  • Calculate frame-to-frame displacement metrics (FD) for scrubbing decisions
  • Compute DVARS to identify abrupt changes in BOLD signal intensity
  • Assess spatial correlation patterns for characteristic motion signatures

• Data Exclusion Protocols:

  • Implement tiered exclusion criteria based on multiple quality metrics
  • Document exclusion rationales to evaluate potential selection biases
  • Report quality metrics alongside results to enable evaluation of motion impacts

The SHAMAN framework enables "reverse stress testing" of neuroimaging conclusions by systematically evaluating how results change under different motion correction strategies:

• Motion Censoring Sweep Analysis:

  • Recompute analyses across a range of FD thresholds (0.1-0.5mm)
  • Track how effect sizes change with increasing stringency of motion removal
  • Identify stability points where results become robust to censoring level

• Quality Covariate Analysis:

  • Include quality metrics (mean FD, SHN) as covariates in statistical models
  • Compare effect sizes with and without quality covariate adjustment
  • Report variance explained by quality metrics alongside primary results

• Subsample Robustness Testing:

  • Replicate analyses in highest-quality data subsets
  • Compare results between full sample and quality-restricted samples
  • Evaluate whether effects persist after excluding potentially problematic data

The following diagram illustrates this comprehensive quality integration framework:

The SHAMAN framework for computing motion impact scores represents a significant advancement in validating trait-FC relationships against systematic motion artifacts. By providing quantitative, trait-specific measures of motion impact, this methodology enables researchers to distinguish robust neurobiological associations from motion-contaminated findings. Evidence from the ABCD Study demonstrates that even with state-of-the-art denoising, residual motion substantially impacts the majority of trait-FC relationships, with differential effects on overestimation versus underestimation of effects.

Integration of motion impact validation into developmental neuroimaging research requires multidimensional quality assessment, reverse stress testing of conclusions across processing variants, and transparent reporting of quality metrics. As large-scale neuroimaging studies increasingly seek to identify subtle brain-behavior relationships, rigorous motion impact validation becomes essential for generating reproducible findings and advancing our understanding of neurodevelopmental disorders.

Head motion is the largest source of artifact in structural and functional MRI (fMRI) signals, introducing systematic bias rather than random noise into neuroimaging data [25]. In developmental, psychiatric, and neurological populations, excessive head motion disproportionately excludes precisely those individuals who may represent more severe or distinct clinical phenotypes, creating a Missing Not at Random (MNAR) problem that limits the generalizability of findings [8]. This creates a fundamental tension in research: removing too much data compromises statistical power, while including poor-quality data introduces systematic error [25]. This technical guide provides evidence-informed protocols to mitigate motion-related bias through integrated behavioral, environmental, and technical adaptations.

Understanding Motion as a Behavioral Phenotype

In populations with psychotic disorders, children, and other clinical groups, head movement during scanning may represent more than mere artifact; it can reflect core behavioral phenotypes of the condition [8]. Patients who struggle to remain still often exhibit higher levels of psychomotor agitation, anxiety, disorganization, or paranoia—symptoms that typically indicate more severe or acute presentation [8]. Excluding these individuals systematically biases study samples toward milder illness presentations and potentially obscures important neurobiological relationships.

The statistical implications are profound. When data missingness relates to the underlying severity of the condition being studied, it violates the assumptions of most standard inferential statistical methods, including t-tests and ANOVAs, potentially yielding biased parameter estimates and invalid inferences [8]. For example, if patients with the most severe hippocampal volume reduction produce unusable scans due to motion, excluding those scans will bias estimated average volume toward larger values, underestimating the true effect [8].

Table 1: Motion as a Source of Systematic Bias in Neuroimaging

Bias Mechanism	Impact on Data	Population Most Affected
Systematic Exclusion	Under-represents severe clinical phenotypes	Psychosis, ADHD, autism, child populations [8]
MNAR Data Patterns	Violates statistical assumptions, biased estimates	All motion-correlated clinical traits [8]
Spurious Brain-Behavior Links	False positive/negative trait-FC relationships	Populations with motion-correlated traits [25]
Measurement Error	Underestimates cortical thickness, overestimates surface area [3]	All populations, but magnitude varies by motion

Comprehensive Adaptation Framework

Behavioral and Environmental Adaptations

Successful protocol design for high-motion populations begins before participants enter the scanner. Evidence supports implementing structured behavioral protocols to enhance compliance and reduce motion:

• Practice Mock Sessions: Expose participants to scanner environment and sounds using simulated scanners to build familiarity and reduce anxiety-driven movement [8]
• Clear, Repeated Instructions: Provide simplified instructions with visual aids, especially for populations with cognitive impairments [8]
• Positive Reinforcement Systems: Implement reward incentives for successful compliance [8]
• Media Engagement: Use child-friendly media content during scan breaks to maintain engagement [8]
• Comfort Optimization: Address physical discomfort with specialized padding and positioning aids [8]

These interventions recognize that motion stems from multiple sources, including difficulty following instructions due to disorganization, restlessness due to psychomotor agitation or anxiety, exaggerated discomfort due to paranoia or claustrophobia, or medication side effects such as akathisia [8].

Technical Acquisition Adaptations

During data acquisition, several technical strategies can mitigate motion effects:

• Prospective Motion Correction (P-MoC): Advanced scanners integrate real-time tracking and correction by updating slice acquisition coordinates based on detected movement [8]
• Real-Time Monitoring Systems: Systems that monitor head motion frame-by-frame can signal operators to pause scans or extend acquisition until sufficient low-motion data are collected [8]
• Sequence Optimization: Implement sequences with inherent motion robustness, such as multi-echo fMRI, which provides additional information for denoising [25]

While these technological solutions show promise, they are not yet widely available due to complexity and hardware limitations [8]. Most studies still rely heavily on behavioral approaches and retrospective correction methods.

Processing and Analytical Adaptations

After data collection, rigorous processing pipelines can address residual motion artifacts:

• Motion Censoring (Scrubbing): Identify and remove volumes with excessive motion using framewise displacement (FD > 0.2-0.5 mm) or DVARS thresholds [8] [25]
• Motion Parameter Regression: Include motion parameters as covariates in group-level statistical models [8]
• Advanced Denoising Algorithms: Implement ICA-based approaches (ICA-AROMA, FIX) to identify and remove motion-related components without discarding entire volumes [8]
• Trait-Specific Motion Impact Assessment: Use methods like SHAMAN (Split Half Analysis of Motion Associated Networks) to quantify motion impact on specific trait-FC relationships [25]

Recent evidence indicates that even with standard denoising pipelines like ABCD-BIDS (which includes global signal regression, respiratory filtering, and motion parameter regression), 23% of signal variance may still be explained by head motion [25]. This underscores the need for complementary approaches.

Table 2: Technical Adaptations Across the Research Workflow

Research Phase	Adaptation Strategy	Implementation Considerations
Participant Preparation	Mock scanner training, clear instructions	Requires specialized equipment, staff time [8]
Data Acquisition	Prospective motion correction, real-time monitoring	Limited hardware availability, increases scan time [8]
Quality Control	Manual rating, automated metrics (e.g., Surface Hole Number)	SHN approximates manual ratings but doesn't eliminate error [3]
Data Processing	Motion censoring, parameter regression, ICA-based denoising	Censoring at FD < 0.2 mm reduces overestimation but not underestimation [25]
Statistical Analysis	Motion impact scores, covariate adjustment	SHAMAN distinguishes overestimation vs. underestimation [25]

Diagram 1: Comprehensive protocol workflow for high-motion populations, integrating adaptations across all research stages

Experimental Protocols for Motion Mitigation

The SHAMAN Methodology for Trait-Specific Motion Impact

The Split Half Analysis of Motion Associated Networks (SHAMAN) methodology addresses the critical need to quantify motion impact on specific trait-functional connectivity relationships [25]. This approach recognizes that traits (e.g., cognitive measures, clinical symptoms) remain stable during scanning, while motion varies second-to-second.

Protocol Implementation:

• Data Splitting: Divide each participant's resting-state fMRI timeseries into high-motion and low-motion halves based on framewise displacement (FD)
• Trait-FC Calculation: Compute correlation between the trait and FC separately for each half
• Impact Score Calculation: Calculate the difference in trait-FC effects between halves
• Directional Interpretation:
  • Motion impact score aligned with trait-FC effect direction indicates overestimation
  • Opposite direction indicates underestimation
• Statistical Testing: Use permutation testing and non-parametric combining across connections to establish significance

Application in the ABCD study revealed that after standard denoising, 42% (19/45) of traits had significant motion overestimation scores, while 38% (17/45) had significant underestimation scores [25]. Censoring at FD < 0.2 mm reduced significant overestimation to 2% (1/45) of traits but did not decrease underestimation, highlighting the complex relationship between motion correction and bias.

Integrated Motion Mitigation Protocol

Participant Preparation Phase:

• Conduct mock scanner session (30-45 minutes) with gradual exposure to scanner sounds and environment
• Provide clear, simplified instructions with visual aids repeated at multiple timepoints
• Establish reward system for successful compliance with age-appropriate incentives

Data Acquisition Phase:

• Implement prospective motion correction if available
• Use real-time motion monitoring with operator alerts for excessive movement
• Acquire multiple shorter runs rather than single extended acquisition
• Include structural sequences less susceptible to motion effects

Quality Control Protocol:

• Manual rating of scan quality using standardized scale (1-4, with 1=minimal correction needed, 4=unusable)
• Supplemental automated quality control using Surface Hole Number (SHN) or similar metrics
• Establish quality thresholds prior to analysis based on pilot data

Data Processing Pipeline:

• Apply standard denoising pipeline (e.g., ABCD-BIDS with global signal regression, respiratory filtering, motion parameter regression)
• Implement motion censoring at FD < 0.2 mm threshold
• Apply ICA-based denoising (ICA-AROMA) to remove motion-related components
• Generate motion summary metrics (mean FD, number of censored volumes) for covariate inclusion

Diagram 2: Data processing pipeline with multiple motion mitigation stages

Table 3: Research Reagent Solutions for Motion Mitigation

Tool/Resource	Primary Function	Implementation Considerations
Mock Scanner	Participant acclimatization to scanner environment	Requires dedicated space and equipment; high initial cost [8]
Framewise Displacement (FD)	Quantifies head movement between successive fMRI volumes	Standard metric; threshold of 0.2-0.5 mm typically used for censoring [8] [25]
ICA-AROMA	Identifies and removes motion-related independent components	Reduces need for volume censoring but may over-clean neural signals [8]
Surface Hole Number (SHN)	Automated quality metric estimating cortical reconstruction imperfections	Approximates manual quality ratings; useful for large datasets [3]
SHAMAN Algorithm	Quantifies trait-specific motion impact on functional connectivity	Distinguishes overestimation vs. underestimation; requires custom implementation [25]
Prospective Motion Correction	Real-time slice acquisition adjustment during scanning	Limited hardware availability; most effective for small movements [8]

Protocol design for high-motion populations requires a multipronged approach that addresses behavioral, environmental, and technical dimensions simultaneously. By recognizing motion as a potential behavioral phenotype rather than mere artifact, researchers can develop more inclusive recruitment and retention strategies that preserve sample representativeness. The integration of rigorous quality control procedures, advanced processing pipelines, and trait-specific motion impact assessment methods like SHAMAN provides a comprehensive framework for reducing systematic bias in developmental neuroimaging. As field standards evolve, explicit reporting of motion mitigation strategies and their potential impacts on results will enhance interpretation and reproducibility across studies of high-motion populations.

