Conquering Motion Artifacts in Behavioral Neuroimaging: From Prevention to AI-Powered Correction

Abstract

This article provides a comprehensive guide for researchers and drug development professionals on addressing motion artifacts in neuroimaging during behavioral tasks. It explores the fundamental physics and origins of motion-related signal corruption in key modalities like fMRI and fNIRS, which are crucial for assessing brain function in active participants. The content details a toolbox of mitigation strategies, from simple patient preparation to advanced deep learning and algorithmic corrections. A comparative analysis validates the performance of various methods, from established techniques like ICA-AROMA to novel generative models, providing a clear framework for selecting optimal motion correction approaches to ensure data integrity in clinical and cognitive neuroscience research.

The Motion Problem: Unraveling the Physics and Impact on Behavioral Data

Why Neuroimaging is Uniquely Sensitive to Subject Motion

Why is motion a particularly critical problem in neuroimaging?

Motion is a paramount concern in neuroimaging because even sub-millimeter movements—smaller than the typical voxel size—are large enough to significantly compromise the quality and reliability of both functional and structural data [1] [2]. Unlike a simple photograph where motion causes blurring, the effects of motion in neuroimaging are complex and can mimic or obscure genuine brain signals, leading to false conclusions in research and clinical assessments [1].

The core reasons for this unique sensitivity include:

• The Physics of Data Acquisition: When a subject moves inside an MRI scanner, it perturbs the spatial frequencies (k-space) used to construct the image. This introduces artifacts that are not localized but propagate throughout the entire image, manifesting as ghosting, ringing, and blurring [1].
• The Challenge of Task-Correlated Motion: In functional MRI (fMRI) studies, participants often move in sync with a task (e.g., a button press in response to a stimulus). This creates a situation where the motion is systematically correlated with the experimental paradigm. Since standard analysis methods are designed to detect signal changes time-locked to the task, this correlated motion can be easily misinterpreted as true brain activation [3] [4].
• Impact on Key Metrics: Motion artifacts have a downstream effect on nearly all common neuroimaging metrics. For structural scans, motion can lead to underestimates of cortical thickness and grey matter volume [1]. For functional scans, it can inflate or distort measures of functional connectivity between brain regions [5].

---

Quantifying the Motion Problem: Key Indicators and Impacts

Understanding which factors predict motion can help in better study planning. The following table summarizes key indicators of in-scanner head motion identified from a large-scale study of 40,969 subjects [2].

Table 1: Subject Indicators of fMRI Head Motion

Indicator Category	Specific Indicator	Association with Head Motion
Anthropometric	Body Mass Index (BMI)	Strongest indicator. A 10-point increase in BMI (e.g., from "healthy" to "obese") corresponds to a 51% increase in motion [2].
Demographic	Age	Motion is higher at the extreme ends of the age distribution (e.g., in children and older adults) [1] [2].
	Ethnicity	A significant association was identified, though the specific reasons are complex and may be socio-economic rather than biological [2].
Clinical & Behavioral	Psychiatric & Neurological Disorders	Populations with ADHD, Autism Spectrum Disorder (ASD), and schizophrenia tend to exhibit significantly increased motion [1].
	Cognitive Task Performance	Performing a cognitive task in the scanner is associated with increased head motion compared to rest [2].
	Executive Functioning	In older adults, poorer performance on tasks of inhibition and cognitive flexibility is correlated with a higher number of motion-corrupted scans [6].

The impact of motion also depends on the type of data being acquired and the analysis method used. The table below outlines how motion affects different neuroimaging domains.

Table 2: Impact of Motion Across Neuroimaging Domains

Neuroimaging Domain	Primary Impact of Motion	Evidence from Literature
Structural MRI	Artificial reduction of grey matter volume and cortical thickness [1].	Motion can create a false, non-linear trajectory of cortical development in youth [1].
fMRI (Task-Based)	Introduction of false positives (artifactual activations) and false negatives (masked true activations), especially with block designs [7] [4].	Motion parameters correlated with a task paradigm (r ~0.5) can cause spurious activations [4].
fMRI (Resting-State)	Inflation of functional connectivity measures, particularly in short-range connections; reduces test-retest reliability [2] [5].	Full correlation is highly sensitive to motion; partial correlation and coherence are more robust [5].
Clinical Group Analysis	Confounding of case-control differences, as patient groups often move more than healthy controls [1].	In Autism studies, inconsistent findings of cortical thickness have been linked to differing motion exclusion criteria [1].

---

A Researcher's Guide to Motion Correction Methodologies

A multi-pronged strategy is required to tackle motion, involving both prospective (during scanning) and retrospective (after scanning) methods.

Prospective Motion Reduction Strategies

These methods aim to minimize motion at its source.

• Subject Preparation: Clear communication, acclimation to the scanner environment, and comfortable positioning are crucial [1] [2].
• Use of Head Stabilization: Personalized head moulds and foam padding can significantly restrict movement [2].
• Real-Time Feedback: Systems that provide visual or tactile feedback to the subject about their head position can help them remain still [2].
• Task Design: For fMRI, using rapid event-related designs instead of block designs can help decouple the task from motion, reducing the risk of correlated artifacts [4].

Retrospective Motion Correction Algorithms

These are standard preprocessing steps applied to the data after acquisition.

• Rigid-Body Realignment: This is the most common method, implemented in tools like FSL's MCFLIRT or SPM's realign. It models head motion as a rigid body with six degrees of freedom (translations and rotations along x, y, z axes) and realigns all volumes in a time-series to a reference volume [3].
• Motion Parameter Regression (MPR): The six motion parameters estimated during realignment are included as nuisance regressors in the statistical model (e.g., General Linear Model) to remove variance associated with motion [4] [8]. Caution: In block designs, if motion is correlated with the task, MPR can remove genuine brain signal along with the artifact [4].
• Motion Scrubbing: Data points (volumes) that are identified as severe motion outliers are removed from the analysis entirely [7].
• Incorporating Motion Covariates in Group Analysis: Including a subject-level measure of motion (e.g., mean framewise displacement) as a covariate in group-level statistics helps control for inter-subject differences in motion [4].

The following diagram illustrates a standard workflow for integrating these retrospective correction methods into an fMRI preprocessing pipeline.

---

Table 3: Key Software Tools and Research Reagents for Motion Correction

Tool / Resource Name	Type	Primary Function	Key Consideration
FSL MCFLIRT [3]	Software Algorithm	Performs rigid-body motion correction on fMRI time-series.	A standard, widely-used tool. Often used as a benchmark.
SPM Realign [4]	Software Algorithm	Performs rigid-body realignment as part of the SPM preprocessing pipeline.	Integrated into the popular SPM software suite.
AIR (Automated Image Registration) [4]	Software Algorithm	An early and influential image registration tool adapted for fMRI motion correction.
Motion Parameters (6+) [4]	Data Output	The translational (x,y,z) and rotational (pitch,roll,yaw) parameters estimated during realignment.	Used for regression and quality control. The "root mean square" (RMS) is a common summary metric [2].
Framewise Displacement (FD)	Quantitative Metric	A scalar value that quantifies volume-to-volume head movement.	Used to identify motion outliers for scrubbing [6].
NPAIRS Framework [8]	Validation Framework	A data-driven framework for evaluating preprocessing pipeline performance using cross-validation.	Helps optimize pipeline choices for a specific dataset.

---

Frequently Asked Questions (FAQs)

Q1: Our patient group moves significantly more than our healthy controls. Should we exclude high-movers, and what are the alternatives? This is a common dilemma in clinical neuroscience [1]. Excluding high-movers risks biasing your sample and losing hard-to-recruit patients. Alternatives include:

• Proactive Mitigation: Invest more time in participant preparation, use improved head stabilization, and consider real-time motion correction techniques [2].
• Statistical Control: Include subject-level motion estimates as a covariate in your group-level analysis [4].
• Robust Analysis Methods: Choose functional connectivity measures that are less sensitive to motion, such as partial correlation instead of full correlation [5].
• Transparent Reporting: Always report motion exclusion criteria and the amount of motion in each group to allow for informed interpretation of results [1].

Q2: We've applied rigid-body motion correction. Is our data now clean? Not necessarily. Motion correction is essential but imperfect [4]. Residual artifacts often remain because:

• Physics-Based Distortions: Head movement causes dynamic changes in the magnetic field, leading to nonlinear distortions that rigid-body models cannot correct [4].
• Intra-Volume Motion: Motion that occurs during the acquisition of a single volume cannot be fully characterized or corrected by standard methods [4].
• Spin-History Effects: Motion affects the longitudinal magnetization of spins, creating signal changes that are not addressed by spatial realignment [4]. Therefore, motion correction should be seen as a necessary first step, but it must often be combined with other methods like MPR or scrubbing.

Q3: For an event-related fMRI design, what is the most effective motion correction strategy? Research indicates that for rapid event-related designs, including the motion parameters as covariates (MPR) in the general linear model is highly effective at increasing sensitivity [4]. Interestingly, in this context, it may matter less whether the motion correction (realignment) was actually applied to the data before this step, as the MPR can account for a large portion of the motion-related variance [4].

Q4: How does motion specifically affect studies of brain development or aging? Motion can profoundly confound studies across the lifespan. In developmental studies, younger children move more, which can create a false impression of exaggerated cortical thinning with age if motion is not controlled [1]. In older adults, higher motion is correlated with poorer executive functioning, meaning that excluding high-movers may systematically remove those with lower cognitive performance, biasing the sample and skewing our understanding of the aging brain [6].

Troubleshooting Guides

Guide 1: Identifying and Resolving Ghosting Artefacts

Problem: My fMRI images show repeating ghost-like duplicates of the brain structure, often in the phase-encoding direction.

Root Cause: Ghosting artefacts primarily arise from inconsistencies between different portions of the k-space data used for image reconstruction. This can be caused by:

• Phase Discontinuities: Constant-phase shifts between echoes, caused by factors like magnetic field inhomogeneities, chemical shift, or receiver-phase misregistrations [9].
• Amplitude Discontinuities: Variations in signal amplitude between k-space lines, for instance, due to T2 decay or periodic motion [9].
• System Imperfections: Timing misregistrations from system filter delays or eddy currents [9].

Solutions:

• Use Reference Scans: Acquire "reference" scans without phase-encoding to measure constant-phase offsets and timing delays. These measurements can then be used for post-processing correction [9].
• Optimize K-space Ordering: For periodic motion (e.g., respiration, cardiac pulsation), use a k-space acquisition order that is synchronized with the motion to minimize inconsistencies [10].
• Apply Post-Processing Algorithms: Implement correction algorithms that apply measured constant-phase shifts and phase rolls to correct for time shifts in the k-space data before Fourier transformation [9].

Guide 2: Addressing Image Blurring

Problem: My fMRI data appears unfocused, with a loss of sharpness at contrast edges.

Root Cause: Blurring is typically the result of slow, continuous motion during the data acquisition period. This is particularly problematic for sequences using interleaved k-space acquisitions, such as T2-weighted Turbo Spin Echo (TSE/FSE) sequences [10]. Unlike sudden motion that causes ghosting, slow drifts violate the assumption that the object is stationary throughout the scan, leading to a smearing of information in k-space.

Solutions:

• Prospective Motion Correction: Use real-time head-tracking systems (e.g., optical tracking, navigator echoes) to adjust the imaging plane during data acquisition [11] [12].
• Increase Acquisition Speed: Use faster imaging sequences to reduce the time window during which motion can occur [10].
• Retrospective Correction: Employ image registration techniques to align all volumes in a time series to a reference volume. This is a common first step in fMRI preprocessing [11] [13].

Guide 3: Managing Residual Motion Artefacts After Correction

Problem: Even after volume realignment (motion correction), my functional connectivity (RSFC) results show motion-related biases.

