Mapping the Landscape: A Comprehensive Guide to the Spatial Distribution of Motion Artifacts in Brain Imaging

Abstract

This article provides a systematic review of the spatial distribution of motion artifacts in brain imaging, a critical challenge for researchers, scientists, and drug development professionals. We first explore the foundational physics and heterogeneous patterns of these artifacts across different neuroimaging modalities, including MRI and fNIRS. The review then transitions to methodological advances, covering both hardware-based and deep-learning-driven correction techniques. We further offer a troubleshooting guide for artifact identification and mitigation in experimental design and data processing. Finally, we present a comparative analysis of validation frameworks and benchmark the performance of leading correction algorithms against standardized datasets. This synthesis aims to equip professionals with the knowledge to improve data fidelity in neuroimaging studies and pharmaceutical research.

The Physics and Patterns: Understanding the Origin and Spatially Heterogeneous Nature of Motion Artifacts

Magnetic resonance imaging (MRI) is a cornerstone of modern biomedical research and clinical diagnostics, providing unparalleled soft-tissue contrast without ionizing radiation. However, its susceptibility to subject motion has been a persistent challenge since its inception. In the specific context of brain research, motion artifacts introduce systematic biases that can compromise the integrity of functional connectivity analyses, structural measurements, and the development of quantitative biomarkers [1]. The prolonged data acquisition times required for high-resolution MRI, often spanning several minutes, far exceed the timescale of most physiological motions, including involuntary head movements, cardiac pulsation, and respiration [2]. This fundamental mismatch between acquisition speed and motion dynamics makes the MRI signal particularly vulnerable to corruption.

Motion during MRI acquisition primarily manifests in the final image as blurring (a loss of sharpness) and ghosting (replicated features of the object appearing as displaced copies) [3] [2]. These artifacts are not merely cosmetic; they can obscure anatomical details, mimic pathological conditions, and, most critically for research, lead to spurious brain-behavior associations [1]. Understanding the genesis of these artifacts requires moving from image space to the raw data domain of MRI, known as k-space. The corruption of k-space data by subject movement is the root cause of the ghosting and blurring observed in the final reconstructed image. This technical guide details the fundamental mechanisms by which motion disrupts k-space, explores the spatial distribution of these artifacts in brain imaging, and summarizes the quantitative frameworks and emerging solutions essential for robust neuroscientific research.

k-Space Fundamentals and the Impact of Motion

k-Space as the Fourier Domain of the Image

In MRI, spatial encoding is not performed directly. Instead, the signal is acquired in the spatial frequency domain, known as k-space. Each point in k-space contains information about the spatial frequency and phase of the entire imaged object, rather than representing a specific location [2]. The center of k-space (low spatial frequencies) determines the overall image contrast and signal intensity, while the periphery (high spatial frequencies) encodes the fine details and edges of the image [2].

The final image is reconstructed through an inverse Fourier transform of the acquired k-space data. This process assumes that the object being imaged has remained perfectly stationary throughout the entire data acquisition. Any violation of this assumption—that is, any motion during the scan—introduces inconsistencies in the k-space data, leading to artifacts in the reconstructed image [2].

How Motion Disrupts k-Space Data Consistency

The MRI signal is acquired sequentially, point-by-point or line-by-line, over time. When a subject moves, the spatial position of the spin protons changes. This movement alters the phase and frequency of the signal they emit, which in turn corrupts the corresponding k-space data points being acquired at that moment.

• Phase Shifts: Motion induces phase errors in the MR signal. In k-space, this is equivalent to a complex multiplication of the true k-space data, which results in a convolution in image space, manifesting as ghosting artifacts [2] [4].
• Inconsistent Encoding: Different k-space lines are acquired at different times. If the object moves between the acquisition of these lines, the data becomes inconsistent. Simple Fourier reconstruction interprets this as the object having multiple, discrete positions, leading to replicated "ghosts" in the image [2].

The specific appearance of the artifact is heavily influenced by the k-space trajectory (the order in which k-space is sampled) and the nature of the motion (e.g., periodic vs. sudden, slow drift vs. rapid oscillation) [2].

Characterizing Motion Artifacts: Ghosting and Blurring

Ghosting Artefacts

Ghosting appears as replicated versions of the main object or parts of it, typically displaced along the phase-encoding direction [3]. This directional preference arises because the phase-encoding gradient is applied sequentially and much more slowly than the frequency-encoding gradient. A single corruption event therefore affects a larger portion of k-space.

• The N/2 Ghost: A particularly famous and structured ghosting artifact is the Nyquist or N/2 ghost, which appears as a replica displaced by exactly half the field of view in the phase-encoding direction. It is often caused by consistent phase errors between odd and even echoes in echo-planar imaging (EPI) sequences, which can be exacerbated by motion and eddy currents from strong diffusion-sensitizing gradients [3] [4].
• Coherent vs. Incoherent Ghosting: Periodic motion (e.g., from cardiac pulsation) that is synchronized with the k-space acquisition rhythm results in coherent ghosting, where a discrete, well-defined number of replicas are visible. In contrast, random or non-periodic motion leads to incoherent ghosting, which appears as a smeared "noise-like" pattern or multiple overlapping stripes along the phase-encoding direction [2].

Blurring Artefacts

Blurring is a diffuse reduction of image sharpness, particularly affecting edges and fine details. It occurs when motion is continuous and slow (e.g., gradual relaxation of neck muscles) during the acquisition. In this scenario, the k-space data for a single object is effectively smeared across multiple spatial locations. Unlike the discrete replication in ghosting, this smearing results in a loss of high-frequency information, which manifests as a blur in the final image [2]. In sequences with interleaved k-space acquisition (e.g., T2-weighted Turbo Spin Echo), even slow drifts can produce significant ghosting rather than simple blurring, because the assumption of consistency between interleaved segments is violated [2].

Table 1: Characteristics of Key Motion-Induced Artefacts in Brain MRI

Artefact Type	Primary Appearance	Common Causes in Brain MRI	k-Space Corruption Mechanism
Ghosting	Replicated structures along phase-encode direction [3] [2]	Periodic motion (cardiac, tremor), sudden jerks [2]	Inconsistent phase between k-space lines [2]
Blurring	Loss of sharpness and edge definition [2]	Slow, continuous drifts [2]	Smearing of high-frequency k-space data [2]
N/2 Ghost	Replica shifted by half the FOV [3]	Eddy currents from diffusion gradients, system delays [4]	Consistent phase error between odd/even echoes [4]
Signal Loss	Localized signal void	In-flow, spin dephasing in gradients	Corruption of signal generation, not just readout [2]

Quantitative Assessment of Ghosting

For quality control and objective comparison, the American College of Radiology (ACR) recommends a standardized metric called "Percent Signal Ghosting" [3]. This method uses signal measurements from both the phantom and the background air to quantify the severity of ghosting artifacts.

The formula for the ghosting ratio (G) is: G = | (T + B) - (L + R) | / (2 × S) [3]

Where:

• S = mean signal intensity from a large region of interest (ROI) in the center of the phantom.
• T, B, L, R = mean signal intensities from ROIs placed in the top, bottom, left, and right background areas outside the phantom.

A lower percent ghosting value indicates better image quality, with generally accepted values being less than 1-3% for diagnostic quality [3].

Table 2: Quantitative Metrics for Motion Artefact Assessment

Metric Name	What It Measures	Application Context
Percent Signal Ghosting [3]	Intensity of signal replicated in background	ACR Phantom QC; system performance
Framewise Displacement (FD) [1]	Magnitude of head movement between volumes	Resting-state fMRI data quality
Peak Signal-to-Noise Ratio (PSNR) [5] [6]	Fidelity of corrected image vs. ground truth	Validating motion correction algorithms
Structural Similarity Index (SSIM) [5] [6]	Perceptual image quality and structure preservation	Validating motion correction algorithms

Experimental Protocols for Motion Artefact Analysis

Simulating Motion for Method Validation

A critical step in developing and testing motion correction algorithms is the use of simulated motion on known data sets. This provides a ground truth for comparison.

Protocol: In-Silico Motion Simulation

• Data Foundation: Start with a high-quality, motion-free brain MRI dataset, often referred to as the "ground truth" [5].
• Motion Corruption: Apply a simulated motion corruption operator to the k-space of the ground truth data. This can involve:
  • Rigid-Body Transformations: Simulating rotations and translations of the head [5] [2].
  • Phase Perturbations: Introducing simulated phase errors to replicate the effects of motion on the MR signal's phase [2].
  • k-Space Line Displacement: Artificially shifting k-space lines to mimic inconsistencies from movement during phase encoding [2].
• Image Reconstruction: Reconstruct the image from the corrupted k-space using a standard Fourier transform (e.g., inverse FFT) to generate the motion-artifacted image for analysis [2].

Protocol for Ghosting Artefact Quantification (ACR Protocol)

For quantitative quality assurance of an MRI system, the following methodology is employed using a standardized phantom [3]:

• Phantom Imaging: Acquire an MRI scan of a large accreditation phantom, typically using a standard 2D spin-echo or gradient-echo sequence.
• Region of Interest (ROI) Placement:
  • Place a large circular or rectangular ROI in the uniform center of the phantom to measure the mean signal (S).
  • Place four smaller ROIs in the background air outside the phantom, positioned at the top, bottom, left, and right edges of the image (representing areas where ghosting artifacts typically appear).
• Signal Measurement: Record the mean pixel intensity from each of the five ROIs.
• Calculation: Compute the Percent Signal Ghosting using the formula provided in Section 3.3.

Diagram 1: Motion to Artefact Pathway. This flowchart illustrates the causal relationship between different types of subject motion, their specific disruptive effects on k-space data, and the final artifacts observed in the reconstructed MR image.

Implications for Brain Research and Corrective Methodologies

Impact on Brain-Behavior Associations

Motion artifacts pose a severe threat to the validity of brain-wide association studies (BWAS). Even after standard denoising procedures, residual motion artifacts can introduce systematic biases in functional connectivity (FC) measurements [1]. For instance, head motion has been shown to systematically decrease long-distance connectivity and increase short-range connectivity, a pattern that can be misattributed to neurological or psychiatric conditions [1]. Studies have shown that in large datasets like the Adolescent Brain Cognitive Development (ABCD) Study, a significant proportion of trait-FC relationships (42% showing overestimation, 38% showing underestimation) can be significantly impacted by motion even after denoising, highlighting the critical need for rigorous motion impact assessment in brain research [1].

Emerging Correction Strategies

The field is actively developing sophisticated methods to mitigate motion artifacts, broadly categorized into prospective and retrospective approaches.

• Prospective Motion Correction: These methods adjust the imaging sequence in real-time to track and compensate for motion during data acquisition. Examples include using optical tracking systems with reflective markers or MR-based navigators (e.g., PROMO, vNavs) to update the scanner's coordinate system [6] [7]. A recent study demonstrated the application of a retrospective technique called alignedSENSE for motion correction in ultralow-field portable MRI, showing clear improvements in brain image quality by jointly estimating motion and image parameters [7].
• Retrospective Motion Correction: These techniques operate on the already-acquired k-space or image data without requiring sequence modification. Traditional methods include rigid or non-rigid image registration. More recently, Deep Learning (DL) has shown remarkable promise [5] [8] [6].
  • Generative Models: Models like Generative Adversarial Networks (GANs) and Denoising Diffusion Probabilistic Models (DDPM) learn a mapping between motion-corrupted and motion-free images. For example, the Res-MoCoDiff model uses a residual-guided diffusion process to correct motion artifacts efficiently, requiring only four reverse diffusion steps for reconstruction [5].
  • Physics-Informed Deep Learning: These models incorporate knowledge of MRI physics, such as k-space corruption models, into the network's learning process to make corrections more robust and reduce the risk of "hallucinations" where the network creates plausible but incorrect image features [8].

Table 3: Research Reagent Solutions for Motion Artefact Investigation

Tool / Reagent	Function / Description	Role in Motion Artefact Research
Digital Phantom (e.g., Modified Shepp-Logan) [2]	Computational model of a human head	Provides a known ground truth for simulating motion artifacts and validating correction algorithms.
ACR Large Phantom [3]	Physical standardized object for MRI QC	Enables quantitative, reproducible measurement of ghosting artifacts on a specific scanner.
Motion Simulation Framework [5] [6]	Software to apply synthetic motion to k-space	Generates paired datasets (corrupted/clean) for training and testing deep learning models.
Navigator Echoes (e.g., Nav1, Nav2) [4]	Additional MR signal acquisitions	Samples system phase information; used to correct for phase errors causing N/2 ghosting.
Deep Learning Models (e.g., U-net, Swin Transformer) [5]	Neural network architectures	Serves as the backbone for image-denoising models that learn to remove motion artifacts.

The corruption of k-space data by subject motion is a fundamental problem in MRI, directly generating the ghosting and blurring artifacts that degrade image quality. In brain research, where quantitative accuracy is paramount, these artifacts are not merely nuisances but significant sources of bias that can lead to spurious scientific conclusions. A deep understanding of the mechanisms—how different motion types create specific phase inconsistencies and data irregularities in k-space—is the first step toward effective mitigation. While the problem is complex and multifaceted, the growing toolbox of prospective tracking, advanced reconstruction, and particularly deep learning-based retrospective correction offers powerful solutions. Future progress will depend on the continued development and rigorous validation of these methods, ensuring that MRI remains a reliable and precise instrument for unlocking the secrets of the brain.

In brain research, particularly in neuroimaging studies using techniques like functional Magnetic Resonance Imaging (fMRI), in-scanner motion presents a significant methodological challenge. Motion artifacts have the potential to systematically bias measures of functional connectivity, especially as in-scanner motion is frequently correlated with key variables of interest such as age, clinical status, and cognitive ability [9]. Understanding the spatial distribution of these artifacts is therefore crucial for interpreting neuroimaging data accurately. This whitepaper explores the biomechanical underpinnings of a consistently observed pattern in motion artifacts: the minimal motion near the atlas vertebra (C1) and the increased motion in frontal cortical regions. We examine the anatomical constraints, present quantitative data on motion distribution, and discuss the implications for researchers and drug development professionals working with neuroimaging data.

Biomechanical Basis of Motion Distribution

Anatomical Constraints of the Upper Cervical Spine

The spatial distribution of head motion within MRI scanners is not random but is fundamentally governed by the biomechanical properties of the cervical spine and the skull's attachment. The upper cervical spine, consisting of the atlas (C1) and axis (C2) vertebrae, forms a highly specialized complex designed for head mobility while protecting the brainstem. However, this mobility is not uniform across all directions.

• Atlanto-occipital Joint (C0-C1): This joint connects the skull to the atlas and primarily allows for flexion and extension (nodding). Its biconvex articular surfaces provide stability but limit other movements [10].
• Atlanto-axial Joint (C1-C2): This joint is specialized for axial rotation (head turning), contributing significantly to the overall rotational capacity of the head [10].

The biomechanical principle of coupling explains why motion is not isolated to a single plane. Coupling refers to the phenomenon where rotation or translation about one axis is consistently associated with simultaneous motion about another axis [10]. In the subaxial cervical spine (below C2), axial rotation inevitably couples with lateral bending in the same direction (ipsilateral coupling), creating complex movement patterns that propagate motion anteriorly.

The Composite Tension Plus (CT+) Model

Cortical expansion and folding, while occurring over a much longer timescale, are also governed by biomechanical principles that inform our understanding of structural constraints. The Composite Tension Plus (CT+) model posits that mechanical tension, mediated by the cytoskeleton of cells, plays a key role in morphogenesis [11]. This model, an update of the earlier Differential Expansion Sandwich Plus (DES+) model, incorporates ten distinct mechanisms to explain cortical expansion and folding. While not directly explaining short-term head motion, the CT+ model underscores that the brain is not a passive entity but possesses its own mechanical properties and internal tensions that may interact with externally imposed motions.

Quantitative Analysis of Motion Distribution

Empirical Evidence from Functional Connectivity Studies

Research specifically investigating motion artifacts in functional connectivity MRI (fcMRI) has quantified the spatial gradient of head motion. Studies using voxel-specific measures of frame displacement (FD) have consistently demonstrated that motion is minimal near the atlas vertebrae and increases with distance from this anchoring point [9]. The following table summarizes key quantitative findings from this research:

Table 1: Quantitative Spatial Distribution of In-Scanner Head Motion

Brain Region	Relative Motion Level	Primary Direction of Motion	Correlation with Global FD
Near Atlas Vertebrae	Minimal	Constrained by biomechanics	High (r ~0.89) [9]
Frontal Cortex	High	Y-axis rotation (nodding) [9]	High (r ~0.89) [9]
Tissue Boundaries	Signal Artifacts	Partial volume effects [9]	Not specified

The high correlation (r = 0.89) between voxel-specific motion measures and global Frame Displacement (FD) indicates that motion, while spatially heterogeneous, is a whole-brain phenomenon with a consistent pattern [9]. The finding that frontal cortex motion is largely driven by y-axis rotation (the nodding movement) directly links the biomechanical capability of the neck to the resulting artifact pattern in the brain image.

Biomechanical Properties of Spinal Segments

The varying range of motion (ROM) across different spinal segments explains why motion is not uniform. The following table summarizes the biomechanical properties of key cervical segments relevant to in-scanner head movement:

Table 2: Biomechanical Properties of Upper Cervical Spinal Segments

Spinal Segment	Primary Motions	Coupled Motions	Biomechanical Constraints
C0-C1 (Atlanto-occipital)	Flexion/Extension	Contralateral lateral bending during axial rotation [10]	Biconvex articular surfaces
C1-C2 (Atlanto-axial)	Axial Rotation	Lateral flexion to opposite side [10]	Transverse ligament, alar ligaments
Subaxial Cervical (C2-C3)	Lateral Bending, Flexion/Extension	Ipsilateral axial rotation during lateral bending [10]	Facet joint orientation, ligamentous constraints

Experimental Protocols for Motion Assessment

Measuring In-Scanner Motion

The standard methodology for quantifying head motion in fcMRI studies involves deriving estimates from the functional time series itself during image preprocessing [9].

• Image Realignment: Each volume in the fMRI time series is rigidly realigned to a reference volume.
• Parameter Estimation: This realignment process generates six realignment parameters (RPs) for each volume: three translations (x, y, z) and three rotations (pitch, roll, yaw).
• Frame Displacement (FD) Calculation: The RPs are summarized into a concise index of volume-to-volume motion. A common method, such as the one implemented in FSL, calculates FD using a matrix root mean squared formulation derived by Jenkinson et al. [9].
• Voxel-Specific FD: For spatial distribution analysis, voxel-specific measures of FD can be computed directly from the image header to understand regional variations [9].

Finite Element Modeling in Biomechanics

Finite element (FE) modeling is a computational technique used to simulate and analyze the biomechanics of complex structures like the spine. The following workflow outlines a typical FE approach, as used in studies of spinal fixation techniques [12]:

Diagram 1: Finite Element Model Workflow

This methodology allows researchers to simulate the Range of Motion (ROM) and stress distribution across spinal segments under various loading conditions (flexion, extension, lateral bending, axial rotation) that mimic in-scanner movements [12].

The Scientist's Toolkit: Research Reagents & Materials

Table 3: Essential Resources for Motion Biomechanics and Artifact Research

Resource Category	Specific Tool / Reagent	Function / Application
Imaging & Analysis Software	FSL (FMRIB Software Library) [9]	Calculating Frame Displacement (FD) from fMRI data.
	Simpleware, Geomagic, Hypermesh [12]	Pipeline for creating and meshing finite element models from CT data.
	Abaqus [12]	Performing finite element analysis on biomechanical models.
Computational Models	Finite Element Model of C0-C3 [12]	Simulating biomechanics and range of motion in the upper cervical spine.
	Composite Tension Plus (CT+) Model [11]	Modeling long-term, tension-based cortical expansion and folding.
Deep Learning Models	Res-MoCoDiff [5]	Efficient diffusion model for MRI motion artifact correction.
	JDAC Framework [13]	Iterative learning framework for joint image denoising and motion artifact correction.

Implications for Neuroimaging Research and Drug Development

The predictable spatial distribution of motion artifacts has profound implications for the interpretation of neuroimaging data in both basic research and clinical trials.

• Systematic Bias: Because motion is often correlated with clinical status (e.g., greater in pediatric, elderly, or neurologically impaired populations), it can introduce systematic bias into group comparisons [9]. The frontal cortex, a region critical for executive function, decision-making, and social cognition, is particularly vulnerable to such artifact-induced bias.
• Denoising Strategies: Understanding that motion follows a biomechanically determined spatial gradient is essential for developing and selecting effective denoising strategies. Techniques like global signal regression (GSR) or censoring (removing high-motion volumes) must account for this non-uniformity [9].
• Advanced Correction Techniques: Emerging deep learning methods, such as Res-MoCoDiff and the JDAC framework, show promise in retrospectively correcting for motion artifacts [5] [13]. These models can be enhanced by incorporating biomechanical priors regarding the expected spatial pattern of motion.

The spatial distribution of in-scanner head motion, characterized by minimal movement near the atlas and increasing motion in frontal regions, is a direct consequence of the biomechanical constraints imposed by the upper cervical spine. The principles of spinal coupling and the specific ranges of motion allowed by the atlanto-occipital and atlanto-axial joints create a predictable gradient of motion that manifests as a corresponding spatial pattern of artifacts in neuroimaging data. For researchers and drug development professionals, acknowledging and accounting for this biomechanically determined artifact profile is not merely a technical consideration but a fundamental requirement for ensuring the validity and interpretability of neuroimaging findings. Future work integrating quantitative biomechanical modeling with advanced artifact correction algorithms presents a promising path forward for mitigating this persistent challenge.

In brain magnetic resonance imaging (MRI), patient motion remains a significant challenge that compromises diagnostic quality and research validity. Motion during acquisition generates distinct spatial artifacts—primarily ghosting, blurring, and signal loss—that vary in appearance and severity across different brain regions. These spatial signatures are not random; their manifestation depends on complex interactions between head motion, k-space sampling strategies, and the unique physiological and structural characteristics of specific brain areas. Understanding these patterns is crucial for researchers and drug development professionals who rely on high-quality neuroimaging data to detect subtle longitudinal changes in brain structure and function, particularly in clinical trials for neurodegenerative diseases.

This technical guide explores the spatial distribution of motion artifacts in brain MRI, providing a detailed analysis of their underlying mechanisms, regional susceptibility, and advanced correction methodologies. By framing this discussion within the broader context of a thesis on the spatial distribution of motion artifacts in brain research, we aim to equip scientists with the knowledge and tools necessary to identify, characterize, and mitigate these confounding factors in neuroimaging studies.

Fundamental Mechanisms Linking Motion to Artifact Formation

K-Space Sampling and Motion Interactions

The appearance of motion artifacts in reconstructed MR images is fundamentally determined by how motion states interact with the specific k-space sampling trajectory used during acquisition. K-space represents the raw data of an MRI scan before Fourier transformation into the final image, with the center of k-space determining image contrast and the periphery determining spatial resolution. When motion occurs during data acquisition, it disrupts the phase consistency of the k-space signal, creating inconsistencies that manifest as artifacts in the image domain.