The predicted age difference (PAD), defined as the difference between an individual's predicted brain age and their chronological age, has emerged as a significant phenotype in developmental neuroimaging research [64]. This metric is increasingly used to characterize how an individual deviates from a healthy brain aging trajectory, with positive PAD (where brain age appears older than chronological age) correlating with neurological degeneration, cognitive impairments, and conditions such as schizophrenia [64]. However, the validity of PAD as a reliable biomarker is fundamentally challenged by systematic biases that arise from both statistical artifacts and motion-related confounding in MRI data acquisition [64].

In developmental neuroimaging research, head motion during scanning represents a particularly pernicious source of systematic bias [8]. This problem is especially pronounced in populations with severe mental illnesses or developmental disorders, where patients exhibit significantly more head movement compared to healthy controls due to factors including psychomotor agitation, anxiety, paranoia, or medication side effects [8]. Excessive head motion creates artifacts that systematically alter structural and functional connectivity measurements, potentially leading to false positive or negative findings in brain age estimation [25]. When researchers exclude scans with excessive motion—a common quality control practice—this introduces missing not at random (MNAR) data, disproportionately removing the most severely affected individuals and systematically biasing the study sample toward less severe clinical presentations [8].

The Nature of Systematic Bias in Brain Age Prediction

Statistical Origins of Bias

The systematic bias observed in brain age prediction arises from fundamental statistical properties of regression models rather than being a limitation of specific algorithms [64]. This phenomenon, known as "regression dilution" or "regression to the mean," occurs because predicted brain age ($\hat{Y}$) and PAD ($\hat{Y}-Y$) are mathematically constrained to be orthogonal [64]. For linear regression models, this forces an angular relationship between PAD and chronological age ($Y$) between 0 and 90 degrees, ensuring they remain correlated [64]. The consequence is consistent over-prediction of age for relatively younger individuals and under-prediction for elderly individuals, regardless of the underlying neural characteristics [64].

Motion-Induced Bias in Neuroimaging

Head motion introduces spatially systematic artifacts into functional connectivity (FC) measurements that are not completely removed by standard denoising algorithms [25]. Motion artifacts consistently decrease long-distance connectivity and increase short-range connectivity, most notably in the default mode network [25]. This creates a distinctive signature in connectivity patterns that can confound brain age predictions, particularly because motion correlates with specific clinical conditions and developmental stages [8] [25]. The problem is especially acute in large-scale developmental studies like the Adolescent Brain Cognitive Development (ABCD) Study, where researchers found that even after denoising, 23% of signal variance was still explained by head motion [25].

Table 1: Characteristics of Systematic Bias in Brain Age Estimation

Bias Type	Statistical Cause	Manifestation in Brain Age Prediction	Impact on PAD Reliability
Regression Dilution	Non-Gaussian distribution of chronological age and mathematical constraints of regression	Over-prediction for younger individuals, under-prediction for older individuals	PAD remains correlated with chronological age, reducing validity as independent biomarker
Motion Artifact	Spatially systematic changes in functional connectivity patterns	Decreased long-distance connectivity, increased short-range connectivity	Introduces confounds that correlate with clinical conditions, potentially creating spurious brain-behavior associations
Selection Bias	Exclusion of high-motion scans creating MNAR data	Underrepresentation of more severe clinical presentations in final sample	Limits generalizability of findings to full patient population

Established Bias Correction Methodologies

Sample-Level Linear Correction Methods

The most widely adopted bias correction approaches operate at the sample level, applying linear adjustments to the entire dataset after brain age prediction [64]. These methods assume a consistent bias pattern across all ages and aim to make the mean PAD across all samples close to zero [64].

Cole's Method and Beheshti's Method represent the two primary linear correction approaches [64]. Both methods involve:

• Calculating uncorrected PAD for all subjects in the dataset
• Fitting a linear regression model between uncorrected PAD and chronological age
• Applying the derived correction factor to remove the linear relationship between PAD and age

These linear methods can be adapted to nonlinear correction by replacing linear regression with quadratic or higher-order polynomial regression, though studies have found similar results between linear and quadratic approaches [64]. Some recent methods incorporate bias correction constraints directly during model training, such as with LASSO regression, which essentially adjusts the degree of linear bias correction after training and provides a balance between Mean Absolute Error (MAE) and PAD bias [64].

Age-Level Bias Correction

Recent research has revealed that even after sample-level correction, significant bias persists in the PAD of samples with the same age [64]. This age-level bias represents a more nuanced form of systematic error that weakens the reliability of PAD as a phenotype in developmental neuroimaging [64].

The age-level bias correction method operates by:

• Grouping participants by chronological age
• Calculating mean PAD bias within each age group
• Applying age-specific corrections to remove the residual bias
• Validating corrected PAD against non-imaging indices to ensure biological relevance

This approach recognizes that bias manifests differently across the developmental spectrum and requires age-specific adjustment strategies rather than a one-size-fits-all correction [64].

Motion-Specific Bias Mitigation Strategies

Given the profound impact of head motion on brain age estimation, several specialized methods have been developed to address motion-related bias:

Split Half Analysis of Motion Associated Networks (SHAMAN) is a novel method that assigns a motion impact score to specific trait-FC relationships [25]. This approach:

• Capitalizes on the stability of traits over time compared to the state-dependent nature of motion
• Measures differences in correlation structure between high- and low-motion halves of each participant's fMRI timeseries
• Distinguishes between motion causing overestimation or underestimation of trait-FC effects
• Provides a statistical framework for determining acceptable versus unacceptable levels of trait-specific motion

Motion censoring (scrubbing) involves identifying and removing individual fMRI volumes with excessive motion, typically using framewise displacement (FD) thresholds (e.g., FD > 0.2mm) [25]. This approach must be carefully calibrated, as overly aggressive censoring can systematically exclude participants with clinical conditions that correlate with higher motion, introducing selection bias [25].

Table 2: Motion Mitigation Strategies in Developmental Neuroimaging

Method Category	Specific Techniques	Mechanism of Action	Limitations
Real-Time Correction	Prospective motion correction, real-time monitoring [8]	Updates slice acquisition coordinates based on detected movement or pauses scan until sufficient low-motion data acquired	Not widely available due to hardware limitations and complexity [8]
Post-Hoc Denoising	ICA-AROMA, FIX, global signal regression, motion parameter regression [8] [25]	Identifies and removes motion-related components from fMRI data without discarding volumes	Cannot fully recover data lost due to larger movements; may remove neural signals along with noise [8]
Quality Control Metrics	Framewise displacement, DVARS, surface hole number [8] [3]	Quantifies motion artifact for exclusion decisions or covariate inclusion	Manual quality control is time-intensive; automated metrics may not capture all relevant quality dimensions [3]

Experimental Protocols for Bias Assessment

Validation Framework for Bias Correction Methods

Comprehensive validation of bias correction schemes requires a multi-faceted approach:

Dataset Diversity: Validation should include multiple independent datasets with varying age ranges and clinical characteristics. Recommended datasets include:

• UK Biobank (n=9,880, age 38-86): Large-scale biomedical database with multi-modal brain imaging [64]
• OASIS (n=3,388, age 42-97): Includes cognitively normal adults and individuals with cognitive decline [64]
• ABIDE (n=1,099, age 6-65): Collection from multiple laboratories for autism research [64]
• ABCD Study (n=9,652+ children age 9-10): Developmental cohort with extensive behavioral assessment [25]

Performance Metrics should extend beyond traditional measures like Mean Absolute Error (MAE) to include:

• Correlation between corrected PAD and chronological age (should approach zero for effective correction)
• Consistency of PAD within age groups after correction
• Association with non-imaging indices (e.g., fluid intelligence, clinical measures) to ensure biological validity [64]

Implementation Protocol for Age-Level Bias Correction

The following step-by-step protocol implements comprehensive age-level bias correction:

• Data Preparation and Sample-Level Correction

  • Obtain brain age predictions using preferred model (deep learning, feature-based, etc.)
  • Apply established sample-level correction (Cole's or Beheshti's method)
  • Calculate initial corrected PAD values

• Age Stratification and Bias Quantification

  • Group participants by chronological age (1-year intervals recommended)
  • Calculate mean PAD within each age group
  • Quantify residual age-level bias patterns

• Age-Level Correction Application

  • Develop age-specific correction factors using flexible modeling approaches
  • Apply corrections to remove age-level bias
  • Verify elimination of bias through correlation analysis

• Biological Validation

  • Test association between corrected PAD and relevant non-imaging indices
  • Ensure correction preserves biologically meaningful variance
  • Compare effect sizes before and after correction

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Resources for Bias-Adjustment Research

Resource Category	Specific Tools & Algorithms	Primary Function	Implementation Considerations
Bias Correction Algorithms	Cole's Method, Beheshti's Method, Age-Level Correction [64]	Remove systematic bias from brain age predictions	Choice of method depends on sample characteristics and age distribution; age-level correction addresses residual bias after sample-level correction
Motion Quantification Metrics	Framewise Displacement (FD), DVARS, Surface Hole Number (SHN) [8] [3]	Quantify head motion artifact for quality control or covariate inclusion	SHN best approximates manual quality ratings; FD thresholds must balance artifact removal with sample representation
Motion Correction Software	ICA-AROMA, FSL's FIX, FSL's MCFLIRT, AFNI's 3dvolreg [8]	Remove motion artifacts from fMRI data through component analysis or volume realignment	ICA-based approaches can remove neural signals along with noise; realignment tools cannot correct for intra-volume motion
Validation Datasets	UK Biobank, OASIS, ABIDE, ABCD Study [64] [25]	Provide diverse age ranges and clinical populations for method validation	Dataset selection should match target population characteristics; large samples needed to detect subtle bias patterns
Statistical Packages	Custom implementations in R/Python, brainageR, NeuroConductor	Implement specialized bias correction algorithms and statistical analyses	Reproducibility requires careful documentation of parameter settings and validation steps

Discussion and Future Directions

The development of effective bias-adjustment schemes represents a critical frontier in developmental neuroimaging research. While current methods like sample-level linear correction and motion censoring provide substantial improvements, the persistence of age-level bias and motion-related confounding underscores the need for more sophisticated approaches [64]. The integration of age-specific correction factors and trait-specific motion impact assessment (SHAMAN) represents promising directions for next-generation bias correction methodologies [64] [25].