Root Cause: Standard 3D volume registration does not fully correct for all motion-induced signal changes. Residual artefacts stem from:

• Partial Volume Effects: The resampling of the target image during realignment causes mixing of signals from different tissue types [12].
• Spin History Effects: Motion moves spins into and out of the imaging plane, altering their excitation history and leading to signal modulations [12].
• Intra-volume Motion: Motion occurring during the acquisition of a single volume cannot be corrected by inter-volume registration [12].

Solutions:

• Incorporate Slice-Level Correction: Use methods like SLOMOCO that perform slice-wise motion correction to account for intra-volume motion [12].
• Apply Nuisance Regressors: Regress out signals derived from motion parameters (e.g., Vol-mopa, Sli-mopa) and proposed regressors like the Partial Volume (PV) regressor to remove residual variance related to motion [12].
• Data Scrubbing: Identify and remove individual volumes that are excessively corrupted by motion [14]. Framewise displacement (FD) is a common metric for this [15].

Frequently Asked Questions (FAQs)

FAQ 1: Why is fMRI particularly sensitive to subject motion compared to other MRI types?

fMRI is exquisitely sensitive to motion because it detects very small signal changes (often 1-2%) related to the BOLD (Blood Oxygenation Level Dependent) effect [9] [15]. The primary data acquisition occurs in k-space (Fourier space), where each sample contains global spatial information about the entire image. Any motion during acquisition creates inconsistencies in k-space, which, upon reconstruction, manifest as artefacts like ghosting and blurring that can obscure or mimic genuine neural activity [10] [13].

FAQ 2: What is the fundamental link between k-space errors and image artefacts?

The final MR image is generated via an inverse Fourier transform of the acquired k-space data. Simple reconstruction assumes the object is perfectly stationary. Motion causes the k-space data to be an inconsistent mix of information from different object positions. This violation of the reconstruction model directly creates artefacts [10]. The specific nature of the motion determines the pattern of k-space corruption and, consequently, the type of artefact:

• Periodic Motion causes coherent modulation, leading to ghosting [10] [9].
• Slow, Continuous Motion causes a broader corruption, leading to blurring [10].

FAQ 3: What are the main categories of motion correction techniques?

Motion correction strategies can be broadly classified into two groups:

• Prospective Correction: Motion is detected and corrected during data acquisition. Examples include optical tracking or navigator echoes that adjust the imaging plane in real-time [11].
• Retrospective Correction: Motion is corrected after data acquisition. This includes volume-based realignment and more advanced slice-based methods, and often involves weighting or "scrubbing" of corrupted data volumes [11] [12] [16].

FAQ 4: How can I quantitatively assess the severity of ghosting artefacts in my data?

The intensity of ghosting artefacts can be mathematically described and quantified. For example, in interleaved EPI, the ghost kernel resulting from an amplitude discontinuity is given by a specific equation that includes the modulation parameters [9]. In quality assurance (QA) protocols, the Signal-to-Ghost Ratio (SGR) is a key metric used to evaluate system performance and can be applied to your data to quantify ghosting severity [15].

Quantitative Data Reference

Table 1: Characteristics of Common fMRI Motion Artefacts

Artefact Type	Primary Cause	K-space Origin	Typical Appearance	Correction Strategy
Ghosting	Periodic motion (respiration, pulsation), system phase errors [10] [9]	Phase/amplitude discontinuities between lines [9]	Replicas of the main image shifted along the phase-encode direction [10]	Reference scans, k-space reordering, post-processing phase correction [9]
Blurring	Slow, continuous motion (drift) [10]	Data inconsistency across entire k-space [10]	Loss of sharpness and fine detail [10]	Prospective motion correction, faster acquisition, registration [10] [11]
Signal Loss	Spin dephasing in magnetic field gradients [10]	Irreversible signal loss from magnetization evolution [10]	Dark regions in the image, often near tissue boundaries	Ensure subject comfort to minimize bulk motion, use spin-prep methods less sensitive to motion [10]

Table 2: Essential Tools for the fMRI Researcher's Toolkit

Tool / Reagent	Category	Primary Function	Example Software/Model
Retrospective Motion Correction	Software Package	Corrects for head motion after data acquisition by aligning volumes.	FSL (MCFLIRT), AFNI, SPM [11] [12]
Advanced Motion Correction Pipelines	Software Pipeline	Corrects for intra-volume motion and spin history effects at the slice level.	SLOMOCO [12]
Nuisance Regressor Regression	Data Cleaning Method	Removes residual motion artefact from the signal after motion correction.	Partial Volume (PV) Regressors, Vol-/Sli-mopa [12]
Data Scrubbing	Data Cleaning Method	Identifies and removes severely motion-corrupted volumes from the time series.	Framewise Displacement (FD) [14] [15]
QA Phantom	Physical Standard	Mimics brain properties to measure scanner stability and artefact levels for QA.	FBIRN Phantom [15]

Experimental Protocols

Protocol 1: Implementing Reference Scan Correction for Ghosting

Purpose: To measure and correct for system-induced phase offsets and timing delays that cause ghosting artefacts in interleaved EPI.

Methodology:

• Acquire Reference Scans: Prior to or following the main fMRI acquisition, run a calibration scan with the phase-encoding gradient turned off.
• Acquire Internal Reference Lines: Collect at least two "internal reference" lines to measure time delays [9].
• Reconstruct with Correction: Integrate the measured phase offsets and time delays into the reconstruction pipeline: a. Fourier transform the raw data in the readout direction. b. Apply the constant-phase shift corrections and phase rolls to the data to correct for the measured time shifts. c. Perform the final Fourier transform in the phase-encoding direction to generate the artefact-corrected image [9].

Protocol 2: A SIMPACE Pipeline for Validating Motion Correction

Purpose: To generate motion-corrupted fMRI data with known, user-defined motion parameters for validating motion correction algorithms.

Methodology:

• Phantom Setup: Use an ex vivo brain phantom fixed in a container to provide a stable, motionless baseline [12].
• Inject Motion: Utilize the Simulated Prospective Acquisition Correction (SIMPACE) sequence to synthetically alter the imaging plane coordinates before each volume and slice acquisition, emulating realistic intervolume and intravolume motion [12].
• Data Analysis: Process the simulated motion-corrupted data with different motion correction pipelines (e.g., VOLMOCO, SLOMOCO) and nuisance regressors. Compare the residual signal (e.g., standard deviation in gray matter) to quantify the efficacy of each correction method [12].

Visualizations

K-space Corruption and Image Artefacts

Motion Correction Pipeline for fMRI Data

Motion artifacts represent one of the most significant challenges in task-based neuroimaging research, potentially compromising data quality and leading to spurious findings. These artifacts originate from two primary sources: physiological motion (e.g., cardiac pulsation, respiration, and tremors) and voluntary motion (e.g., head movements, swallowing, and task-related movements). In functional magnetic resonance imaging (fMRI), even small movements can cause significant signal changes that may be misinterpreted as neural activity, particularly because the blood oxygen level-dependent (BOLD) signal change related to neuronal activity is typically only 1-2% [17]. The problem is especially pronounced in task-based paradigms where subjects must perform voluntary movements as part of the experimental design, creating a dual challenge of capturing intended motor activity while eliminating confounding motion artifacts.

The sensitivity of neuroimaging techniques to motion varies considerably. MRI is particularly vulnerable to motion due to prolonged acquisition times and the encoding of spatial information in frequency space (k-space), where motion causes inconsistencies that manifest as ghosts, blurring, and signal loss in reconstructed images [10]. Functional near-infrared spectroscopy (fNIRS), while more resilient to motion than fMRI, still faces significant artifact challenges from optode-scalp decoupling during head movements [18]. Understanding, mitigating, and correcting these artifacts is therefore essential for maintaining data integrity in behavioral tasks research.

Frequently Asked Questions (FAQs)

Q1: Why is MRI particularly sensitive to subject motion compared to other imaging modalities?

MRI's sensitivity to motion stems from its sequential data acquisition process in k-space (frequency space). Unlike photographic techniques that capture image data directly, MRI encodes spatial information through a series of measurements in k-space that are later transformed into images using Fourier transformation. When motion occurs during this sequential acquisition, it creates inconsistencies in k-space data that manifest as various artifacts in the reconstructed images, including blurring, ghosting, and signal loss [10]. This problem is exacerbated in longer acquisitions and techniques with inherent low signal-to-noise ratio like fMRI and diffusion tensor imaging.

Q2: What are the main types of motion artifacts encountered in task-based neuroimaging?

Motion artifacts generally fall into three categories:

• Rigid body motion: Movement of the head or body as a single unit, causing shifted or rotated imaging data
• Non-rigid body motion: Deformation or shape changes during scanning, causing more complex artifacts
• Physiological motion: Motion from breathing, heartbeat, blood flow, or other biological processes [19]

In task-based paradigms, voluntary movements required by the experimental design introduce additional rigid and non-rigid motions that can be difficult to distinguish from the neural signals of interest.

Q3: How do motion artifacts affect statistical analysis in task-based fMRI?

Motion artifacts can induce signal changes that confound statistical analysis in multiple ways. In worst-case scenarios, motion-related signal changes may become correlated with task activation patterns, leading to false positives or inflated effect sizes. Motion also introduces structured noise that violates the assumption of independent and identically distributed Gaussian noise in standard statistical models [17]. Even after motion correction, residual artifacts can reduce statistical power and validity.

Q4: Can motion artifacts mimic genuine brain network activity?

Yes, research has demonstrated that motion artifacts can produce spatial patterns resembling genuine functional connectivity networks. Carefully mapped head movement artifacts have been shown to create patterns similar to the default mode network, potentially leading to misinterpretation of network connectivity in resting-state and task-based studies [17]. This is particularly problematic for studies of populations with different movement characteristics, such as children, elderly individuals, or patients with movement disorders.

Troubleshooting Guides

Prevention Strategies During Data Acquisition

Subject Preparation and Positioning

• Comfort Optimization: Use foam padding and comfortable head restraints to minimize voluntary movement. Ensure the subject is properly positioned and comfortable before scanning begins.
• Comprehensive Pre-scan Briefing: Clearly explain the importance of remaining still, providing specific examples of problematic movements (e.g., swallowing, head adjustments). For task-based paradigms, practice the movement outside the scanner to minimize excessive motion.
• Physiological Monitoring: Implement cardiac and respiratory monitoring for later gating or noise correction, particularly for paradigms sensitive to physiological noise.

Sequence Optimization

• Acquisition Parameter Adjustment: Utilize faster imaging sequences (e.g., parallel imaging, multiband acquisitions) to reduce motion sensitivity by shortening acquisition windows [10].
• K-space Reordering Strategies: Employ motion-resistant k-space sampling trajectories such as radial, PROPELLER, or spiral sequences that are less sensitive to motion artifacts than standard Cartesian sampling [10].
• Gating and Triggering: Implement cardiac gating for sequences sensitive to pulsatility (e.g., diffusion imaging) and respiratory gating for abdominal or thoracic imaging.

Motion Tracking and Monitoring

Hardware-Based Motion Tracking

• Camera-Based Systems: Optical tracking systems using cameras to monitor head position in real-time, allowing for prospective motion correction [19].
• Sensor-Based Systems: Inertial measurement units (IMUs), accelerometers, or gyroscopes attached to the subject to provide continuous motion data [19].
• MR-Compatible Optical Tracking: Systems specifically designed to operate within the MR environment, providing real-time head position data for prospective correction.

Image-Based Motion Tracking

• Navigator Echoes: Brief additional MR acquisitions interspersed throughout the sequence to detect subject position.
• Image-Based Registration: Real-time image registration to detect inter-volume motion for prospective correction.