Research by Schauman et al. demonstrates that the severity and nature of motion artifacts depend heavily on the k-space distribution of motion states [14]. Through motion-sampling plots that map motion states directly onto k-space, they identified nine distinct categories of motion artifacts. Their work revealed that artifacts are especially pronounced when motion discontinuities occur near the center of k-space or align with slow phase-encoding directions [14]. This explains why certain motion patterns produce severe artifacts while others have minimal impact on image quality.

Physics of Motion-Induced Signal Changes

The interaction between motion and MRI physics extends beyond simple displacement. Rigid-body head motion in the presence of magnetic field inhomogeneities introduces additional complexity through motion-induced phase shifts. The complete forward model for motion-affected k-space data can be represented as:

Where the motion transform U_t encompasses not only rotation (R_t) and translation (T_t) but also phase shifts induced by position-dependent B0 inhomogeneities (ω_t) that increase with echo time (TE_n):

This formulation explains why T2*-weighted sequences are particularly vulnerable to motion, as the impact of motion-induced B0 inhomogeneity changes increases with longer echo times [15]. The resulting signal loss can be misinterpreted as pathological findings in susceptibility-weighted imaging or quantitative BOLD applications, potentially compromising research conclusions in functional and vascular brain imaging studies.

Regional Susceptibility to Motion Artifacts

Different brain regions exhibit varying susceptibility to motion artifacts due to their unique structural characteristics, proximity to tissue interfaces, and functional roles. The table below summarizes the spatial signatures of motion artifacts across key brain areas.

Table 1: Regional Vulnerability to Motion Artifacts in Brain MRI

Brain Region	Ghosting Patterns	Blurring Effects	Signal Loss	Primary Causes
Frontal Cortex	Moderate ghosting in phase-encode direction	Mild to moderate blurring of gyral patterns	Minimal	Proximity to sinuses, susceptibility gradients
Brainstem	Severe ghosting along multiple axes	Significant blurring of fine structures	Pronounced in T2*-weighted	Pulsatile motion, proximity to bone
Hippocampal Formation	Direction-specific ghosting	High impact on subfield delineation	Moderate	Complex geometry, pulsation from adjacent vessels
Corpus Callosum	Anisotropic ghosting patterns	White matter boundary blurring	Minimal in conventional, severe in diffusion	Central location, fiber orientation dependence
Visual Cortex	Posterior-anterior ghosting	Moderate impact on layer differentiation	Calcarine fissure specific	Tissue-air interfaces near occipital pole
Cerebellum	Multi-directional ghosting	Severe folia pattern degradation	Moderate to severe	Continuous physiological motion, posterior fossa anatomy

The medial temporal lobe, particularly hippocampal subregions, demonstrates unique vulnerability due to its complex architecture and proximity to cerebrospinal fluid spaces. Studies investigating functional connectivity in these regions note that motion artifacts can significantly alter apparent connectivity measures, potentially confounding research on cognitive dysfunction [16]. Similarly, the brainstem and cerebellum are highly susceptible to both cardiac and respiratory-induced motion, often exhibiting severe blurring that obscures fine structural details essential for neurodegenerative disease monitoring.

Research by Giocomo et al. reveals that spatial memory networks centered on the medial entorhinal cortex show age-related instability in grid cell function, which may compound motion-related artifacts in functional studies of navigation and memory [17]. This intersection of biological vulnerability and technical artifact presents particular challenges for studies of aging and neurodegeneration where both motion and biological signals coexist.

Methodologies for Characterizing Motion Artifacts

Motion-Sampling Plot Framework

Schauman et al. introduced motion-sampling plots as a systematic framework for predicting and classifying motion artifacts based on k-space sampling properties [14]. This methodology enables researchers to map motion states directly onto k-space coordinates and assess their relationship to artifact appearance. The experimental protocol involves:

• Acquisition of 3D MRI data using multiple sampling trajectories (Cartesian, stack-of-stars, kooshball)
• Introduction of controlled motion through healthy volunteers mimicking patient motion while wearing real-time pose-tracking devices
• Systematic variation of motion direction, magnitude, and timing relative to k-space sampling
• Generation of motion-sampling plots that visualize the relationship between motion states and k-space coordinates

This approach has revealed that the k-space distribution of motion states is more predictive of artifact appearance than motion magnitude alone, providing researchers with a powerful tool for optimizing acquisition strategies for specific brain regions [14].

Physics-Informed Motion Detection

PHIMO+ represents an advanced methodology for T2* quantification that leverages physics principles to detect motion-corrupted k-space lines [15]. The technique employs a physics-informed loss function based on the empirical correlation coefficient between reconstructed signal intensities and mono-exponentially fitted intensities:

This loss function capitalizes on motion-related B0 inhomogeneity changes that disrupt the expected mono-exponential signal decay in multi-echo GRE sequences [15]. The experimental workflow involves:

• Acquisition of multi-echo GRE data for T2* quantification
• Self-supervised optimization of k-space line exclusion masks based on physics loss
• Data-consistent reconstruction using only motion-free k-space lines
• Central k-space preservation by assuming minimal change in mean image intensity despite motion

This methodology is particularly valuable for mqBOLD applications where accurate T2* quantification is essential for assessing oxygen metabolism, especially in deep gray matter structures susceptible to motion-induced signal loss [15].

Table 2: Advanced Motion Correction Methods in Brain MRI

Method	Underlying Principle	Best-Suited Brain Regions	Limitations
JDAC Framework [18]	Joint iterative denoising and artifact correction	Whole-brain, especially gray-white matter interfaces	Requires extensive training data
PHIMO+ [15]	Physics-informed k-space line detection and exclusion	Brainstem, basal ganglia (T2*-weighted)	Limited to rigid-body motion
DIMA [19]	Diffusion models for unsupervised artifact correction	Cortical regions, hippocampal formation	Computational intensity
Leverage-Score Sampling [20]	Functional connectome feature selection	Large-scale networks, connectivity studies	Restricted to fMRI applications

Experimental Protocols for Motion Artifact Characterization

Protocol for Spatial Mapping of Motion Artifacts

Comprehensive characterization of motion artifact spatial signatures requires a standardized experimental approach:

• Participant Selection: Include healthy volunteers capable of mimicking controlled head movements during scanning. Sample size justification should consider expected effect sizes for motion-artifact correlations.

• Motion Simulation: Implement both continuous (slow drift) and abrupt (sudden jerks) motion patterns using real-time pose tracking [14]. Motion parameters should include translation (x, y, z) and rotation (pitch, roll, yaw) across a range of magnitudes.

• Multi-Protocol Acquisition: Acquire data using:

  • T1-weighted and T2-weighted structural sequences
  • T2*-weighted GRE for susceptibility-weighted contrast
  • Diffusion-weighted imaging with multiple b-values
  • Resting-state fMRI for functional connectivity assessment

• Reference Standard: Include motion-free scans for each participant as a baseline for artifact quantification.

• Artifact Assessment: Utilize quantitative metrics including:

  • Image intensity variance in homogeneous regions
  • Edge sharpness at tissue boundaries
  • Signal-to-fluctuation noise ratio in time-series data
  • Functional connectivity stability in network hubs

Protocol for Validation of Correction Algorithms

Robust validation of motion correction methods requires carefully designed experiments:

• Dataset Curation: Utilize public datasets with paired motion-corrupted and motion-free data where available [21]. The Diff5T dataset provides valuable k-space data for method development and benchmarking [21].

• Performance Metrics: Evaluate using quantitative measures including:

  • Peak signal-to-noise ratio (PSNR)
  • Structural similarity index (SSIM)
  • Mean squared error (MSE) in region-of-interest analyses
  • Anatomical fidelity using validated segmentation pipelines

• Regional Analysis: Assess performance separately for:

  • Cortical gray matter
  • White matter tracts
  • Deep gray matter structures
  • Posterior fossa contents

• Clinical Validation: Correlate motion-corrected image quality with quantitative biomarkers relevant to drug development, such as hippocampal volume in Alzheimer's trials or lesion load in multiple sclerosis studies.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Resources for Motion Artifact Research

Resource	Function	Example Applications
Diff5T Dataset [21]	5.0 Tesla diffusion MRI with raw k-space data	Method benchmarking, artifact simulation
Cam-CAN Dataset [20]	Multi-modal aging study with functional connectomes	Aging-related motion vulnerability studies
Motion-Sampling Plots [14]	Framework for predicting artifact appearance	Sequence optimization, protocol design
JDAC Framework [18]	Joint denoising and artifact correction	Processing of low-quality clinical scans
PHIMO+ Algorithm [15]	Physics-informed motion correction for T2* mapping	Quantitative susceptibility mapping, mqBOLD
DIMA Framework [19]	Unsupervised correction using diffusion models	Clinical data with no motion-free references
Leverage-Score Feature Selection [20]	Identification of motion-resistant connectome features	Functional connectivity studies in aging

Visualization of Motion Artifact Characterization Workflows

Motion Artifact Characterization and Correction Workflow

Computational Frameworks for Motion Correction

The spatial signatures of motion artifacts in brain MRI—ghosting, blurring, and signal loss—follow predictable patterns across brain regions based on interactions between motion parameters, k-space sampling strategies, and regional anatomical characteristics. Understanding these patterns is essential for researchers and drug development professionals seeking to maximize data quality and validity in neuroimaging studies.

Methodologies such as motion-sampling plots, physics-informed motion detection, and joint denoising-artifact correction frameworks provide powerful tools for characterizing and mitigating these artifacts. By incorporating these approaches into standardized experimental protocols and leveraging emerging datasets and algorithms, the neuroscience community can significantly improve the reliability of neuroimaging biomarkers in both basic research and clinical trials.

Future directions in this field include the development of region-specific correction algorithms, real-time motion detection and compensation systems, and standardized quality control pipelines that account for the spatial distribution of motion artifacts. Such advances will be particularly valuable for longitudinal studies tracking subtle changes in brain structure and function, where consistent image quality is essential for detecting treatment effects.

Motion artifacts represent a significant source of noise in neuroimaging, critically impacting the fidelity of data acquisition and interpretation in brain research. These artifacts manifest in distinct patterns across different imaging modalities, each with unique implications for spatial distribution analysis and functional connectivity mapping. Understanding these modality-specific manifestations is paramount for developing effective correction strategies and ensuring the validity of neuroscientific findings, particularly in clinical and drug development contexts where accurate spatial localization of brain activity is essential.

The spatial distribution of motion artifacts is intrinsically linked to the fundamental physical principles underlying each imaging technology. Functional Magnetic Resonance Imaging (fMRI) measures the blood-oxygen-level-dependent (BOLD) signal, which is highly sensitive to head displacement due to the requirement for precise magnetic field homogeneity [22]. Functional Near-Infrared Spectroscopy (fNIRS) detects changes in hemoglobin concentrations through optode-scalp coupling, where motion affects signal quality through optode displacement and altered light transmission paths [23] [24]. Structural MRI, while not measuring dynamic function, suffers from spatial blurring and reconstruction artifacts that compromise anatomical accuracy [14] [18]. These differential mechanisms result in characteristic artifact patterns that necessitate tailored correction approaches for each modality, particularly in research focusing on the spatial organization of brain networks.

Fundamental Principles and Artifact Generation Mechanisms

Physical Basis of Artifact Formation

The generation of motion artifacts in neuroimaging modalities stems from their distinct physical measurement principles. fMRI relies on detecting minute changes in magnetic susceptibility caused by variations in blood oxygenation. Head motion disrupts the static magnetic field homogeneity, leading to spin history effects and phase inconsistencies during k-space sampling [14]. The spatial manifestation of these artifacts depends critically on the interaction between the motion trajectory and k-space sampling pattern, with motion discontinuities near the center of k-space or aligned with slow phase-encoding directions producing particularly severe artifacts [14].

fNIRS operates on different principles, utilizing near-infrared light (650-1000 nm wavelengths) to measure concentration changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) in cortical tissue [23]. Motion artifacts in fNIRS primarily occur through two mechanisms: optode displacement relative to the scalp, which alters the light-coupling efficiency, and changes in pressure on the scalp, which modulates peripheral blood flow in superficial tissues [24]. Unlike fMRI, fNIRS artifacts are typically channel-specific rather than affecting the entire image, though their impact can be substantial due to the relatively small signal changes associated with neural activity.

Characterization of Motion Artifact Types

Motion artifacts can be categorized based on their temporal characteristics and spatial extent. In fNIRS, researchers have classified artifacts into four distinct types: Type A (sharp spikes with standard deviation >50 from mean within 1 second), Type B (peaks with standard deviation >100 lasting 1-5 seconds), Type C (gentle slopes with standard deviation >300 over 5-30 seconds), and Type D (slow baseline shifts >30 seconds with standard deviation >500) [23]. This classification provides a framework for developing and evaluating targeted correction algorithms.

In MRI and fMRI, artifacts are often characterized by their appearance in reconstructed images, including ghosting (replication of structures along the phase-encoding direction), spin-history effects (signal loss due to intra-slice motion), and blurring (loss of spatial resolution) [14] [18]. The specific manifestation depends on the sampling strategy (e.g., Cartesian, stack-of-stars, or kooshball trajectories) and the nature of the motion, with recent research identifying nine distinct categories of motion artifacts in 3D MRI [14].

Table 1: Comparative Characteristics of Motion Artifacts Across Neuroimaging Modalities

Characteristic	fMRI	fNIRS	Structural MRI
Primary Mechanism	Magnetic field inhomogeneity, spin history effects	Optode-scalp decoupling, pressure changes	K-space inconsistencies, sampling errors
Spatial Pattern	Whole-image distortions, ghosting along phase-encode direction	Channel-specific artifacts, superficial cortical regions	Volumetric blurring, global resolution loss
Temporal Classes	Low-frequency drift, sudden jumps	Type A (spikes), B (peaks), C (slopes), D (baseline shifts)	Transient vs. persistent motion effects
Depth Sensitivity	Whole-brain (cortical and subcortical)	Superficial cortex (1-1.5 cm depth)	Whole-brain
Sampling Interaction	Strong dependence on k-space trajectory	Minimal sampling rate dependence	Strong dependence on k-space trajectory

Modality-Specific Artifact Profiles and Spatial Representations

fMRI Artifact Patterns and Spatial Distribution

Functional MRI exhibits distinctive motion artifact patterns that directly impact spatial localization accuracy. The blood-oxygen-level-dependent (BOLD) contrast mechanism underlying fMRI is particularly vulnerable to motion-induced magnetic field fluctuations. Rigid head motion interacts with k-space sampling strategies to produce characteristic artifacts whose appearance can be predicted using motion-sampling plots [14]. These artifacts are especially pronounced when motion discontinuities occur near the center of k-space or align with slow phase-encoding directions [14].

The spatial specificity of fMRI is further complicated by the phenomenon of spatial misregistration between functional scans and anatomical reference images. Even submillimeter movements can cause significant signal variations that mimic true neural activation patterns, particularly in regions near tissue boundaries with strong magnetic susceptibility gradients (e.g., orbitofrontal cortex, temporal lobes) [22]. Recent studies utilizing simultaneous fNIRS-fMRI have revealed that motion artifacts in fMRI can reduce the positive predictive value for detecting true brain activation to as low as 41.5% in within-subject analyses [25], highlighting the critical impact of motion on spatial accuracy.

fNIRS Artifact Patterns and Spatial Considerations

Functional NIRS exhibits a different artifact profile dominated by signal spikes and baseline shifts resulting from optode movement relative to the scalp. The spatial distribution of fNIRS artifacts is inherently linked to probe design and placement. Unlike fMRI, where motion affects the entire image, fNIRS artifacts typically impact specific channels, though their effects can propagate through analysis pipelines [23] [24]. The limited penetration depth of near-infrared light (approximately 1-1.5 cm) confines fNIRS measurements to superficial cortical regions, making them particularly vulnerable to systemic physiological noise from scalp blood flow that can be motion-induced [26].

The reproducibility of fNIRS spatial measurements is significantly affected by variations in optode placement across sessions. Studies have demonstrated that increased shifts in optode position correlate with reduced spatial overlap of detected activation across multiple sessions [27]. This has important implications for longitudinal studies and clinical applications where consistent spatial localization is essential. Research has shown that oxyhemoglobin (HbO) measurements are significantly more reproducible across sessions than deoxyhemoglobin (HbR) measurements [27], suggesting that HbO may be preferable for studies requiring high spatial reliability.

Structural MRI Artifact Patterns

Structural MRI, while not measuring dynamic function, faces unique motion artifact challenges that compromise anatomical accuracy. Motion during acquisition causes k-space inconsistencies that manifest as blurring, ghosting, and resolution loss in reconstructed images [14] [18]. The specific appearance of these artifacts depends on the interaction between the motion trajectory and the k-space sampling pattern, with certain sampling strategies (e.g., Cartesian) being more vulnerable to particular motion types than others (e.g., radial or spiral trajectories) [14].

In 3D MRI acquisitions, motion artifacts demonstrate complex spatial dependencies based on the direction and timing of movement relative to the phase-encoding sequence. Recent research has shown that the severity and nature of artifacts depend heavily on the k-space distribution of motion states, which can be visualized and interpreted using motion-sampling plots [14]. These artifacts are particularly problematic in clinical populations who may have difficulty remaining still, such as children, elderly patients, or individuals with neurological disorders affecting motor control.

Table 2: Quantitative Impact of Motion on Signal Quality and Spatial Accuracy

Metric	fMRI	fNIRS	Structural MRI
Temporal Resolution	0.3-2 Hz [22]	Up to 10 Hz (typically 5-10 Hz) [24]	N/A (structural acquisition)
Spatial Resolution	1-3 mm (high) [22]	1-3 cm (moderate) [28]	0.5-1 mm (very high)
Spatial Overlap with Ground Truth	47.25% within-subject [25]	Varies with optode placement consistency [27]	Qualitative assessment of anatomical accuracy
Positive Predictive Value	41.5% within-subject [25]	Dependent on artifact correction method [29]	N/A
Reproducibility	High with motion correction	HbO more reproducible than HbR [27]	High with motion correction

Methodologies for Motion Artifact Correction

fMRI Motion Correction Techniques

fMRI motion correction primarily relies on image registration algorithms that realign sequential volumes to a reference image. Tools such as FSL MCFLIRT and AFNI 3dVolReg calculate rigid-body transformation parameters (three translations and three rotations) to minimize variance between volumes [28]. More advanced approaches include prospective motion correction, which adjusts the imaging field of view in real-time based on head tracking, and sampling strategies less sensitive to motion, such as multiband acquisitions and radial sampling trajectories [14].

Recent innovations have explored the use of deep learning approaches for joint image denoising and motion artifact correction. The Joint Image Denoising and Artifact Correction framework employs iterative learning with an adaptive denoising model and an anti-artifact model to progressively improve image quality [18]. This approach incorporates a novel gradient-based loss function designed to maintain the integrity of brain anatomy throughout the correction process, addressing the limitation of traditional methods that often treat denoising and motion correction as separate tasks [18].

fNIRS Motion Correction Algorithms

Multiple algorithmic approaches have been developed to address motion artifacts in fNIRS data, each with distinct strengths and limitations. Wavelet-based methods utilize multiscale decomposition to identify and remove artifact components in the wavelet domain, demonstrating particular efficacy for spike artifacts [23] [29]. Temporal Derivative Distribution Repair employs a robust statistical approach to identify and downweight outliers in the temporal derivative of fNIRS signals, making it effective for both spike and baseline shift artifacts [29].

Other commonly used techniques include spline interpolation (MARA), which identifies artifact segments and replaces them with spline interpolants [29], principal component analysis, which removes components with variance patterns characteristic of motion [23], and correlation-based signal improvement, which exploits the negative correlation between HbO and HbR to identify and remove motion artifacts [29]. Comparative studies have found that the moving average and wavelet methods yield the best outcomes for pediatric data [23], while TDDR and wavelet filtering are most effective for functional connectivity analysis [29].

Multimodal Integration for Enhanced Correction

The integration of multiple neuroimaging modalities has enabled innovative approaches to motion artifact correction. Simultaneous fNIRS-fMRI recordings allow the use of high-temporal-resolution motion information derived from fMRI to correct fNIRS data. The AMARA-fMRI algorithm reconstructs high-resolution motion traces from slice-level acquisition times in simultaneous multislice fMRI, enabling effective motion correction in fNIRS without requiring MR-compatible accelerometers [28].

This multimodal approach capitalizes on the complementary strengths of each technique - fMRI's high spatial resolution and fNIRS' superior temporal resolution and portability [22]. Studies have demonstrated that such integration improves the detection of activation in deoxyhemoglobin and shows high overlap with fMRI activation when considering both HbO and HbR signals [28]. The spatial correspondence between fNIRS and fMRI detection of task-related activity has been shown to be good in terms of true positive rate, with fNIRS overlapping up to 68% of fMRI activation for group analyses [25].

Experimental Protocols for Artifact Assessment

Protocol Design for Motion Artifact Characterization

Systematic assessment of motion artifacts requires carefully designed experimental protocols that incorporate controlled motion conditions. For fNIRS studies, protocols should include task paradigms known to elicit motion, such as motor execution tasks or language production tasks, particularly in challenging populations like children [23]. These should be combined with periods of rest to establish baseline signal characteristics. The use of structured motion artifact classification (Types A-D) enables standardized quantification and comparison across studies [23].

For fMRI studies, protocols should incorporate various k-space sampling strategies (Cartesian, stack-of-stars, kooshball) to evaluate their interaction with different motion types [14]. The use of real-time pose tracking devices during acquisition allows precise correlation between motion parameters and artifact appearance [14]. Experimental designs should include both task-based and resting-state conditions to evaluate motion effects on both activation maps and functional connectivity measures.

Validation Metrics and Performance Assessment

Rigorous validation of motion correction methods requires multiple complementary metrics. For quantitative comparison, the receiver operating characteristic analysis provides sensitivity-specificity curves that evaluate the ability of correction methods to preserve true signals while removing artifacts [29]. Spatial overlap metrics quantify the correspondence between activation maps or functional networks derived from corrected data and ground truth references [25].

Additional important metrics include temporal signal-to-noise ratio, which measures the stability of the signal over time, and graph theory metrics (e.g., network degree, clustering coefficient) for evaluating the impact on functional connectivity patterns [29]. Reproducibility across multiple sessions provides a crucial measure of reliability for longitudinal studies [27]. For structural MRI, qualitative assessment by expert radiologists remains an important validation approach alongside quantitative metrics [18].