Future research should prioritize:

• Development of integrated correction frameworks that simultaneously address statistical and motion-related biases
• Standardized validation protocols using multiple independent datasets with diverse age ranges and clinical characteristics
• Open-source implementation of bias correction algorithms to enhance reproducibility
• Methodological transparency in published research, with explicit reporting of bias correction approaches and validation results

As brain age estimation continues to evolve as a biomarker in developmental neuroimaging and drug development, rigorous attention to bias-adjustment schemes will be essential for producing valid, reliable, and clinically meaningful results.

Beyond Correction: Validating Findings and Comparing Method Efficacy

In-scanner head motion represents the most substantial source of artifact in functional MRI (fMRI) data, introducing systematic bias into resting-state functional connectivity (FC) measurements that persists despite denoising algorithms [25]. This challenge is particularly acute in developmental neuroimaging research involving children, older adults, and clinical populations with neurological or psychiatric disorders, as these groups often exhibit higher motion levels that correlate with the very traits under investigation [25]. The complexity of MRI physics creates non-linear artifacts that resist complete removal during standard post-processing, with motion particularly affecting resting-state FC due to the unknown timing of underlying neural processes [25]. Early studies have demonstrated that motion artifact systematically decreases FC between distant brain regions while increasing short-range connectivity, potentially leading to spurious conclusions about conditions like autism where increased motion may be misinterpreted as decreased long-distance connectivity [25].

Within this context, the Split Half Analysis of Motion Associated Networks (SHAMAN) framework was developed to address a critical methodological gap: the need for trait-specific motion impact quantification [25]. Traditional motion quantification approaches remain agnostic to research hypotheses and cannot establish thresholds for acceptable motion levels for specific trait-FC relationships [25]. SHAMAN provides researchers with a method to assign motion impact scores to specific trait-FC relationships, distinguishing between motion causing overestimation or underestimation of effects, thereby addressing a fundamental challenge in brain-wide association studies (BWAS) involving large cohorts such as the Adolescent Brain Cognitive Development (ABCD) Study [25].

SHAMAN Methodological Framework

Theoretical Foundation and Core Principle

The SHAMAN framework capitalizes on a fundamental physiological observation: traits such as cognitive ability or clinical measures remain stable over the timescale of an MRI scan, while head motion represents a state that varies from second to second [25]. This temporal dissociation enables the detection of motion-related artifacts specifically affecting trait-FC relationships. The method operates by measuring differences in correlation structure between split high-motion and low-motion halves of each participant's fMRI timeseries [25]. When trait-FC effects remain independent of motion, the difference in connectivity between high- and low-motion halves will be non-significant due to trait stability. A significant difference indicates that state-dependent motion variations impact the trait's connectivity measures.

SHAMAN introduces a crucial directional component to motion artifact characterization. A motion impact score aligned with the direction of the trait-FC effect indicates motion causing overestimation of the true effect ("motion overestimation score"). Conversely, a motion impact score opposite the direction of the trait-FC effect indicates motion causing underestimation ("motion underestimation score") [25]. This directional discrimination provides researchers with specific information about how motion may be biasing their results, enabling more informed interpretations of neuroimaging findings.

Computational Implementation and Statistical Testing

The SHAMAN algorithm employs permutation testing of the timeseries and non-parametric combining across pairwise connections to generate a motion impact score with an associated p-value distinguishing significant from non-significant motion impacts on trait-FC effects [25]. The method can accommodate one or more resting-state fMRI scans per participant and can be adapted to model covariates, enhancing its flexibility across different experimental designs [25]. This computational approach provides a rigorous statistical framework for quantifying motion artifacts while controlling for false positives.

Table: Key Analytical Features of the SHAMAN Framework

Feature	Description	Advantage
Split-Half Analysis	Compares high-motion and low-motion halves of fMRI timeseries	Capitalizes on trait stability versus motion variability
Directional Scoring	Distinguishes overestimation versus underestimation effects	Provides specific bias direction information
Permutation Testing	Non-parametric statistical approach	Robust statistical inference without distributional assumptions
Covariate Accommodation	Adaptable to include relevant covariates	Enhances specificity of motion impact assessment
Trait-Specific Focus	Quantifies motion impact for specific trait-FC relationships	Moves beyond global motion metrics to hypothesis-relevant effects

Experimental Validation and Quantitative Findings

Application in the ABCD Study

The SHAMAN framework was rigorously validated using data from the Adolescent Brain Cognitive Development (ABCD) Study, which collected up to 20 minutes of resting-state fMRI data on 11,874 children ages 9-10 years [25]. From this cohort, 7,270 participants with sufficient data quality were included in the SHAMAN analysis, which assessed 45 different behavioral and demographic traits [25]. The traits examined spanned cognitive, emotional, and clinical domains relevant to developmental neuroimaging research.

After standard denoising with the ABCD-BIDS pipeline (which includes global signal regression, respiratory filtering, spectral filtering, despiking, and motion parameter timeseries regression) and without additional motion censoring, SHAMAN analysis revealed widespread motion impacts [25]. Specifically, 42% (19/45) of traits exhibited significant (p < 0.05) motion overestimation scores, while 38% (17/45) demonstrated significant underestimation scores [25]. These findings demonstrate that residual motion artifact substantially impacts trait-FC relationships even after application of comprehensive denoising procedures.

Efficacy of Motion Censoring Approaches

The SHAMAN framework was further employed to evaluate the effectiveness of motion censoring strategies. Implementing framewise displacement (FD) censoring at a threshold of < 0.2 mm dramatically reduced significant overestimation to just 2% (1/45) of traits [25]. This finding supports the utility of rigorous motion censoring for mitigating false positive associations resulting from motion-induced overestimation of trait-FC effects.

However, the same censoring approach did not decrease the number of traits with significant motion underestimation scores [25]. This important nuance demonstrates that motion censoring strategies differentially affect overestimation versus underestimation artifacts, highlighting the value of SHAMAN's directional assessment capabilities for optimizing motion correction strategies specific to research contexts.

Table: Motion Impact Across 45 Traits in ABCD Study at Different Processing Stages

Processing Stage	Traits with Significant Overestimation	Traits with Significant Underestimation	Total Traits Impacted
After ABCD-BIDS denoising (no censoring)	19/45 (42%)	17/45 (38%)	36/45 (80%)
With FD < 0.2 mm censoring	1/45 (2%)	17/45 (38%)	18/45 (40%)
Reduction with censoring	95% reduction	No reduction	50% reduction

Research Reagent Solutions: Essential Methodological Components

Table: Essential Research Components for SHAMAN Implementation

Component	Function/Role	Implementation Example
High-Quality fMRI Data	Foundation for reliable FC and motion estimation	ABCD Study protocol: 20 minutes rs-fMRI, multiband acquisition [25]
Motion Quantification Metric	Quantifies frame-by-frame head motion	Framewise displacement (FD) calculated from rigid-body head realignment parameters [25]
Denoising Pipeline	Removes non-neural signal contaminants	ABCD-BIDS: global signal regression, respiratory filtering, motion parameter regression, despiking [25]
Trait Measures	Behavioral, clinical, or cognitive measures for association testing	45 diverse traits in ABCD: psychiatric symptoms, cognitive performance, demographic factors [25]
Computational Framework	Implements SHAMAN statistical methodology	Permutation testing with non-parametric combining across connections [25]

Experimental Protocol for SHAMAN Implementation

Data Acquisition and Preprocessing Requirements

Successful implementation of the SHAMAN framework requires careful attention to data acquisition and preprocessing stages. For resting-state fMRI data, acquisition of sufficient volumes is critical, with the ABCD study including participants with at least 8 minutes of data (n = 9,652) in initial motion characterization analyses [25]. The preprocessing pipeline should incorporate comprehensive denoising approaches; the ABCD-BIDS pipeline achieved a 69% relative reduction in motion-related signal variance compared to minimal processing (motion-correction only), though 23% of signal variance remained explained by motion after denoising [25].

Researchers should implement quality control procedures to identify excessive motion, with the ABCD study employing framewise displacement (FD) calculations for each volume. The strong negative correlation (Spearman ρ = -0.58) between motion-FC effects and average FC matrices highlights the systematic nature of motion artifacts, which manifest as decreased connectivity in high-motion participants [25]. This systematic pattern persists even after rigorous censoring (FD < 0.2 mm), though the correlation is somewhat reduced (Spearman ρ = -0.51) [25].

Analytical Execution Workflow

The core SHAMAN analysis follows a structured workflow: (1) segmentation of each participant's timeseries into high-motion and low-motion halves based on median framewise displacement; (2) calculation of trait-FC relationships within each half; (3) comparison of effect sizes between halves using permutation testing; (4) directional classification of significant differences as overestimation or underestimation scores; and (5) multiple comparisons correction across connections using non-parametric combining [25]. This workflow enables trait-specific motion impact assessment while controlling for false positives.

SHAMAN Analytical Workflow

Interpreting SHAMAN Output in Research Context

Motion Impact Score Interpretation

The motion impact score generated by SHAMAN provides specific quantitative guidance for interpreting trait-FC relationships. A significant positive motion impact score aligned with the trait-FC effect direction indicates that motion artifact is causing overestimation of the true effect size, potentially leading to false positive conclusions [25]. Conversely, a significant negative motion impact score opposite to the trait-FC effect direction indicates underestimation, which may obscure true relationships and increase false negative risk [25]. This directional information enables researchers to determine whether motion is likely inflating or masking the effects they observe.

The p-value associated with each motion impact score indicates whether the observed motion effect reaches statistical significance, with p < 0.05 suggesting substantial motion impact in the ABCD application [25]. Researchers should consider both the magnitude and direction of significant motion impact scores when determining whether trait-FC findings reflect genuine neural relationships or motion-induced artifacts. This interpretative framework represents a significant advancement over generic motion assessment approaches that cannot distinguish between these fundamentally different bias types.

Integration with Existing Methodological Approaches

SHAMAN complements rather than replaces existing motion mitigation approaches. The framework operates effectively on data processed with standard denoising pipelines like ABCD-BIDS, providing an additional layer of artifact quantification [25]. Researchers should implement SHAMAN alongside comprehensive denoising procedures including global signal regression, motion parameter regression, and spectral filtering to maximize data quality [25].

The finding that motion censoring at FD < 0.2 mm dramatically reduces overestimation but not underestimation artifacts provides crucial guidance for methodological optimization [25]. This differential impact suggests that researchers concerned about false positives may prioritize rigorous censoring, while those focused on avoiding false negatives may need alternative approaches for addressing underestimation artifacts. SHAMAN thus enables evidence-based selection of motion mitigation strategies tailored to specific research questions and analytical priorities.