Post-Processing Correction Methods

Retrospective Motion Correction

• Rigid-Body Registration: Standard volume realignment using six-parameter (three translation, three rotation) rigid body transformation to correct for inter-volume motion [17].
• Slice-Based Correction: Methods like SLOMOCO that perform slice-wise motion correction to address intra-volume motion, which is particularly effective for spin-echo sequences [12].
• Advanced Registration Algorithms: Non-linear registration techniques that can correct for non-rigid motion and deformations.

Nuisance Regression Techniques

• Motion Parameter Regression: Including the six motion parameters (and their derivatives) as regressors in general linear model analysis to account for motion-related variance.
• Spike Regression: Identifying and regressing out volumes with excessive motion ("spikes") to minimize their influence on statistical results.
• Partial Volume Regressors: Voxel-wise motion nuisance regressors that account for partial volume effects at tissue boundaries [12].

Data-Driven Cleaning Approaches

• Independent Component Analysis: ICA-based methods like ICA-AROMA that automatically identify and remove motion-related components from fMRI data without requiring manual classification [20].
• Principal Component Analysis: Using PCA to identify and remove major sources of variance attributable to motion.
• CompCor: A component-based noise correction method that identifies noise regions of interest (e.g., white matter, CSF) and removes their signal contributions.

Table 1: Comparison of Motion Correction Techniques

Technique	Principle	Advantages	Limitations	Best Suited For
Volume-Based Registration	3D rigid-body transformation between volumes	Widely available, computationally efficient	Cannot correct intra-volume motion	Block-design fMRI with minimal movement
Slice-Based Correction (SLOMOCO)	Motion correction at slice acquisition level	Addresses spin history effects, handles intra-volume motion	More computationally intensive	Sequences with long TR, high-resolution fMRI
ICA-AROMA	Automatic classification and removal of motion components	No manual classification needed, preserves temporal structure	May remove neural signal in some cases	Resting-state and task-based fMRI
Prospective Motion Correction	Real-time adjustment of imaging plane	Prevents artifacts rather than correcting them	Requires specialized hardware/sequences	Populations prone to movement (children, patients)
Retrospective Correction with SIMPACE	Uses simulated motion data for optimization	Provides gold-standard correction validation	Currently limited to research settings	Methodological development and validation

Experimental Protocols for Motion Management

Protocol for Combined fMRI-fNIRS in Task Paradigms

Rationale: Combining fMRI's high spatial resolution with fNIRS's superior temporal resolution and motion resilience provides complementary data streams, with fNIRS serving as a validation tool for fMRI findings in moving subjects [21].

Equipment Setup

• Synchronize fMRI and fNIRS acquisition systems using a common trigger signal
• Position fNIRS optodes to cover regions of interest while minimizing interference with MRI head coils
• Use MRI-compatible fNIRS systems with non-magnetic components and filtered cabling
• Implement motion tracking for both modalities (camera-based for fNIRS, navigators for fMRI)

Data Acquisition Parameters

• fMRI: Use multiband EPI sequences with in-plane acceleration to reduce TR and motion sensitivity
• fNIRS: Employ dual-wavelength systems (e.g., 760 nm and 850 nm) with sampling rates ≥10 Hz
• Motion Tracking: Record camera-based optode movement tracking for fNIRS and volumetric navigators for fMRI simultaneously

Processing Pipeline

• Preprocessing: Apply standard preprocessing separately to fMRI (motion correction, smoothing) and fNIRS (filtering, motion artifact correction) data
• Motion Detection: Identify motion-contaminated epochs in both modalities using threshold-based approaches
• Temporal Alignment: Precisely align fMRI and fNIRS time series using recorded trigger pulses
• Hemodynamic Response Comparison: Compare motion-corrected hemodynamic responses across modalities to validate correction efficacy

Protocol for Residual Motion Artifact Removal Using SIMPACE

Rationale: The SIMPACE (Simulated Prospective Acquisition Correction) method uses ex vivo brain phantoms to generate gold-standard motion-corrupted data, enabling optimization of motion correction pipelines with known ground truth [12].

Phantom Preparation

• Prepare formalin-fixed ex vivo brain phantom soaked in Fomblin to remove bubbles
• Position phantom in 3D-printed holder within standard head coil
• Ensure temperature stabilization before data acquisition

SIMPACE Data Acquisition

• Acquire reference 2D EPI data without motion simulation
• Program SIMPACE sequence with predefined motion patterns (intervolume and intravolume)
• Inject varying degrees of motion (translation, rotation, and combined movements)
• Repeat for different k-space sampling trajectories (linear, interleaved)

Motion Correction Pipeline Optimization

• Apply standard volume-based correction (e.g., FSL MCFLIRT) and calculate residual artifacts
• Implement slice-wise correction (SLOMOCO) addressing intravolume motion
• Apply partial volume regressors to address residual boundary artifacts
• Compare performance against ground truth phantom data
• Iteratively refine parameters to minimize residual motion in gray matter

Validation Metrics

• Standard deviation of residual time series in gray matter regions
• Quantitative comparison with ground truth reference data
• Reduction in motion-to-signal ratio across different motion patterns

Table 2: Quantitative Performance of Motion Correction Methods on SIMPACE Data [12]

Correction Method	Motion Parameters Used	Residual Noise Reduction (1× motion)	Residual Noise Reduction (2× motion)	Computational Time
VOLMOCO (Volume-based)	6 Vol-mopa + PV regressors	Baseline	Baseline	Fastest
Original SLOMOCO	14 voxel-wise regressors	12% improvement	14% improvement	Moderate
Modified SLOMOCO	12 Vol-/Sli-mopa + PV regressors	29% improvement	45% improvement	Longest
ICA-AROMA	Automated component removal	22% improvement	31% improvement	Fast-Moderate

Visualization of Motion Correction Workflows

Motion Artifact Correction Pipeline

Multimodal Integration for Motion Resilience

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Tools for Motion Management in Neuroimaging Research

Tool/Category	Specific Examples	Function	Implementation Considerations
Motion Tracking Hardware	Camera-based systems (MRC Systems), inertial measurement units (IMUs), MR-compatible optical tracking	Real-time monitoring of subject movement	Compatibility with imaging environment, sampling rate, accuracy
Post-Processing Software	FSL (MCFLIRT), SPM, AFNI, SLOMOCO, ICA-AROMA	Retrospective motion correction and artifact removal	Compatibility with data format, computational demands, ease of use
Phantom Systems	SIMPACE-compatible phantoms, custom motion platforms	Validation and optimization of correction methods	Reproducibility of motion patterns, tissue-like properties
Multimodal Integration Platforms	NIRS-KIT, Homer2, NIRS-ICA, SPM-fNIRS	Co-registration and joint analysis of multiple data modalities	Data synchronization, coordinate system alignment
Reference-Based Noise Correction	ECG, respiratory belt, carbon wire loops, component-based noise correction (CompCor)	Identification and removal of physiological noise	Signal quality, temporal precision relative to imaging data
Accelerated Imaging Sequences	Multiband fMRI, compressed sensing, parallel imaging	Reduction of acquisition time to minimize motion window	Signal-to-noise tradeoffs, hardware requirements
Motion-Resistant Acquisition	PROPELLER, radial, spiral k-space trajectories	inherently motion-resistant data acquisition	Reconstruction complexity, sequence availability
2-(Prop-2-ynyloxy)ethyl acetate	2-(Prop-2-ynyloxy)ethyl acetate, CAS:39106-97-3, MF:C7H10O3, MW:142.15 g/mol	Chemical Reagent	Bench Chemicals
Vinyl oleate	Vinyl oleate, CAS:3896-58-0, MF:C20H36O2, MW:308.5 g/mol	Chemical Reagent	Bench Chemicals

Core Problem: How Motion Artifacts Compromise Neuroimaging Data Integrity

Motion artifacts introduce non-neuronal noise that systematically biases neuroimaging data, leading to false conclusions about brain function and connectivity. Even minor, unavoidable movements—from breathing, cardiac cycles, or small head shifts—generate signal changes that can mimic, mask, or distort true neural activity. In functional MRI (fMRI), these artifacts are particularly problematic because the blood-oxygen-level-dependent (BOLD) signal changes reflecting neural activity are very small (typically 1–2%), making them highly susceptible to contamination by motion-induced noise [17] [22]. This contamination manifests as two primary threats to data integrity: signal loss (reduced sensitivity to detect true effects) and spurious findings (increased false positives in connectivity and activation maps).

The fundamental challenge is that motion artifacts are not random noise. They introduce spatiotemporally structured patterns that can be misinterpreted as biologically plausible neural processes. For instance, studies have shown that motion artifacts can create spatial patterns that resemble the brain's default mode network—a key resting-state network—during functional connectivity analysis [17] [22]. This occurs because motion often causes signal changes at tissue boundaries (e.g., at the borders between gray matter, white matter, and cerebrospinal fluid), creating structured artifacts that confound statistical analyses [10].

Table 1: Primary Motion Artifact Types and Their Direct Impacts on Data

Artifact Type	Primary Cause	Impact on Data Integrity
Ghosting/Blurring	Periodic motion (respiration, cardiac pulse) [23]	Reduces spatial accuracy; creates false replicas of anatomy [10]
Spin History Effects	Movement altering proton excitation history [12]	Causes local signal loss or gain that mimics activation/deactivation [12]
Magnetic Field Distortions	Head movement in B0 field [24] [12]	Introduces geometric distortions and voxel misplacement [12]
Physiological Noise	Cardiorespiratory cycles, blood flow [17]	Creates periodic signal changes that confound connectivity analyses [17]

Quantitative Impacts: From Signal Loss to Spurious Connectivity

The consequences of motion are quantifiable and severe. In fMRI, head motion has been identified as one of the main sources of bias, with residual motion artifacts persisting even after standard correction procedures [12]. The impact is particularly pronounced in studies requiring high precision, such as cortical thickness measurements, where motion can create the false appearance of cortical thinning, mimicking disease-related atrophy [25].

The relationship between motion severity and data corruption follows a dose-response pattern. For example:

• In diffusion tensor imaging (DTI), longer acquisition times (4-5 minutes for 20-60 gradient directions vs. 2 minutes for 3 directions) significantly increase sensitivity to motion, resulting in misaligned data and introduced noise that compromises fiber tracking reliability [17] [22].
• In functional connectivity analyses, motion artifacts induce spurious correlations between brain regions that do not actually share functional relationships. This occurs because motion affects different brain regions in a coordinated manner, creating the illusion of connected networks [17].
• In activation maps for task-based fMRI, motion-related signal changes may accidentally correlate with the task timing, leading to false positive activations. Conversely, motion can also mask true activations, resulting in false negatives [22].

Table 2: Impact of Motion on Different Neuroimaging Modalities

Imaging Modality	Key Vulnerability	Consequence for Data Interpretation
Resting-State fMRI	Low signal-to-noise ratio; sensitive to slow drifts [17] [22]	Spurious functional connectivity; corrupted network maps [17] [26]
Task-Based fMRI	Temporal correlation with task design [22]	False positive/negative activation; biased group comparisons [22]
Diffusion MRI (DTI)	Long acquisition times; sensitivity to misalignment [17]	Inaccurate fiber tracking; compromised white matter integrity measures [17]
Structural MRI (T1-weighted)	High spatial resolution demands [25]	Compromised cortical surface reconstructions; inaccurate volumetry [25]
Simultaneous EEG-fMRI	Electromagnetic induction from motion [24]	Neuronal signals masked by imaging/ballistocardiogram artifacts [24]

Diagram 1: Motion artifact impact pathway on data integrity.