Table 3: Performance Comparison of fNIRS Motion Correction Algorithms

Algorithm	Best For	Advantages	Limitations
Wavelet Filtering	Spike artifacts (Type A), functional connectivity analysis	Automatic, no additional hardware required	Can exacerbate baseline shifts
Temporal Derivative Distribution Repair	Mixed artifact types, online processing	Robust statistical foundation, real-time capability	Assumptions about derivative distribution
Spline Interpolation	Isolated motion artifacts	Preserves signal shape outside artifacts	Requires accurate artifact detection
Principal Component Analysis	Global artifact patterns	Effective for structured noise	May remove neural signal in low-channel setups
Correlation-Based Signal Improvement	Co-occurring HbO/HbR artifacts	Physiologically motivated model	Assumes perfect negative HbO-HbR correlation
Kalman Filtering	Progressive motion trends	Model-based approach	Requires parameter tuning

The Scientist's Toolkit: Essential Research Reagents and Materials

• Homer2 Software Package: A comprehensive MATLAB-based analysis suite for fNIRS data preprocessing, including implementation of multiple motion correction algorithms such as wavelet filtering and spline interpolation [23].

• TechEN-CW6 fNIRS System: A continuous-wave fNIRS system operating at 690 and 830 nm wavelengths, suitable for pediatric and adult studies with configurable source-detector arrangements [23].

• NIRSport2 Portable fNIRS System: A continuous-wave portable fNIRS device with 16 LED light sources (760 and 850 nm) and 15 silicon photodiode detectors, enabling flexible experimental setups outside laboratory environments [26].

• Multimodal fNIRS-fMRI Probes: MR-compatible optode assemblies integrated into phased-array RF coils, allowing simultaneous acquisition while maintaining precise alignment between modalities [28].

• FSL MCFLIRT: FMRIB Software Library module for motion correction in fMRI time-series data using rigid-body transformation with trilinear interpolation [28].

• BrainVoyager QX: Comprehensive software package for fMRI analysis including preprocessing, statistical analysis, and visualization, with specialized tools for motion detection and correction [26].

• ADNI Dataset: Large-scale public database of MRI scans including T1-weighted structural images used for training and validation of denoising and motion correction algorithms [18].

• Real-Time Pose Tracking Systems: MR-compatible motion tracking devices that provide continuous head position data during scanning for prospective motion correction and artifact analysis [14].

Motion artifacts manifest in fundamentally distinct patterns across MRI, fMRI, and fNIRS neuroimaging modalities, each requiring specialized correction approaches tailored to their specific mechanisms and spatial characteristics. The spatial distribution of these artifacts is inextricably linked to the underlying physical principles of each technology, from k-space sampling interactions in fMRI to optode-scalp coupling in fNIRS. Effective artifact mitigation necessitates a comprehensive understanding of these modality-specific manifestations, particularly for research focusing on the spatial organization of brain function.

Advancements in motion correction increasingly leverage multimodal integration and machine learning approaches that jointly address multiple artifact types while preserving neural signals. The development of standardized evaluation metrics and experimental protocols for assessing correction efficacy remains crucial for validating these methods across diverse populations and experimental contexts. As neuroimaging continues to expand into more naturalistic settings and clinical applications, robust motion artifact management will play an increasingly critical role in ensuring the spatial accuracy and reliability of neuroscientific findings.

From Proactive to Post-Hoc: Cutting-Edge Techniques for Motion Artifact Correction

Motion artifacts represent a fundamental challenge in non-invasive brain imaging, significantly impacting the data quality and validity of neuroscientific findings. These artifacts arise from subject movement during data acquisition, causing a range of deleterious effects from signal distortion to complete data loss. In the context of studying the spatial distribution of neural activity, motion introduces systematic biases that can obscure true brain network dynamics and lead to spurious interpretations of functional specialization. The problem is particularly acute in clinical populations where motion control is difficult, potentially confounding research on neurological and psychiatric disorders. Understanding and mitigating these artifacts is therefore not merely a technical exercise but a prerequisite for accurate brain mapping. Hardware-based solutions—including accelerometers, inertial measurement units (IMUs), and optical tracking systems—provide a direct means to quantify and correct for motion effects by capturing head movement with high temporal precision. This whitepaper examines these critical technologies, their implementation methodologies, and their role in preserving the spatial fidelity of brain activity maps.

The Spatial Distribution of Motion Artifacts: A Brain Research Perspective

Motion artifacts are not uniform in their impact across the brain; their spatial distribution is influenced by imaging modality, hardware configuration, and neuroanatomy. In functional near-infrared spectroscopy (fNIRS), optode-skin decoupling caused by head movements produces signal anomalies that vary in amplitude and morphology—from rapid spikes to slow baseline shifts—depending on the nature and severity of motion [30]. These artifacts can create false patterns of functional connectivity if uncorrected, particularly in resting-state studies where subtle correlations define network architecture.

For magnetic resonance imaging (MRI), motion disrupts the sequential k-space data acquisition, leading to ghosting, blurring, and signal loss in reconstructed images [2]. The appearance and location of these artifacts depend on the interaction between the motion vector, the pulse sequence, and k-space trajectory. Critically, motion-induced signal changes can systematically vary across brain regions, potentially mimicking disease-related atrophy or activation patterns [31]. This is especially problematic for large-scale brain initiatives and clinical data warehouses, where automated analysis of images with varying motion contamination can introduce profound biases [31].

Table 1: Classification and Spatial Manifestation of Motion Artifacts in Brain Imaging

Artifact Type	Primary Cause	Spatial Manifestation in Brain Data	Commonly Affected Brain Regions
Spikes (fNIRS)	Sudden optode decoupling	High-amplitude, transient signal deviations	Regions under mobile optodes (prefrontal, motor cortex)
Baseline Shifts (fNIRS)	Slow optode displacement	Low-frequency signal drift	All measured regions, particularly problematic for long channels
Ghosting (MRI)	Periodic motion	Replicated structures along phase-encode direction	Cortical boundaries, high-contrast interfaces
Blurring (MRI)	Continuous motion	Loss of structural definition and edge contrast	Fine structures (hippocampus, brainstem)
Signal Loss (MRI)	Spin dephasing	Focal signal dropout	Tissue-bone interfaces, medial temporal lobe

Understanding this spatial dimension is crucial for developing targeted correction strategies. Hardware-based motion tracking provides the essential kinematic data needed to disambiguate true neural signals from motion-induced artifacts, thereby preserving the topographic integrity of brain maps.

Core Hardware Technologies for Motion Tracking

Accelerometers and Inertial Measurement Units (IMUs)

Operating Principle: Accelerometers measure proper acceleration along one or more axes, while IMUs typically combine triaxial accelerometers, gyroscopes (measuring angular velocity), and magnetometers (sensing orientation relative to Earth's magnetic field) to provide comprehensive motion data [32] [33]. Through sensor fusion algorithms, IMUs can track head position and orientation with high temporal resolution (>100 Hz).

Integration in Brain Imaging: In fNIRS and EEG, these sensors are directly mounted on the head cap or optode holders to capture head movement [33]. The acquired motion data serves two primary functions: (1) as a reference signal for adaptive filtering to remove motion artifacts from physiological signals, and (2) as a quality metric to identify and exclude motion-corrupted epochs from analysis.

Key Methodologies:

• Active Noise Cancellation (ANC): Uses accelerometer output as a reference signal in an adaptive filter to subtract motion artifacts from fNIRS/EEG data [33].
• Accelerometer-Based Motion Artifact Removal (ABAMAR): Identifies motion-contaminated segments based on accelerometer signal magnitude and applies signal correction [33].
• BLind Source Separation, Accelerometer-based Artifact Rejection, and Detection (BLISSA2RD): Combines blind source separation with accelerometer guidance to isolate and remove motion components [33].

Optical Motion Tracking Systems

Operating Principle: Optical tracking systems use multiple cameras to track the 3D position of reflective or active markers placed on the subject's head [32]. By triangulating marker positions from different camera viewpoints, these systems provide high-precision spatial coordinates (sub-millimeter accuracy) for rigid-body motion.

Integration in Brain Imaging: In MRI research, optical tracking is often implemented for prospective motion correction, where real-time head position data actively adjusts the imaging sequence to maintain consistent spatial encoding [2]. This approach is particularly valuable for high-resolution structural and functional scans where even sub-millimeter movement degrades image quality.

Performance Validation: Studies comparing optical tracking to inertial sensors show that optical systems generally provide superior accuracy for spatial displacement measurements, while IMUs excel at capturing high-frequency vibrations and tremors. The choice between technologies depends on the specific motion profile of interest and practical constraints of the imaging environment.

Table 2: Technical Comparison of Motion Tracking Technologies for Brain Imaging

Parameter	Accelerometers/IMUs	Optical Tracking Systems	Practical Implications for Brain Research
Spatial Accuracy	Moderate (mm-cm range)	High (sub-mm range)	Optical preferred for precise spatial mapping
Temporal Resolution	Very High (>100 Hz)	High (50-200 Hz)	IMUs better for capturing tremor, micro-movements
Measurement Volume	Self-contained, unlimited	Limited by camera field-of-view	IMUs suitable for unconstrained movement paradigms
Line-of-Sight Requirement	None	Required for all markers	IMUs advantageous for prone positions, head turns
Setup Complexity	Low (wearable sensors)	High (camera calibration)	IMUs facilitate rapid subject setup
Susceptibility to Interference	Magnetic fields, metallic objects	Ambient light, marker occlusion	Consideration for specific imaging environments (MRI)
Integration with Brain Data	Direct synchronization with physiological signals	May require specialized interface hardware	IMUs offer simpler temporal alignment

Experimental Protocols and Implementation

IMU Integration for fNIRS Motion Correction

Equipment and Sensor Placement:

• IMU Selection: Choose miniaturized IMUs (e.g., MTx, Xsens) with appropriate sampling rates (≥100 Hz) [32].
• Mounting: Securely attach IMUs to the head using double-sided adhesive tape with additional elastic straps to minimize sensor-skin movement [32]. Optimal placement is on the forehead or posterior head regions with minimal hair interference.
• Synchronization: Implement hardware or software triggering to synchronize IMU data acquisition with fNIRS recording using a common pulse signal [32].

Data Processing Workflow:

• Motion Data Acquisition: Record triaxial acceleration, angular velocity, and orientation data throughout the imaging session.
• Motion Artifact Identification: Apply threshold-based detection to accelerometer magnitude to flag motion-contaminated epochs.
• Adaptive Filtering: For ANC approaches, use the accelerometer signal as a reference input to an adaptive filter (e.g., LMS, RLS) to estimate and subtract motion artifacts from fNIRS signals [33].
• Validation: Compare motion-corrected fNIRS signals with simultaneously acquired quality metrics (e.g., short-separation channels) to verify artifact reduction.

Diagram 1: IMU-Based Motion Correction Workflow

Optical Tracking for MRI Motion Correction

System Configuration:

• Camera Setup: Position multiple infrared cameras around the scanner bore to maintain line-of-sight to head-mounted markers throughout the examination.
• Marker Configuration: Apply reflective markers in a non-collinear arrangement on a rigid frame attached to the head coil or directly to a custom head cap [31].
• Calibration: Perform system calibration using a reference object to establish the transformation between camera coordinates and scanner coordinates.

Prospective Motion Correction Implementation:

• Real-Time Tracking: Continuously monitor head position relative to the magnet isocenter during sequence execution.
• Coordinate Transformation: Convert optical marker positions to scanner coordinate system using a predetermined transformation matrix.
• Sequence Adjustment: Dynamically update gradient orientations, RF frequencies, and slice positions to compensate for detected motion [2].
• Quality Assurance: Implement motion thresholds to trigger scan reacquisition if correction limits are exceeded.

Diagram 2: Prospective Motion Correction in MRI

Validation Protocols for Motion Correction Systems

Ground Truth Establishment:

• Conduct simultaneous recordings with multiple motion tracking modalities (e.g., IMU + optical) to establish convergent validity [32].
• Perform controlled motion experiments with known movement patterns (e.g., metronome-guided head rotations) to quantify system accuracy.

Performance Metrics:

• Temporal Accuracy: Measure synchronization precision between motion data and brain imaging signals.
• Spatial Accuracy: Quantify positional error against known displacement standards.
• Correction Efficacy: Evaluate improvement in standard image quality metrics (SNR, CNR) and functional connectivity measures after motion correction [30].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Research Materials for Hardware-Based Motion Tracking

Component	Function/Purpose	Example Products/Specifications
Miniature IMU	Measures linear acceleration and angular velocity for motion tracking	Xsens MTx (3D accelerometer, 3D gyroscope, 3D magnetometer) [32]
Inertial Sensor System	Multi-sensor platform for full-body motion capture	Xsens Motion Tracking System (multiple synchronized MTx units) [32]
Optical Motion Capture	High-precision 3D position tracking using reflective markers	Codamotion system with active markers [32]
Head-Mounted Marker Set	Rigid attachment of reflective markers to the head for optical tracking	Custom helmet or headband with non-metallic markers for MRI compatibility [31]
Synchronization Interface	Temporal alignment of motion data with brain imaging signals	Digital I/O trigger box or dedicated sync pulse generator [32]
Motion Simulation Platform	Validation of tracking accuracy using controlled movements	Robotic or manual positioning stages with known displacement [31]

Discussion and Future Directions

Hardware-based motion tracking represents an essential methodology for protecting the spatial fidelity of brain mapping data against the corrupting influence of subject movement. The complementary strengths of IMUs (high temporal resolution, wearability) and optical tracking (high spatial accuracy) provide researchers with flexible options tailored to their specific experimental needs. As brain research increasingly focuses on distributed network properties and fine-grained topographic organization, the accurate characterization and correction of motion artifacts becomes ever more critical.

Future developments will likely focus on multimodal sensor fusion combining inertial and optical data to leverage their respective advantages, and deep learning approaches that use motion sensor inputs to guide more intelligent artifact correction [34]. Furthermore, the miniaturization and wireless integration of these sensors will facilitate their application in naturalistic imaging environments and special populations where motion is most problematic. By rigorously implementing these hardware solutions, researchers can ensure that the spatial patterns observed in their data reflect genuine brain organization rather than motion-induced artifacts, thereby advancing more accurate models of brain function in health and disease.

Motion artifacts in brain Magnetic Resonance Imaging (MRI) represent a significant challenge in both clinical diagnostics and neuroscientific research. These artifacts, primarily stemming from rigid head motion, degrade image quality and can hinder downstream applications such as image segmentation, registration, and accurate target tracking in MR-guided radiation therapy [5]. The spatial distribution of these artifacts is not random; they can alter the B0 field, resulting in susceptibility artifacts, and disrupt k-space readout lines, potentially violating the Nyquist criterion and causing characteristic ghosting and ringing patterns [5] [35]. From a research perspective, understanding this spatial distribution is paramount, as motion artifacts may compromise the validity of studies investigating fine-scale brain anatomy and function. The pursuit of robust correction methods is not merely a technical exercise in image enhancement but a fundamental prerequisite for ensuring the accuracy and reliability of brain research findings. Traditional mitigation strategies, including repeated acquisitions or prospective motion tracking, impose substantial workflow burdens and are often impractical in time-sensitive clinical or research settings [36] [35]. This context has catalyzed the deep learning revolution, with Convolutional Neural Networks (CNNs) establishing a strong foundation and novel approaches like Residual-Guided Diffusion Models (Res-MoCoDiff) pushing the boundaries of what is possible in artifact correction.

Convolutional Neural Networks: The Foundational Bedrock

Convolutional Neural Networks (CNNs) have emerged as a dominant class of artificial neural networks for processing data with a grid-like topology, such as images. Their design, inspired by the organization of the animal visual cortex, is built to automatically and adaptively learn spatial hierarchies of features from low- to high-level patterns [37]. A typical CNN architecture is composed of multiple building blocks: convolution layers, pooling layers, and fully connected layers. The convolution and pooling layers are responsible for feature extraction, while the fully connected layers map these extracted features to the final output, such as a classification or segmentation map [37]. The convolution operation itself uses small arrays of learnable parameters called kernels, which are applied across the entire input image to create feature maps. Key features of this process include weight sharing, which makes the network translation invariant and dramatically increases model efficiency compared to fully connected networks [37].

In the specific domain of brain MRI analysis, CNNs have been extensively applied. An overview of deep CNNs for brain image analysis on MRI highlighted their use in segmenting brain lesions, tissues, and sub-cortical structures [38]. The review detailed how these models leverage building blocks like convolution, pooling, and non-linear activation functions (e.g., ReLU) to achieve state-of-the-art performance. CNNs have become the first choice for many problems in computer vision, and their success in brain MR image analysis is evidenced by their overwhelming acceptance within the community, particularly in challenges organized by the Medical Image Computing and Computer-Assisted Intervention (MICCAI) society [38]. A bibliometric analysis of CNNs in medical imaging further underscored their impact, noting that ResNet and U-Net were among the most cited architectures, with brain-related research being the most prevalent disease topic [39]. This established CNN backbone provides the essential context for understanding the evolution towards more sophisticated, generative models like diffusion models for tasks such as motion artifact correction.

Residual-Guided Diffusion Models (Res-MoCoDiff): A Paradigm Shift

Core Principles and Problem Formulation

Denoising Diffusion Probabilistic Models (DDPMs) represent a class of generative models that have recently revolutionized image synthesis. Conventional DDPMs involve a forward diffusion process, where a Markov chain gradually corrupts an input image with Gaussian noise over hundreds or thousands of steps until it becomes pure noise. A reverse (denoising) process is then learned to reconstruct the original image from the noise [5] [35]. While powerful, this approach has a critical drawback for medical image correction: its reliance on a pure Gaussian prior and the need for many iterative steps make it computationally burdensome and potentially sub-optimal for ill-posed inverse problems like motion artifact correction (MoCo) [35].

Res-MoCoDiff introduces a fundamental shift in this paradigm. It is an efficient denoising diffusion probabilistic model specifically engineered for MRI motion artifact correction. Its core innovation lies in explicitly incorporating the residual error between the motion-free image x and the motion-corrupted image y (i.e., r = y - x) directly into the forward diffusion process [36] [5] [35]. This residual-guided mechanism allows the model to simulate noise evolution with a probability distribution that closely matches the actual corrupted data. Consequently, the reverse diffusion process, which reconstructs the clean image, can be dramatically accelerated, requiring only four steps instead of the hundreds or thousands needed in conventional DDPMs [36] [40]. This approach offers two significant advantages: 1) enhanced reconstruction fidelity by avoiding the restrictive and potentially unrealistic purely Gaussian prior, and 2) a substantial reduction in computational overhead, making clinical deployment feasible [35].

Architectural and Methodological Innovations

The Res-MoCoDiff framework incorporates several key technical innovations that contribute to its performance:

• Residual Error Shifting Mechanism: This novel noise scheduler uses the residual (r) to guide the diffusion process, ensuring a more precise transition between steps and a closer match to the distribution of motion-corrupted data [40].
• U-net with Swin Transformer Blocks: The model employs a U-net backbone, a common architecture in image-to-image tasks, but with a significant modification: the standard attention layers are replaced by Swin Transformer blocks. This replacement enhances the model's robustness and effectiveness across different image resolutions [36] [35].
• Combined L1 + L2 Loss Function: The training process integrates a hybrid loss function that sums the L1 (mean absolute error) and L2 (mean squared error) losses. This combination promotes image sharpness while simultaneously reducing pixel-level errors [5] [35].

The following diagram illustrates the core workflow and logical relationship of the Res-MoCoDiff diffusion process.

Experimental Protocols and Quantitative Performance

Evaluation Methodologies

The performance of Res-MoCoDiff was rigorously evaluated using established experimental protocols on two primary types of datasets [36] [5] [35]:

• In-silico Dataset: Generated using a realistic motion simulation framework to create brain MRI images with controlled, known levels of motion artifacts (minor, moderate, and heavy distortion).
• In-vivo Movement-Related Artifacts (MR-ART) Dataset: Comprised of clinical data with real motion artifacts, used to validate the model's performance in real-world conditions.

The model was compared against several established baseline and state-of-the-art methods, including:

• Cycle Generative Adversarial Network (CycleGAN)
• Pix2Pix
• A diffusion model with a vision transformer backbone (MT-DDPM)

Evaluation was conducted using standard quantitative metrics for image quality and fidelity:

• Peak Signal-to-Noise Ratio (PSNR): Measures the ratio between the maximum possible power of a signal and the power of corrupting noise. Higher is better.
• Structural Similarity Index Measure (SSIM): Assesses the perceptual similarity between two images. Higher is better (max 1.0).
• Normalized Mean Squared Error (NMSE): Quantifies the normalized average squared difference between pixels. Lower is better.

Performance Results and Analysis

The following table summarizes the key quantitative results demonstrating Res-MoCoDiff's superior performance in motion artifact correction.

Table 1: Quantitative Performance of Res-MoCoDiff vs. Established Methods [36] [35]

Method	Distortion Level	PSNR (dB) ↑	SSIM ↑	NMSE ↓
Res-MoCoDiff	Minor	41.91 ± 2.94	Highest	Lowest
Res-MoCoDiff	Moderate	Superior	Highest	Lowest
Res-MoCoDiff	Heavy	Superior	Highest	Lowest
CycleGAN	Various	Lower	Lower	Higher
Pix2Pix	Various	Lower	Lower	Higher
MT-DDPM (ViT backbone)	Various	Lower	Lower	Higher

Beyond correction quality, Res-MoCoDiff's efficiency is a critical advancement. The average sampling time was reduced to 0.37 seconds per batch of two image slices, compared with 101.74 seconds for conventional DDPM approaches [36] [35]. This represents a speed-up of over 270 times, making it practical for integration into fast-paced clinical workflows.

The Scientist's Toolkit: Essential Research Reagents and Materials

For researchers aiming to implement or build upon models like Res-MoCoDiff for brain image analysis, a specific set of computational "reagents" and tools is essential. The following table details these key components and their functions as derived from the cited research.

Table 2: Key Research Reagent Solutions for Deep Learning in Brain MRI

Research Reagent / Tool	Function & Explanation	Application in Res-MoCoDiff/CNNs
U-net Architecture	An encoder-decoder CNN architecture with skip connections, highly effective for image-to-image tasks like segmentation and restoration.	Serves as the foundational backbone for the diffusion model [35].
Swin Transformer Blocks	A type of transformer architecture that uses a shifted windows mechanism for efficient computation and cross-resolution feature modeling.	Replaces standard attention layers in the U-net to enhance robustness across resolutions [36] [35].
Synthetic (In-silico) Motion Datasets	Datasets generated by simulating physical corruption processes (e.g., motion in k-space) on clean medical images.	Used for initial training and controlled evaluation where ground truth is known [5] [35].
In-vivo Motion Artifact (MR-ART) Datasets	Datasets comprising clinical scans with real, naturally occurring motion artifacts.	Essential for validating model performance and generalizability in real-world scenarios [5].
Combined L1 + L2 Loss Function	A hybrid loss function that promotes image sharpness (L1) while minimizing large pixel-level errors (L2).	Used during model training to optimize the network parameters for high-fidelity output [5] [35].
Denoising Diffusion Probabilistic Model (DDPM) Framework	A generative framework that learns data distributions by progressively denoising from a noisy input.	The core generative framework upon which Res-MoCoDiff is built and optimized [35].