SHAMAN Result Interpretation Logic

The SHAMAN framework represents a significant methodological advancement for addressing the persistent challenge of motion artifacts in developmental neuroimaging research. By providing trait-specific motion impact scores that distinguish between overestimation and underestimation effects, SHAMAN enables researchers to quantify and characterize motion bias in functional connectivity analyses with unprecedented specificity [25]. The application of this framework to the large-scale ABCD dataset demonstrates that residual motion artifacts substantially impact trait-FC relationships even after comprehensive denoising, with 80% of traits showing significant motion effects before censoring [25].

The differential impact of motion censoring on overestimation versus underestimation artifacts reveals the complex relationship between motion mitigation strategies and specific bias types, highlighting the need for tailored approaches based on research goals [25]. As neuroimaging studies continue to expand in scale and scope, with initiatives like ABCD, HCP, and UK Biobank encompassing thousands of participants, methodologies like SHAMAN will play an increasingly critical role in ensuring the validity and reproducibility of brain-behavior associations [25]. By integrating SHAMAN into analytical workflows, researchers can strengthen causal inference in developmental neuroimaging and advance our understanding of brain development and its relationship to behavioral and clinical traits.

Motion artifacts represent a significant source of systematic bias in developmental neuroimaging research, potentially compromising data quality, reproducibility, and the validity of scientific conclusions. This technical review synthesizes current evidence on benchmarking retrospective denoising pipelines, with a particular focus on addressing motion-related biases in neurodevelopmental populations. We systematically evaluate performance metrics across 14 different correction strategies, detailing their methodologies, effectiveness in mitigating motion confounds, and implications for developmental cognitive neuroscience. Evidence indicates that pipelines combining multiple correction strategies—including volume censoring, signal regression, and advanced noise modeling—can reduce motion-contaminated functional connectivity edges to less than 1%, significantly improving data quality and reliability. The findings underscore the critical importance of pipeline selection in minimizing spurious effects and enhancing the detection of true neurodevelopmental signals.

The characterization of human brain development, particularly during the fetal, infant, and toddler (FIT) period, relies heavily on non-invasive neuroimaging techniques such as functional magnetic resonance imaging (fMRI). However, the very populations central to developmental neuroscience—fetuses, infants, children, and individuals with neurodevelopmental disorders—are often the most prone to in-scanner motion. This creates a fundamental methodological challenge: motion artifacts can introduce systematic biases that threaten the internal validity of developmental trajectories and the detection of meaningful individual differences [65] [8].

In individuals with psychotic disorders, for example, in-scanner head motion is significantly higher than in healthy controls, often due to factors like psychomotor agitation, anxiety, or difficulty following instructions [8]. When data from these participants are excluded—a common practice to ensure data quality—it introduces a form of selection bias that systematically removes the most severely affected individuals from the study sample. This results in a dataset that is no longer representative of the full clinical population, limiting the generalizability of the findings and potentially underestimating the true effect sizes of neurobiological alterations [8]. This problem is conceptualized as "Missing Not At Random" (MNAR) data, where the probability of data being missing is directly related to the severity of the condition under investigation.

Furthermore, motion artifacts do not merely add random noise; they can introduce structured, systematic noise that corrupts functional connectivity measures and volumetric estimates. In fMRI, movement can shift the location of the signal from a given voxel, creating artifactual patterns in connectivity measures [8]. The severity and nature of these artifacts in 3D MRI have been found to depend heavily on the k-space distribution of motion states, with particularly pronounced effects when motion discontinuities occur near the center of k-space or align with slow phase-encoding directions [66]. Consequently, the choice of data-processing pipeline is not merely a technical consideration but a fundamental decision that can determine whether a study yields valid biological insights or spurious, motion-driven results.

Methodology for Pipeline Benchmarking

Benchmarking Framework and Evaluation Criteria

The systematic evaluation of denoising pipelines requires a multi-faceted approach that assesses performance across several critical dimensions. A comprehensive framework, as demonstrated in a large-scale study of 768 data-processing pipelines, should incorporate the following evaluation criteria [67]:

• Minimization of Motion Confounds: A primary criterion is the pipeline's ability to reduce the association between motion and functional connectivity measures. This is typically quantified using Quality Control-Functional Connectivity (QC-FC) correlations, where lower correlations indicate better motion denoising.
• Test-Retest Reliability: The pipeline should minimize spurious discrepancies in network topology across repeated scans of the same individual. This is crucial for ensuring that observed differences reflect true biological variation rather than measurement error.
• Sensitivity to Biological Signals: Beyond noise reduction, an optimal pipeline must preserve sensitivity to meaningful experimental effects, inter-individual differences, and clinical contrasts of interest.

This multi-criterion evaluation ensures that pipelines do not merely suppress all variability but selectively reduce noise while preserving signal. To ensure generalizability, benchmarking should be performed across multiple independent datasets spanning different time intervals (e.g., minutes, weeks, and months) and acquisition parameters [67].

Performance Metrics and Analytical Approaches

The benchmarking of denoising pipelines employs several key metrics and analytical approaches:

• Portrait Divergence (PDiv): An information-theoretic measure of dissimilarity between networks that simultaneously considers all scales of organization, from local structure to large-scale connectivity patterns, providing a comprehensive assessment of network topology [67].
• Fraction of Contaminated Edges: The proportion of functional connectivity edges that remain significantly correlated with motion metrics after denoising.
• Network Topology Metrics: Evaluation of network properties such as modular structure, global efficiency, and small-world characteristics to ensure that denoising does not distort fundamental organizational properties of brain networks.

These metrics collectively provide a robust framework for comparing the performance of different denoising approaches and identifying optimal strategies for developmental neuroimaging applications.

Performance Comparison of 14 Retrospective Motion Correction Pipelines

A systematic comparison of 14 retrospective motion correction pipelines revealed substantial variability in their effectiveness at mitigating motion artifacts in functional MRI data. The evaluation demonstrated that pipelines combining various strategies of signal regression and volume scrubbing were most effective, achieving a remarkable reduction in motion-related contamination [8].

Table 1: Performance Metrics of Select Motion Correction Pipelines

Pipeline Category	Key Components	Fraction of Contaminated Edges	Test-Retest Reliability	Network Modularity Recovery
Basic Motion Correction	Volume realignment (e.g., MCFLIRT, 3dvolreg)	>50%	Low	Poor
Scrubbing-Based	Framewise displacement-based censoring (FD > 0.5mm)	15-25%	Moderate	Moderate
ICA-Based Noise Removal	ICA-AROMA, FIX	10-20%	Moderate	Moderate-High
Combined Pipelines	Multiple strategies (scrubbing + regression + ICA)	<1%	High	High

The most effective pipelines integrated multiple complementary approaches, including volume censoring (scrubbing), motion parameter regression, and ICA-based noise component removal. These combination pipelines successfully reduced the fraction of connectivity edges contaminated by motion to less than 1%, a significant improvement compared to basic motion correction alone, where most edges remained biased by motion artifacts [8]. Furthermore, these advanced pipelines demonstrated superior recovery of network modular structure and improved performance on other metrics used to evaluate movement noise contamination, such as QC-FC correlation [8].

It is important to note that no model completely eliminated motion-related variance, indicating that scans with excessive movement remain a potential source of bias despite advanced processing. This underscores the importance of combining rigorous acquisition protocols with sophisticated post-processing approaches, particularly in developmental populations prone to movement [8].

Research Reagent Solutions: Essential Tools for Motion Correction

The implementation of effective motion correction pipelines requires a suite of specialized software tools and algorithmic approaches. The following table details key "research reagents" essential for benchmarking and implementing denoising pipelines in developmental neuroimaging research.

Table 2: Essential Research Reagents for Motion Correction Pipelines

Tool/Algorithm	Type	Primary Function	Application Notes
FSL MCFLIRT [8]	Volume Realignment Tool	Corrects for small movements between fMRI volumes through rigid-body registration.	Standard preprocessing step; ineffective for large movements or intra-volume motion.
Framewise Displacement (FD) [8]	Motion Metric	Quantifies head movement between successive fMRI volumes in millimeters.	Used for motion scrubbing (conventional threshold: FD > 0.5mm).
ICA-AROMA [8]	ICA-Based Denoising	Automatically identifies and removes motion-related independent components from fMRI data.	Preserves neural signals while removing motion artifacts without discarding entire volumes.
FSL FIX [8]	ICA-Based Denoising	Classifies and removes noise components from fMRI data using trained classifiers.	Particularly effective for high-motion data; requires classifier training.
Portrait Divergence (PDiv) [67]	Network Comparison Metric	Quantifies dissimilarity between brain networks across all organizational scales.	Evaluates test-retest reliability beyond specific graph metrics.
Efficiency Cost Optimization (ECO) [67]	Network Filtering	Data-driven method to define network edges by optimizing efficiency-cost balance.	Alternative to fixed-threshold edge filtering.
Orthogonal Minimum Spanning Trees (OMST) [67]	Network Filtering	Network filtering approach that optimizes the balance between network efficiency and wiring cost.	Data-driven alternative to density-based thresholding.

These research reagents form the foundational toolkit for constructing and evaluating the denoising pipelines benchmarked in contemporary developmental neuroimaging research. Their appropriate implementation and combination are critical for mitigating the systematic biases introduced by head motion.

Experimental Protocols and Workflows

Pipeline Construction and Evaluation Workflow

The systematic construction and evaluation of denoising pipelines follows a structured workflow that encompasses data preprocessing, network construction, and comprehensive validation. The following diagram illustrates the key decision points and processing stages in this workflow:

Diagram 1: Pipeline Construction and Evaluation Workflow

This workflow illustrates the combinatorial nature of pipeline construction, where choices at each step (such as the use of global signal regression, parcellation scheme, edge definition method, and filtering approach) collectively determine the final network topology and its susceptibility to motion artifacts [67]. The systematic evaluation of pipelines constructed through different combinations of these choices enables researchers to identify optimal strategies for specific research contexts and populations.

Motion Severity Assessment and Correction Protocol

Advanced motion correction approaches incorporate sophisticated protocols for assessing motion severity and applying appropriate correction strategies. The following diagram outlines the protocol for a novel unsupervised learning method that integrates motion severity assessment with artifact correction:

Diagram 2: Motion Severity Assessment and Correction Protocol

This protocol exemplifies the trend toward integrated correction approaches that simultaneously assess motion severity and apply tailored corrections. By replacing one of the CycleGAN generators with an MR physics-based motion artifact simulator and regularizing the cycle-consistency loss with a motion severity prior, this approach significantly reduces training complexity and addresses the inherent ill-posed problem of unsupervised motion correction [68]. The result is a more stable training process and superior performance in both motion correction and motion severity measurement compared to traditional approaches.