Methodological Guide: Motion Mitigation Protocols

Acquisition-Based Prevention Strategies

Participant Preparation and Positioning: Meticulous attention to participant comfort and stabilization is the first line of defense. Use vacuum cushions and adjustable head restraints to minimize head movement. Provide clear instructions about the importance of staying still, and consider practice sessions in a mock scanner to acclimatize participants. For special populations (children, patients with movement disorders), appropriate sedation or anesthesia may be necessary [23].

Sequence Optimization and Hardware Selection: Implement fast imaging sequences (e.g., gradient echo, multiband EPI) to reduce acquisition time and motion probability [10] [22]. Utilize cardiac gating for pulsation artifacts and breath-hold timing or respiratory triggering for abdominal/chest imaging [23]. Higher channel count phased-array coils improve signal-to-noise ratio, potentially mitigating some motion effects [17]. For EEG-fMRI simultaneously, consider carbon-wire loop systems to specifically capture MR-induced artifacts for later removal [24].

Prospective Motion Correction: These real-time approaches adjust the imaging plane during acquisition based on detected motion. Optical tracking systems monitor head position and orientation, while navigator echoes or real-time image registration can be embedded in sequences to prospectively correct for motion [12]. The SIMPACE (simulated prospective acquisition correction) sequence represents an advanced approach that alters the imaging plane coordinates before each volume and slice acquisition to emulate motion-free conditions [12].

Post-Processing Correction Techniques

Volume-Based Registration: The most common retrospective correction method involves 3D rigid body transformation with six parameters (three translational, three rotational) to realign each volume to a reference volume [22] [12]. While essential, this approach alone is insufficient as it cannot address spin history effects or intravolume motion [12].

Slice-Level Correction: Advanced methods like SLOMOCO (slice-oriented motion correction) address intravolume motion by applying slice-wise rigid motion parameters, significantly outperforming volume-based methods alone. As demonstrated in studies using the SIMPACE sequence, combining volume-wise and slice-wise motion parameters with partial volume regressors reduced residual motion artifacts by 29-45% compared to traditional approaches [12].

Nuisance Regression and Data Cleaning: After motion correction, residual motion artifacts must be addressed through nuisance regression. This includes:

• Expanding motion parameters (including temporal derivatives and squared terms) to capture nonlinear relationships [12]
• Implementing spike regression (or "scrubbing") to remove severely motion-corrupted volumes [12]
• Incorporating tissue-based regressors (white matter and CSF signals) to remove non-neuronal signals [12]
• Applying data-driven approaches like ICA-AROMA to automatically identify and remove motion-related components [24]

Diagram 2: Motion mitigation protocol workflow.

The Scientist's Toolkit: Essential Research Reagents & Solutions

Table 3: Key Research Reagents and Computational Tools for Motion Management

Tool/Reagent	Function/Purpose	Application Context
SIMPACE Sequence	Simulates motion-corrupted data by altering imaging plane coordinates [12]	Method validation; algorithm testing
SLOMOCO Pipeline	Implements slice-wise motion correction for intravolume motion [12]	fMRI preprocessing; motion correction
3D CNN Correction	Deep learning approach for retrospective motion correction in structural MRI [25]	T1-weighted image enhancement
Carbon-Wire Loops (CWL)	Reference system capturing MR-induced artifacts in EEG-fMRI [24]	Artifact removal in simultaneous EEG-fMRI
WCBSI Algorithm	Combined wavelet and correlation-based signal improvement for fNIRS [27]	Motion artifact correction in fNIRS data
Point-Process Analysis	Sparse representation of BOLD signals as discrete events [26]	Dimensionality reduction; noise filtering
Rigid Body Transformation	Six-parameter (3 translation, 3 rotation) volume realignment [22]	Standard motion correction across modalities
Zilantel	Zilantel, CAS:22012-72-2, MF:C26H38N2O6P2S4, MW:664.8 g/mol	Chemical Reagent
5,5-Dibutyldihydrofuran-2(3H)-one	5,5-Dibutyldihydrofuran-2(3H)-one|High-Purity|RUO	Get 5,5-Dibutyldihydrofuran-2(3H)-one for your lab. This high-purity lactone is For Research Use Only. Not for human consumption.

Frequently Asked Questions (FAQs)

Q1: Our group comparisons show significant cognitive network differences, but motion was greater in our patient group. Are these results valid? This is a critical confound. Motion artifacts can create spurious group differences that mimic disease effects. You must:

• Quantify motion metrics (mean framewise displacement, DVARS) for both groups [22]
• Include motion as a covariate in group analyses
• Consider implementing matched motion between groups through subgroup selection
• Apply rigorous motion correction (slice-level correction + nuisance regression) [12] Without these steps, your network differences may reflect motion artifacts rather than true neural effects.

Q2: We're studying a population with inherent movement disorders. What acquisition strategies are most effective? Prioritize fast acquisition sequences to minimize motion probability [10]. Consider:

• Multiband EPI for reduced scan times
• Prospective motion correction (PACE) if available [12]
• Smaller voxel sizes can sometimes reduce spin history effects
• For structural imaging, use sequences less sensitive to motion (e.g., MP-RAGE) Always acquire multiple runs to ensure some usable data if others are severely corrupted.

Q3: After volume realignment, why do we still see motion artifacts in our connectivity matrices? Volume-based correction alone cannot address spin history effects or intravolume motion [12]. The signal changes from these artifacts persist as structured noise. Implement:

• Slice-level motion correction (e.g., SLOMOCO) [12]
• Additional nuisance regression using:
  • Expanded motion parameters (including derivatives and squares)
  • Tissue-based regressors (white matter, CSF signals)
  • Spike regression for severely corrupted volumes
  • Data-driven approaches like ICA-AROMA

Q4: How does motion specifically create spurious functional connectivity? Motion creates coordinated signal changes across the brain that are not neuronal in origin. Specifically:

• Motion affects signal intensity most prominently at tissue boundaries [10]
• These signal changes occur simultaneously across distant brain regions
• Standard correlation metrics interpret these coordinated artifacts as "functional connections"
• The resulting patterns can resemble biologically plausible networks like the default mode network [17] [22]

Q5: For cortical thickness analysis, how much motion is acceptable before data should be excluded? There's no universal threshold, but studies show that even subtle motion can compromise cortical surface reconstructions [25]. Implement quality control metrics such as:

• Quantitative image quality metrics (e.g., CNR, SNR)
• Visual inspection for gray-white boundary integrity
• Cortical surface reconstruction success rates Consider deep learning-based correction methods, which have shown success in recovering usable data from motion-corrupted structural scans [25].

The Correction Toolbox: Proven and Novel Mitigation Strategies

FAQs: Addressing Motion Artifacts in Neuroimaging Research

1. What are the most common causes of motion artifacts in neuroimaging? Motion artifacts are primarily caused by subject movement during the scan. This includes gross involuntary movements, but also physiological motion from cardiac pulsation, respiration, and tremors. In the context of behavioral tasks, even small movements like button presses can introduce artifacts that degrade image quality [10].

2. Why is neuroimaging particularly sensitive to patient motion? Magnetic resonance imaging (MRI) is highly sensitive to motion because it requires a long time to collect sufficient data to form an image. This acquisition time is often far longer than the timescale of most physiological motions. The process of spatial encoding in k-space means that even small, transient movements can cause inconsistencies in the data, resulting in blurring, ghosting, or signal loss in the final image [10].

3. How can we proactively manage patient anxiety to reduce motion? Non-pharmacological interventions are the first line of defense. Creating a calm environment by minimizing noise and light is recommended. Furthermore, providing patients with clear information about the scanning procedure can reduce anxiety. For some patients, relaxation techniques or music therapy may be effective [28].

4. When is sedation considered, and what are the current best practices? Sedation may be necessary for patients who cannot remain still, such as certain pediatric populations or patients with conditions that cause involuntary movements. Recent 2025 clinical practice guidelines from the Society of Critical Care Medicine conditionally recommend using dexmedetomidine over propofol for sedation in adults, as it may promote more favorable outcomes. The guidelines strongly emphasize using the lightest effective level of sedation to keep patients more awake and alert, which is associated with better recovery [29] [28].

5. What are the key considerations for immobilization? While physical immobilization using head straps and padding is common and useful, it must be balanced with patient comfort. Discomfort from excessive restraint can itself lead to movement. The goal of immobilization is to minimize motion without causing stress or anxiety, which requires careful setup and communication with the patient [10].

Troubleshooting Guide: Identifying and Mitigating Motion Artifacts

Problem Symptom	Potential Cause	Proactive Prevention Strategy	Corrective Action
Ghosting or replication of structures in the phase-encoding direction [10]	Periodic motion (e.g., respiration, cardiac pulsation)	Use prospective motion correction (MoCo) sequences or cardiac/respiratory gating where available [10].	Consider re-acquiring the sequence with a different phase-encoding direction to change the artifact's orientation.
General blurring of image details [10]	Slow, continuous patient drift or gross involuntary movement	Optimize patient comfort with padding, use vacuum immobilization mats, and provide clear instructions on the importance of staying still.	Implement post-processing motion correction algorithms or use a sequence less sensitive to motion (e.g., single-shot EPI).
Signal loss in specific areas [10]	Spin dephasing due to movement during diffusion-sensitizing gradients or other contrast preparation.	Ensure secure head immobilization and consider sedation protocols for at-risk populations [29] [28].	Re-scan with a reduced echo time (TE) or a sequence with reduced motion sensitivity, if diagnostically acceptable.
Anxiety and agitation in the scanner, leading to motion	Claustrophobia, scanner noise, or underlying medical condition.	Conduct a pre-scan rehearsal, use a mirror for visual exit, provide earplugs/headphones, and employ non-pharmacological anxiety reduction techniques [28].	Follow institutional sedation protocols, which may include short-acting agents like dexmedetomidine [29] [28].

Experimental Protocols for Motion Mitigation

Protocol 1: Pre-Scan Patient Preparation and Comfort Optimization

Objective: To minimize motion at its source by reducing patient anxiety and maximizing physical comfort.

• Screening and Communication: Prior to the scan, screen patients for claustrophobia and anxiety. Explain the entire procedure, including the sounds and duration, and emphasize the critical importance of remaining still.
• Immobilization Setup: Use comfortable but firm foam padding around the patient's head within the head coil. Secure the head with a strap without causing discomfort.
• Environmental Adjustments: Offer the patient earplugs or noise-canceling headphones to reduce acoustic noise. Ensure the scanner bore is well-ventilated. For longer scans, consider placing a mirror to give the patient a view outside the bore.
• Task Practice: If the session involves a behavioral task, have the patient practice the task outside the scanner to build familiarity and reduce movement associated with task uncertainty [30].

Protocol 2: Implementation of Sedation for Motion-Prone Populations

Objective: To safely administer sedation to ensure scan viability in patients who cannot otherwise remain still. Note: This protocol must be conducted by qualified medical personnel following institutional regulations.

• Patient Selection and Assessment: Identify patients requiring sedation (e.g., young children, patients with movement disorders). Obtain informed consent and perform a pre-sedation health assessment.
• Sedation Protocol: Based on the latest 2025 guidelines, a protocol using Dexmedetomidine is preferred for its sedative and analgesic properties with a lower delirium risk profile compared to other sedatives [29] [28].
  • Loading Dose: Administer 0.5–1 mcg/kg over 10 minutes.
  • Maintenance Infusion: Titrate between 0.2–1 mcg/kg/hr to achieve a light sedation level where the patient is sleepy but rousable.
• Monitoring: Continuously monitor vital signs (heart rate, blood pressure, oxygen saturation) throughout the scan. Use capnography if available.
• Recovery: Monitor the patient in a dedicated recovery area until standard discharge criteria are met.