The deep learning revolution in brain MRI research is characterized by a continuous evolution from foundational CNNs to highly specialized generative models like Res-MoCoDiff. While CNNs provided the essential building blocks for automatic spatial feature learning, the residual-guided diffusion approach represents a significant leap forward for the specific and critical problem of motion artifact correction. By intelligently incorporating the spatial structure of artifacts through residual shifting and leveraging advanced architectures like the Swin Transformer, Res-MoCoDiff achieves a remarkable combination of high fidelity and computational efficiency. This progress directly addresses the core challenge of spatial artifact distribution in brain research, ensuring that corrected images retain the anatomical and functional details necessary for robust scientific discovery. As these tools become more refined and integrated into research pipelines, they hold the promise of unlocking more reliable and precise analyses of the brain's intricate architecture from MRI data.

In brain research, functional imaging techniques like functional Magnetic Resonance Imaging (fMRI) and functional Near-Infrared Spectroscopy (fNIRS) are indispensable for exploring neural activity and connectivity. However, a significant challenge confounds these measurements: motion artifacts. These artifacts, caused by subject movement, introduce signal changes that can mimic or obscure genuine neural activity, compromising data integrity [41] [42]. The spatial distribution of these artifacts across the brain is not random; recent studies suggest that the geometric layout of the cortex itself plays a role in the observable patterns of task-related activation and deactivation [43]. Consequently, correcting for motion artifacts is not merely a signal cleaning exercise but a crucial step for accurate spatial and functional analysis.

Classical signal processing techniques form the first line of defense against these artifacts. This whitepaper provides an in-depth technical guide to three pivotal methods: spline interpolation, wavelet analysis, and Kalman filtering. We frame their utility within a broader thesis on the spatial distribution of motion artifacts, providing researchers and drug development professionals with the experimental protocols and quantitative data needed to implement these corrections effectively.

Core Principles and Mathematical Foundations

Motion artifacts in neuroimaging data are typically characterized by their high amplitude and abrupt changes, which distinguish them from the slower, hemodynamic response of neural origin [41]. The fundamental objective of motion artifact correction is to isolate and remove these contaminating components while preserving the underlying physiological signals of interest.

The Spatial Context of Motion Artifacts

The spatial layout of the brain imposes constraints on functional activity. Research using spatial regression techniques like kriging has demonstrated that the spatial distribution of task-induced decreases in brain activity can predict the layout of regions showing activity increases, and vice versa [43]. This suggests that the cortex operates as a system with topographically determined, antagonistic relationships. Motion artifacts can disrupt these spatial relationships, leading to misinterpretations of functional specialization and connectivity. Therefore, effective motion correction is essential not just for cleaner time-series data, but for accurately mapping the brain's intrinsic spatial architecture.

Methodological Deep Dive

Spline Interpolation

The spline interpolation method, often implemented as the Motion Artifact Reduction Algorithm (MARA), operates on the principle that motion-corrupted segments can be identified and modeled independently [41] [42].

Workflow:

• Artifact Detection: A moving window calculates the standard deviation (MSD) of the signal. Segments where the MSD exceeds a predefined threshold (SDThresh) are flagged as containing motion artifacts [41] [42].
• Spline Modeling: A cubic spline is fitted to the identified artifact segments. This spline models the shape of the motion artifact.
• Artifact Removal: The fitted spline is subtracted from the original signal, effectively removing the artifact.
• Level Correction: To account for baseline shifts caused by the artifact, the mean value of the corrected segment is adjusted to align with adjacent non-corrupted segments [42].

Table 1: Key Parameters for Spline Interpolation (MARA)

Parameter	Typical Value	Function
SDThresh	20	Threshold for standard deviation change to detect artifacts [41]
AMPThresh	0.5	Threshold for amplitude change to detect artifacts [41]
tMotion	0.5 s	Duration of the moving window for detection [41]
tMask	2 s	Period following a detected artifact that is also masked [41]

Diagram 1: Spline interpolation workflow for motion artifact correction.

Wavelet Analysis

Wavelet filtering leverages time-frequency analysis to separate motion artifacts, which are typically transient and localized in time, from the physiological signal [41] [42].

Workflow:

• Signal Decomposition: The original signal ( y(t) ) is decomposed using a Discrete Wavelet Transform (DWT) into wavelet coefficients across multiple frequency scales. This is represented as: ( y(t) = \sumk c{j0,k} \phi{j0,k}(t) + \sumk \sum{j=j0}^{\infty} d{j,k} \psi{j,k}(t) ) where ( c{j0,k} ) are the approximation coefficients, ( d_{j,k} ) are the detail coefficients, and ( \phi ) and ( \psi ) are the scaling and wavelet functions, respectively [42].
• Coefficient Thresholding: The detail coefficients, which often contain the high-frequency motion artifacts, are analyzed. Those exceeding a statistically derived threshold are identified as artifact-dominated and set to zero.
• Signal Reconstruction: The inverse Discrete Wavelet Transform (iDWT) is applied to the thresholded coefficients to reconstruct the motion-corrected signal in the time domain [41] [42].

Table 2: Key Parameters for Wavelet Analysis

Parameter	Function
Wavelet Family	Choice of wavelet basis function (e.g., Daubechies). Defines the shape used for signal decomposition.
Decomposition Level	The number of frequency scales (levels) for the decomposition.
Thresholding Method	The algorithm for determining the threshold (e.g., universal, minimax).
Thresholding Rule	The rule for applying the threshold (e.g., hard, soft thresholding).

Diagram 2: Wavelet analysis process for motion artifact removal.

Kalman Filtering

Kalman filtering treats motion artifact correction as a state estimation problem. It uses a probabilistic model to predict the clean signal state and updates this prediction based on new measurements [42].

Workflow:

• Model Definition: The artifact-free fNIRS signal ( x(n) ) is modeled as an Autoregressive (AR) process: ( \phi(n) = A\phi(n-1) + \omegan ) where ( \phi(n) = [x(n) \dots x(n-p+1)]^T ) represents the state vector, ( A ) is the AR model coefficients, and ( \omegan ) is the process noise [42].
• State Prediction: The filter predicts the current state (clean signal) based on the previous state estimate.
• Measurement Update: The filter incorporates the new, motion-corrupted measurement ( y(n) ) to update its state prediction. The Kalman gain, which is dynamically computed, balances the trust between the model's prediction and the new measurement.
• Iteration: This predict-update cycle recurs for each time point, resulting in a real-time estimate of the clean signal. The measurement noise covariance is typically set to the variance of the entire data series, representing the motion artifacts [42].

Table 3: Key Parameters for Kalman Filtering

Parameter	Function
AR Model Order (p)	The number of previous samples used to predict the current state.
Process Noise Covariance	Estimates the uncertainty in the state transition model (from motion-free segments).
Measurement Noise Covariance	Estimates the uncertainty in the measurements (from the entire data series).

Diagram 3: Kalman filtering cycle for recursive state estimation.

Experimental Protocols and Performance Benchmarking

Standardized Evaluation Protocol

To objectively compare the efficacy of these algorithms, a consistent evaluation framework is critical. A common protocol involves using real resting-state fNIRS datasets known to contain motion artifacts. A simulated hemodynamic response function (HRF) is added to this data, creating a ground truth for neural activation [41]. The motion correction algorithms are then applied, and their performance is quantified by how accurately they recover the simulated HRF.

Key Metrics:

• Mean-Squared Error (MSE): Measures the accuracy of the recovered HRF shape and amplitude compared to the ground truth.
• Contrast-to-Noise Ratio (CNR): Quantifies the ability to detect the activation signal against the background noise.
• Receiver Operating Characteristic (ROC) Analysis: Evaluates the sensitivity and specificity of functional connectivity analysis after motion correction [42].

Quantitative Performance Comparison

Table 4: Performance Benchmarking of Classical Motion Correction Algorithms

Algorithm	Key Principle	Reported Performance (MSE Reduction)	Reported Performance (CNR Increase)	Strengths	Weaknesses
Spline Interpolation	Models and subtracts artifact segments.	55% (Highest average reduction) [41]	Not Specified	High accuracy in HRF recovery [41].	Sensitive to accurate artifact detection; requires offline processing [42].
Wavelet Analysis	Thresholds artifact-related coefficients in frequency domain.	Not Specified	39% (Highest average increase) [41]	Effective denoising for functional connectivity; preserves signal integrity [42].	Performance depends on wavelet choice and threshold tuning.
Kalman Filtering	Recursive state estimation using an AR model.	Significant reduction [41]	Significant increase [41]	Suitable for real-time application [42].	Requires estimation of noise covariances and AR model parameters [42].
PCA	Removes first few principal components with high variance.	Significant reduction [41]	Significant increase [41]	Simple and fast to implement.	Risk of removing physiological signal of interest along with artifacts.
TDDR	Repairs signal based on temporal derivative distribution.	N/A	N/A	Superior for functional connectivity and topology analysis [42].	Assumes non-motion fluctuations are normally distributed [42].

Summary of Findings: Systematic comparisons reveal that all major motion correction techniques yield a significant improvement in MSE and CNR compared to no correction or simple trial rejection [41]. Spline interpolation has been shown to produce the largest average reduction in MSE (55%), whereas wavelet analysis produces the highest average increase in CNR (39%) [41]. In the context of functional connectivity and network topology analysis, Temporal Derivative Distribution Repair (TDDR) and wavelet filtering have been identified as the most effective methods, demonstrating superior denoising ability and an enhanced capacity to recover original FC patterns [42].

The Scientist's Toolkit: Research Reagent Solutions

Table 5: Essential Tools and Software for Motion Artifact Correction Research

Tool / Reagent	Function / Description	Example Use in Field
HOMER2 Software Package	A widely used open-source platform for fNIRS data processing.	Contains implementations of hmrMotionArtifact for detection and several correction algorithms, including spline interpolation [41].
MATLAB with Signal Processing Toolbox	A high-level technical computing language with specialized toolboxes.	Provides built-in functions for wavelet decomposition (wavedec), Kalman filtering (kalman), and spline fitting, enabling custom algorithm development.
R with ggplot2	A statistical programming language with a powerful graphics package.	Used for generating publication-quality visualizations of time-series data before and after motion correction, as well as performance metrics [44].
fNIRS Data (Resting-State)	Real datasets containing motion artifacts, crucial for validation.	Serves as a realistic baseline for adding simulated hemodynamic responses to test correction algorithms [41] [42].
Accelerometer / Reference Signal	An external hardware sensor to measure head motion.	Provides an independent signal highly correlated with motion artifacts, which can be used as a reference in adaptive filtering techniques [41].

Spline interpolation, wavelet analysis, and Kalman filtering represent three robust, classical approaches for mitigating the confounding effects of motion in functional brain imaging. The choice of algorithm involves a trade-off between computational efficiency, required accuracy, and the specific analytical goal (e.g., activation detection vs. functional connectivity). Spline interpolation excels in accurate HRF recovery, wavelet filtering is particularly effective for preserving connectivity patterns, and Kalman filtering offers a pathway for real-time correction. As the field moves towards integrating these methods with modern deep-learning approaches [5] [13], their principles remain foundational. A rigorous, standardized application of these techniques is paramount for ensuring the validity of neuroscientific findings and the development of reliable biomarkers in drug development.

In brain magnetic resonance imaging (MRI) research, head motion systematically alters functional connectivity (FC), biasing results in brain-behavior association studies and introducing spurious findings [1]. The spatial distribution of these motion artifacts is non-random, preferentially reducing long-distance connections and increasing short-range connectivity, most notably within the default mode network [1]. This paper explores the critical challenge of mitigating these spatially systematic artifacts within the constraints of clinical and research workflows. We examine the innovative Res-MoCoDiff algorithm, a residual-guided denoising diffusion probabilistic model that achieves high-fidelity motion correction with a dramatically reduced number of sampling steps [5]. By framing this technical advancement within the broader context of neuroimaging research, we demonstrate how enhanced algorithmic efficiency not only improves image quality but also strengthens the validity of scientific inferences drawn from brain data.

Motion artifacts (ARTs) represent a fundamental confound in brain MRI, degrading image quality and hindering downstream clinical and research applications [5]. In functional MRI (fMRI), even sub-millimeter head movements systematically alter the B0 field and disrupt k-space readout lines, leading to ghosting and ringing artifacts that violate the core assumptions of data analysis [5] [1].

The impact on functional connectivity (FC) is particularly severe and spatially systematic. Motion induces a characteristic pattern of decreased long-distance connectivity and increased short-range connectivity, which can create false positives in brain-behavior association studies [1]. For instance, early studies concluding that autism decreases long-distance FC were later found to have their results confounded by increased head motion in the autistic participant groups [1]. This is especially problematic for studies of populations prone to greater movement, such as children, older adults, and individuals with psychiatric or neurological disorders [1].

Traditional mitigation strategies, including repeated acquisitions, motion tracking, navigator echoes, and k-space correction algorithms, impose significant workflow burdens and often require raw data not routinely stored in clinical archives [5]. While deep learning approaches have shown promise, many generative adversarial network (GAN)-based models suffer from mode collapse and unstable training [5]. Consequently, there is a pressing need for efficient, robust, and accurate correction methods that operate directly on reconstructed magnitude images, which are universally available and can be seamlessly integrated into existing analytical pipelines without vendor-specific modifications [5].

The Res-MoCoDiff Framework: Core Methodology

Res-MoCoDiff is an efficient denoising diffusion probabilistic model (DDPM) specifically designed for MRI motion artifact correction. Its core innovation lies in a novel residual error shifting mechanism that fundamentally rethinks the diffusion process for inverse problems [5].

Problem Formulation and Residual Guidance

In conventional DDPMs, the reverse diffusion process typically starts from pure Gaussian noise, x_N ~ N(0,I), and requires hundreds or thousands of iterative steps to reconstruct an image. This is computationally expensive and can be suboptimal for motion correction, as the Gaussian prior might encourage unrealistic reconstructions or hallucinations [5].

Res-MoCoDiff reformulates this problem by explicitly incorporating the residual error r = y - x between the motion-corrupted image y and the motion-free image x into the forward diffusion process. This allows the model to generate noisy images at step N with a probability distribution that closely matches the motion-corrupted data, specifically p(x_N) ~ N(x; y, γ²I) [5]. This approach offers a more informed starting point for the reverse process, enabling both higher reconstruction fidelity and drastic reductions in computational overhead.

Table 1: Key Architectural Components of Res-MoCoDiff

Component	Description	Function
Residual Error Shifting	Incorporates r = y - x into the forward diffusion process.	Aligns the prior distribution with the corrupted data, enabling a shorter reverse process [5].
U-net Backbone	Standard network architecture for diffusion models.	Serves as the primary model for iterative denoising [5].
Swin Transformer Blocks	Replaces standard attention layers in the U-net.	Enhances model robustness and performance across different image resolutions [5].
Combined L1+L2 Loss	A joint loss function used during training.	Promotes image sharpness while simultaneously reducing pixel-level errors [5].

Workflow and Efficiency

The following diagram illustrates the streamlined, four-step reverse diffusion process of Res-MoCoDiff.

Experimental Protocol and Validation

The performance of Res-MoCoDiff was rigorously evaluated using a multi-faceted experimental protocol [5]:

• Datasets: The model was trained and tested on both an in-silico dataset, generated using a realistic motion simulation framework, and an in-vivo dataset featuring movement-related artifacts.
• Distortion Levels: Evaluations covered minor, moderate, and heavy motion-induced distortion levels to test robustness across a range of artifact severity.
• Comparative Analysis: Res-MoCoDiff was benchmarked against established methods, including CycleGAN, Pix2Pix, and a diffusion model with a vision transformer backbone.
• Quantitative Metrics: Performance was assessed using standard image quality metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Normalized Mean Squared Error (NMSE).

Quantitative Performance and Comparative Analysis

Res-MoCoDiff demonstrated superior performance in removing motion artifacts across all levels of distortion. The quantitative results, summarized below, highlight its efficacy and efficiency.

Table 2: Quantitative Performance of Res-MoCoDiff vs. Established Methods

Method	Key Metric	Performance on Minor Distortions	Performance Across All Distortions	Computational Efficiency
Res-MoCoDiff	PSNR	Up to 41.91 ± 2.94 dB [5]	Consistently highest SSIM and lowest NMSE [5]	0.37 s per batch (2 slices) [5]
Conventional DDPM	PSNR	Lower than Res-MoCoDiff [5]	Lower SSIM, higher NMSE than Res-MoCoDiff [5]	101.74 s per batch [5]
CycleGAN / Pix2Pix	PSNR	Lower than Res-MoCoDiff [5]	Lower SSIM, higher NMSE than Res-MoCoDiff [5]	Not specified

The data shows that Res-MoCoDiff achieves a speed-up of over 270 times compared to a conventional DDPM approach while maintaining superior image restoration quality [5]. This balance of high fidelity and computational efficiency is critical for integration into time-sensitive clinical workflows.

The Scientist's Toolkit: Research Reagent Solutions

The following table details key computational tools and frameworks essential for research in this field.

Table 3: Essential Research Tools for Algorithmic Motion Correction

Research Reagent / Tool	Function / Description	Relevance to Motion Correction
U-net Architecture	A convolutional neural network architecture with a contracting and expansive path.	Commonly used as the backbone in diffusion models like Res-MoCoDiff for high-resolution image generation [5].
Swin Transformer	A transformer architecture with shifted windows for hierarchical feature maps.	Replaces standard attention layers to enhance robustness across resolutions in Res-MoCoDiff [5].
Proper Orthogonal Decomposition (POD)	A dimensionality reduction method that identifies a low-dimensional subspace.	A core component in non-intrusive Reduced Order Models (ROMs); foundational for handling high-dimensional data efficiently [45] [46].
Active Subspace Methodology (ASM)	A supervised input dimensionality reduction technique.	Identifies a low-dimensional linear subspace that best describes output variability, helping to combat the curse of dimensionality in complex models [45].
Adaptive Sampling Algorithms	Strategies for optimal placement of samples in a parameter space.	Improves the efficiency and accuracy of ROMs by focusing computational resources on regions of interest, reducing the total samples needed [47] [46].

Broader Context: Implications for Brain-Behavior Research

The development of efficient correction algorithms like Res-MoCoDiff has profound implications for the validity of brain-behavior research. Residual motion artifact, even after standard denoising pipelines, remains a significant source of spurious findings.

As demonstrated in large-scale studies like the Adolescent Brain Cognitive Development (ABCD) Study, head motion has a systematic and spatially defined impact on functional connectivity. After standard denoising, the motion-FC effect matrix shows a strong negative correlation (Spearman ρ = -0.58) with the average FC matrix, meaning participants who moved more showed systematically weaker connections overall [1]. This effect was often larger than the trait-FC effects under investigation [1]. One analysis found that after standard denoising, 42% (19/45) of behavioral traits showed significant motion-induced overestimation of trait-FC effects, while 38% (17/45) showed significant underestimation [1]. While stringent motion censoring can reduce overestimation, it does not fully address underestimation and can bias sample distributions by systematically excluding high-motion individuals [1].

The following diagram contextualizes the role of efficient correction algorithms within the broader workflow of neuroimaging research, highlighting how they help safeguard against spurious associations.

The pursuit of algorithmic efficiency in motion correction is not merely a technical exercise in computational optimization. It is a fundamental requirement for ensuring the scientific validity of neuroimaging research, particularly in studies of motion-correlated traits and populations. The Res-MoCoDiff framework exemplifies a significant leap forward, demonstrating that high-fidelity correction of spatially distributed motion artifacts can be achieved with a drastic reduction in sampling steps. By integrating such efficient, robust, and accurate correction tools into the analytical pipeline, researchers can better isolate true neural signals from motion-induced confounds, thereby strengthening the foundation for discoveries in brain function and their application in drug development and clinical neuroscience.

A Practical Framework for Identifying, Quantifying, and Mitigating Spatial Artifacts

Within the context of investigating the spatial distribution of motion artifacts in brain research, ensuring data integrity through rigorous quality control (QC) is paramount. Motion artifacts during magnetic resonance imaging (MRI) acquisition introduce spatially varying noise, blurring, and ghosting that can systematically bias the analysis of neural structures and their spatial relationships [31]. This technical guide provides an in-depth comparison of two predominant methodological paradigms for automated quality assessment: Automated Deep Learning Detection and Image Quality Metrics (IQMs). We detail their principles, experimental protocols, and performance to inform researchers and drug development professionals in selecting and implementing appropriate QC frameworks for their studies.

Automated Deep Learning Detection of Motion Artifacts

Deep learning (DL), particularly convolutional neural networks (CNNs), offers a powerful, data-driven approach for directly identifying motion artifacts from image content.

Core Principles and Architectures

These models are typically trained on large datasets of both clean and motion-corrupted images to learn the complex, often subtle, features associated with motion artifacts. The goal is to classify images or patches as artifact-free or corrupted, sometimes even rating the severity of motion [48] [31]. A common strategy involves using transfer learning, where a CNN pre-trained on a large natural image dataset (e.g., ImageNet) is fine-tuned on a smaller medical imaging dataset, effectively leveraging pre-learned features like edge and texture detectors [48] [31]. Another key technique is training on synthetic motion, where artifacts are digitally introduced into high-quality scans to create a large, labeled training set where the ground truth is known [31].

Table 1: Key Deep Learning Architectures for Motion Artifact Detection

Architecture	Description	Application Context	Key Findings
2D CNN [49] [50]	A 2D convolutional network for slice-wise classification.	Detection of rigid motion in T1-weighted brain images.	Achieved 85% precision, 80% recall on simulated data; 93% agreement with radiologists.
ShallowNet [48]	A shallower version of standard architectures (e.g., ResNet).	Fine-grained motion detection in image patches.	Achieved 90.4% average accuracy on patches, outperforming the original, deeper networks.
Ensemble ANN [48]	A classifier combining multiple shallower architectures.	Providing a final motion probability for a whole MRI volume.	Achieved 100% accuracy on the test set for volume-based classification.
Vision Transformer (ViT) [31]	Transformer-based architecture for image classification.	Detection of motion in clinical data warehouses.	Compared against other architectures like ResNet in a large-scale clinical validation.

Detailed Experimental Protocol

A typical protocol for developing a DL-based motion detector is outlined below and in Figure 1.

1. Data Preparation and Preprocessing:

• Source Data: Gather a large set of volumetric T1-weighted brain MRIs. These can come from public research cohorts (e.g., ADNI, UK Biobank) or clinical data warehouses [48] [31].
• Synthetic Motion Generation (for training): Two primary methods exist:
  • k-space simulation: A high-quality image is Fourier-transformed to k-space. The motion is simulated by introducing phase errors, applying translations or rotations, and removing lines of k-space data before reconstructing the corrupted image [31].
  • Image-based simulation: Motion artifacts like blurring or ghosting are applied directly to the magnitude image through convolutional operations [31].
• Annotation (for real data): For images with real motion, expert radiologists or technicians manually label the data using Likert scales (e.g., 1-5) for motion severity or binary labels (e.g., pass/fail) [51] [52] [31].
• Preprocessing: Skull-stripping is often performed using tools like BET (FSL). Images may be co-registered to a standard space [52].