Implications for Developmental Neuroimaging Research

The benchmarking of denoising pipelines has profound implications for developmental neuroimaging research, particularly for studies investigating individual differences in brain development. The FIT period is characterized by rapid, profound, and heterogeneous changes in brain structure and function, creating both opportunities and challenges for identifying meaningful individual differences that may predict long-term outcomes [65]. Motion-related artifacts can obscure these subtle but developmentally significant variations, potentially leading to erroneous conclusions about neurodevelopmental trajectories.

The implementation of optimal denoising pipelines is particularly crucial for large-scale multi-site studies, such as the HEALthy Brain and Child Development (HBCD) Study, which aims to examine brain, cognitive, behavioral, social, and emotional development from the prenatal period through early childhood [32]. The success of such initiatives depends on the harmonization of acquisition protocols and processing pipelines across sites to minimize non-biological sources of variability and maximize the detection of true biological signals [32] [69]. The integration of motion-robust imaging techniques with sophisticated retrospective denoising pipelines represents a promising strategy for addressing the unique challenges of developmental neuroimaging.

Furthermore, the systematic evaluation of denoising pipelines helps to establish best practices that enhance the reproducibility and comparability of findings across studies and populations. As the field moves toward more precise characterization of neurodevelopmental trajectories and their relationship to cognitive and behavioral outcomes, the adoption of validated, optimized processing pipelines will be essential for generating reliable, interpretable, and clinically meaningful results.

The systematic benchmarking of 14 retrospective motion correction pipelines reveals that combinations of multiple denoising strategies—particularly those integrating scrubbing, regression, and ICA-based approaches—achieve superior performance in reducing motion artifacts while preserving biological signals. These pipelines can reduce motion-contaminated functional connectivity edges to less than 1%, a significant improvement over basic correction methods. For developmental neuroimaging research, where motion-related biases can systematically distort findings and exclude the most clinically informative participants, the implementation of these optimized pipelines is essential for generating valid, reliable, and generalizable results. Future advances will likely integrate prospective acquisition strategies with sophisticated retrospective corrections, further enhancing our ability to characterize neurodevelopmental trajectories and identify meaningful individual differences from infancy through childhood.

In-scanner head motion remains a paramount source of systematic bias in neuroimaging, particularly in developmental and clinical populations. Despite the proliferation of standardized denoising pipelines, significant motion-related variance persists, confounding functional connectivity estimates and morphometric analyses. This technical review synthesizes evidence on the nature and assessment of these residual confounds, detailing benchmarking methodologies that reveal the heterogeneous efficacy of common denoising strategies. We provide a framework for quantifying residual motion artifacts, emphasizing that even state-of-the-art processing leaves substantial, spatially structured variance that can inflate or obscure true brain-behavior relationships. The persistence of these artifacts necessitates rigorous, post-denoising quality assessment to ensure the validity of neuroimaging findings, especially in studies of developmental populations where motion is inherently correlated with traits of interest.

Head motion is the largest source of artifact in structural and functional MRI signals, producing systematic biases that persist despite extensive denoising efforts [25]. In functional connectivity studies, motion introduces spurious correlations that can completely obscure neuronally-driven connections, while in structural imaging, subtle, sub-millimeter movements bias automated measurements of brain anatomy [70] [71]. The problem is particularly acute in developmental neuroimaging, where motion is inherently greater in children and often correlated with clinical traits such as ADHD symptoms [33]. Critically, motion artifacts do not represent random noise but rather introduce systematic bias that can inflate effect sizes and generate false positives in large-scale datasets [3].

While numerous denoising strategies have been developed to mitigate these artifacts, evidence indicates that standard approaches leave substantial residual variance. After applying the default denoising pipeline for the Adolescent Brain Cognitive Development (ABCD) Study, which includes global signal regression and motion parameter regression, 23% of signal variance remained explainable by head motion—a substantial reduction from 73% with minimal processing, but nevertheless a significant residual confound [25]. This persistent artifact manifests differently across pipelines and population characteristics, demanding specialized assessment frameworks to detect and quantify its influence on scientific conclusions.

Quantifying Residual Motion Artifacts: Benchmarks and Metrics

Functional Connectivity Benchmarks

Residual motion artifacts in functional connectivity data exhibit distinct spatial and statistical signatures that can be quantified through specific benchmark measures. Research comparing 14 participant-level denoising pipelines revealed marked heterogeneity in performance across four key benchmarks [72]:

Table 1: Benchmark Measures for Residual Motion in Functional Connectivity

Benchmark	Description	Interpretation
Motion-FC Relationship	Residual correlation between head motion and connectivity estimates	Lower values indicate better artifact removal
Distance-Dependent Effects	Strength of motion-correlated artifacts as a function of distance between brain regions	Presence indicates spatially specific artifacts
Network Identifiability	Ability to identify known functional networks after denoising	Higher values indicate better preservation of neural signals
Degrees of Freedom Lost	Number of statistical degrees consumed by denoising regressors	Lower values indicate more efficient denoising

The most effective approaches for functional connectivity include aCompCor (a principal component-based method) and global signal regression (GSR), though each involves trade-offs. GSR minimizes the relationship between connectivity and motion but introduces distance-dependent artifacts, while censoring methods (removing high-motion volumes) mitigate both motion artifact and distance-dependence but at the cost of reduced statistical power and network identifiability [70] [72].

Structural MRI Benchmarks

In structural imaging, subtle in-scanner motion—even at levels that don't produce visible artifacts—systematically biases measurements of brain morphology. Analysis of 127 children, adolescents, and young adults revealed that motion during scanning correlates with reduced cortical gray matter volume and thickness, and increased mean curvature [71]. These effects are anatomically heterogeneous, persist across different automated processing pipelines, and exhibit convergent validity with effects of more pronounced motion. This demonstrates that standard structural processing fails to completely eliminate motion-related bias, particularly concerning for developmental studies where motion is inversely correlated with age.

A study of over 10,000 structural scans from the ABCD study found that incorporating lower-quality images consistently underestimated cortical thickness and overestimated cortical surface area [3]. As scan quality decreased, these measurement errors increased, demonstrating how residual motion introduces systematic bias rather than random noise. When analyzing only the highest-quality scans (n=4,600), group differences in cortical volume appeared in 3 brain regions comparing children with versus without aggressive behaviors. Including moderate-quality scans increased this to 21 significant regions, while pooling all scans inflated the number to 43 regions—demonstrating how residual motion artifacts can dramatically inflate effect sizes [3].

Experimental Protocols for Assessing Residual Variance

The SHAMAN Framework for Trait-Specific Motion Impact

The Split Half Analysis of Motion Associated Networks (SHAMAN) provides a novel method for computing trait-specific motion impact scores, operating on one or more resting-state fMRI scans per participant [25]. This approach capitalizes on the observation that traits (e.g., cognitive measures) are stable over the timescale of an MRI scan, while motion is a state that varies second-to-second.

Table 2: SHAMAN Protocol Steps

Step	Procedure	Output
Data Preparation	Split each participant's fMRI timeseries into high-motion and low-motion halves based on framewise displacement	Paired datasets for comparison
Connectivity Calculation	Compute functional connectivity matrices for both high-motion and low-motion halves	Paired connectivity maps
Trait-FC Effect Estimation	Calculate correlation between trait scores and connectivity across participants for each half	Trait-FC effect sizes for high/low motion conditions
Motion Impact Score	Compare trait-FC effects between high-motion and low-motion halves	Directional measure of motion impact (overestimation/underestimation)
Statistical Testing	Permutation testing with non-parametric combining across connections	p-values for significance of motion impact

Application of SHAMAN to 45 traits from 7,270 participants 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 [25]. Censoring at framewise displacement < 0.2 mm reduced significant overestimation to 2% (1/45) of traits but did not decrease the number of traits with significant motion underestimation scores, demonstrating the complex relationship between denoising strategies and trait-specific bias.

Distance-Dependence Analysis

A robust method for detecting residual motion artifacts involves examining the distance-dependence of motion-correlated effects on connectivity. This protocol quantifies how the relationship between motion and connectivity varies with the distance between brain regions:

• Parcellate the brain into regions of interest using a standardized atlas
• Calculate mean connectivity between each region pair as correlation coefficients
• Compute physical distance between regions as Euclidean distance between centroids
• Calculate motion-connectivity correlation for each region pair across participants
• Plot motion-connectivity correlations against physical distance
• Fit a regression model to quantify the distance-dependent relationship

Studies consistently show that motion increases short-distance connectivity and decreases long-distance connectivity, creating a characteristic distance-dependent pattern [73] [25]. Effective denoising should attenuate this pattern, though most strategies only partially succeed. For example, global signal regression reduces global motion artifacts but fails to eliminate distance-dependent effects [72].

Signaling Pathways: How Motion Artifacts Persist After Denoising

The flow of residual motion artifacts through neuroimaging data can be conceptualized as a pathway from physical movement to statistical confounding. The diagram below illustrates this process and the points of intervention for assessment methods:

Pathway of Residual Motion Artifacts Through Neuroimaging Data. Physical head motion creates image artifacts through voxel displacement and spin history effects, leading to structured signal changes that bias connectivity estimates. While denoising pipelines partially mitigate these effects, residual artifacts persist and can generate spurious brain-behavior associations. Specialized assessment methods intercept this pathway by quantifying residual variance before false conclusions are drawn.

The mechanisms underlying persistent artifacts include:

• Partial Volume Effects: When motion-corrected images are resampled to a reference space, the interpolation process creates residual signal changes due to mixing of signals from different tissues [74].
• Spin History Effects: Head motion alters the steady-state of magnetization by changing the time between RF excitations for different slices, creating signal modulation that cannot be corrected by spatial realignment alone [74].
• Distance-Dependent Artifacts: Motion-correlated signal changes tend to be more similar for nearby voxels than distant ones, creating spurious correlations that decrease with distance between brain regions [73].
• Global Signal Changes: Motion produces widespread signal fluctuations across gray matter, white matter, and cerebrospinal fluid that mimic globally synchronized neural activity [70].

Table 3: Research Reagent Solutions for Motion Assessment

Tool/Resource	Function	Application Context
Framewise Displacement (FD)	Quantifies head motion between consecutive volumes	Primary motion metric for quality control and censoring
SHAMAN Algorithm	Computes trait-specific motion impact scores	Assessing confounding in brain-behavior associations
Surface Hole Number (SHN)	Estimates imperfections in cortical reconstruction	Automated quality proxy for structural MRI
aCompCor	Noise component extraction via principal component analysis	Denoising strategy effective for task-based connectivity
ICA-AROMA	Automatic removal of motion components via independent component analysis	Denoising method preserving neural signals
SLOMOCO	Slice-oriented motion correction addressing intravolume motion	Advanced correction for spin history effects
QC-FC Plots	Visualizes correlation between motion and connectivity across participants	Diagnostic tool for residual motion artifacts
Distance-Dependence Analysis	Quantifies motion-connectivity relationships as function of distance	Benchmarking denoising efficacy

Implementation Guidelines

Framewise Displacement should be calculated as the sum of absolute derivatives of the six motion parameters (three translations, three rotations), with rotational components converted to millimeters by assuming a brain radius of 50 mm [25] [33]. For censoring, a threshold of FD < 0.2 mm significantly reduces motion overestimation effects, though it may not address underestimation artifacts [25].