Research Reagent Solutions for Motion Management

Reagent / Material	Function / Application in Research
Dexmedetomidine	A sedative and analgesic used in research protocols to facilitate motion-free scanning in awake or lightly sedated subjects, valued for its minimal impact on respiratory drive [29].
Foam Padding & Vacuum Immobilization Mats	Essential non-invasive tools for comfortably stabilizing the subject's head and body within the scanner, reducing motion from muscle relaxation and discomfort.
Electroencephalography (EEG) Cap with Motion Sensors	Integrated systems (e.g., with accelerometers) used to quantitatively measure head motion in real-time, providing data for prospective or retrospective motion correction algorithms [31].
Multi-dimensional Experience Sampling (mDES)	A validated questionnaire battery administered to subjects during task performance to collect introspective reports on psychological state, which can be correlated with motion-prone brain states [32].
Deep Convolutional Neural Network (CNN)	A class of deep learning models that can be trained to automatically rate motion artifacts in neuroimages, enabling rapid quality control of large datasets [33].

Workflow for a Proactive Motion Prevention Strategy

The diagram below outlines a logical workflow for implementing a comprehensive strategy to prevent motion artifacts, integrating comfort, immobilization, and sedation.

Technical Troubleshooting Guide: FAQ on Acquisition-Based Motion Reduction

Q1: What are the primary acquisition-based strategies for mitigating motion artifacts in MRI? The main strategies can be categorized into three groups: (1) Prospective Motion Correction, which actively adjusts the imaging sequence in real-time based on detected motion (e.g., using navigator echoes or external tracking systems); (2) Gating, which acquires data only during specific phases of a periodic motion (e.g., cardiac or respiratory cycle); and (3) Ultrafast Imaging, which uses very short acquisition times to "freeze" motion [34] [35].

Q2: When should I use navigator echoes versus external optical tracking for prospective motion correction? The choice depends on your experimental needs. Navigator echoes are integrated into the pulse sequence and do not require additional hardware; they are particularly well-suited for tracking periodic motions like diaphragm movement during respiration [36]. External optical motion tracking systems use a camera to track markers placed on the subject, correcting for arbitrary rigid-body motion. They are ideal for imaging freely moving objects or for neuroimaging studies where even small head movements can degrade image quality [37].

Q3: My cardiac-triggered coronary artery images still show blurring. What gating parameter should I check? This is often due to respiratory motion. You should implement respiratory gating in addition to cardiac triggering. Use a navigator echo placed on the right hemidiaphragm and set a narrow gating window (e.g., 5 mm). The scanner will only acquire data when the diaphragm position falls within this window, significantly reducing respiratory blurring. Be aware that this will reduce scanning efficiency and increase total scan time [36].

Q4: What is a major limitation of using a simple linear model for diaphragm-based correction of heart motion? Research using multiple navigators has shown that the relationship between diaphragm and heart motion is not perfectly linear and often exhibits patient-dependent hysteresis. This means that for the same diaphragm position, the actual position of the heart can differ between the inspiration and expiration phases. A simple linear model can therefore lead to residual motion artifacts; a more complex, calibrated model is recommended for high-precision applications [36].

Q5: For an uncooperative patient, should I use sedation or an ultrafast sequence? Whenever ethically and medically feasible, ultrafast sequences (e.g., HASTE, EPI) should be attempted first. These sequences can acquire images in 2-5 seconds, often fast enough to freeze bulk motion without the need for sedation, which simplifies clinical workflow and reduces patient risk [35].

The following tables summarize key performance metrics and parameters for the acquisition-based solutions discussed.

Table 1: Performance Metrics of Motion Reduction Techniques

Technique	Reported Accuracy/Performance	Primary Application Context	Key Limitations
Navigator Echo (for real-time gating)	Gating window of 5 mm provided reproducible image quality for coronary arteries [36].	Respiratory motion compensation in cardiac imaging [36].	Reduces gating efficiency (20-60%), increasing scan time [36].
External Optical Motion Tracking	Enabled imaging of freely moving objects without motion-related artefacts [37].	Prospective correction of arbitrary rigid body motion in neuroimaging [37].	Requires additional external hardware and camera system setup [37].
Ultrafast Sequences (e.g., HASTE, EPI)	Acquisition times of 2-5 seconds can freeze bulk motion [35].	Imaging uncooperative patients or reducing specific artifacts (e.g., in abdominal imaging) [35].	May have lower signal-to-noise ratio or contrast compared to conventional sequences [35].
Deep CNN for Motion Rating	100% acquisition-based accuracy on test set; 90.3% on generalization epilepsy dataset [33].	Reference-free automated quality evaluation of MR images for motion artifact rating [33].	Performance can drop (e.g., 63.6%) on data from different domains/scanners without adaptation [33].

Table 2: Key Parameters for Navigator Echo Implementation

Parameter	Typical Value / Setting	Explanation and Impact
Pencil Beam Diameter	~25 mm [36]	Defines the spatial region being monitored. A larger diameter averages over a larger area.
Navigator Total Duration	~20 ms [36]	Includes excitation, acquisition, and evaluation time. Determines how frequently motion can be sampled.
Displacement Accuracy	<1 mm [36]	The precision with which the navigator can detect position changes, achieved via sub-pixel interpolation.
Gating Window	5 mm (example for coronary imaging) [36]	The range of motion within which data is accepted. A smaller window improves quality but lowers efficiency.
Gating Efficiency	20% - 60% [36]	The percentage of accepted data acquisitions. Patient-dependent and inversely related to gating window tightness.

Experimental Protocol: Implementing Navigator Echoes for Respiratory Gating

Objective: To integrate a navigator echo for respiratory gating in a cardiac-triggered 3D coronary MR angiography sequence to mitigate respiratory motion artifacts during free breathing.

Materials and Equipment:

• 1.5 T or 3 T whole body MRI scanner.
• ECG triggering equipment.
• The pulse sequence must support the integration of a navigator echo pulse.

Step-by-Step Methodology:

• Subject Setup: Position the subject in the scanner in a supine position. Attach the ECG electrodes for cardiac triggering.
• Localizer and Planning: Acquire standard survey scans (e.g., coronal, sagittal, transverse) to locate the heart and the diaphragm.
• Navigator Positioning: On a coronal localizer image, graphically position the pencil beam navigator so that it passes through the dome of the right hemidiaphragm, creating a clear lung-liver tissue interface [36]. The Fourier transform of the signal from this beam creates a "navigator profile" with a sharp edge.
• Sequence Integration: Integrate the navigator pulse into the coronary MRA sequence. A typical navigator pulse sequence (excitation, acquisition, and evaluation) has a duration of about 20 ms [36].
• Reference Acquisition: Before starting the main scan, acquire a reference navigator profile.
• Set Gating Parameters: Define the acceptance window (e.g., 5 mm). During scanning, the system will use a cross-correlation algorithm in real-time to compare each new navigator profile to the reference. Data is only accepted if the measured displacement of the diaphragm is within the specified 5 mm window [36].
• Data Acquisition: Start the ECG-triggered, segmented k-space 3D gradient echo sequence. The scanner will continuously monitor the diaphragm position and only acquire imaging data when the respiratory position is within the gating window.
• Monitoring: Monitor the "gating efficiency" (the percentage of accepted acquisitions) throughout the scan. Be prepared for longer total scan times, as efficiency can vary between 20% and 60% depending on the subject's breathing pattern [36].

Visual Workflows

The following diagram illustrates the logical workflow and relationship between the primary acquisition-based solutions for motion artifact reduction.

The Scientist's Toolkit: Research Reagent Solutions

This table details the essential "research reagents" – the key hardware, software, and sequence components required for implementing the featured motion reduction techniques.

Table 3: Essential Materials for Motion Artifact Experiments

Item Name / Solution	Category	Primary Function in Motion Reduction
2D RF Pulse (Spiral Trajectory)	Pulse Sequence Component	Creates a spatially selective "pencil beam" for exciting the navigator echo, which is used to monitor tissue position [36].
External Optical Motion Tracking System	Hardware	Tracks the position of markers placed on the subject in real-time, allowing the scanner to prospectively correct the imaging volume position prior to each excitation [37].
ECG Triggering Device	Hardware	Detects the cardiac cycle (R-wave) to prospectively trigger the start of data acquisition during a specific, stable phase of the heart cycle (e.g., diastole) [34] [35].
Respiratory Belt or Bellows	Hardware	Monitors the expansion and contraction of the chest wall for use in respiratory triggering or gating [35].
Ultrafast Sequences (e.g., HASTE, EPI)	Pulse Sequence	Acquires images rapidly (in seconds or less) to minimize the time during which motion can occur, effectively "freezing" motion [35].
Radial/Spiral k-space Trajectories	Pulse Sequence Design	Disperses motion artifacts throughout the image rather than concentrating them as discrete ghosts, which is common with Cartesian trajectories [35].
N-(4-Bromobenzyl)-N-ethylethanamine	N-(4-Bromobenzyl)-N-ethylethanamine|4885-19-2	N-(4-Bromobenzyl)-N-ethylethanamine (CAS 4885-19-2), a tertiary amine for synthetic chemistry. For Research Use Only. Not for human or veterinary use.
Pyrrolo[1,2-c]pyrimidin-1(2H)-one	Pyrrolo[1,2-c]pyrimidin-1(2H)-one|CAS 223432-96-0	High-purity Pyrrolo[1,2-c]pyrimidin-1(2H)-one for research. A key heterocyclic scaffold in medicinal chemistry. For Research Use Only. Not for human consumption.

Troubleshooting Guides

Guide 1: Troubleshooting ICA-AROMA for Task-Based fMRI

Problem 1: Poor Component Classification After Running ICA-AROMA

• Symptoms: The classifier fails to identify obvious motion-related components, or too many neural components are mistakenly flagged as noise.
• Potential Causes & Solutions:
  • Cause: Incorrect feature calculation due to misalignment between high-resolution brain extraction and functional data.
  • Solution: Ensure the brain mask derived from your high-resolution structural image is accurately registered to your functional space. Visually inspect the overlap.
  • Cause: The motion parameters used for feature extraction are not representative of the true head displacement.
  • Solution: Verify the quality of your motion parameter estimation by plotting frame-wise displacement. Check for outliers or failed registrations.
  • Cause: Unusual motion patterns that fall outside the classifier's trained feature space.
  • Solution: Manually inspect the classified components. ICA-AROMA allows for manual intervention to override the automatic classification if necessary [20] [38].

Problem 2: Over-Aggressive Denoising Leading to Loss of Neural Signal

• Symptoms: A significant reduction in task-based activation, particularly in regions known to be susceptible to motion (e.g., prefrontal cortex).
• Potential Causes & Solutions:
  • Cause: The threshold for the classifier's decision boundary is too sensitive for your specific dataset.
  • Solution: While ICA-AROMA is largely automatic, some implementations allow for a "non-aggressive" denoising option. This regresses out the time courses of noise components rather than fully removing them, helping to preserve the variance of the signal of interest [20].
  • Cause: The data has a very low signal-to-noise ratio (SNR), making it difficult to distinguish motion from neural activity.
  • Solution: This is a fundamental data quality issue. Consider acquiring more data points or using a multi-echo sequence to improve ICA component estimation in future studies.

Guide 2: Troubleshooting Structured Low-Rank Matrix Completion

Problem 1: High Computational Demand and Memory Usage

• Symptoms: The algorithm runs very slowly or fails due to insufficient memory, especially with high-resolution datasets.
• Potential Causes & Solutions:
  • Cause: The constructed Hankel matrix is extremely large, leading to high computational complexity [39].
  • Solution: The original paper proposes a variable splitting strategy to decouple the problem into simpler sub-problems. Ensure you are using an implementation that employs this or similar optimization (e.g., alternating direction method of multipliers) for memory efficiency [39].
  • Cause: The data dimensions are too high.
  • Solution: As a preprocessing step, consider modest spatial smoothing or down-sampling of the data, balancing the trade-off between resolution and computational feasibility.