2. Model Training and Fine-Tuning:

• Patch-based vs. Slice-based: The 3D volume is often processed as a series of 2D slices or patches from the three orthogonal axes (axial, coronal, sagittal) [48].
• Transfer Learning: A pre-trained CNN (e.g., ResNet) is fine-tuned on the MRI patches or slices. The lower, more general layers may be frozen, while the upper layers are trained to recognize motion-specific features [48].
• Clinical Fine-Tuning: A model pre-trained on research data with synthetic motion is further fine-tuned on a smaller set of labeled, real clinical data to bridge the domain gap and improve generalizability to heterogeneous hospital data [31].

3. Validation and Interpretation:

• Performance Metrics: Models are tested on held-out datasets using accuracy, precision, recall, and balanced accuracy. Performance is often compared against human rater agreement [49] [31].
• Explainability: Techniques like Grad-CAM are used to generate heatmaps highlighting the image regions that most influenced the model's decision, aiding in failure mode analysis and building trust [49] [50].

Figure 1: Workflow for developing a deep learning model for motion artifact detection, incorporating synthetic data and clinical fine-tuning.

Image Quality Metrics (IQMs) for Motion Assessment

IQMs are mathematical models that quantify image quality without necessarily understanding the image content. They are divided into reference-based and reference-free categories.

Taxonomy and Performance of IQMs

Table 2: Classification and Performance of Common Image Quality Metrics

Metric Type	Metric Name	Principle	Correlation with Radiological Evaluation
Reference-Based	SSIM [52]	Measures similarity in luminance, contrast, and structure between two images.	Strong correlation, but performance varies across studies.
	PSNR [52]	Ratio between the maximum possible signal power and corrupting noise power.	Strong correlation, but less sensitive to motion than other metrics.
	FSIM [52]	Uses phase congruency and gradient magnitude to evaluate feature similarity.	Strong correlation, reported to outperform SSIM.
	VIF [52]	Evaluates quality based on information fidelity for the human visual system.	Strong correlation, can indicate improvement over reference (>1).
	LPIPS [52]	Measures distance between features extracted by a pre-trained deep neural network.	Strong correlation.
Reference-Free	Average Edge Strength (AES) [51] [52]	Calculates the average gradient magnitude across detected edges.	Most promising reference-free metric, strong and consistent correlation.
	Tenengrad (TG) [52]	Averages gradient magnitudes across the image to assess sharpness.	Moderate correlation.
	Image Entropy (IE) [52]	Quantifies the randomness in the distribution of pixel intensities.	Lower values indicate higher quality; correlation varies.

Detailed Experimental Protocol for IQM Validation

The following protocol, based on Marchetto et al. (2025), outlines how to rigorously evaluate IQMs against expert radiological assessment [51] [53] [52].

1. Data Acquisition:

• Acquire a unique dataset with paired scans: one with no motion (serving as the reference) and one with intentional, real head motion (e.g., nodding, shaking) from the same subject [51] [52].
• Include different sequence types (e.g., 3D MP-RAGE, 2D TSE) to test metric robustness [52].

2. Pre-processing and Metric Calculation:

• Skull-stripping: Apply a brain extraction tool (e.g., BET from FSL) to the reference image to create a brain mask [52].
• Alignment: Co-register the motion-corrupted image to its reference using a rigid registration tool (e.g., FLIRT) [52].
• Normalization: Apply intensity normalization. Percentile normalization (e.g., to the 1st and 99th percentiles) within the brain mask has been shown to yield stronger correlations with human scores than min-max or no normalization [51] [52].
• Metric Calculation: Compute the various reference-based and reference-free IQMs. For reference-based metrics, the motion-free scan is the reference. Calculations should be restricted to the skull-stripped brain region [52].

3. Radiological Evaluation and Correlation:

• Expert Rating: Present the motion-corrupted images (anonymized and randomized) to radiologists and/or radiographers for scoring on a 1-5 Likert scale for overall image quality [52].
• Statistical Analysis: Calculate the Spearman's rank correlation coefficient between each IQM value and the average expert score across all subjects and sequences [51] [52].

Figure 2: Experimental workflow for validating Image Quality Metrics (IQMs) against radiological evaluation.

Table 3: Key Tools and Datasets for Motion Artifact Research

Tool / Resource	Type	Function in Research
Synthetic Motion Generators [31]	Software Algorithm	Creates large-scale training data for deep learning models by introducing realistic motion artifacts into clean images.
Clinical Data Warehouse (CDW) [31]	Dataset	Provides large, heterogeneous datasets of clinical MRIs for model validation and fine-tuning in real-world conditions.
Pre-trained CNNs (e.g., ResNet) [48]	Deep Learning Model	Serves as a base for transfer learning, reducing the need for vast annotated medical datasets.
Grad-CAM [49] [50]	Explainable AI Tool	Generates visual explanations for DL model decisions, crucial for debugging and clinical trust.
Average Edge Strength (AES) [51] [52]	Reference-free IQM	Provides a robust, automated quality score that strongly correlates with human perception of motion.
Skull-Stripping (BET) & Registration (FLIRT) [52]	Pre-processing Software	Essential pre-processing steps for accurate calculation and comparison of IQMs.

The choice between deep learning and IQM approaches is context-dependent, guided by the specific requirements of the research project, as summarized in Table 4.

Table 4: Comparative Analysis of Deep Learning and IQM Approaches

Characteristic	Deep Learning Detection	Image Quality Metrics
Principle	Data-driven, learns features directly from images.	Model-based, uses predefined mathematical models.
Data Needs	Requires large, labeled datasets for training.	No training required; validation needs expert-rated images.
Reference Image	Not required.	Required for reference-based metrics; not for reference-free.
Interpretability	Lower; requires XAI tools (e.g., Grad-CAM) for insight.	Higher; the calculation is transparent and based on known principles.
Generalizability	Can struggle with new scanner protocols or pathologies without fine-tuning.	Generally robust across domains, though performance may vary.
Primary Strength	High accuracy and ability to detect complex, subtle artifacts.	Simplicity, speed, and well-understood methodology.
Best Suited For	Large-scale, automated QC in heterogeneous datasets (e.g., CDWs).	Rapid, quantitative benchmarking of reconstruction or correction algorithms.

For research focused on the spatial distribution of motion artifacts, both approaches offer unique insights. DL models, especially when combined with explainable AI, can potentially learn to identify which brain regions are most affected by specific motion types, information that could be foundational for developing spatially-informed correction techniques. Conversely, IQMs like AES provide a global quality measure that could be correlated with spatial analyses to determine how overall motion severity impacts specific anatomical measurements.

In conclusion, Automated Deep Learning Detection and Image Quality Metrics are complementary tools in the modern neuroimaging QC pipeline. DL offers powerful, automated classification suited for large-scale clinical data, while IQMs provide a fast, interpretable, and quantitative measure of image integrity. The integration of both methods, leveraging their respective strengths, presents the most robust strategy for ensuring data quality in studies of brain morphology and the spatial characteristics of artifacts.

In neuroimaging, particularly in magnetic resonance imaging (MRI), motion artifacts (ARTs) represent a significant challenge, degrading image quality and hindering downstream analytical applications such as image segmentation and target tracking in MR-guided radiation therapy [5]. These artifacts primarily arise from rigid head motion, which can alter the B0 field, result in susceptibility artifacts, and disrupt k-space readout lines, potentially violating the Nyquist criterion and causing ghosting and ringing artifacts [5]. The spatial distribution of these motion-induced distortions is not random; it is influenced by the geometric principles of cortical organization and the nature of the movement itself. This whitepaper provides an in-depth technical guide to experimental design strategies—encompassing participant instruction, motion-constrained hardware, and sequence choice—framed within the broader thesis that the spatial layout of motion artifacts in the brain is a predictable, topographically determined phenomenon. Understanding this spatial distribution is paramount for developing robust correction algorithms and designing experiments that are inherently resilient to motion-related degradation.

The Spatial Distribution of Motion and Neural Activity

The brain's functional architecture adheres to fundamental geometric principles. Recent research utilizing kriging, a spatial regression technique from earth sciences, has demonstrated that the spatial distribution of regions showing task-induced decreases in neural activity can accurately predict the layout of regions showing activity increases, and vice versa [43]. This suggests that antagonistic relationships between brain networks are topographically determined. This principle can be extended to motion artifacts: the corruption caused by movement is not uniformly distributed across the image but is instead structured by the underlying spatial processes of the cortex. The spatial configuration of artifact-free data can, therefore, be used to predict and correct the spatial distribution of motion-corrupted data [43]. Advanced motion correction models, such as Res-MoCoDiff, exploit this spatial predictability by incorporating residual error information to guide the restoration process, effectively learning the topographic relationship between corrupted and clean images [5].

Figure 1: Spatial Relationships in Neural Activity and Motion Artifacts. This diagram illustrates how geometric principles of cortical organization underpin both the spatial distribution of antagonistic neural networks and motion artifacts, enabling predictive correction models.

Experimental Design Strategy 1: Participant Instruction

Effective participant instruction is the first line of defense against motion artifacts. The goal is to minimize voluntary movement and manage participant anxiety, which can manifest as involuntary motion.

Detailed Methodology for Participant Instruction

• Pre-Scanning Session Briefing: Prior to the scan, provide a comprehensive, layperson-accessible explanation of the importance of remaining still. Use analogies, such as comparing head movement to a camera shake in a photograph, to make the concept tangible. For pediatric populations or studies involving participants with neurodevelopmental disorders, this briefing should be tailored and may involve visual aids or practice sessions in a mock scanner [54].
• Reinforcement During Scan: Implement a real-time communication protocol. Provide participants with an emergency stop button to alleviate anxiety and use the intercom between sequences to offer positive reinforcement and gentle reminders to maintain head position.
• Instruction Scripting for Task-Based fMRI: For experiments involving cognitive tasks, instructions must be meticulously scripted and practiced beforehand. This minimizes confusion-induced movement and ensures that task-related motions (e.g., button presses) are isolated and minimal. Studies on human-robot interaction highlight that participant behavior is highly variable; standardized instructions are crucial for reducing this uncertainty and its associated motion [55].

Experimental Design Strategy 2: Motion-Constrained Hardware

Motion-constrained hardware physically limits the participant's ability to move, directly addressing the source of motion artifacts.

Detailed Methodology for Hardware Setup

• Head Stabilization System: Utilize a combination of foam padding, vacuum cushions, and a head coil locking mechanism. The objective is to create a comfortable yet firm restraint that minimizes degrees of freedom for head movement without causing discomfort, which could itself lead to motion.
• Bite Bar Systems (for high-precision studies): For studies requiring extreme stability, a custom-molded dental bite bar can be used. This apparatus, attached to the head coil, mechanically couples the participant's head to the scanner, drastically reducing translational and rotational movement. Its use is typically reserved for specialized research due to participant burden.
• Environmental Considerations: The dataset on sequential human motions highlights the impact of the external environment, such as occlusions from moving equipment, on participant behavior and data quality [55]. While not a direct constraint on the participant, optimizing the scanner environment to be less claustrophobic and ensuring clear sightlines can reduce anxiety-induced motion.

Research Reagent Solutions: Essential Materials

Table 1: Key Research Reagent Solutions for Motion-Constrained Hardware

Item	Function	Specification Notes
Vacuum Cushion	Conforms to the head and neck shape, then is evacuated to become rigid, providing custom immobilization.	Ensure compatibility with the RF head coil and availability of multiple sizes.
MRI-Compatible Foam Padding	Fills gaps between the head, coil, and cushion to prevent minor shifts.	Hypoallergenic, firm density foam is recommended.
Head Coil with Locking Mechanism	The RF receiver coil that must be securely fixed to the scanner bed to provide a stable reference frame.	Modern designs often incorporate integrated padding and comfort features.
Mirror and Display System	Allows participant to see outside the bore, reducing claustrophobia and anxiety-related movement.	Critical for task-based fMRI to present visual stimuli.

Experimental Design Strategy 3: Sequence Choice

The choice of imaging sequence can determine a dataset's inherent vulnerability to motion. Prospective versus retrospective correction strategies represent a fundamental trade-off between workflow burden and corrective power.

Detailed Methodology for Sequence Selection

• Prospective Motion Correction (PMC): This approach actively adjusts the imaging sequence in real-time to account for measured head motion.
  • Implementation: An external tracking system (e.g., camera-based with fiducial markers) or a navigator echo sequence is used to continuously monitor head position. This tracking data is fed back to the scanner, which updates the gradient fields and RF pulses to maintain a consistent frame of reference with the moving head [5].
  • Protocol: Integrate a camera-based tracking system (e.g., Moiré Phase Tracking, Optical Tracking). Calibrate the system to the scanner's coordinate system before the scan. The sequence must be designed to accept and apply positional updates with minimal latency.
• Retrospective Motion Correction: This approach applies corrections after data acquisition, during image reconstruction.
  • Implementation: This includes algorithms that operate on the raw k-space data (e.g., phase shift correction, motion trajectory estimation) or on the reconstructed image domain. Deep learning-based methods, such as Res-MoCoDiff, fall into this category [5].
  • Protocol: Acquire standard anatomical (e.g., T1-weighted, T2-weighted) or functional sequences. For k-space methods, ensure raw k-space data is saved, which is not routinely stored in clinical archives and can be vendor-specific [5]. For image-based methods, use magnitude images, which are universally available.

Advanced Sequence Choice: Deep Learning-Based Correction

The Res-MoCoDiff model represents a state-of-the-art retrospective, image-domain approach. Its methodology is as follows [5]:

• Objective: To recover an unknown motion-free image (x) from a motion-corrupted image (y), where y = A(x) + n, and A is an unknown motion corruption operator and n is additive noise.
• Approach: A denoising diffusion probabilistic model (DDPM) that uses a novel residual error shifting mechanism (r = y - x). This residual is incorporated into the forward diffusion process, ensuring the noisy data at step N has a distribution closely matching the motion-corrupted data. This allows the reverse diffusion process to reconstruct the image in only four steps, drastically reducing inference time from 101.74 seconds to 0.37 seconds per batch [5].
• Network Architecture: Employs a U-net backbone where attention layers are replaced by Swin Transformer blocks to enhance robustness across resolutions. The training uses a combined ℓ1+ℓ2 loss function to promote image sharpness and reduce pixel-level errors [5].

Figure 2: Motion Correction Sequence Decision Workflow. This flowchart guides the selection between prospective and retrospective motion correction strategies, highlighting the advanced Res-MoCoDiff path.

Quantitative Data Synthesis

The performance of motion correction strategies can be quantitatively evaluated using standardized image quality metrics. The following tables synthesize quantitative data from the evaluation of the Res-MoCoDiff model against established methods [5].

Table 2: Quantitative Performance Comparison of Motion Correction Methods

Method	Peak Signal-to-Noise Ratio (PSNR)	Structural Similarity Index (SSIM)	Normalized Mean Squared Error (NMSE)
Res-MoCoDiff (Proposed)	Up to 41.91 ± 2.94 dB (for minor distortions)	Consistently Highest	Consistently Lowest
CycleGAN	Lower than Res-MoCoDiff	Lower than Res-MoCoDiff	Higher than Res-MoCoDiff
Pix2Pix	Lower than Res-MoCoDiff	Lower than Res-MoCoDiff	Higher than Res-MoCoDiff
Diffusion Model with ViT	Lower than Res-MoCoDiff	Lower than Res-MoCoDiff	Higher than Res-MoCoDiff

Table 3: Computational Efficiency of Res-MoCoDiff Versus Conventional DDPM

Model	Number of Diffusion Steps	Average Sampling Time
Res-MoCoDiff	4	0.37 seconds (per batch of 2 slices)
Conventional DDPM	Hundreds to Thousands	101.74 seconds

A Synthesized Experimental Protocol

For a robust fMRI study on the spatial distribution of cognitive processes, the following integrated protocol is recommended:

• Participant Preparation: Conduct a detailed pre-scan briefing and practice task in a mock scanner environment. Screen for claustrophobia and anxiety.
• Hardware Setup: Position the participant in the scanner using a vacuum cushion and foam padding for optimal, comfortable immobilization. Calibrate the prospective motion tracking camera if available.
• Sequence Acquisition:
  • Acquire structural and functional scans with a prospective motion correction sequence enabled, if available.
  • If prospective correction is not available, acquire standard sequences, ensuring raw k-space data or magnitude images are saved for retrospective correction.
• Post-Processing: Apply the Res-MoCoDiff model or a similar advanced, image-based deep learning algorithm to the magnitude images for final artifact removal [5].

The spatial distribution of motion artifacts in brain research is not a random nuisance but a phenomenon governed by the underlying geometry of the cortex. This whitepaper has detailed three pillars of experimental design to combat this issue: comprehensive participant instruction to minimize voluntary motion, sophisticated motion-constrained hardware to physically restrict movement, and intelligent sequence choice that leverages either real-time prospective correction or highly efficient retrospective deep learning models like Res-MoCoDiff. By integrating these strategies, researchers can preserve the integrity of the brain's spatial activity patterns, ensuring that the observed neural dynamics are a true reflection of cognitive function and not topographic artifacts of movement. This holistic approach is essential for advancing the precision and reliability of neuroimaging in both basic neuroscience and applied drug development.

Motion artifacts represent one of the most significant methodological challenges in neuroimaging, systematically biasing both morphometric and functional connectivity estimates. This technical guide examines the spatial distribution of motion artifacts in brain research, detailing how in-scanner movement induces spurious signal fluctuations that can confound statistical inferences about brain structure and function. Particularly problematic is the fact that motion correlates with key demographic and clinical variables—children, older adults, and psychiatric populations tend to move more—creating systematic bias rather than random noise. This whitepaper synthesizes current research on motion artifact characteristics, presents quantitative data on their impact, and provides detailed methodologies for mitigation, with special emphasis on implications for multisite studies and drug development trials.

Motion artifacts have been problematic since the introduction of magnetic resonance imaging (MRI) as a clinical and research modality [2]. Unlike other imaging modalities, MRI is particularly sensitive to subject motion due to the prolonged time required to collect sufficient data for image formation—far longer than the timescale of most physiological motions [2]. The problem persists despite technological advancements; while improved hardware enables faster imaging, the push for higher resolution and signal-to-noise ratio has maintained sensitivity to motion. In functional connectivity MRI (fc-MRI), the recognition of motion's confounding effects came somewhat late, with three seminal studies in 2012 systematically demonstrating that motion artifacts can markedly alter fc-MRI measures [9].

The clinical and research implications are substantial. Motion affects up to a third of clinical MRI sequences, with approximately 20% of MRI studies requiring scan repetition due to motion corruption [56]. This increases scan times, reduces patient comfort, and causes additional costs. In research settings, motion artifact has been shown to inflate effect sizes systematically and potentially invalidate findings, particularly in developmental studies, clinical populations, and investigations of individual differences [57] [58].

Spatial Characteristics of Motion Artifacts

Spatial Distribution of Head Motion

Head motion during MRI acquisition follows predictable spatial patterns based on biomechanical constraints. Motion is minimal near the atlas vertebrae (where the skull attaches to the neck) and increases with distance from this anchor point [9]. Frontal cortex shows particularly high motion, likely due to the preponderance of y-axis rotation associated with nodding movements [9]. Despite this spatial heterogeneity, voxel-specific measures of motion remain highly correlated with global measures (r = 0.89), indicating that motion produces widespread effects [9].

Table 1: Spatial Patterns of Motion Artifacts in Functional Connectivity

Artifact Type	Spatial Pattern	Effect on Connectivity	Primary Mitigation Strategies
Type 1: Local Effects	Homogeneous signal changes in proximal voxels	Spuriously inflated correlations among nearby regions	ICA, PCA, censoring approaches
Type 2: Global Effects	Widespread, homogeneous signal fluctuations throughout brain	Induces widespread inflation of correlations	Global signal regression (GSR)
Type 3: Heterogeneous Effects	Spatially varied signal fluctuations across brain	Disrupts correlations, particularly between distal regions	Censoring approaches, spike regression

Distance-Dependent Bias in Functional Connectivity

Motion artifacts exhibit a marked distance-dependent effect on functional connectivity estimates. Motion typically inflates short-distance connections while weakening long-distance connections [57] [58]. This pattern creates particular problems for developmental neuroimaging, as it mimics genuine maturational processes—originally reported as strengthening of long-range connections and weakening of short-range connections during development [58]. When motion is rigorously controlled, the prevalence of age-related change and the strength of distance-related effects are substantially reduced, though significant age effects remain [58].

Figure 1: Spatial Taxonomy of Motion Artifacts and Their Effects on Connectivity Estimates

Quantifying Motion and Its Impacts

Motion Quantification Metrics

In-scanner motion is typically estimated from the functional time series itself during preprocessing. Each volume is rigidly realigned to a reference volume, producing six realignment parameters (3 translations, 3 rotations) that describe how much a given volume must be moved to achieve alignment [9]. These parameters are often summarized as Framewise Displacement (FD), which provides a concise index of volume-to-volume motion [57] [9].

Table 2: Quantitative Metrics for Motion Assessment in fMRI

Metric	Calculation	Interpretation	Implementation
Framewise Displacement (FD)	Root mean square of translational and rotational realignment parameters	Estimates head movement between consecutive volumes	FSL: fsl_motion_outliers; XCP: fd.R
DVARS	Temporal derivative of root mean square intensity	Measures frame-to-frame signal change across brain	FSL: fsl_motion_outliers; XCP: dvars
Outlier Count	Number of voxelwise outlier values within each frame	Identifies volumes with extreme signal abnormalities	AFNI: 3dToutcount
FD-DVARS Correlation	Correlation between FD and DVARS	Quantifies relationship between movement and signal fluctuations	XCP: featureCorrelation.R
Voxelwise Displacement	Estimate of each voxel's movement between frames	Provides spatially specific motion estimates	Custom implementations

Different FD calculations exist and, while highly correlated, may scale differently. For example, FD as calculated by Power et al. is approximately twice the magnitude of FD provided by Jenkinson et al. [9]. Work by Yan et al. has shown that the matrix root mean squared formulation derived by Jenkinson et al. aligns best with voxel-specific measures of displacement [9].

Quantitative Impact on Functional Connectivity

Motion artifacts have demonstrable quantitative effects on functional connectivity measures. In a large developmental sample (n=780, ages 8-22), motion artifact inflated both overall estimates of age-related change and specific distance-related changes in connectivity [58]. When motion was rigorously controlled, the distribution of absolute t-statistics for age effects was significantly reduced, with the 99th percentile decreasing from t=10.2 to t=7.1 after improved preprocessing and group-level motion covariation [58].