Surface Hole Number serves as an automated proxy for manual quality ratings in structural MRI. Higher SHN values indicate poorer image quality and can be used to "stress-test" conclusions by examining how effect sizes change as lower-quality scans are progressively excluded from analysis [3].

Developmental Considerations: Special Challenges in Neurodevelopment

Motion artifacts present particular challenges in developmental neuroimaging due to systematic relationships between age, clinical status, and in-scanner motion. Longitudinal data reveal that head motion decreases with age during both diffusion and resting-state fMRI in typically developing children, while children with ADHD display consistently higher motion levels across ages 9-14 years [33]. Crucially, children in remission from ADHD continue to show elevated motion compared to controls, suggesting that head movement may represent a persistent neurodevelopmental trait rather than merely reflecting current symptom levels.

These developmental patterns create inherent confounds because motion is systematically related to both the neurodevelopmental processes under investigation and clinical status. In studies comparing clinical and control groups, unequal motion distribution can completely obscure or artificially create group differences in brain structure and function. For example, early studies falsely attributed motion-related decreases in long-distance connectivity to autism-specific hypoconnectivity [25].

Developmental studies require specialized approaches:

• Age-Matched Motion Comparisons: Ensure motion levels are balanced across age groups in cross-sectional designs
• Longitudinal Motion Modeling: Explicitly model age-related motion changes in growth curve analyses
• Remitted Group Designs: Include clinical remittance groups to disentangle state versus trait aspects of motion
• Motion Interaction Testing: Statistically test for interactions between motion and developmental variables

Residual motion artifacts represent a persistent threat to validity in developmental neuroimaging despite sophisticated denoising approaches. The assessment frameworks detailed herein—including trait-specific motion impact scores, distance-dependence analyses, and multimodal quality metrics—provide essential tools for quantifying and addressing these confounds. No single denoising strategy completely eliminates motion-related variance, and each involves trade-offs between artifact removal and signal preservation.

Moving forward, the field must adopt more rigorous reporting standards for residual motion assessment, including quantification of post-denoising motion-connectivity relationships and transparency about trait-specific motion impacts. As large-scale datasets become increasingly central to developmental neuroscience, ensuring that findings reflect neural phenomena rather than motion artifacts is paramount for valid inference about brain-behavior relationships across development.

Large-scale neuroimaging datasets such as the Adolescent Brain Cognitive Development (ABCD) Study, the Human Connectome Project (HCP), and the Healthy Brain Network (HBN) Initiative provide unprecedented resources for brain-wide association studies. However, ensuring the generalizability of findings across these datasets presents significant methodological challenges. This technical guide examines sources of systematic bias, particularly motion artifacts, that threaten the validity and reproducibility of cross-dataset findings. We synthesize evidence-based methodologies for quality control, data processing, and analytical frameworks that mitigate these biases, providing researchers and drug development professionals with practical tools to enhance the reliability of neuroimaging research.

Large-scale neuroimaging datasets have transformed neuroscience research by enabling the detection of subtle brain-behavior relationships. However, the promise of generalizable findings is undermined by systematic biases that vary across datasets. Among these, head motion represents the most pervasive source of systematic error, particularly in developmental populations and those with neuropsychiatric conditions [75] [25]. Motion artifacts introduce non-random noise that correlates with participant characteristics, potentially creating spurious associations or masking true effects [62] [3].

The Adolescent Brain Cognitive Development (ABCD) Study, the Human Connectome Project (HCP), and the Healthy Brain Network (HBN) Initiative each employ different data acquisition protocols, participant recruitment strategies, and quality control procedures. Understanding these methodological differences is essential for interpreting cross-dataset findings. This technical guide examines the specific challenges in integrating data from these initiatives and provides evidence-based solutions to enhance generalizability, with particular focus on mitigating motion-related bias in developmental neuroimaging research.

Dataset Profiles and Methodological Heterogeneity

Understanding the distinct characteristics of each major neuroimaging dataset is fundamental to assessing generalizability. The ABCD Study, HCP, and HBN differ significantly in their acquisition parameters, participant populations, and primary research foci.

Table 1: Key Characteristics of Major Neuroimaging Datasets

Dataset	Primary Population	Sample Size	Key Structural Sequence	Motion Mitigation	Primary Research Focus
ABCD	Children aged 9-10 at baseline	~11,874 [25]	MPRAGE with prospective motion correction (PMC) [75]	Prospective motion correction with volumetric navigators (vNav) [75]	Developmental trajectories, substance use, mental health
HCP Development (HCP-D)	Ages 5-21 [75]	~800+	MPRAGE with PMC [75]	Prospective motion correction [75]	Lifespan brain connectivity development
HCP Young Adult (HCP-YA)	Healthy adults	~1,200	Traditional MPRAGE [75]	Standard acquisition without PMC	Normative brain connectivity in adulthood
HBN	Developmental community sample across lifespan	348 in comparative study [75]	Both traditional MPRAGE and MPRAGE with PMC [75]	Comparative design evaluating PMC efficacy [75]	Mental health across development, naturalistic sampling

Table 2: Structural Sequence Parameters Across Datasets

Parameter	ABCD	HCP Young Adult	HCP Aging/Development	HBN (MPRAGE+PMC)
Sequence Type	MPRAGE+PMC [75]	MPRAGE [75]	MPRAGE+PMC [75]	MPRAGE+PMC [75]
Voxel Size (mm)	1.0×1.0×1.0 [75]	0.7×0.7×0.7 [75]	0.8×0.8×0.8 [75]	1.0×1.0×1.0 [75]
Matrix Size	256×256 [75]	320×320 [75]	320×300 [75]	256×256 [75]
TI/TR (ms)	1060/2500 [75]	1000/2400 [75]	1000/2500 [75]	1060/2500 [75]
Bandwidth (Hz/Pz)	240 [75]	210 [75]	740 [75]	240 [75]

Technical differences in structural sequences significantly impact morphometric measurements. The transition from traditional MPRAGE to MPRAGE with prospective motion correction (MPRAGE+PMC) in newer studies reflects recognition of motion's confounding effects [75]. Intraclass correlation coefficients (ICC) for morphometric measurements show highest intra-sequence reliability with two MPRAGE+PMC sequences compared to traditional MPRAGE, particularly valuable for hyperkinetic populations [75].

Motion Artifacts: A Fundamental Threat to Generalizability

Systematic Bias from Motion

Head motion introduces systematic bias rather than random noise into neuroimaging data, disproportionately affecting studies of developmental populations and those with psychiatric conditions [25] [3]. In functional MRI, motion reduces long-distance connectivity while increasing short-range connectivity, particularly affecting default mode network integrity [25]. For structural MRI, motion artifacts manifest as underestimated cortical thickness and overestimated cortical surface area [62] [3].

Evidence from the ABCD Study demonstrates that including lower-quality scans dramatically alters findings. In one analysis, significant group differences in cortical volume between children with and without aggressive behaviors appeared in only 3 brain regions when using highest-quality scans, but this increased to 43 regions when all scans were included—demonstrating how motion artifacts inflate effect sizes and potentially create false positives [3].

Quantifying Motion Impact

The Split Half Analysis of Motion Associated Networks (SHAMAN) framework provides a method to quantify motion's impact on specific trait-FC relationships [25]. SHAMAN distinguishes between motion overestimation (impact aligned with trait-FC effect direction) and motion underestimation (impact opposite trait-FC effect direction) [25]. Analyses of ABCD data revealed that after standard denoising, 42% of traits showed significant motion overestimation and 38% showed significant underestimation [25].

Quality Control Frameworks and Methodological Solutions

Quality Control Metrics and Their Limitations

Robust quality control is essential for generalizable findings, yet approaches vary across datasets. Manual quality assessment of over 10,000 ABCD scans revealed that 55% were suboptimal (rating ≥2 on a 4-point scale), with 2% deemed unusable [3]. These findings challenge assumptions that large sample sizes automatically overcome noisy data.

Table 3: Quality Control Metrics and Their Efficacy

Quality Control Approach	Detection Capability	Limitations	Implementation in Large Datasets
Manual Quality Rating	High accuracy for identifying motion artifacts [3]	Time-intensive; subjective components	Impractical for datasets >10,000 scans [3]
Surface Hole Number (SHN)	Good specificity for identifying lower-quality scans [62]	Does not eliminate error as effectively as manual rating [3]	Automated; suitable for large datasets [62]
Framewise Displacement (FD)	Standard metric for motion quantification in fMRI [25]	Does not address trait-specific motion effects [25]	Widely implemented; enables censoring approaches
Intraclass Correlation (ICC)	Quantifies reliability across repeated scans [75]	Requires additional scan time	Used in methodological studies [75]

Surface hole number (SHN), an automated index of topological complexity, shows promise as a scalable quality metric. While controlling for SHN doesn't eliminate error as effectively as manual ratings, it provides a practical approach for "stress-testing" conclusions by examining how effect sizes change as quality thresholds vary [62] [3].

Prospective Motion Correction in Structural Imaging

The adoption of prospective motion correction (PMC) sequences represents a technological solution to motion artifacts. PMC uses volumetric navigators (vNav) to periodically collect fast-acquisition images that estimate head motion, with sequence parameters adjusted in real-time to nullify this motion [75]. In MPRAGE sequences, navigators collected during inversion recovery time allow updating readout orientation for the current line of k-space [75].

Comparative studies in the HBN dataset demonstrate MPRAGE+PMC's advantages: higher intra-sequence reliability and robustness to head motion, though it is "not impervious to high head motion" [75]. Interestingly, traditional MPRAGE outperformed MPRAGE+PMC on 5 of 8 quality control metrics, highlighting tradeoffs in sequence selection [75].

Analytical Mitigation Strategies

Beyond acquisition improvements, analytical approaches can mitigate motion effects:

• Motion Censoring: Excluding high-motion fMRI frames reduces spurious findings, with censoring at framewise displacement (FD) < 0.2 mm reducing significant motion overestimation from 42% to 2% of traits in ABCD data [25]. However, this approach doesn't decrease motion underestimation and may systematically exclude participants with clinical conditions [25].

• Multivariate Methods: Machine learning approaches using multimodal neuroimaging data (combining structural MRI, functional MRI, and diffusion imaging) show superior prediction of cognitive abilities (r = 0.54) compared to univariate approaches [76]. These methods better capture distributed brain-behavior relationships less susceptible to localized motion artifacts.

• Cross-Validation Strategies: Proper cross-validation that accounts for family structure prevents data leakage that inflates prediction performance. Family leakage slightly increases prediction performance for some measures (e.g., attention problems), while feature leakage and subject duplication cause dramatic performance inflation [77].