Problem 2: Ineffective Recovery of Censored Time Points

• Symptoms: The recovered time series still shows clear discontinuities or introduces strange patterns at the censored time points.
• Potential Causes & Solutions:
  • Cause: The low-rank prior assumption is violated, which can happen if the motion is very severe and affects a large portion of the data.
  • Solution: The method works best when the number of censored time points is not excessive. Increase the stringency of your motion censoring (scrubbing) to remove only the most severely corrupted volumes, leaving a sufficient number of "clean" time points for the matrix completion to work effectively [39].
  • Cause: The linear recurrence relation (LRR) model order is inappropriate for your data.
  • Solution: The window length L for the Hankel matrix is a key parameter. Try adjusting this parameter, as an overestimated window length can lead to a higher-than-expected rank and poor performance [39].

Guide 3: Troubleshooting Wavelet-Based Filters

Problem 1: Signal Distortion Following Artifact Removal

• Symptoms: The denoised signal appears overly smoothed, and key features of the physiological signal (e.g., SCR peaks in EDA) are attenuated [40].
• Potential Causes & Solutions:
  • Cause: The thresholding of wavelet coefficients is too aggressive, removing not only artifacts but also signal components of interest.
  • Solution: The artifact proportion parameter δ controls the threshold. Reduce the value of δ to be more conservative in what is considered an artifact. Use adaptive thresholding that calculates thresholds within local time windows to account for dynamic changes in the signal amplitude [40].
  • Cause: The choice of mother wavelet is not well-matched to the signal morphology.
  • Solution: The Haar wavelet is often preferred for its edge-detection capabilities with motion artifacts. However, if your signal of interest has specific characteristics, testing other wavelets (e.g., Daubechies) might yield better results [40].

Problem 2: Residual Motion Artifacts Remain

• Symptoms: Clear spike artifacts are still visible in the data after filtering.
• Potential Causes & Solutions:
  • Cause: The artifact has a frequency profile that overlaps significantly with the signal of interest.
  • Solution: This is a fundamental limitation. Consider a hybrid approach: first use a wavelet-based method to detect the location of artifacts, then use an interpolation method (like cubic spline) specifically for the identified contaminated segments, rather than relying solely on coefficient thresholding [40].
  • Cause: The decomposition level of the stationary wavelet transform (SWT) is insufficient to isolate the artifact.
  • Solution: Increase the number of decomposition levels (e.g., from 8 to 10) to ensure the artifact is captured in the detail coefficients where it can be effectively thresholded [40].

Frequently Asked Questions (FAQs)

FAQ 1: Under what conditions should I choose ICA-AROMA over a simpler motion regression model (e.g., 24-parameter model)?

ICA-AROMA is particularly advantageous when you are concerned about preserving the temporal degrees of freedom and autocorrelation structure of your data. Simple regression models remove motion-induced signal variations at the cost of destroying this autocorrelation, which can invalidate subsequent statistical tests. ICA-AROMA overcomes this drawback. Furthermore, it has been shown to remove motion-related noise to a larger extent than 24-parameter regression or spike regression and can increase sensitivity to group-level activation in both resting-state and task-based fMRI [20] [38].

FAQ 2: My research involves infant populations with frequent, large head movements. Which algorithm is most suitable?

For populations like infants who exhibit infrequent but large motions, a combination of strategies is often best. The JumpCor technique was specifically designed for this scenario. It identifies large "jumps" in motion and models separate baselines for the data segments between these jumps, effectively accounting for signal intensity changes caused by the head moving into a different part of the RF coil [41]. This can be combined with structured matrix completion to recover the censored time points, as this method has been validated to improve functional connectivity estimates even in the presence of large motions by exploiting the underlying structure of the fMRI time series [39] [41].

FAQ 3: How does structured low-rank matrix completion compare to simple interpolation for filling in censored ("scrubbed") volumes?

Simple interpolation (e.g., linear or spline) replaces missing data using only immediately adjacent time points, which can create smooth but unrealistic transitions and does not account for the global spatio-temporal structure of the brain's signals. In contrast, structured matrix completion uses a low-rank prior, formalized by constructing a Hankel matrix from the time series. This model leverages information from the entire dataset and across voxels to recover the missing entries in a physiologically more plausible way, leading to functional connectivity matrices with lower errors in pair-wise correlation compared to methods using only censoring and interpolation [39].

FAQ 4: Can wavelet-based filters be applied to real-time artifact correction?

The stationary wavelet transform (SWT), which is time-invariant and performs no down-sampling, is a prerequisite for any real-time application. Its structure makes it more suitable than the standard discrete wavelet transform (DWT). However, the adaptive thresholding step, which often involves estimating a Gaussian mixture model for the wavelet coefficients within a sliding window, can be computationally demanding. While promising for near-real-time applications, its true real-time capability depends on a highly optimized implementation that can perform the SWT and statistical estimation within the sampling interval of the acquired signal [40].

Table 1: Performance Metrics of Featured Motion Correction Algorithms

Algorithm	Modality	Key Performance Findings	Comparative Advantage
ICA-AROMA [20] [38]	fMRI (Task & Resting-state)	Increased sensitivity to group-level activation; More effective motion removal than 24-parameter regression or spike regression.	Preserves temporal degrees of freedom (tDOF); No need for classifier re-training; Fully automatic.
Structured Matrix Completion [39]	rsfMRI	Resulted in functional connectivity matrices with lower errors in pair-wise correlation; Improved delineation of the default mode network.	Recovers censored data using global information; Also performs slice-timing correction.
Wavelet-Based Filter (SWT) [40]	EDA	Achieved >18 dB attenuation in motion artifact energy; Induced less than -16.7 dB distortion in artifact-free regions.	Adaptive thresholding retains valid signal; Effective for spike-like artifacts in continuous data.
JumpCor [41]	fMRI (Infant studies)	Significantly reduced motion-related signal changes from infrequent large motions; Improved functional connectivity estimates.	Specifically designed for large, discrete head "jumps"; Simple regression-based approach.

Experimental Protocols

Protocol 1: Implementing ICA-AROMA for fMRI Preprocessing

This protocol outlines the steps to integrate ICA-AROMA into a standard fMRI preprocessing pipeline for motion artifact removal [20] [38].

• Standard Preprocessing: Begin with standard steps including slice-timing correction, realignment (motion correction), and spatial smoothing. Co-registration to a high-resolution structural image is also required.
• ICA-AROMA Execution:
  • Input: The realigned (and optionally smoothed) 4D functional data and the corresponding motion parameters obtained from realignment.
  • Process: Run the ICA-AROMA classifier. The algorithm performs:
    • Melodic ICA: Decomposes the functional data into independent components.
    • Feature Extraction: For each component, it calculates four robust features: high-frequency content, correlation with motion parameters, edge fraction (spatial), and CSF fraction (spatial).
    • Classification: Uses these features to automatically classify components as either "motion" or "non-motion."
• Denoising: Regress out the time courses of the classified motion components from the original functional data. The "non-aggressive" method is recommended to preserve signal variance.
• Downstream Analysis: Proceed with standard analysis, such as general linear model (GLM) for task-fMRI or functional connectivity analysis for rs-fMRI, using the denoised data.

Protocol 2: Validating a Wavelet-Based Motion Filter for EDA Data

This protocol describes a method to validate the performance of a stationary wavelet transform (SWT) filter for removing motion artifacts from electrodermal activity (EDA) data [40].

• Data Collection & Labeling:
  • Record EDA signals (e.g., skin conductance) alongside accelerometer data (actigraph) to provide an objective measure of movement.
  • Have multiple expert reviewers manually label portions of the EDA signal that are contaminated by motion artifacts. This serves as the ground truth for validation.
• Wavelet Denoising Procedure:
  • Decomposition: Perform an 8-level decomposition of the EDA signal using the Stationary Wavelet Transform (SWT) with the Haar mother wavelet.
  • Modeling: For each level of detail coefficients, model their distribution as a Gaussian mixture model (GMM) using an Expectation-Maximization (EM) algorithm. The GMM represents coefficients from the underlying EDA signal.
  • Adaptive Thresholding: Divide the wavelet coefficients at each level into time windows. For each window, calculate upper and lower thresholds based on the GMM and a predefined artifact proportion parameter (δ). Coefficients exceeding these thresholds are set to zero.
  • Reconstruction: Apply the inverse SWT to the thresholded coefficients to reconstruct the denoised EDA signal.
• Performance Validation:
  • Quantitative Metrics: Calculate the attenuation of artifact energy (in dB) in the manually labeled artifact segments. Calculate the normalized mean-square error (NMSE) in artifact-free segments to measure signal distortion.
  • Qualitative Inspection: Visually compare the raw and denoised signals to ensure physiological patterns (e.g., SCR peaks) are preserved.

Method Visualization

Diagram 1: ICA-AROMA Workflow

Diagram 2: Structured Matrix Completion for rsfMRI

Diagram 3: Wavelet-Based Filtering Logic

The Scientist's Toolkit

Table 2: Essential Research Reagents & Computational Tools

Item Name	Function / Purpose	Example Use Case
FSL MELODIC	Performs Independent Component Analysis (ICA) on fMRI data.	Core decomposition engine for the ICA-AROMA pipeline [20] [38].
Hankel Matrix	A structured matrix formed from time-series data where each descending diagonal is constant.	Used in structured low-rank completion to enforce linear recurrence relations in fMRI signals [39].
Stationary Wavelet Transform (SWT)	A time-invariant wavelet transform that does not use down-sampling.	Essential for wavelet-based filters to avoid artifacts during the decomposition and reconstruction of signals like EDA [40].
Gaussian Mixture Model (GMM)	A probabilistic model for representing normally distributed subpopulations within data.	Used to statistically model the distribution of wavelet coefficients, separating valid signal from motion artifacts [40].
Ex Vivo Brain Phantom	A physical model of the brain used for controlled MRI experiments.	Provides a motion-free ground truth for validating motion correction algorithms like SIMPACE and SLOMOCO [42].
Simulated Prospective Acquisition Correction (SIMPACE)	An MRI sequence that injects user-defined motion into the acquisition of a static phantom.	Generates motion-corrupted fMRI data with a known ground truth for rigorous algorithm testing and development [42].
6-Chloro-1H-benzimidazol-1-amine	6-Chloro-1H-benzimidazol-1-amine|CAS 63282-63-3	6-Chloro-1H-benzimidazol-1-amine (CAS 63282-63-3). A benzimidazole-based building block for antimicrobial and anticancer research. For Research Use Only. Not for human or veterinary use.
1-Methylquinolinium methyl sulfate	1-Methylquinolinium methyl sulfate, CAS:38746-10-0, MF:C11H13NO4S, MW:255.29 g/mol	Chemical Reagent

Frequently Asked Questions (FAQs)

General Concepts

Q1: What are the main types of deep learning models used for MRI motion artifact correction? The main types are Generative Adversarial Networks (GANs), including Conditional GANs (cGANs), and Diffusion Models (such as Denoising Diffusion Probabilistic Models - DDPMs). These are deep generative models trained to learn the mapping between motion-corrupted and motion-free images. They are particularly powerful for capturing complex image priors and correcting non-linear distortions, outperforming earlier methods like simple Convolutional Neural Networks (CNNs) or Autoencoders in many scenarios [43] [44].

Q2: Why is motion artifact a particularly critical problem in neuroimaging studies on behavior? During behavioral tasks, participants often cannot avoid making small head movements. These motions introduce spurious signal fluctuations in functional MRI (fMRI) data that can confound measures of functional connectivity. If not mitigated, this artifact can bias statistical inferences about relationships between brain connectivity and individual behavioral differences, potentially leading to false conclusions in your research [45].