The impact extends to network-level analyses as well. Motion tends to obscure age-related changes in connectivity associated with segregation of functional brain modules; improved preprocessing techniques increase sensitivity to detect increased within-module connectivity occurring with development [58]. Despite these confounding effects, functional connectivity remains a valuable phenotype when motion is adequately addressed—multivariate patterns of connectivity can still accurately predict subject age even after controlling for motion [58].

Methodological Strategies for Motion Mitigation

Confound Regression Approaches

Confound regression aims to separate signal from noise by modeling artefactual processes as time series, which might include frame-to-frame estimates of head movement or signals from noise-prone tissue compartments like ventricles [57]. These "artefact" time series are fit to the observed BOLD time series using a general linear model, with residuals representing "cleaned" data [57].

Benchmarking studies have yielded remarkably convergent results regarding optimal denoising approaches [57]. Models relying exclusively on frame-to-frame head movement estimates fail to correct adequately for motion artifact, while models incorporating global signal regression (GSR) and/or time series from signal decomposition techniques (PCA or ICA) show markedly improved performance [57]. These can be augmented with temporal censoring operations (spike regression, scrubbing) that specifically remove volumes corrupted by artefact [57].

Figure 2: Comprehensive Motion Mitigation Pipeline for fMRI Data

Emerging Deep Learning Approaches

Recent advances in deep learning have produced promising approaches for motion correction. Res-MoCoDiff is an efficient denoising diffusion probabilistic model that leverages residual information to correct motion artifacts with only four diffusion steps, substantially accelerating reconstruction times compared to traditional DDPMs [5]. This method exploits a novel residual error shifting mechanism during forward diffusion to incorporate information from motion-corrupted images, enabling a reverse diffusion process that requires minimal steps [5].

The JDAC framework takes a different approach, jointly performing image denoising and motion artifact correction through iterative learning [13]. This method employs an adaptive denoising model with noise level estimation based on gradient map variance, combined with an anti-artifact model that uses a gradient-based loss to maintain brain anatomical integrity [13]. Experimental results demonstrate effectiveness on motion-affected MRIs with severe noise, outperforming methods that handle denoising and artifact correction separately [13].

Data Augmentation Strategies

For AI models applied to MRI data, data augmentation strategies can enhance robustness to motion artifacts. A study evaluating lower limb segmentation found that models trained with standard nnU-net augmentations maintained better segmentation quality under increasing artifact severity compared to baseline (DSCdefault = 0.72±0.22 vs. DSCbaseline = 0.58±0.22 for severe artifacts) [56]. MRI-specific augmentations offered minimal additional benefit over general-purpose augmentations [56].

Table 3: Research Reagent Solutions for Motion Correction

Tool/Category	Specific Examples	Function/Purpose	Implementation Considerations
Software Pipelines	XCP Engine, FSL, AFNI, ANTs	Implements denoising and diagnostic procedures	XCP combines FSL, AFNI, ANTs with new pipeline software
Motion Estimation Tools	FSL: mcflirt, fsl_motion_outliers; AFNI: 3dVolreg	Calculate motion parameters and identify outlier volumes	Different FD calculations (Power vs. Jenkinson) scale differently
Confound Regression Tools	FSL: fsl_glm; AFNI: 3dTfitter	Fit confound models to BOLD time series	Model efficacy depends on included regressors (movement parameters, physiological signals, etc.)
Deep Learning Models	Res-MoCoDiff, JDAC framework	Correct motion artifacts using advanced neural networks	Res-MoCoDiff uses residual-guided diffusion; JDAC employs iterative denoising and artifact correction
Quality Assessment Tools	FSL: fsl_motion_outliers; XCP: fd.R, dvars	Quantify motion and its effects on data quality	FD-DVARS correlation indexes relationship between movement and signal fluctuations

Experimental Protocols for Motion Correction

High-Performance Denoising Protocol for Functional Connectivity

This protocol implements a validated, high-performance denoising strategy combining physiological signals, motion estimates, and mathematical expansions to target both widespread and focal effects of subject movement [57]:

• Data Preparation: Begin with minimally preprocessed fMRI data that has undergone slice-time correction, motion realignment, and normalization to standard space.

• Confound Model Specification: Construct a comprehensive confound model including:

  • 6 rigid-body motion parameters (3 translation, 3 rotation)
  • Their temporal derivatives (6 additional parameters)
  • Quadratic terms of above parameters (24 total additional terms)
  • Physiological signals (heart rate, respiration when available)
  • Average signals from noise-prone regions (white matter, CSF)

• Global Signal Regression: Include global signal as a confound when scientifically appropriate, particularly effective for mitigating Type 2 motion artifacts [57].

• Signal Decomposition: Apply data-driven decomposition (ICA or PCA) to identify and remove noise components. A common approach uses aCompCor to extract noise components from white matter and CSF [57].

• Temporal Censoring: Identify and scrub motion-corrupted volumes using Framewise Displacement (FD) and DVARS thresholds. Common thresholds include FD > 0.2-0.5 mm and DVARS > 2-3% [57].

• Residual Processing: Apply high-pass filtering (typically 0.008-0.01 Hz) to remove slow drifts after denoising.

This protocol requires 40 minutes to 4 hours of computing per image depending on model specifications and data dimensionality [57].

Deep Learning Motion Correction Protocol

For implementing Res-MoCoDiff motion correction [5]:

• Data Preparation: Gather motion-free and motion-corrupted image pairs for training. For clinical data without paired motion-free images, simulate motion using realistic motion simulation frameworks.

• Forward Diffusion Process: Apply novel residual error shifting mechanism to incorporate information from motion-corrupted images during forward diffusion, enabling noisy images at step N with probability distribution matching motion-corrupted data.

• Model Architecture: Implement U-Net backbone with attention layers replaced by Swin Transformer blocks to enhance robustness across resolutions.

• Training Procedure: Utilize combined loss function (ℓ1+ℓ2) to promote image sharpness and reduce pixel-level errors.

• Reverse Diffusion: Execute efficient 4-step reverse diffusion process for rapid inference (0.37 s per batch of two image slices compared to 101.74 s for conventional approaches).

Validation metrics should include PSNR (up to 41.91±2.94 dB for minor distortions), SSIM, and NMSE across minor, moderate, and heavy distortion levels [5].

Motion artifacts present a complex challenge that requires multifaceted solutions tailored to specific research contexts and imaging modalities. The spatial distribution of motion artifacts follows predictable patterns that systematically bias morphometric and functional connectivity estimates, particularly problematic when motion correlates with variables of interest. While no single solution eliminates motion artifacts in all situations, current methodologies—from rigorous confound regression to emerging deep learning approaches—can substantially mitigate these effects when appropriately implemented. For drug development professionals and clinical researchers, implementing comprehensive motion mitigation strategies is essential for ensuring the validity of neuroimaging biomarkers and treatment effects. The protocols and tools detailed in this whitepaper provide a foundation for robust motion correction across diverse research applications.

In functional magnetic resonance imaging (fMRI), the blood oxygen level-dependent (BOLD) signal reflects both neuronal activations and confounding physiological fluctuations [59]. Motion artifacts are now recognized as a major methodological challenge for studies of functional connectivity, with the potential to introduce systematic bias, particularly when in-scanner motion correlates with variables of interest such as clinical status or age [9]. The spatial distribution of motion artifacts is not uniform across the brain; artifacts exhibit distinct patterns, with signal drops affecting entire brain parenchyma while areas at the brain's edge show signal increases due to partial volume effects [9]. Understanding these spatial characteristics is fundamental to selecting and optimizing denoising pipelines for specific research objectives in studies examining the spatial distribution of motion artifacts.

The pervasive impact of head motion on data quality is particularly pronounced in clinical populations. Patients with brain pathologies, including brain tumors, stroke, and various neurological disorders, exhibit significantly greater head motion than healthy controls [60]. This vulnerability underscores the critical need for robust, tailored denoising strategies to ensure accurate and reliable analysis of fcMRI data, especially when investigating spatially distributed artifacts [60].

Spatial and Spectral Characteristics of Motion Artifacts

Spatial Distribution of Motion Artifacts

Motion in the scanner follows biomechanical constraints of the neck, resulting in minimal movement near the atlas vertebrae (where the skull attaches to the neck) and increasing with distance from the atlas [9]. Frontal cortex often experiences high motion due to the preponderance of y-axis rotation associated with nodding movements [9]. Despite this spatial heterogeneity in motion itself, the artifacts resulting from motion produce global signal changes, including a substantial drop in signal intensity across the entire brain parenchyma immediately following movement events [9].

Areas at the edge of the brain demonstrate different artifact characteristics, showing large signal increases likely due to partial volume effects [9]. Similar partial volume effects occur at tissue class boundaries throughout the brain. These spatially distinct artifact patterns must be considered when evaluating denoising pipeline efficacy for specific research goals focused on spatial distribution.

Spectral Properties of Physiological Noise

Physiological processes contaminate BOLD signals with distinct spectral signatures. Using ultrafast fMRI (TR = 400 ms), researchers have spectrally separated three primary physiological noise sources [59]:

• Cardiac pulsation (0.8-1.0 Hz) predominantly affects the base of the brain, where high density of arteries exists
• Respiration (0.3-0.4 Hz) influences prefrontal and occipital areas, suggesting motion associated with breathing contributes to noise
• Physiological low-frequency oscillations (pLFOs) (<0.2 Hz) dominate many prominent independent components and represent the most influential physiological noise source in BOLD fMRI

With typical TR values (≥2 seconds), high-frequency physiological signals alias into the low-frequency band, complicating their identification and removal [59]. This spectral overlapping is particularly problematic for studies of spatial distribution, as different physiological processes affect distinct brain regions.

Table 1: Spatial Distribution of Physiological Noise Sources in BOLD fMRI

Physiological Process	Frequency Range	Primary Spatial Distribution	Probable Mechanism
Cardiac pulsation	0.8-1.0 Hz	Base of the brain	Proximity to major arteries
Respiration	0.3-0.4 Hz	Prefrontal and occipital areas	Motion associated with breathing
Physiological LFOs	<0.2 Hz	Widespread, network-specific	Systemic signals traveling with blood

Denoising Methodologies and Experimental Protocols

Common Denoising Strategies

Multiple denoising approaches have been developed to address motion-related and physiological artifacts in fMRI data:

• Independent Component Analysis-based Automatic Removal of Motion Artifacts (ICA-AROMA): Data-driven approach identifying motion-related components for removal [60]
• Anatomical Component Correction (CompCor): Regression of noise components derived from principal component analysis of signals from white matter and cerebrospinal fluid regions [60]
• Head Motion Parameter (HMP) regression: Inclusion of 6 or 24 rigid-body parameters derived from volume realignment as nuisance regressors [60]
• Spike Regression (SpikeReg): Identification and correction of extreme motion outliers [60]
• Scrubbing (Scr): Censoring of high-motion volumes identified by framewise displacement metrics [60]
• Global Signal (GS) Regression: Removal of whole-brain average signal, though this approach remains controversial [9]

Experimental Protocol for Pipeline Evaluation

To systematically evaluate denoising pipeline performance, researchers should implement a standardized assessment protocol incorporating multiple quality metrics:

• Data Acquisition: Acquire resting-state fMRI data using parameters appropriate for the research question. Multiband sequences with short TR (<1 second) enable better separation of physiological signals [59]

• Pipeline Application: Implement multiple denoising pipelines in parallel for comparison. Include strategies such as those listed in Table 2

• Quality Metric Calculation: Compute multiple quality control metrics for each pipeline:

  • QC-FC correlations: Correlations between head motion and functional connectivity measures [60]
  • tDOF-loss: Loss in temporal degrees of freedom due to denoising [60]
  • RSN identifiability: Ability to identify known resting-state networks [60]
  • RSN reproducibility: Consistency of network identification across datasets [60]

• Performance Evaluation: Use a summary performance index accounting for both noise removal and biological signal preservation [61]

Table 2: Denoising Strategies and Their Combinations for Systematic Evaluation

Pipeline	Denoising Strategies	Primary Applications
1	6HMP	Basic motion correction
2	24HMP	Extended motion parameter regression
3	SpikeReg	Extreme outlier correction
4	Scr	High-motion volume removal
5	ICA	Automatic motion component removal
6	CC	Physiological noise from WM/CSF
7	ICA + 24HMP	Combined motion and physiological noise
8	ICA + SpikeReg	Motion components with spike correction
9	ICA + Scr	Motion components with scrubbing
10	CC + 24HMP	Physiological noise with motion parameters
11	CC + SpikeReg	Physiological noise with spike correction
12	CC + Scr	Physiological noise with scrubbing
13	ICA + SpikeReg + 24HMP	Comprehensive motion correction
14	ICA + Scr + 24HMP	Motion components with scrubbing and parameters
15	CC + SpikeReg + 24HMP	Physiological and motion correction
16	CC + Scr + 24HMP	Physiological noise with scrubbing and parameters

Pipeline Selection for Specific Research Contexts

Clinical Population Considerations

Denoising pipeline effectiveness varies significantly across clinical populations. A key finding from recent research indicates that at comparable levels of head motion, optimal denoising strategy varies depending on the nature of the brain disease [60]:

• Non-lesional encephalopathic conditions: Combinations involving ICA-AROMA denoising strategies were most effective [60]
• Lesional conditions (glioma, meningioma): Combinations including anatomical component correction (CC) yielded the best results [60]

This population-specific efficacy underscores the importance of tailoring denoising approaches to the particular subject group being studied, especially when investigating spatial distributions of artifacts that may interact with pathological brain changes.

Multi-Metric Performance Evaluation

A comprehensive multi-metric approach comparing denoising techniques for resting-state fMRI reveals performance heterogeneity across different quality metrics [61]. Studies favor denoising strategies including regression of mean signals from white matter and cerebrospinal fluid brain areas along with global signal regression as providing the best compromise between artifact removal and preservation of resting-state network information [61].

When implementing such evaluation frameworks, researchers should consider:

• Metric selection incorporating both artifact removal and signal preservation measures
• Summary performance indices that balance competing quality objectives
• Dataset-specific optimization as optimal pipelines may vary across acquisition parameters and subject populations

Diagram 1: Pipeline selection workflow for spatial artifact research

Implementation Framework and Research Reagents

Essential Research Reagents and Tools

Table 3: Essential Research Reagents and Computational Tools for Denoising Research

Tool/Reagent	Function	Implementation Considerations
FSL (FMRIB Software Library)	Comprehensive fMRI analysis suite including MELODIC for ICA	Standard for ICA-AROMA implementation; provides robust motion parameter estimation
HALFpipe Software	Unified framework for comparing multiple denoising pipelines	Facilitates standardized comparison of pipeline performance
MATLAB with SPM	Flexible scripting environment for customized pipeline development	Enables implementation of novel denoising strategies
MRIQC	Automated image quality metrics extraction	Provides quantitative quality assessment for pipeline evaluation
MR-ART Dataset	Matched motion-corrupted and clean structural MRI scans	Enables validation of denoising approaches with ground truth reference
Ultrafast fMRI Sequences (Multiband)	High-temporal resolution acquisition	Enables spectral separation of physiological noise sources

Practical Implementation Considerations

When implementing denoising pipelines for spatial distribution research, several practical factors require attention:

• Temporal Degrees of Freedom (tDOF) Preservation: Aggressive denoising strategies substantially reduce tDOF, potentially impacting statistical power. Strategies incorporating scrubbing particularly affect tDOF through volume censoring [60]

• Spatial Specificity Considerations: Global denoising approaches (e.g., global signal regression) may remove biologically relevant signals along with artifacts, potentially altering spatial patterns of connectivity [9]

• Population-Specific Optimization: As optimal denoising strategies differ between lesional and non-lesional conditions [60], researchers should validate pipeline performance on representative datasets

• Metric Complementarity: No single quality metric fully captures pipeline performance; a multi-metric approach is essential for comprehensive evaluation [61]

Diagram 2: Comprehensive denoising pipeline with evaluation metrics

Optimizing denoising pipelines for research on spatial distribution of motion artifacts requires careful consideration of research goals, subject populations, and analytical priorities. No single pipeline universally outperforms others across all contexts and metrics. The most effective approach involves systematic evaluation of multiple strategies using complementary quality metrics tailored to the specific research question.

Future developments in denoising will likely include more sophisticated population-specific pipelines, deep learning approaches that adapt to individual data characteristics, and integrated acquisition-denoisng frameworks that optimize both data collection and processing. As the field moves toward greater standardization, multi-metric evaluation frameworks will be essential for validating new methodologies and ensuring reproducible research findings in studies of spatial artifact distributions.

Benchmarks and Standards: Evaluating Correction Efficacy with Paired Datasets and Metrics

The spatial distribution of motion artifacts in brain MRI is not random; it follows patterns dictated by the underlying physics of the acquisition process, primarily affecting the phase encoding direction and manifesting as blurring, ghosting, or ringing. This technical guide provides a comprehensive framework for leveraging matched motion-corrupted and clean data from public repositories to build robust validation datasets for motion correction algorithms. Such datasets are critical for advancing research in neuroimaging and for developing reliable tools for drug development, where image quality can significantly impact the assessment of treatment outcomes. We detail methodologies for generating realistic synthetic motion, present quantitative performance metrics of state-of-the-art correction models, and provide essential protocols and resources to equip researchers with the tools necessary for rigorous, reproducible validation.

In brain MRI research, patient movement during the inherently long acquisition times introduces spatially systematic artifacts that can compromise data integrity. These artifacts predominantly manifest in the phase encoding direction due to the faster readout acquisition compared to repetition time, resulting in blurring, ghosting, and ringing [31]. The consequences are particularly severe in clinical and drug development contexts, where motion has been shown to significantly reduce diagnostic accuracy, especially for critical conditions like intracranial hemorrhage [62]. Furthermore, in resting-state functional MRI (fMRI), head motion introduces systematic bias to functional connectivity (FC), potentially leading to spurious brain-behavior associations [1].

The development of reliable motion correction algorithms hinges on access to high-quality validation datasets comprising matched motion-corrupted and clean images. Such datasets enable the quantitative evaluation of a model's ability to restore image fidelity without introducing hallucinations or distorting underlying anatomy. This guide outlines the methodologies for creating and utilizing these datasets, with a focus on leveraging public repositories and synthetic data generation techniques to overcome the scarcity of perfectly matched clinical data.

Public data repositories are invaluable sources for benchmarking motion correction algorithms. The table below summarizes key datasets used in recent state-of-the-art studies.

Table 1: Public Repositories for Motion Correction Research

Repository Name	Key Characteristics	Utilized In	Primary Use Case
IXI	Contains nearly 600 T1-weighted MRI scans from healthy subjects [63].	PI-MoCoNet [63]	Development and validation of correction models; source for synthetic motion.
ADNI	A large-scale repository with T1-weighted MRI scans; over 9,544 used in one study [18].	JDAC Framework [18]	Training and validation of denoising and artifact correction models.
MR-ART	Specifically includes data with motion artifacts, providing real-world examples [63] [18].	PI-MoCoNet, JDAC [63] [18]	Testing model performance on real, rather than simulated, motion artifacts.
ABCD-BIDS	Pre-processed resting-state fMRI data from over 11,874 children [1].	SHAMAN Analysis [1]	Large-scale analysis of motion's impact on functional connectivity and traits.

Methodologies for Generating Matched Datasets

A primary challenge in motion correction research is obtaining pairs of motion-corrupted and pristine images from the same subject. Two principal methodologies address this challenge.

Synthetic Motion Generation

This approach applies simulated motion artifacts to high-quality scans, creating a perfectly matched pair. The two main techniques are:

• k-Space Corruption: This physics-informed method directly perturbs the k-space data to mimic motion. A common approach involves applying random rigid transformations to a subset of phase encoding lines, which disrupts the data consistency necessary for a clear image reconstruction [63]. This method is valued for its realism, as it directly replicates the source of motion artifacts in the MR signal acquisition process.
• Image-Based Simulation: This technique operates in the spatial domain, applying transformations or filters to the reconstructed image to create artifacts that resemble motion. While potentially less physically accurate than k-space manipulation, it can be more computationally efficient and easier to implement [31].

Transfer Learning from Research to Clinical Data

Models pre-trained on research data with synthetic motion can be fine-tuned for clinical application. One effective framework involves:

• Pre-training: A convolutional neural network (CNN) classifier is trained on large, publicly available research databases (e.g., IXI, ADNI) using synthetically generated motion [31].
• Fine-tuning: The pre-trained model is subsequently adapted and validated on a smaller set of manually annotated clinical images from a data warehouse. This process helps the model generalize to the high heterogeneity and varied artifact patterns found in real-world clinical data [31].

Quantitative Performance of State-of-the-Art Models

Rigorous validation on standardized datasets is key to assessing algorithmic performance. The following table summarizes the quantitative results of recent advanced models on the IXI and MR-ART datasets, demonstrating the effectiveness of modern approaches.

Table 2: Performance Metrics of Motion Correction Models on Public Datasets

Model / Dataset	Artifact Level	PSNR (dB) (Higher is Better)	SSIM (Higher is Better)	NMSE (%) (Lower is Better)
PI-MoCoNet [63] (IXI Dataset)	Minor Motion	45.95	1.00	0.04
	Moderate Motion	42.16	0.99	0.09
	Heavy Motion	36.01	0.97	0.36
PI-MoCoNet [63] (MR-ART Dataset)	Low Level	33.01	0.87	6.24
	High Level	31.72	0.83	8.32
Joint Denoising and Artifact Correction (JDAC) [18]	Iterative Framework	Effective on motion-affected MRIs with severe noise, outperforming state-of-the-art methods in both tasks.

Abbreviations: PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index Measure), NMSE (Normalized Mean Square Error).

These results highlight that physics-informed models like PI-MoCoNet, which leverage both spatial and k-space information, achieve superior artifact correction and soft-tissue contrast preservation [8] [63]. Furthermore, frameworks like JDAC that jointly address noise and motion artifacts can progressively improve image quality through iterative learning, which is crucial for handling low-quality clinical scans where multiple sources of degradation coexist [18].

Experimental Protocols for Validation

Protocol 1: k-Space Corruption and Dual-Network Correction

This protocol is based on the PI-MoCoNet framework, which uses a physics-informed deep learning approach [63].

Diagram Title: k-Space Corruption and Correction Workflow

Workflow Steps:

• Input: Start with a clean reference image from a repository like IXI.
• k-Space Perturbation: Apply realistic motion artifacts by simulating disruptions in the phase encoding lines of the image's k-space data using random rigid transformations [63].
• Motion Detection: The motion-corrupted image is fed into a U-Net-based detection network. This network identifies the corrupted regions in k-space and outputs a binary mask, aided by a spatial averaging module to reduce prediction uncertainty [63].
• Motion Correction: The corrupted image and the predicted mask are processed by a correction network. This network uses a U-Net backbone with Swin Transformer blocks to enhance feature representation and is trained using a combination of:
  • Reconstruction loss (L1) for data fidelity.
  • LPIPS loss for perceptual similarity.
  • Data consistency loss in k-space to enforce fidelity to the measured data [63].
• Output: The final motion-corrected image.