The Researcher's Toolkit: Essential Methods for Enhanced Generalizability

Table 4: Research Reagent Solutions for Motion Mitigation

Tool Category	Specific Solution	Function	Implementation Considerations
Acquisition Sequences	MPRAGE with prospective motion correction (vNav) [75]	Real-time correction of head motion during structural acquisition	Recommended for hyperkinetic populations; tradeoffs in some quality metrics [75]
Quality Control Metrics	Surface Hole Number (SHN) [62] [3]	Automated identification of topological errors in cortical reconstruction	Good specificity; useful for stress-testing findings across quality thresholds [3]
Motion Impact Assessment	SHAMAN framework [25]	Quantifies trait-specific motion effects on functional connectivity	Distinguishes overestimation vs. underestimation; requires sufficient temporal data [25]
Analytical Frameworks	Multivariate predictive modeling with multimodal data [76]	Enhances prediction robustness by integrating complementary data sources	Superior to univariate methods; opportunistic stacking handles missing data [76]
Data Processing	Appropriate cross-validation schemes [77]	Prevents data leakage and inflation of prediction performance	Must account for family structure; feature selection within training folds only [77]

Generalizability across neuroimaging datasets requires meticulous attention to methodological heterogeneity and systematic bias sources. Motion artifacts represent a fundamental challenge that differentially affects datasets based on their participant populations, acquisition protocols, and processing pipelines. The adoption of prospective motion correction, implementation of robust quality control metrics like surface hole number, application of motion impact assessments such as SHAMAN, and utilization of appropriate analytical frameworks that prevent data leakage collectively enhance the validity and reproducibility of findings across ABCD, HCP, and HBN datasets.

For drug development professionals utilizing neuroimaging biomarkers, these methodological considerations are particularly crucial. Systematic bias from motion may obscure true treatment effects or create spurious associations, potentially derailing development pipelines. Implementing the rigorous quality assessment and motion mitigation strategies outlined in this guide provides a pathway toward more reliable, generalizable neuroimaging biomarkers that can accelerate therapeutic development for neurological and psychiatric disorders.

The search for biologically based diagnostics in youth psychiatry is a critical endeavor. However, a fundamental methodological challenge systematically biases this research: in-scanner motion. Motion artifacts in neuroimaging are not merely random noise but a source of systematic bias that disproportionately affects data from youth and individuals with psychiatric symptoms, potentially confounding the identification of true neurobiological markers [8] [3]. In psychiatric research, participants with more severe symptoms—such as psychomotor agitation, anxiety, or disorganization—often exhibit greater difficulty remaining still during scans [8]. When these participants' data are excluded for quality control, the resulting sample no longer represents the full spectrum of the clinical population, a phenomenon known as Missing Not at Random (MNAR) [8] [58]. This bias can lead to underestimating the true effect sizes of brain abnormalities and threatens the validity and generalizability of findings intended to distinguish disorders [8]. This technical guide examines motion-related artifacts as both a confound and a potential behavioral phenotype, providing methodologies for analyzing motion patterns across youth psychiatric disorders.

Motion as a Confound and Phenotype: Empirical Evidence

Evidence of Systematic Bias from Major Datasets

Large-scale neuroimaging studies provide compelling evidence that data quality issues introduce systematic bias. An analysis of over 10,000 structural MRI scans from the Adolescent Brain Cognitive Development (ABCD) Study found that more than half were of suboptimal quality [3]. Crucially, including these lower-quality scans did not simply add random noise; it introduced directional bias, consistently underestimating cortical thickness and overestimating cortical surface area [3].

The impact on research findings is substantial. One analysis demonstrated that when comparing cortical volume in children with and without aggressive behaviors, the number of significant brain regions inflated from 3 to 43 as lower-quality scans were added to the analysis, dramatically altering the apparent results [3]. Similarly, research on resting-state fMRI confirms that exclusion of participants due to motion is related to a broad spectrum of behavioral, demographic, and health-related variables, inevitably biasing brain-behavior relationship analyses [58].

Motion Patterns Across Psychiatric Populations

Table 1: Motion-Related Findings Across Youth Psychiatric Populations

Disorder or Population	Nature of Motion Artifact	Implied Behavioral Phenotype	Key Supporting Evidence
Psychosis Spectrum Disorders (e.g., schizophrenia)	Significantly increased head motion during MRI scans [8].	Proxy for psychomotor agitation, disorganization, anxiety, or paranoia [8].	Patients exhibit significantly more head movement compared to healthy controls; exclusion biases samples toward less severe cases [8].
Major Depressive Disorder (MDD) with Psychotic Features	Specific eye movement abnormalities during dual-task tracking paradigms [78] [79].	Impaired attention allocation and multitasking abilities [78].	MDD with psychosis group showed greater number and total excursion of eye movements during dual-task vs. controls [78] [79].
MDD (with and without psychosis)	Altered visual exploration during free-viewing tasks (increased saccades and fixations) [78] [79].	Compensatory cognitive resource allocation; general cognitive dysfunction in depression [78].	Both patient groups exhibited higher saccade and fixation counts than healthy controls [78] [79].
Transdiagnostic Youth Samples	Motion acts as a systematic confound correlated with sociodemographic and clinical characteristics [58].	Lower inhibitory control, higher psychopathology, and association with environmental factors like socioeconomic status [58].	ABCD study data shows motion-related exclusion related to behavioral, demographic, and health variables [58].

Eye Movement as a Complementary Biomarker

Oculomotor metrics provide a valuable window into cognitive dysfunction with less vulnerability to the artifact problems of MRI. A 2024 study of adolescents and young adults with Major Depressive Disorder (MDD) revealed distinct eye movement patterns that differentiated diagnostic subgroups [78] [79]. While both MDD groups shared alterations in free-viewing patterns, only the MDD with psychotic features (MDDwP) group showed specific abnormalities in the dual-task tracking paradigm, suggesting this measure may tap into unique aspects of cognitive control impaired in psychotic depression [78] [79]. This underscores the potential of eye-tracking to provide objective biomarkers for psychiatric subtyping.

Methodological Approaches and Experimental Protocols

Core Experimental Paradigms for Motion Assessment

A. In-Scanner Head Motion Protocol

• Purpose: To quantify head movement during structural and functional MRI acquisition.
• Setup: Participants undergo scanning in a 3T MRI scanner. Physical restraints (foam padding) are used to minimize head motion [8].
• Acquisition: For fMRI, a standard resting-state sequence is run (e.g., 5-10 minutes). Structural scans (T1-weighted) are also collected [8].
• Quantification:
  • Framewise Displacement (FD): Calculates the head movement (in mm) between successive fMRI volumes. Volumes with FD > 0.5 mm are typically flagged as high-motion [8].
  • DVARS: Measures the change in BOLD signal variance across the entire brain between volumes [8].
  • Exclusion Criteria: Participants with >20% of volumes scrubbed or maximum displacement >3 mm are often excluded, though this risks bias [8].

B. Eye Movement Assessment Protocol

• Purpose: To evaluate oculomotor function as a potential neurocognitive biomarker.
• Setup: Participants sit comfortably in front of an eye-tracking monitor. Head is stabilized with a chin rest.
• Paradigms (as used in Huang et al., 2024 [78] [79]):
  • Smooth Pursuit: Participant tracks a slowly moving visual target on the screen. Metrics: gain, velocity.
  • Dual-Task Tracking: Participant performs a smooth pursuit task while simultaneously engaging in a secondary cognitive task. Metrics: number of excursions, total excursion amplitude [78].
  • Free Viewing: Participant freely views complex images or scenes. Metrics: saccade count, fixation count, fixation duration [78].
• Analysis: Group comparisons (e.g., ANOVA) of eye movement metrics between diagnostic groups and healthy controls.

Data Analysis and Mitigation Strategies

Diagram Title: Decision Workflow for Motion-Related Data

Table 2: Mitigation Strategies for Motion-Related Bias

Strategy Category	Specific Methods	Brief Explanation & Function	Key Considerations
Acquisition	Prospective Motion Correction (POC) [8].	Real-time updating of slice acquisition coordinates based on detected head motion.	Not yet widely available due to hardware and complexity limitations [8].
	Mock Scanner Practice [8].	Habituates participants, especially children, to the scanner environment to reduce anxiety and motion.	Can improve data quality but does not eliminate motion in clinical populations [8].
Preprocessing	Motion Scrubbing (Censoring) [8].	Removal of individual fMRI volumes with excessive motion (e.g., FD > 0.5 mm).	Can lead to loss of data; if >20% of volumes are removed, the entire subject is often excluded, causing bias [8].
	Volume Realignment [8].	Corrects for small movements between volumes by registering each to a reference.	Does not correct for large movements or intra-volume motion [8].
Advanced Denoising	ICA-AROMA / FIX [8].	Uses Independent Component Analysis to identify and remove motion-related noise components from the data.	Effective but may remove neural signal if motion is severe; requires careful validation [8].
	aCompCor [58].	Regresses out signal from noise regions of interest (e.g., white matter, CSF) rather than global signal.	Helps remove physiological noise without relying on global signal regression.
Statistical Correction	Motion Covariates [8].	Including mean FD or the number of scrubbed volumes as nuisance regressors in group-level models.	Adjusts for residual motion effects but assumes a linear relationship [8].
	Multiple Imputation [58].	Statistical technique to account for data missing not at random (MNAR) by creating several plausible datasets.	Helps address bias introduced by non-random exclusion but is not yet common practice [58].

The Scientist's Toolkit: Key Research Reagents

Table 3: Essential Materials and Tools for Motion Research

Item/Reagent	Function/Application	Example/Specification
3T MRI Scanner with Head Coil	Acquisition of high-resolution structural and functional neuroimaging data.	Essential for collecting BOLD signal for fMRI and T1-weighted anatomical images [8].
Framewise Displacement (FD)	A quantitative metric to measure head movement between successive fMRI volumes.	Critical for quality control; volumes with FD > 0.5 mm are typically censored [8].
Video-Based Eye Tracker	Records eye movements (saccades, fixations, pursuit) with high temporal resolution.	Used to administer smooth pursuit, free-viewing, and dual-task paradigms [78] [79].
ICA-AROMA Algorithm	A preprocessing tool for automatic removal of motion artifacts from fMRI data.	Identifies motion-related independent components in BOLD data for regression [8].
FSL MCFLIRT Tool	Performs volume realignment (motion correction) on fMRI data.	Standard tool for correcting for small head movements during an fMRI time series [8].
Structured Clinical Interview	Ensures accurate diagnostic classification of participants, crucial for diagnostic specificity.	e.g., MINI International Neuropsychiatric Schedule, used to establish clinical groups [79].
Surface Hole Number (SHN)	An automated quality-control metric that estimates imperfections in cortical reconstruction.	Proposed as a proxy for manual quality ratings to "stress-test" conclusions in large datasets [3].