Q3: My dataset is small and I lack paired data (motion-corrupted and clean images from the same subject). What are my options? This is a common challenge. The field has developed several effective approaches:

• Synthetic Motion Generation: You can artificially induce realistic motion artifacts into your existing clean images to create a paired dataset for training supervised models like cGANs or U-Nets [43] [46].
• Unpaired Learning with Diffusion Models: Unconditional diffusion models can be trained solely on a collection of clean, motion-free images. For correction, the motion-affected image is used to initialize the reverse denoising process, guiding the model to reconstruct a clean version without needing a paired clean counterpart [46].
• Unpaired Learning with CycleGANs: Cycle-Consistent GANs can learn to translate images from a "motion-corrupted" domain to a "clean" domain without requiring one-to-one paired examples [44].

Implementation & Troubleshooting

Q4: I am getting blurry results from my cGAN model. How can I improve the output sharpness? Blurry outputs often stem from the generator network prioritizing pixel-wise loss (like L1 or L2 distance) at the expense of perceptual quality. To address this:

• Refine your Loss Function: Combine the traditional L1/L2 loss with an adversarial loss and a perceptual loss (e.g., VGG-based loss). The adversarial loss encourages the generator to produce images that are indistinguishable from real, clean images, thereby enhancing visual authenticity [43].
• Inspect your Training Data: Ensure that the simulated motion artifacts in your training data are realistic and varied enough. The model can only learn to correct patterns it has seen during training [43].

Q5: When using a diffusion model for correction, the output sometimes "hallucinates" features not present in the original. How can I control this? Hallucination occurs due to the strong generative nature of diffusion models. The key is to control the starting point of the denoising process.

• Do not start denoising from pure noise (t = T). Instead, add a smaller amount of noise to your corrupted input image and start the reverse process from an intermediate timestep n (where n < T) [46].
• There is a trade-off: starting at a very high n (adding more noise) gives the generative model more "creativity" to hallucinate details. Starting at a very low n (adding less noise) may be insufficient to remove artifacts. You must experimentally find the optimal n that provides a balance for your specific dataset [46].

Q6: How can I make my 3D diffusion model feasible to train with limited GPU memory? Training diffusion models on full 3D volumes is notoriously memory-intensive. A effective strategy is the PatchDDM approach:

• Train on Patches, Test on Volumes: During training, the model is trained on randomly sampled small 3D patches from your larger volumes. This drastically reduces memory consumption.
• At test time (inference), the fully trained model can be applied to the entire 3D volume, processing it in a single pass or via a sliding window, to produce a high-quality, coherent output without the memory constraints of training [47].

Troubleshooting Guides

Problem: The Model Fails to Generalize to New Data with Different Artifact Patterns

Possible Causes & Solutions:

• Cause 1: Narrow Training Data. The model was trained on a dataset with a limited variety of motion types, severities, or anatomical appearances.
  • Solution: Augment your training dataset. Incorporate simulated motion in both phase-encoding directions (e.g., horizontal and vertical). Using a training set that combines both directions has been shown to create a more robust model that performs well regardless of the artifact's orientation [43].
• Cause 2: Domain Shift. The new data comes from a different scanner, acquisition protocol, or patient population than the training data.
  • Solution: Employ domain adaptation techniques. If retraining is possible, fine-tune your pre-trained model on a small set of data from the new domain. Alternatively, use unpaired methods like CycleGAN or test-time adaptation strategies to bridge the domain gap [44].

Problem: Over-smoothing and Loss of Fine Anatomical Detail in Corrected Images

Possible Causes & Solutions:

• Cause: Over-reliance on Pixel-wise Loss Functions. Models trained primarily with Mean Squared Error (MSE) loss tend to produce averages of possible outputs, leading to blurred images.
  • Solution: Integrate adversarial and perceptual loss functions into your training. The adversarial component of a cGAN pushes the solution towards the manifold of natural images, helping to preserve textural details and edges that are critical for diagnostic accuracy [43] [44].
• Cause: Excessive Denoising in Diffusion Models. Using too many denoising steps (starting from a timestep n that is too high) can cause the model to generate an image that deviates from the original underlying structure.
  • Solution: Carefully tune the starting timestep n in your diffusion-based correction algorithm. Use a separate validation set to find the value of n that best removes artifacts while faithfully preserving the true anatomical information [46].

Comparative Performance Data

The following tables summarize quantitative results from key studies, providing benchmarks for expected performance.

Table 1: Performance of cGAN vs. Other Models on Simulated Head MRI Motion Artefacts (T2-weighted)

Model	SSIM	PSNR (dB)	Key Findings
cGAN	> 0.9	> 29	Achieved ~26% improvement in SSIM and ~7.7% in PSNR. Best image reproducibility. [43]
U-Net	Lower than cGAN	Lower than cGAN	Outputs can be too smooth, lacking visual authenticity. [43]
Autoencoder	Lower than cGAN	Lower than cGAN	Less effective at capturing and correcting complex artifact patterns. [43]

Table 2: Performance of Diffusion Model vs. U-Net on Brain MRI (MR-ART Dataset)

Model / Approach	SSIM	NMSE	PSNR (dB)	Key Findings
DDPM (n=150)	0.858*	Data not fully shown	Data not fully shown	Effective correction but risk of hallucination if n is not tuned properly. [46]
U-Net (Trained on Synthetic Data)	Slightly lower than DDPM*	Data not fully shown	Data not fully shown*	Robust performance, less risk of hallucination than diffusion models, but requires good synthetic data. [46]
U-Net (Trained on Real Paired Data)	0.858	Data not fully shown	Data not fully shown	Serves as an upper-bound benchmark; often unavailable in practice. [46]

Note: Specific values for DDPM and U-Net on synthetic data were not fully listed in the source, but the study concluded that performance was comparable, with a trade-off between diffusion's accuracy and its potential for hallucination [46].

Table 3: Supervised Conditional Diffusion Model for Knee MRI (Real-World Paired Data)

Metric	Result	Comparison
RMSE	11.44 ± 3.21	Smallest among compared methods (ESR, EDSR, ESRGAN) [48]
PSNR (dB)	33.05 ± 2.90	Highest among compared methods [48]
SSIM	0.97 ± 0.02	Highest among compared methods [48]
Subjective Rating	No significant difference from ground-truth rescans	Outperformed the input artifacted images significantly [48]

Experimental Protocols

Protocol 1: Training a cGAN for Motion Artefact Reduction using Simulated Data

This protocol is based on the methodology described in [43].

1. Data Preparation & Motion Simulation:

• Input: Collect a dataset of high-quality, motion-free T2-weighted axial head MRI images (e.g., 5500 images).
• Artefact Simulation: For each clean image, generate a motion-corrupted version using a k-space simulation method.
  • Apply a combination of translational (±10 pixels) and rotational (±5°) shifts to the original image, creating multiple displaced versions.
  • Compute the Fourier transform (k-space) of these displaced images.
  • Randomly select lines from these corrupted k-spaces and combine them to form a new, motion-affected k-space data.
  • Apply the inverse Fourier transform to generate the final simulated motion-artefact image.
  • Crucially, simulate artefacts in both the horizontal and vertical phase-encoding directions to build a robust model.
• Splitting: Split the paired dataset (corrupted vs. clean) into training (90%), validation (for hyperparameter tuning, 10% of training data), and test sets (10%).

2. Model Training:

• Architecture: Implement a cGAN with a U-Net as the generator and a convolutional neural network (CNN) as the discriminator.
• Conditioning: The generator takes the motion-corrupted image as the input condition and learns to generate the clean image.
• Loss Function: Use a composite loss function: Loss = L1_Loss(Generated, Clean) + λ * Adversarial_Loss, where λ controls the balance. The L1 loss encourages structural similarity, while the adversarial loss improves perceptual realism.
• Training: Train the model until the validation loss plateaus.

3. Evaluation:

• Use standard image quality metrics on the test set: Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR).
• Perform qualitative assessment by visual inspection for blurring or residual artefacts.

The workflow for this protocol can be summarized as follows:

Protocol 2: Correcting Artefacts with an Unconditional Diffusion Model

This protocol is based on the approach outlined in [46].

1. Model Training (Unsupervised, on Clean Data Only):

• Data: Gather a dataset of clean, motion-free MRI images. No paired corrupted data is needed for this stage.
• Process: Train an unconditional Denoising Diffusion Probabilistic Model (DDPM) on these clean images. The model learns the underlying probability distribution of motion-free neuroimages.

2. Artefact Correction (Inference):

• Input: A novel motion-corrupted image that needs correction.
• Noising: Instead of starting from pure Gaussian noise, take the corrupted image and add Gaussian noise to it up to a pre-determined, intermediate timestep n (e.g., n=150 out of T=500).
• Denoising: Execute the learned reverse denoising process from this timestep n back to timestep 0. The model will progressively "denoise" the image, guided by its knowledge of what a clean image looks like, thereby removing the motion artefacts.

3. Critical Parameter Tuning:

• The choice of the starting timestep n is critical. It represents a trade-off:
  • n too high: The final output may contain hallucinations (features not present in the original).
  • n too low: The correction will be insufficient, with residual artefacts remaining.
• Use a separate validation set of corrupted images to empirically determine the optimal n for your data.

The inference process for this diffusion-based correction is illustrated below:

The Scientist's Toolkit: Research Reagent Solutions

Table 4: Essential Components for Deep Learning-based Motion Correction

Component	Function / Description	Examples / Notes
Deep Learning Framework	Software library for building and training neural networks.	PyTorch, TensorFlow. Used in all cited studies [46].
cGAN (Conditional GAN)	Generative model architecture that maps a conditioned input (corrupted image) to a target output (clean image).	Consists of a Generator (e.g., U-Net) and a Discriminator. Provides high image reproducibility [43].
DDPM (Denoising Diffusion Probabilistic Model)	Generative model that learns data distribution by progressively denoising from noise. Can be adapted for correction.	Enables unpaired learning. Risk of hallucination requires careful tuning of the starting noise step [46].
U-Net	CNN architecture with a contracting encoder and expansive decoder, often used in image-to-image tasks.	Commonly used as the generator in cGANs or as a standalone supervised model [43] [46].
Synthetic Motion Simulator	Algorithm to generate realistic motion artefacts in clean images for creating training data.	K-space simulation with translational/rotational shifts is a common and effective method [43].
Image Quality Metrics	Quantitative measures to evaluate the performance of the correction model.	SSIM (structural similarity), PSNR (peak signal-to-noise ratio), NMSE (normalized mean squared error) [43] [46].
Public MRI Datasets	Benchmark datasets for training and validation.	MR-ART (contains paired motion-free and motion-affected brain scans) [46].
2-Hydroxy-5-isopropylbenzoic acid	2-Hydroxy-5-isopropylbenzoic Acid|CAS 31589-71-6
(2-Methyl-benzylamino)-acetic acid	(2-Methyl-benzylamino)-acetic acid|CAS 702629-73-0	High-purity (2-Methyl-benzylamino)-acetic acid (CAS 702629-73-0) for laboratory research. This N-substituted glycine derivative is for Research Use Only (RUO). Not for human or veterinary use.

Optimizing Your Pipeline: A Practical Guide for Robust Data

Motion artifacts represent a significant and pervasive challenge in neuroimaging, capable of severely degrading data quality and compromising the validity of research findings, particularly in studies involving behavioral tasks. Effectively mitigating these artifacts is not a one-size-fits-all endeavor; the optimal strategy is highly dependent on the imaging modality. This guide provides a structured approach to selecting and implementing motion correction techniques tailored for functional Magnetic Resonance Imaging (fMRI) and functional Near-Infrared Spectroscopy (fNIRS), enabling researchers to make informed decisions for their specific experimental contexts.