Protocol 2: Iterative Joint Denoising and Artifact Correction

This protocol addresses the common scenario where motion artifacts and noise are present simultaneously, using the JDAC framework [18].

Diagram Title: Iterative Joint Denoising and Correction

Workflow Steps:

• Input: A brain MRI degraded by both motion artifacts and noise.
• Noise Level Estimation: A novel strategy estimates the image's noise level by calculating the variance of its gradient map [18].
• Adaptive Denoising: A U-Net model, conditioned on the estimated noise variance, performs feature normalization to adaptively reduce the noise [18].
• Motion Artifact Correction: The denoised image is passed to an anti-artifact U-Net, which uses a gradient-based loss function to eliminate motion artifacts while preserving the integrity of brain anatomical details [18].
• Iteration: Steps 3 and 4 are repeated iteratively. An early stopping strategy, dependent on the noise level estimation, is applied to accelerate the process until image quality converges to a satisfactory level [18].

The Scientist's Toolkit: Research Reagent Solutions

The following table catalogs essential computational tools and data resources for building and validating motion correction models.

Table 3: Essential Research Reagents for Motion Correction Studies

Reagent / Resource	Type	Function / Application	Example/Reference
Synthetic Motion Generator	Software Script	Creates matched motion-corrupted data from clean images for supervised training.	k-space perturbation with rigid transformations [63].
Pre-trained CNN Classifier	Algorithm	Provides a foundation model for detecting motion artifacts, transferable to clinical data.	Model pre-trained on IXI/ADNI, fine-tuned on clinical data [31].
Dual-Network Framework (PI-MoCoNet)	Model Architecture	Simultaneously detects corrupted k-space lines and corrects motion artifacts in a physics-informed manner.	[8] [63]
Iterative JDAC Framework	Model Architecture	Jointly performs denoising and motion artifact correction, handling co-occurring degradations.	[18]
Public Datasets (IXI, ADNI, MR-ART)	Data	Provides foundational clean and/or artifact-laden images for training, validation, and benchmarking.	[63] [18]
Quality Metrics (PSNR, SSIM, NMSE)	Analytical Tool	Quantifies the performance of correction algorithms against a known ground truth.	Standard evaluation metrics [63] [18].
Motion Impact Score (SHAMAN)	Analytical Method	Quantifies the trait-specific impact of residual motion on functional connectivity in fMRI studies.	Detects spurious brain-behavior associations [1].

In neuroimaging research, particularly in studies investigating the spatial distribution of motion artifacts in the brain, the quantitative evaluation of image quality is paramount. Magnetic resonance imaging (MRI) is exceptionally susceptible to motion artifacts due to its relatively long acquisition time compared to other modalities such as computed tomography [64]. These artifacts, which primarily manifest as blurring, ghosting, or ringing in the phase-encoding direction, can significantly degrade image quality and introduce systematic biases in morphometric analyses [65] [31]. For instance, head motion has been shown to mimic signs of cortical atrophy by causing a reduced estimation of grey matter volume and cortical thickness [31]. Therefore, robust and quantitative metrics are essential for characterizing these artifacts, evaluating correction algorithms, and ensuring the reliability of subsequent analyses. This technical guide provides an in-depth examination of four fundamental image quality metrics—Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Normalized Mean Square Error (NMSE), and Contrast-to-Noise Ratio (CNR)—framed within the context of motion artifact research in brain imaging.

The assessment of motion artifacts has traditionally relied on subjective evaluation by expert readers, such as radiologists or technologists. However, this approach is poorly reproducible, time-consuming, and costly [65]. The development of objective, quantitative metrics provides a crucial toolset for the consistent and automated evaluation of image quality, which is especially valuable when dealing with large datasets from clinical data warehouses where manual inspection is infeasible [31]. This guide will detail the principles, calculations, and applications of these core metrics, provide structured experimental protocols for their use, and visualize their role in a typical research workflow.

Core Metric Definitions and Mathematical Foundations

Peak Signal-to-Noise Ratio (PSNR)

PSNR is a classical, pixel-wise fidelity metric that measures the ratio between the maximum possible power of a signal and the power of corrupting noise. It is most commonly calculated between a reference (ideal) image and a processed or degraded image. For two images, a reference image ( A ) and a distorted image ( B ), the PSNR (in dB) is defined as:

( \text{PSNR} = 10 \cdot \log{10} \left( \frac{\text{MAX}A^2}{\text{MSE}} \right) )

where ( \text{MAX}_A ) is the maximum possible pixel value of image ( A ) (e.g., 255 for 8-bit images), and MSE is the Mean Squared Error between ( A ) and ( B ), given by:

( \text{MSE} = \frac{1}{MN} \sum{i=1}^{M} \sum{j=1}^{N} [A(i,j) - B(i,j)]^2 )

for images of dimensions ( M \times N ) [66]. A higher PSNR value generally indicates a closer resemblance to the reference image, implying less distortion or noise. However, a key limitation is that its pixel-based nature does not always align with human visual perception [65] [66].

Structural Similarity Index (SSIM)

The SSIM index was developed to improve upon traditional metrics like PSNR by modeling image degradation as a change in structural information, which is more aligned with human perception [66]. Rather than comparing pixel values directly, SSIM assesses the similarity between two images based on three key comparisons: luminance (( l )), contrast (( c )), and structure (( s )).

The SSIM between two image patches, ( x ) and ( y ), is computed as:

( \text{SSIM}(x, y) = [l(x, y)]^\alpha \cdot [c(x, y)]^\beta \cdot [s(x, y)]^\gamma )

The default form simplifies to:

( \text{SSIM}(x, y) = \frac{(2\mux\muy + C1)(2\sigma{xy} + C2)}{(\mux^2 + \muy^2 + C1)(\sigmax^2 + \sigmay^2 + C_2)} )

where:

• ( \mux ) and ( \muy ) are the mean intensities of ( x ) and ( y ).
• ( \sigmax^2 ) and ( \sigmay^2 ) are the variances of ( x ) and ( y ).
• ( \sigma_{xy} ) is the covariance between ( x ) and ( y ).
• ( C1 ) and ( C2 ) are small constants added to stabilize the division [65] [66].

The SSIM index yields a value between -1 and 1, where 1 indicates perfect structural similarity. It has become a benchmark metric in medical image analysis for tasks including super-resolution and artifact detection [67] [65].

Normalized Mean Square Error (NMSE)

The Normalized Mean Square Error is a variant of the MSE that is normalized by the energy of the reference image. This normalization makes it a dimensionless quantity and can facilitate comparisons across different datasets or imaging conditions. The NMSE is defined as:

( \text{NMSE} = \frac{\sum{i=1}^{M} \sum{j=1}^{N} [A(i,j) - B(i,j)]^2}{\sum{i=1}^{M} \sum{j=1}^{N} [A(i,j)]^2} )

where ( A ) is the reference image and ( B ) is the test image [66]. Unlike PSNR, a lower NMSE value indicates better agreement with the reference. NMSE provides a direct measure of the total error relative to the signal strength of the original data.

Contrast-to-Noise Ratio (CNR)

Contrast-to-Noise Ratio is a critical metric for assessing the ability to distinguish between different tissue types or regions of interest (ROIs) in an image. It quantifies the difference in signal intensity between two regions relative to the background noise. A common definition for CNR between two ROIs is:

( \text{CNR} = \frac{|\mu{\text{ROI}1} - \mu{\text{ROI}2}|}{\sigma_{\text{background}}} )

where ( \mu{\text{ROI}1} ) and ( \mu{\text{ROI}2} ) are the mean signal intensities of the two regions, and ( \sigma_{\text{background}} ) is the standard deviation of the signal in a background area (e.g., air or a noise-only region) [31]. A high CNR indicates that the features of interest are clearly distinguishable from each other and from the background, which is vital for accurate diagnosis and segmentation. Motion artifacts can significantly degrade CNR by introducing blurring and noise that reduce contrast and increase apparent noise levels.

Table 1: Summary of Core Image Quality Metrics

Metric	Principle	Calculation	Value Range	Interpretation
PSNR	Signal fidelity vs. noise	( 10 \cdot \log{10} \left( \frac{\text{MAX}A^2}{\text{MSE}} \right) )	( 0 ) dB to ( \infty )	Higher is better
SSIM	Perceived structural similarity	( \frac{(2\mux\muy + C1)(2\sigma{xy} + C2)}{(\mux^2 + \muy^2 + C1)(\sigmax^2 + \sigmay^2 + C_2)} )	-1 to 1	Closer to 1 is better
NMSE	Normalized energy of error	( \frac{\sum (A-B)^2}{\sum A^2} )	( 0 ) to ( \infty )	Lower is better
CNR	Tissue contrast vs. background noise	( \frac{	\mu{\text{ROI}1} - \mu{\text{ROI}2}	}{\sigma_{\text{background}}} )	( -\infty ) to ( \infty )	Higher magnitude is better

The Scientist's Toolkit: Essential Research Reagents and Materials

To conduct rigorous research on motion artifacts, a specific set of data, software, and computational tools is required. The following table details key components of the research toolkit.

Table 2: Essential Research Reagents and Materials for Motion Artifact Studies

Item Name	Type	Function/Description	Example/Reference
Matched Motion Dataset	Dataset	Paired motion-free and motion-corrupted scans from the same participant for controlled evaluation.	MR-ART dataset [64]
Clinical Data Warehouse (CDW)	Dataset	Large-scale repository of real-world clinical images for validation under heterogeneous conditions.	AP-HP CDW [31]
Synthetic Motion Generator	Software	Generates motion-corrupted images from clean data for algorithm training and testing.	k-space merging simulation [65]
Convolutional Neural Network (CNN)	Software/Model	Deep learning model for tasks like artifact detection, quantification, or image restoration.	Conv5FC3, ResNet [65] [31]
Image Quality Control Pipeline	Software	Automated tool to extract a wide range of IQMs from structural scans.	MRIQC [31] [64]
Structural Similarity Index (SSIM)	Metric	Quantifies perceived image quality and structural preservation.	Wang et al. 2004 [66]
Full-Reference IQA Metrics	Metric	Quality measures (PSNR, SSIM) that require a clean reference image for comparison.	PSNR, CCC, SSIM [65]

Experimental Protocols for Metric Evaluation

Protocol 1: Quantifying Motion Artifact Severity with Full-Reference Metrics

This protocol is designed to objectively quantify the severity of motion artifacts in structural brain MRI (e.g., T1-weighted scans) using full-reference metrics, requiring a paired dataset with motion-free and motion-corrupted images.

• Data Preparation: Obtain a dataset containing matched motion-corrupted and motion-free (reference) T1-weighted structural brain MRIs. This can be achieved through:
  • Controlled Acquisition: Instructing participants to remain still for a "standard" scan and to perform predefined nodding motions for "motion" scans, as in the MR-ART dataset [64].
  • Synthetic Generation: Simulate motion artifacts on clean images. One method involves: a. Applying rigid transformations (rotation and translation) to the original image. b. Transforming both the original and motion-transformed images to k-space using a 2D Fast Fourier Transform (2D FFT). c. Merging the two k-space datasets by exchanging specific, randomly selected data points. d. Applying an inverse 2D FFT to the merged k-space to produce the final motion-corrupted image [65].
• Image Registration: Rigidly or affinely register all motion-corrupted images to their corresponding motion-free reference to ensure spatial alignment before metric computation. Misalignment can artificially inflate error metrics.
• Region of Interest (ROI) Definition: Define ROIs for CNR calculation. Common choices in brain MRI include grey matter (GM), white matter (WM), and cerebrospinal fluid (CSF). A background region (e.g., air outside the brain) should also be selected for noise estimation.
• Metric Computation:
  • PSNR: Calculate globally across the entire brain volume or a specific slice using the formulas in Section 2.1.
  • SSIM: Compute the SSIM index using a sliding window approach across the image and report the mean SSIM (MSSIM) for the whole image or a specific ROI to evaluate structural preservation [66].
  • NMSE: Compute the normalized error across the entire image or within a specific mask (e.g., brain mask).
  • CNR: Calculate the CNR between key tissue types (e.g., GM and WM) using the formula in Section 2.4. The mean intensities are taken from the segmented tissue maps, and the noise is derived from the background ROI.
• Statistical Analysis: Correlate the computed metric values with subjective quality scores provided by expert radiologists (e.g., on a 1-5 scale of artifact severity) to validate their clinical relevance [65] [64].

Protocol 2: Evaluating Super-Resolution and Denoising Models in Biomedical Imaging

This protocol outlines a comprehensive framework for evaluating image enhancement models, such as super-resolution (SR) or denoising networks, assessing both image quality and downstream clinical task performance.

• Model Training: Train or obtain state-of-the-art SR or denoising models. Relevant architectures for comparison include:
  • CNN-based: SRCNN, EDSR, SRResNet [67].
  • Attention-based: RCAN, SwinIR [67].
  • GAN-based: SRGAN, ESRGAN [67].
  • Multi-task Networks: Bi-task deep learning models that use auxiliary information (e.g., MRI) to enhance another modality (e.g., PET) [68].
• Image Enhancement: Apply the trained models to low-resolution or noisy input images (e.g., low-dose PET, low-resolution CT) to generate enhanced output images [67] [68].
• Image Quality Assessment:
  • Full-Reference Phase: If ground-truth high-quality images are available, compute PSNR, SSIM, and NMSE between the model's output and the ground truth.
  • No-Reference Phase: For real-world data without a clean reference, employ no-reference metrics or a pre-trained CNN to predict FR-IQA scores like SSIM without a reference [65].
• Downstream Task Evaluation: Quantify the impact of image enhancement on critical clinical tasks to assess practical utility.
  • Segmentation: Use a standard segmentation model (e.g., U-Net) on both the original and enhanced images. Evaluate segmentation accuracy using the Dice coefficient or Intersection over Union (IoU) against a manual ground truth [67].
  • Classification: Train a classifier (e.g., ResNet) to perform a diagnostic task (e.g., disease detection) using features from the original and enhanced images. Evaluate performance using metrics like the area under the receiver operating characteristic curve (AUC) or F1-score [67].
• Generalization Testing: Evaluate the model's performance on a cross-domain dataset (e.g., from a different scanner or patient population) to test robustness [67].

Workflow Visualization of Motion Artifact Research

The following diagram illustrates the logical workflow and key decision points in a comprehensive motion artifact research pipeline, integrating the metrics and protocols described.

Discussion and Interpretation of Metrics

Interpreting Metric Values in Context

The values of PSNR, SSIM, NMSE, and CNR are not absolute and must be interpreted within the specific context of the study, including the imaging modality, dataset characteristics, and research objective. For example, a study on super-resolution for lung CT scans reported PSNR and SSIM values for various models, with SwinIR achieving among the best results, demonstrating its effectiveness in preserving diagnostic features [67]. In motion artifact detection, a CNN-predicted SSIM value was able to classify images requiring rescanning with a sensitivity of 89.5% and a specificity of 78.2% [65]. Furthermore, studies have shown that metrics like CNR, derived from pipelines like MRIQC, can effectively demonstrate the degradation in image quality between motion-free and motion-corrupted acquisitions in a matched dataset [64]. It is critical to note that performance metrics can vary considerably depending on cohort characteristics such as age range; therefore, direct comparison of metric values across different studies should be done with caution [69].

Strengths, Limitations, and Complementary Use

No single metric provides a complete picture of image quality. A practical evaluation strategy should involve the complementary use of multiple metrics.

• PSNR is computationally simple and provides a clear global measure of error, but it is often criticized for its poor correlation with human perception of quality [66].
• SSIM was developed to address the shortcomings of PSNR and MSE by focusing on structural information, leading to better alignment with human vision. It has become a standard benchmark in medical image analysis [67] [66]. However, it may be less sensitive to certain types of distortions that do not heavily affect structural content.
• NMSE, like PSNR, is a fundamental error measure. Its normalized nature can be advantageous for comparing results across different datasets or when the absolute signal intensity varies.
• CNR is indispensable for evaluating the practical utility of an image for diagnostic tasks where distinguishing between tissues is crucial. It directly measures a property that is fundamental to radiological reading.

In conclusion, a robust assessment of image quality in motion artifact research requires a multi-faceted approach. By leveraging the strengths of PSNR, SSIM, NMSE, and CNR in a complementary fashion and validating findings against subjective expert scores or downstream task performance, researchers can generate reliable and clinically relevant conclusions about the spatial distribution and impact of motion artifacts in the brain.

Motion artifacts (ARTs) in brain magnetic resonance imaging (MRI) represent a significant challenge in both clinical diagnostics and neuroscientific research. These artifacts, primarily stemming from rigid head motion, degrade image quality and can hinder downstream applications such as image segmentation, target tracking in MR-guided radiation therapy, and automated disease classification [5] [6]. The spatial distribution of these artifacts is not random; they can alter the B0 field, inducing susceptibility artifacts, and disrupt k-space readout lines, which may violate the Nyquist criterion and cause characteristic ghosting and ringing artifacts [5]. Understanding this spatial distribution is critical, as artifact patterns can mimic or obscure pathological features, directly impacting the validity of structural and functional brain research.

The imperative for robust motion correction (MoCo) algorithms is underscored by the clinical and economic burden of motion-corrupted scans. It is estimated that 15–20% of neuroimaging exams require repeat acquisitions, potentially incurring additional annual costs exceeding $300,000 per scanner [6]. While prospective methods (e.g., motion tracking, navigator echoes) aim to prevent artifacts during acquisition, they often require hardware modifications and can struggle with fast or non-rigid motion [6]. Consequently, retrospective, image-domain correction methods that operate on reconstructed magnitude images have gained prominence. These methods are universally applicable across scanner vendors and can be integrated into existing clinical workflows without changes to acquisition hardware or reconstruction software [5] [6].

The field has witnessed a rapid evolution in artificial intelligence (AI)-driven approaches for MoCo. This review performs a comparative analysis of three key algorithmic families—Traditional Machine Learning (ML), Generative Adversarial Networks (GANs), and Denoising Diffusion Probabilistic Models (DDPMs)—benchmarking their performance on real-world data and within the specific context of correcting spatially distributed motion artifacts in brain MRI.

Traditional Machine Learning and Generative Adversarial Networks (GANs)

Traditional ML models in medical imaging often rely on hand-crafted features and are less common for direct image restoration tasks like motion correction. However, they provide a foundational understanding. More relevantly, deep learning models, particularly Convolutional Neural Networks (CNNs), have been successfully applied to learn direct mappings between motion-corrupted and motion-free images [6]. These supervised models are trained to minimize pixel-wise losses (e.g., Mean Absolute Error - MAE), but can sometimes produce blurry outputs that lack high-frequency details.

GANs represent a significant advancement in generative modeling. A GAN consists of two neural networks: a generator that creates images from input noise or corrupted data, and a discriminator that distinguishes between generated and real "clean" images. This adversarial training process encourages the generator to produce outputs that are perceptually realistic [6]. Models like cycle-consistent GANs (CycleGAN) and conditional GANs (cGANs) have been widely adopted for MRI motion artifact removal, as they can learn from unpaired or paired datasets, respectively [5] [6]. Despite their success, GAN-based approaches frequently encounter practical limitations, including mode collapse and unstable training dynamics [5].

Denoising Diffusion Probabilistic Models (DDPMs)

Diffusion models have recently revolutionized image generation and restoration. Inspired by non-equilibrium thermodynamics, DDPMs employ a forward diffusion process and a reverse denoising process [5] [70]. The forward process is a fixed Markov chain that gradually adds Gaussian noise to a clean image over a large number of steps (e.g., 1000), until the image becomes pure noise. The reverse process is trained to iteratively denoise the image, recovering the underlying clean data [5].

For motion correction, conventional DDPMs often concatenate the motion-corrupted image with Gaussian noise and perform backward diffusion to reconstruct the motion-free image [5]. While achieving promising results, their reliance on hundreds or thousands of iterative steps substantially increases inference time, limiting clinical applicability [5].

Recent innovations have focused on enhancing the efficiency and fidelity of diffusion models. For instance, Res-MoCoDiff introduces a novel residual-guided diffusion mechanism. Instead of starting from pure Gaussian noise, it exploits the residual error (( r = y - x )) between the motion-corrupted image (( y )) and the motion-free image (( x )) during the forward process. This allows the model to generate a noisy prior that more closely matches the probability distribution of the corrupted data, enabling a dramatically shortened reverse process of only four steps [5]. Architecturally, it also replaces standard attention layers with Swin Transformer blocks to enhance robustness across resolutions and employs a combined ( \ell 1 + \ell 2 ) loss function [5].

Another approach, MU-Diff, employs a mutual learning-based adversarial diffusion framework for multi-contrast MRI synthesis, which can address corruption in specific sequences. It uses two denoising networks: one for comprehensive structural information and another for fine-grained texture details, with a shared critic network ensuring consistency [70].

The following diagram illustrates the core architectural and workflow differences between these model families.

Experimental Protocols and Benchmarking Methodology

Dataset Curation and Motion Simulation

A critical aspect of benchmarking MoCo models is the use of datasets with known ground truth. Evaluations typically employ both in-silico (simulated) and in-vivo (clinical) movement-related artifact datasets [5].

• In-Silico Datasets: These are generated using realistic motion simulation frameworks that apply known motion corruption operators to clean MRI volumes. This allows for precise quantitative comparison against a known ground truth. Simulations can model various artifact severities (minor, moderate, heavy) and types (ghosting, ringing) by introducing perturbations in k-space or image domain [5] [6].
• In-Vivo Datasets: These consist of clinical scans with real motion artifacts, often identified by expert radiologists. While essential for validating real-world performance, the absence of a perfect ground truth makes quantitative assessment more challenging, often necessitating qualitative evaluation by experts [5].

Performance Metrics and Evaluation Criteria

To ensure a comprehensive comparison, studies utilize a suite of quantitative metrics that assess different aspects of image quality and fidelity [5] [6] [70]:

• Peak Signal-to-Noise Ratio (PSNR): Measures the ratio between the maximum possible power of a signal and the power of corrupting noise. Higher values indicate better quality.
• Structural Similarity Index Measure (SSIM): Assesses the perceptual similarity between two images based on luminance, contrast, and structure. Values range from 0 to 1, with 1 indicating perfect similarity.
• Normalized Mean Squared Error (NMSE): Quantifies the normalized average squared difference between pixel values. Lower values are better.
• Mean Absolute Error (MAE): Measures the average absolute pixel-wise difference between images.

Beyond pixel-level accuracy, the computational efficiency of each model, particularly the average sampling or inference time, is a critical metric for assessing clinical utility [5].

Benchmarking Results and Comparative Analysis

The table below summarizes the quantitative performance of different model families as reported in recent studies, including the advanced Res-MoCoDiff model.