Analyzing motion patterns in youth psychiatric research is a complex challenge that sits at the intersection of methodology and clinical science. The evidence is clear that motion introduces systematic, non-random bias that can distort brain-behavior relationships and impede the search for valid diagnostic markers [8] [3] [58]. To advance the field, researchers must move beyond simply excluding "noisy" data. A multimodal approach is essential, combining rigorous mitigation strategies at every stage—from acquisition and preprocessing to statistical analysis—with an appreciation that motion itself may carry valuable clinical information as a behavioral phenotype [8]. Acknowledging and correcting for this bias is not merely a technical exercise but a fundamental prerequisite for achieving the diagnostic specificity needed to understand and treat psychiatric disorders in youth.

Conclusion

Systematic bias from motion is not merely a technical nuisance but a fundamental challenge that can distort our understanding of neurodevelopment and psychiatric disorders. The evidence confirms that motion introduces structured, non-random noise that persists even after standard denoising, inflating effect sizes and potentially leading to spurious brain-behavior associations. Successfully mitigating this bias requires a multi-faceted approach: rigorous acquisition protocols tailored to developmental populations, thoughtful implementation of correction algorithms, and—crucially—systematic validation of findings using frameworks like SHAMAN to quantify motion's impact. Future directions must include the development and widespread adoption of real-time correction technologies, standardized reporting of motion metrics, and analytical methods that account for the 'missing not at random' problem inherent in excluding high-motion participants. For drug development and clinical translation, ensuring that neural biomarkers are free from motion confounds is paramount for building reliable, reproducible foundations for diagnostic and therapeutic innovation.

References

• 1. Motion artifact in studies of functional connectivity [https://pmc.ncbi.nlm.nih.gov/articles/PMC5930165/]
• 2. Associations Between Head Motion, Age, and Psychiatric ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12495084/]
• 3. Poor image quality introduces systematic bias into MRI data [https://www.thetransmitter.org/spectrum/poor-image-quality-introduces-systematic-bias-into-large-neuroimaging-datasets/]
• 4. Changes in MRI head motion across development [https://link.springer.com/article/10.1007/s11682-024-00910-w]
• 5. Heterogeneous impact of motion on fundamental patterns ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811913006824]
• 6. Motion Artefacts in MRI: a Complex Problem with Many Partial ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4517972/]
• 7. Basic Theory of Motion Artifacts in MRI [https://cds.ismrm.org/protected/21MProceedings/PDFfiles/E5763.html]
• 8. The Bias Associated with Movement-related Data Exclusion in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12558892/]
• 9. Head motion during fMRI tasks is reduced in children and ... [https://www.sciencedirect.com/science/article/pii/S1878929320300517]
• 10. Best Practices in Structural Neuroimaging of ... [https://link.springer.com/article/10.1007/s11065-021-09496-2]
• 11. Variations in structural MRI quality significantly impact ... [https://braininformatics.springeropen.com/articles/10.1186/s40708-021-00128-2]
• 12. Uncovering and mitigating bias in large, automated MRI ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10002762/]
• 13. Subtle in‐scanner motion biases automated measurement ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5110234/]
• 14. Retrospective motion artifact correction of structural MRI ... [https://www.sciencedirect.com/science/article/pii/S1053811921000331]
• 15. A perspective on enhancing representative samples in ... - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC9480848/]
• 16. Equity, diversity, and inclusion in developmental ... [https://www.frontiersin.org/journals/integrative-neuroscience/articles/10.3389/fnint.2022.1007249/full]
• 17. Demographic reporting across a decade of neuroimaging [https://pmc.ncbi.nlm.nih.gov/articles/PMC9712398/]
• 18. Bias and Justice in Neurological AI [https://link.springer.com/chapter/10.1007/978-3-032-11596-6_7]
• 19. Socio-Economic Disparities in Access to Diagnostic ... [https://www.mdpi.com/1660-4601/18/20/10633]
• 20. Quantitative Assessment of Systematic Bias: A Guide for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10714380/]
• 21. Quantative Bias Analysis [https://www.iseepi.org/quantative_bias_analysis.php]
• 22. Quantitative Assessment of Systematic Bias: A Guide for ... [https://pubmed.ncbi.nlm.nih.gov/37786916/]
• 23. Missing data are associated with clinical status and ... [https://www.sciencedirect.com/science/article/pii/S0022395624002425]
• 24. Motion sensing superpixels (MOSES) is a systematic ... [https://elifesciences.org/articles/40162]
• 25. Motion impact score for detecting spurious brain-behavior ... [https://www.nature.com/articles/s41467-025-63661-2]
• 26. Representing Brain-Behavior Associations by Retaining High ... [https://pubmed.ncbi.nlm.nih.gov/39921132/]
• 27. BrainEffeX: A web app for exploring fMRI effect sizes [https://apertureneuro.org/article/146251]
• 28. Brief mock-scan training reduces head motion during real ... [https://www.sciencedirect.com/science/article/pii/S187892932300049X]
• 29. Brief mock-scan training reduces head motion during real ... [https://pubmed.ncbi.nlm.nih.gov/37062244/]
• 30. Low-motion fMRI data can be obtained in pediatric ... [https://www.nature.com/articles/s41598-020-78885-z]
• 31. Real-time motion monitoring improves functional MRI data ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9157440/]
• 32. Quantifying brain development in the HEALthy Brain and ... [https://www.sciencedirect.com/science/article/pii/S1878929324001130]
• 33. Changes in MRI head motion across development [https://pmc.ncbi.nlm.nih.gov/articles/PMC11582210/]
• 34. Combining Prospective Acquisition CorrEction (PACE) with... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6710196/]
• 35. Beyond motion : correction – Alessandro... Framewise displacement [https://crimia.wordpress.com/2016/02/07/beyond-motion-correction-framewise-displacement/]
• 36. Prospective motion correction with volumetric navigators ... [https://pubmed.ncbi.nlm.nih.gov/26654788/]
• 37. Comparing data-driven physiological denoising ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2024.1223230/full]
• 38. Advancing motion denoising of multiband resting-state ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9057309/]
• 39. fMRIPrep: A Robust Preprocessing Pipeline for fMRI Data ... [https://fmriprep.org/]
• 40. FuNP (Fusion of Neuroimaging Preprocessing) Pipelines [https://pmc.ncbi.nlm.nih.gov/articles/PMC6378808/]
• 41. A Systematic Comparison of Motion Artifact Correction ... - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC3468891/]
• 42. Prospective motion correction of high-resolution magnetic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3146240/]
• 43. Integrated fMRI Preprocessing Framework Using Extended ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2018.00268/full]
• 44. Motion artifacts in functional near-infrared spectroscopy [https://www.sciencedirect.com/science/article/abs/pii/S1053811913004291]
• 45. TARGETED PRINCIPLE COMPONENT ANALYSIS: A NEW ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4211632/]
• 46. Noise reduction in brain magnetic resonance imaging using ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11757293/]
• 47. Prof. Rihui Li's team documented the impact of motion ... [https://ccbs.ici.um.edu.mo/prof-rihui-lis-team-documented-the-impact-of-motion-artifact-correction-on-brain-functional-connectivity-analysis-in-fnirs/]
• 48. Supervised Deep Learning for Head Motion Correction in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10725740/]
• 49. Deep learning based motion correction in ultrasound ... [https://www.nature.com/articles/s41598-025-02728-y]
• 50. Iterative Learning for Joint Image Denoising and Motion ... [https://www.sciencedirect.com/science/article/abs/pii/S1361841525004189]
• 51. Iterative Learning for Joint Image Denoising and Motion ... [https://arxiv.org/abs/2403.08162]
• 52. Iterative Learning for Joint Image Denoising and Motion ... [https://arxiv.org/html/2403.08162v1]
• 53. [Literature Review] Iterative Learning for Joint Image Denoising ... [https://www.themoonlight.io/en/review/iterative-learning-for-joint-image-denoising-and-motion-artifact-correction-of-3d-brain-mri]
• 54. Laser Tracking and Positioning System Analysis 2025 ... [https://www.datainsightsmarket.com/reports/laser-tracking-and-positioning-system-1385091]
• 55. USC Researchers Develop First Wireless, Wearable, Real- ... [https://viterbischool.usc.edu/news/2025/10/usc-researchers-develop-first-wireless-wearable-real-time-camera-less-motion-tracker/]
• 56. Validation sampling can reduce bias in healthcare ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3911878/]
• 57. A Practical Guide to Identifying Robust Clusters in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12406288/]
• 58. Balancing Data Quality and Bias: Investigating Functional ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11717557/]
• 59. Impact of model selection, sample size and evaluation criteria [https://www.sciencedirect.com/science/article/pii/S1053811923000137]
• 60. Benchmarking methods for mapping functional connectivity ... [https://www.nature.com/articles/s41592-025-02704-4]
• 61. Quantitatively validating the efficacy of artifact suppression ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6688933/]
• 62. Addressing artifactual bias in large, automated MRI analyses ... [https://pubmed.ncbi.nlm.nih.gov/40595447/]
• 63. Motion impact score for detecting spurious brain-behavior ... [https://pubmed.ncbi.nlm.nih.gov/41022827/]
• 64. Age-level bias correction in brain age prediction - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC9860514/]
• 65. Measuring and interpreting individual differences in fetal ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11930173/]
• 66. An Exploration of Motion-Sampling Interactions in 3D MRI ... [https://pubmed.ncbi.nlm.nih.gov/41204061/]
• 67. Systematic evaluation of fMRI data-processing pipelines for ... [https://www.nature.com/articles/s41467-024-48781-5]
• 68. Unsupervised learning for motion correction and ... [https://www.sciencedirect.com/science/article/abs/pii/S0952197624021377]
• 69. Harmonization in Magnetic Resonance Imaging: A Survey ... [https://arxiv.org/html/2507.16962v1]
• 70. Evaluation of denoising strategies for task‐based ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7978116/]
• 71. Subtle in-scanner motion biases automated measurement of ... [https://pubmed.ncbi.nlm.nih.gov/27004471/]
• 72. Benchmarking of participant-level confound regression ... [https://www.sciencedirect.com/science/article/pii/S1053811917302288]
• 73. Evaluation of Denoising Strategies to Address Motion- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5105353/]
• 74. Estimation and Removal of Residual Motion Artifact in ... [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]
• 75. Is it time to switch your T1W sequence? Assessing the impact ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7898192/]
• 76. Relationship between cognitive abilities and mental health as ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12618009/]
• 77. Data leakage inflates prediction performance in ... [https://www.nature.com/articles/s41467-024-46150-w]
• 78. Abnormal eye movement patterns in adolescent and early ... [https://pubmed.ncbi.nlm.nih.gov/40946409/]
• 79. Abnormal eye movement patterns in adolescent and early ... [https://www.sciencedirect.com/science/article/abs/pii/S0925492725001192]