Table 1: Quantitative Benchmarking of MoCo Models on Brain MRI Data

Model Type	Representative Model	PSNR (dB)	SSIM	NMSE	Inference Time	Key Advantages	Key Limitations
GAN-based	CycleGAN, Pix2Pix [5]	~39.0 [5]	~0.95 [5]	Moderate [5]	Moderate	High perceptual quality, can use unpaired data	Unstable training, risk of hallucination [5]
Standard DDPM	ViT-Diff [5]	~41.0 [5]	~0.96 [5]	Lower [5]	High (e.g., 101.74 s) [5]	High reconstruction fidelity, stable training	Slow inference, many steps required [5]
Efficient DDPM	Res-MoCoDiff [5]	41.91 ± 2.94 [5]	Highest [5]	Lowest [5]	0.37 s (per 2 slices) [5]	High speed & fidelity, preserves fine details	Requires paired data for residual guidance

The meta-analysis of AI-driven MRI motion correction confirms that while GANs and traditional DDPMs show promise, they face challenges. GANs are noted for risks of visual distortions and limited generalizability, while standard DDPMs are hampered by computational demands [6]. Res-MoCoDiff demonstrates a breakthrough by achieving superior SSIM and the lowest NMSE while reducing sampling time by three orders of magnitude compared to conventional DDPM approaches [5].

The following experimental workflow visualizes the typical pipeline for training and evaluating these models, from data preparation to final benchmarking.

Implementing and benchmarking MoCo algorithms requires a suite of computational tools and datasets. The table below details key resources essential for research in this field.

Table 2: Essential Research Reagents and Resources for AI-based Motion Correction

Resource Category	Specific Item / Tool	Function and Application in Research
Public Datasets	BraTS (Brain Tumor Segmentation) [70] [71]	Provides multi-contrast brain MRI data; used for training and evaluating synthesis and correction models, often with simulated artifacts.
	ISLES (Ischemic Stroke Lesion Segmentation) [70]	Offers data for challenging lesion modeling; tests model robustness on pathological brains with artifacts.
Software Libraries	TorchIO [72]	A Python library for medical image augmentation and preprocessing; used to generate realistic motion artifacts at weak and strong scales for model training and evaluation.
Deep Learning Frameworks	PyTorch, TensorFlow	Core frameworks for implementing, training, and evaluating deep learning models such as CNNs, GANs, and Diffusion models.
Evaluation Metrics	PSNR, SSIM, NMSE, MAE [5] [6] [70]	Standard quantitative metrics implemented in code to objectively compare the performance of different motion correction algorithms.
Architectural Components	U-Net [5] [71]	A common backbone network for image-to-image tasks, frequently used in GANs and Diffusion models for medical image restoration.
	Swin Transformer Block [5]	An attention mechanism that replaces standard transformers in models like Res-MoCoDiff to enhance robustness across image resolutions.
	Denoising Network [70]	The core neural network within a DDPM that is trained to reverse the diffusion process by predicting and removing noise.

The comparative analysis reveals a clear trajectory in the development of MoCo algorithms. While GANs opened the door to perceptually realistic image correction, their instability and tendency to hallucinate pose risks in clinical settings. Standard diffusion models offer remarkable fidelity but at a computational cost that is often prohibitive for time-sensitive applications like radiotherapy guidance [5] [6].

The emergence of efficient, residual-guided diffusion models like Res-MoCoDiff represents a significant step forward. By leveraging the residual error between corrupted and clean images, these models bridge the gap between speed and quality, achieving state-of-the-art correction in a fraction of the time [5]. This makes them particularly promising for integration into clinical workflows.

Future research directions should focus on several key areas:

• Enhanced Generalizability: Developing models that are robust to a wider range of motion patterns, scanner types, and anatomical regions beyond the brain [6].
• Reduced Data Dependency: Exploring self-supervised and unsupervised techniques to reduce reliance on large, paired datasets (motion-corrupted and motion-free pairs of the same subject), which are difficult to acquire [6].
• Standardized Benchmarking: The community would benefit from comprehensive public benchmarks with standardized datasets and evaluation protocols to ensure fair and meaningful comparisons across studies [6] [72].
• Integration with Prospective Correction: Combining powerful retrospective correction methods with prospective motion tracking data could lead to hybrid systems that offer unprecedented robustness against patient motion [6].

In conclusion, for research focused on the spatial distribution of motion artifacts in the brain, efficient diffusion models currently offer the best balance of correction accuracy, structural preservation, and practical deployment potential. Their ability to faithfully restore fine details without introducing spurious hallucinations is essential for ensuring the integrity of downstream quantitative analyses in neuroscience and clinical research.

Within research on the spatial distribution of motion artifacts in the brain, a critical translational step involves determining whether processed images retain their diagnostic value for clinical practice. Expert radiologist scoring serves as a bridge between technical innovation and clinical application, providing a validated method for assessing whether motion-corrected or motion-impacted images are usable for diagnostic purposes. This guide details the methodologies, protocols, and analytical frameworks for robustly integrating radiologist scoring into the evaluation of neuroimaging data, particularly within studies investigating motion artifacts.

The reliability of expert evaluation itself is a key consideration. Studies of inter-rater reliability in fields like usability engineering and radiology reveal that tasks requiring significant judgment can exhibit substantial variability [73]. This underscores the necessity of structured protocols, explicit scoring criteria, and statistical analysis of inter-rater concordance in imaging studies to ensure that the resulting usability assessments are valid and meaningful.

Foundations of Expert Scoring in Imaging

Expert radiologist scoring is a form of qualitative assessment that quantifies human perception of image quality and diagnostic utility. In the context of motion artifacts, which can manifest as blurring, ghosting, or streaks [74] [75] [64], this scoring provides a direct measure of clinical usability that purely quantitative metrics may not fully capture.

The Role of Scoring in Clinical Translation

The primary goal of incorporating expert scoring is to determine if an image, despite the presence of artifacts, is sufficient for accurate diagnosis. This is a critical endpoint for validating any motion correction algorithm or for understanding the practical impact of different artifact types and severities. A study on functional connectivity, for instance, demonstrated that motion-related changes in data could reflect a combination of genuine neural activity and artifacts, requiring careful interpretation [76]. Expert scoring helps to disentangle such confounding effects from a clinical perspective.

Furthermore, the process of scoring itself can be subject to variability, much like diagnostic interpretations in radiology where different radiologists may see "pending death to a clean bill of health" in the same image [73]. This inherent "perception problem" necessitates rigorous study design to ensure that the scoring of diagnostic usability is as consistent and unbiased as possible.

Experimental Protocols for Scoring Studies

Implementing a robust expert scoring system requires careful planning, from the initial acquisition of images to the final statistical analysis of scores. The following protocols provide a framework for generating reliable and clinically relevant data on diagnostic usability.

Image Acquisition and Dataset Design

A well-designed dataset is the foundation of a valid assessment. The ideal design incorporates matched scans from the same participant, allowing for direct, within-subject comparison of different states.

• Controlled Motion Acquisition: To systematically study the impact of motion, images can be acquired under different motion conditions. A proven protocol involves collecting T1-weighted 3D MPRAGE sequences in three conditions [64]:
  • Standard (STAND): Participants are instructed to remain still.
  • Low Head Motion (HM1): Participants perform a prescribed motion, such as nodding their head a limited number of times (e.g., 5 times) during the acquisition when cued.
  • High Head Motion (HM2): Participants perform the same motion more frequently (e.g., 10 times) during the acquisition.
• Motion Patterns: Nodding along the sagittal plane is recommended as it is one of the most prominent types of head motion responsible for artifacts [64]. Participants should be instructed to keep their heads in contact with the scanner table and return to the original position after each motion.

This matched design, as implemented in the Movement-Related Artefacts (MR-ART) dataset, allows researchers to isolate the effect of motion from other inter-individual variability [64].

Radiologist Scoring Procedure

The scoring process must be standardized to minimize variability and maximize the reliability of the results.

• Rater Selection and Training: Scoring should be performed by experienced radiologists (e.g., with ten or more years of experience) [64]. Before formal scoring, raters should harmonize their assessments on a common set of training images not included in the main study, discussing ambiguous cases to align their scoring criteria.
• Blinding: Radiologists should be blinded to the acquisition condition (e.g., STAND, HM1, HM2) and the purpose of the experiment to prevent bias [64].
• Scoring Scale: A simple, ordinal scale focused on clinical utility is most effective. The following 3-point scale is an example [64]:
  • Score 1 (Good): Image is of clinically good quality and usable for diagnostic purposes.
  • Score 2 (Medium): Image is of medium quality; it may have artifacts that complicate but do not wholly prevent diagnostic interpretation.
  • Score 3 (Bad): Image is of bad quality and considered unusable for clinical diagnostics due to severe artifacts.
• Image Evaluation: Radiologists visually inspect the structural volumes, noting artifacts such as blurring, ghosting, or streaks that compromise anatomical clarity [75] [64].

Quantitative Image Quality Metrics (IQMs)

To complement subjective expert scores, objective quantitative metrics should be extracted from each scan. These provide a continuous, unbiased measure of image degradation and can help validate the radiologist scores. Common IQMs calculated by tools like MRIQC include [64]:

• Total Signal-to-Noise Ratio (SNR): Measures the overall clarity of the image; motion typically reduces SNR.
• Entropy Focus Criterion (EFC): Quantifies the ghosting and blurring in an image; higher EFC can indicate more severe motion artifacts.
• Coefficient of Joint Variation (CJV): Measures the intensity homogeneity across tissue types; motion can increase CJV.

Including these metrics allows for a more granular analysis of how specific technical aspects of image quality correlate with clinical usability judgments.

Table 1: Core Components of an Expert Scoring Study Protocol

Component	Description	Example from Literature
Study Design	Matched within-subject design	STAND, HM1, HM2 scans from the same participant [64]
Scoring Scale	Ordinal scale focused on diagnostic use	3-point scale (1=Good, 2=Medium, 3=Bad) [64]
Rater Profile	Experienced clinical experts	Neuroradiologists with >10 years of experience [64]
Blinding	Raters are unaware of experimental condition	Blinded to STAND/HM1/HM2 acquisition label [64]
Quality Control	Use of objective Image Quality Metrics (IQMs)	SNR, EFC, CJV from MRIQC tool [64]

Data Analysis and Interpretation

Once data is collected, analysis focuses on linking subjective scores and objective metrics to draw conclusions about diagnostic usability.

Statistical Analysis of Scores

• Descriptive Statistics: Tabulate the distribution of scores for each acquisition condition. For example, a study might find that 100% of STAND scans received a score of 1 (Good), while HM2 scans were distributed across scores 2 and 3 [64].
• Inter-Rater Reliability: Calculate statistical measures of agreement between different radiologists, such as Cohen's Kappa or Intraclass Correlation Coefficient (ICC). This quantifies the consistency of the scoring system.
• Association with IQMs: Use statistical tests (e.g., ANOVA, correlation analysis) to determine if significant differences exist in objective IQMs (SNR, EFC, CJV) across the different expert score categories. A strong, significant relationship validates that the subjective scores are capturing objective image degradation.

Presentation of Findings

The results of the scoring study can be effectively summarized in tables and figures that illustrate the impact of motion and the relationship between different measures.

Table 2: Illustrative Data Linking Motion, Expert Scores, and Quantitative Metrics

Acquisition Condition	Typical Expert Score Distribution	Representative Change in SNR (vs. STAND)	Impact on Diagnostic Usability
STAND (No Motion)	100% Score 1 (Good)	Reference Value	Uncompromised
HM1 (Low Motion)	Mixed Scores 1 & 2	Slight Decrease	Mildly to Moderately Compromised
HM2 (High Motion)	Majority Score 3 (Bad)	Large Decrease	Severely Compromised or Unusable

The Scientist's Toolkit

Table 3: Essential Research Reagents and Resources

Item	Function/Description	Relevance to Scoring Studies
Matched MRI Dataset	A set of images from the same subject with and without motion artifacts.	Essential for controlled, within-subject comparisons of motion effects. The MR-ART dataset is a public example [64].
Structured Scoring Scale	A predefined ordinal scale (e.g., 3-point) for rating diagnostic usability.	Standardizes qualitative assessments from experts, enabling quantitative analysis [64].
MRIQC Software	An open-source tool for extracting no-reference image quality metrics (IQMs).	Provides objective, quantitative data (SNR, EFC) to validate and complement expert scores [64].
Statistical Analysis Plan	A pre-specified plan for analyzing scores and IQMs, including reliability tests.	Ensures robust and unbiased interpretation of results, accounting for inter-rater variability [73].

Integrating expert radiologist scoring into the assessment of motion artifacts provides an indispensable, clinically grounded measure of diagnostic usability. By adhering to rigorous protocols for image acquisition, blinded scoring, and statistical analysis that incorporates both subjective scores and objective quality metrics, researchers can robustly translate technical findings into outcomes with clear relevance for patient care. This methodology ensures that advancements in understanding and correcting motion artifacts are evaluated against the ultimate benchmark: their impact on the radiologist's ability to provide an accurate diagnosis.

Conclusion

The spatial distribution of motion artifacts in the brain is not random but follows predictable patterns dictated by biomechanics and imaging physics. Addressing this challenge requires a multi-faceted toolbox, as no single solution is effective in all situations. The emergence of efficient deep learning models, such as Res-MoCoDiff, and standardized, paired datasets for validation marks a significant leap forward, offering robust correction with dramatically reduced computational overhead. Future directions must focus on developing even more efficient, real-time correction algorithms that can be seamlessly integrated into clinical and research workflows, improving the diagnostic accuracy of MRI and the reliability of functional and structural biomarkers in both basic neuroscience and pharmaceutical development.

References

• 1. Motion impact score for detecting spurious brain-behavior ... [https://www.nature.com/articles/s41467-025-63661-2]
• 2. Motion Artefacts in MRI: a Complex Problem with Many Partial ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4517972/]
• 3. Ghosting? [https://mriquestions.com/ghosting.html]
• 4. Reducing Eddy Current Artifacts With Dummy Gradients ... [https://i-mri.org/DOIx.php?id=10.13104/imri.2025.0013]
• 5. Res-MoCoDiff: residual-guided diffusion models for motion ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12532098/]
• 6. Systematic Review and Meta-analysis of AI-driven MRI ... [https://arxiv.org/html/2509.05071v1]
• 7. Motion-corrected brain MRI at ultralow field (64 mT) [https://pubmed.ncbi.nlm.nih.gov/40152515]
• 8. A Physics-Informed Deep Learning Method for Correcting ... [https://archive.ismrm.org/2025/3844.html]
• 9. Motion artifact in studies of functional connectivity [https://pmc.ncbi.nlm.nih.gov/articles/PMC5930165/]
• 10. Biomechanics of the Spine: the Coupled Motion of the Spine [https://www.anatomystandard.com/biomechanics/spine/coupled-motions.html]
• 11. Biomechanical models and mechanisms of cellular ... [https://www.sciencedirect.com/science/article/pii/S1084952122002178]
• 12. The biomechanical study of different fixation techniques for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11921601/]
• 13. Iterative Learning for Joint Image Denoising and Motion ... [https://www.sciencedirect.com/science/article/pii/S1361841525004189?dgcid=rss_sd_all]
• 14. An Exploration of Motion-Sampling Interactions in 3D MRI ... [https://pubmed.ncbi.nlm.nih.gov/41204061/]
• 15. Motion‐robust T2∗ quantification from low‐resolution ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12620166/]
• 16. Hippocampal subregions functional connectivity alterations in ... [https://bmcneurol.biomedcentral.com/articles/10.1186/s12883-025-04485-x]
• 17. Researchers uncover why mental maps fade with age [https://med.stanford.edu/news/all-news/2025/10/aging-mental-map.html]
• 18. Iterative Learning for Joint Image Denoising and Motion ... [https://www.sciencedirect.com/science/article/abs/pii/S1361841525004189]
• 19. DIMA: DIffusing Motion Artifacts for unsupervised correction ... [https://arxiv.org/abs/2504.06767]
• 20. Characterizing changes to individual-specific brain ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12263593/]
• 21. Diff5T: Benchmarking human brain diffusion MRI with an ... [https://www.nature.com/articles/s41597-025-05640-2]
• 22. Applications and advances of combined fMRI-fNIRs ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11958174/]
• 23. Comparison of motion correction techniques applied to ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9900395/]
• 24. Optimizing spatial specificity and signal quality in fNIRS [https://www.frontiersin.org/journals/neuroergonomics/articles/10.3389/fnrgo.2024.1286586/full]
• 25. Spatial correspondence of cortical activity measured with ... [https://www.sciencedirect.com/science/article/pii/S1053811924000648]
• 26. Multimodal assessment of the spatial correspondence ... [https://www.nature.com/articles/s41598-023-29123-9]
• 27. Reproducibility of fNIRS within Subject for Visual and Motor ... [https://www.sciencedirect.com/science/article/pii/S1053811925004951]
• 28. An Automatic Motion-Based Artifact Reduction Algorithm ... [https://www.mdpi.com/1999-4893/16/5/230]
• 29. Disentangling the impact of motion artifact correction ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11498316/]
• 30. Comparing different motion correction approaches for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11448702/]
• 31. Automatic motion artefact detection in brain T1-weighted ... [https://www.sciencedirect.com/science/article/abs/pii/S136184152300333X]
• 32. A portable system for collecting anatomical joint angles during stair... [https://pmc.ncbi.nlm.nih.gov/articles/PMC2684094/]
• 33. Motion artifacts removal and evaluation techniques for ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2022.878750/full]
• 34. Outlook for Noninvasive Brain-Computer Interfaces for AI [https://etcjournal.com/2025/09/21/outlook-for-noninvasive-brain-computer-interfaces-for-ai/]
• 35. 1 Introduction [https://arxiv.org/html/2505.03498v2]
• 36. Res-MoCoDiff: residual-guided diffusion models for motion ... [https://pubmed.ncbi.nlm.nih.gov/41061726/]
• 37. Convolutional neural networks: an overview and application in ... [https://insightsimaging.springeropen.com/articles/10.1007/s13244-018-0639-9]
• 38. Deep convolutional neural networks for brain image ... [https://www.sciencedirect.com/science/article/abs/pii/S0933365716305206]
• 39. Application of convolutional neural networks in medical images [https://pmc.ncbi.nlm.nih.gov/articles/PMC11074758/]
• 40. Res-MoCoDiff: Residual-guided diffusion models for ... [https://arxiv.org/abs/2505.03498]
• 41. A Systematic Comparison of Motion Artifact Correction ... - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC3468891/]
• 42. Disentangling the impact of motion artifact correction ... [https://www.spiedigitallibrary.org/journals/Disentangling-the-impact-of-motion-artifact-correction-algorithms-on-functional/volume-11/issue-04/045006/Disentangling-the-impact-of-motion-artifact-correction-algorithms-on-functional/10.1117/1.NPh.11.4.045006.full]
• 43. The spatial layout of antagonistic brain regions is ... [https://www.nature.com/articles/s42003-025-08295-2]
• 44. 1 Data visualization - R for Data Science (2e) - Hadley Wickham [https://r4ds.hadley.nz/data-visualize.html]
• 45. A Multi-Fidelity Methodology for Reduced Order Models ... [https://arxiv.org/html/2402.17061]
• 46. Adaptive sampling strategies for reduced-order modeling [https://link.springer.com/article/10.1007/s13272-022-00574-6]
• 47. Application and comparison of several adaptive sampling ... [https://www.sciencedirect.com/science/article/pii/S2405844024109590]
• 48. Automatic MR image quality evaluation using a Deep CNN [https://www.sciencedirect.com/science/article/abs/pii/S0895611121000458]
• 49. of Automated in detection MR images using... motion artifacts brain [https://pubmed.ncbi.nlm.nih.gov/39439086/]
• 50. [2402.08749] Automated of detection in motion MR... artifacts brain [https://arxiv.org/abs/2402.08749]
• 51. Agreement of image quality metrics with radiological ... [https://pubmed.ncbi.nlm.nih.gov/40493331/]
• 52. Agreement of image quality metrics with radiological ... [https://link.springer.com/article/10.1007/s10334-025-01266-y]
• 53. Agreement of Image Quality Metrics with Radiological ... [https://arxiv.org/abs/2412.18389]
• 54. A Systematic Review of Discrete Choice Experiments [https://pmc.ncbi.nlm.nih.gov/articles/PMC10517687/]
• 55. Sequential Human Assembly and Disassembly Motions in ... [https://www.nature.com/articles/s41597-025-06042-0]
• 56. AI in motion: the impact of data augmentation strategies on ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12559149/]
• 57. Mitigating head motion artefact in functional connectivity MRI [https://pmc.ncbi.nlm.nih.gov/articles/PMC8161527/]
• 58. Heterogeneous impact of motion on fundamental patterns ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811913006824]
• 59. Studying the Spatial Distribution of Physiological Effects on ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2014.00196/full]
• 60. Denoising strategies of functional connectivity MRI data in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12580809/]
• 61. Multi-Metric Approach for the Comparison of Denoising ... [https://pubmed.ncbi.nlm.nih.gov/40309965/]
• 62. Motion artifacts and image quality in stroke MRI [https://pubmed.ncbi.nlm.nih.gov/40664863/]
• 63. A Physics-Informed Deep Learning Model for MRI Brain ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11844622/]
• 64. Movement-related artefacts (MR-ART) dataset of matched ... [https://www.nature.com/articles/s41597-022-01694-8]
• 65. Evaluation of motion artifacts in brain magnetic resonance ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8782596/]
• 66. Prospects of Structural Similarity Index for Medical Image ... [https://www.mdpi.com/2076-3417/12/8/3754]
• 67. Evaluating Super-Resolution Models in Biomedical Imaging [https://pmc.ncbi.nlm.nih.gov/articles/PMC12027580/]
• 68. High-quality PET image synthesis from ultra-low-dose PET/MRI using... [https://qims.amegroups.org/article/view/102636/html]
• 69. Mind the gap: Performance metric evaluation in brain‐age ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9188975/]
• 70. MU-Diff: a mutual learning diffusion model for synthetic MRI ... [https://www.nature.com/articles/s44387-025-00016-8]
• 71. Deep LearnIng and Machine Learning for Brain Tumor ... [https://link.springer.com/article/10.1007/s11831-025-10416-3]
• 72. Understanding the robustness of vision-language models ... [https://www.nature.com/articles/s41746-025-02108-w]
• 73. Usability Evaluators: Reliable as Radiologists? [https://measuringu.com/usability-radiology/]
• 74. Overview of Functional Magnetic Resonance Imaging - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC3073717/]
• 75. Artifacts in spiral CT protocols: The importance of ... [https://www.sciencedirect.com/science/article/abs/pii/S1756117509000214]
• 76. Does motion -related brain functional connectivity reflect both artifacts ... [https://pubmed.ncbi.nlm.nih.gov/24999036/]