Frequently Asked Questions (FAQs)

1. Why can't I use the same motion correction method for both my fMRI and fNIRS data? fMRI and fNIRS are fundamentally different technologies. fMRI measures the blood-oxygen-level-dependent (BOLD) signal, and its artifacts often arise from disruptions in a large, static magnetic field and precise k-space sequencing. fNIRS, in contrast, measures light absorption by hemoglobin, and its artifacts are primarily caused by physical disruptions in the optode-scalp coupling [49] [50]. The underlying physics of the noise demands distinct correction strategies.

2. For fNIRS, is it better to correct for motion artifacts or to completely reject contaminated trials? The prevailing consensus, supported by comparative studies, is that correction is almost always superior to rejection. Trial rejection can lead to a significant loss of statistical power, especially in populations where motion is frequent (e.g., children, patients with movement disorders). Research has shown that it is "always better to correct for motion artifacts than reject trials" [51] [52].

3. What is the most robust method for correcting motion artifacts in fNIRS? While the "best" method can depend on the specific artifact type and data characteristics, wavelet-based filtering has been repeatedly identified as one of the most powerful and effective techniques. One key study found it to be the most effective approach for correcting motion artifacts induced by a cognitive task involving speech, reducing the artifact area under the curve in 93% of cases [51] [52]. Other strong contenders include Temporal Derivative Distribution Repair (TDDR) and correlation-based signal improvement (CBSI) [53] [54].

4. How are AI and deep learning changing the landscape of motion correction in fMRI? Deep learning (DL), particularly generative models, is revolutionizing motion correction by learning direct, non-linear mappings from motion-corrupted images to clean images. This includes models like Generative Adversarial Networks (GANs) and Denoising Diffusion Probabilistic Models (DDPMs) [55] [56]. These methods can achieve high-fidelity corrections without always requiring a precise mathematical model of the motion degradation process, making them highly adaptable. However, challenges remain, including limited generalizability across diverse datasets and the risk of introducing unrealistic "hallucinated" image features [55].

Troubleshooting Guides

Guide 1: Selecting an fNIRS Motion Correction Algorithm

Follow this decision workflow to choose the most appropriate correction method for your fNIRS data.

Summary of fNIRS Correction Methods:

Method	Principle	Best For	Limitations
Wavelet Filtering [51] [52]	Identifies and removes motion artifacts in the wavelet domain.	General-purpose use; highly effective for various artifact types.	Can be computationally complex.
TDDR [53] [54]	Uses temporal derivatives to correct shifts & spikes automatically.	Automated correction without need for user-defined parameters.	May not handle all complex, slow drifts.
Spline Interpolation [51] [52]	Identifies artifact segments and replaces them with a spline fit.	Clear, identifiable artifact segments in otherwise good data.	Requires manual selection of artifact periods.
CBSI [51] [53]	Leverages negative correlation between HbO and HbR.	Artifacts that are correlated with the hemodynamic response.	Assumes a specific physiological relationship.
PCA [53]	Removes components of variance associated with motion.	Large, global motion artifacts affecting many channels.	Risk of removing physiological signal of interest.

Guide 2: Implementing Motion Correction for fMRI

The approach to mitigating motion artifacts in fMRI can be broadly categorized into prospective (during scan) and retrospective (after scan) methods. The following workflow outlines the decision process.

Comparison of fMRI Motion Correction Strategies:

Strategy	Type	Key Examples	Key Considerations
Prospective [55]	During Acquisition	MoCo sequences, navigators (PROMO, vNavs), optical tracking.	Prevents k-space inconsistencies. Requires specific hardware/sequence support.
Image Registration [55]	Retrospective	Rigid-body volume realignment.	Standard first step; corrects for head movement but not for spin history effects.
Deep Learning (Generative) [55] [56]	Retrospective	3D UNet, GANs, Diffusion Models, Mean-Reverting SDE.	Powerful for severe artifacts; risk of hallucinations; requires significant computational resources.
k-Space/Model-Based [55]	Retrospective	Compressed Sensing, joint reconstruction.	Can directly correct k-space inconsistencies; often computationally intensive.

Experimental Protocols

Protocol 1: Implementing Wavelet Motion Correction for fNIRS Data

This protocol outlines the steps for applying wavelet-based motion artifact correction, a method validated to be highly effective in real cognitive data [51] [52].

• Data Preparation: Begin with your raw light intensity data. Convert it to optical density (OD) units using the standard formula: ( OD = \log{10}(I0 / I) ), where ( I_0 ) is the emitted light intensity and ( I ) is the detected intensity [53] [50].
• Algorithm Selection: Choose and initialize the wavelet filtering algorithm. In the HOMER3 software package, this function is typically called hmrR_MotionCorrectWavelet [53].
• Parameter Configuration: Set the critical input parameters. These often include:
  • iqr: The interquartile range threshold for artifact detection. A common starting value is 1.5.
  • x_scale: The wavelet scale that determines the frequency band targeted for correction.
  • iterations: The number of cycles for the iterative correction process.
• Execution: Run the algorithm on your optical density data. The method will identify outlier segments based on their wavelet coefficients and reconstruct the signal using the remaining coefficients.
• Validation: Visually inspect the corrected data by comparing plots before and after correction. Validate the correction by ensuring that the characteristic spike or baseline shift is removed without distorting the underlying physiological signal. Calculate and compare the Signal-to-Noise Ratio (SNR) before and after processing.

Protocol 2: A Framework for Deep Learning-Based Motion Correction in fMRI

This protocol provides a general framework for implementing a DL-based retrospective motion correction method, such as a 3D UNet or a Generative Adversarial Network (GAN) [57] [55].

• Data Preparation & Simulation:
  • Ideal Scenario: Use a dataset of paired motion-corrupted and motion-free 3D MRI volumes.
  • Common Practice: If paired data is unavailable, simulate motion artifacts. This can be done by applying realistic, random rigid-body transformations (rotations, translations) to high-quality, motion-free volumes, effectively creating your own corrupted-clean pairs [55].
• Model Selection & Architecture:
  • Select a model architecture. A 3D UNet is a strong baseline, effective for capturing volumetric context [57] [55].
  • For more complex artifact patterns, consider a Generative Adversarial Network (GAN) or a Score-Based Generative Model (SGM) to better model the distribution of artifact-free images [55] [56].
• Training Configuration:
  • Loss Function: Use a combination of loss functions. A common choice is a weighted sum of L1 loss (e.g., Mean Absolute Error) to enforce pixel-wise accuracy and a perceptual loss (e.g., Structural Similarity Index - SSIM) to improve image quality [55].
  • Optimizer & Hyperparameters: Use the Adam optimizer. A typical initial learning rate is 1x10⁻⁴, with a batch size as large as your GPU memory allows (e.g., 1-4 for 3D volumes).
• Training & Validation:
  • Train the model on your prepared dataset, using a separate validation set to monitor for overfitting.
  • Use quantitative metrics like Peak Signal-to-Noise Ratio (PSNR) and SSIM to evaluate performance on the validation set [55].
• Inference & Evaluation:
  • Apply the trained model to novel motion-corrupted data.
  • Perform a final evaluation using both quantitative metrics and qualitative assessment by an expert radiologist or neuroscientist to ensure the corrected images are diagnostically/reliably usable.

The Scientist's Toolkit: Key Research Reagents & Materials

Item	Function in Motion Correction
Accelerometers / Motion Sensors [49]	Small hardware devices attached to the subject or imaging apparatus (e.g., fNIRS cap) to provide direct, real-time measurements of motion dynamics, which can be used as a reference signal for correction.
Short-Separation Detectors [50]	In fNIRS, these are optodes placed very close (e.g., 8 mm) to a source. They measure physiological noise from the scalp, which can be regressed out to isolate the deeper brain signal and mitigate motion-related superficial artifacts.
Optical Tracking Systems [55]	Used in fMRI, these external camera systems track reflective markers on the subject's head. The real-time motion data can be fed back to the scanner for prospective slice-position correction.
Collodion-Fixed Optodes [49]	A method for securing fNIRS optodes to the scalp using collodion, which forms a strong, rigid bond, significantly reducing motion artifacts by minimizing optode movement relative to the head.
HOMER3 / MNE-NIRS Software [53] [54]	Open-source software packages that provide integrated pipelines for fNIRS data preprocessing, including a suite of motion correction algorithms (wavelet, TDDR, CBSI, PCA, spline) for direct implementation.
1-(3-Iodo-4-methylphenyl)ethanone	1-(3-Iodo-4-methylphenyl)ethanone|CAS 52107-84-3
2,3-Dioxoindoline-5-carbonitrile	2,3-Dioxoindoline-5-carbonitrile|CAS 61394-92-1

Troubleshooting Guides

Guide 1: Addressing Poor Activation Maps After Motion Correction

Problem Description Researchers frequently obtain weak or non-significant activation maps in first-level fMRI analyses after participants move during a behavioral task. The chosen artifact correction method may be overly aggressive, removing genuine neural signals along with motion artifacts.

Diagnosis and Solution This issue requires a diagnostic approach to evaluate and refine the artifact correction strategy.

• Step 1: Quantify Data Corruption Calculate the following metrics from your first-level fMRI dataset to objectively identify corrupted volumes [58]:

  • Δ%D-var (fast variance component of DVARS): Measures the rate of change of the signal.
  • Scan-to-Scan Total Displacement (STS): Quantifies the degree of head movement between consecutive scans.
  • Explained Variance (R²): Indicates how well a general linear model fits each scan.

• Step 2: Identify Outlying Volumes Use a data-driven, multidimensional approach that combines the three indicators from Step 1. Employ a balanced criterion like Akaike's corrected criterion (AICc) to automatically identify outlying datapoints without overcorrection [58].

• Step 3: Compare Correction Strategies The core of the problem is choosing between censoring (removing bad volumes) or interpolating (estimating replacement values). The effects are distinct and complex, and the optimal choice can depend on your specific data [58]. The table below summarizes the trade-offs:

Table: Comparison of Censoring and Interpolation Strategies

Feature	Censoring (Volume Removal)	Interpolation (Data Replacement)
Principle	Complete removal of identified outlying volumes from the time series [58].	Estimation of replacement data for outlying volumes based on surrounding good data [58].
Best For	Many settings; when the number of bad volumes is relatively low [58].	Datasets where maintaining a continuous time series is critical for analysis.
Data Retention	Lower. Directly removes data, potentially shortening the usable time series.	Higher. Retains the original data structure and length by filling gaps.
Risk of Introducing Bias	Lower, as no new data is synthesized.	Higher, as the method may interpolate noise or create spurious temporal correlations.
Impact on Temporal Structure	Disrupts the equidistant timing of scans, which can complicate some time-series analyses.	Preserves the temporal structure and timing of the scan sequence.

Figure 1: Decision Workflow for Motion Correction Strategies

Guide 2: Mitigating Motion Artifact in Functional Connectivity Studies

Problem Description In-scanner motion is a major confound in functional connectivity analyses. Even small movements can systematically bias connectivity estimates, which is particularly problematic when motion correlates with variables of interest like age, clinical status, or task performance [59].

Diagnosis and Solution A multi-stage denoising pipeline is required to mitigate these artifacts without removing neurobiologically meaningful signal.

• Step 1: Implement Robust Preprocessing Begin with standard preprocessing steps (realignment, normalization) and include motion parameters as regressors in your model.

• Step 2: Apply Data Cleaning Techniques Integrate motion artifact correction methods into your pipeline. The choice of method involves a direct trade-off between the amount of data retained and the potential for artifact leakage.

  • Censoring (e.g., "Motion Scrubbing"): Removes volumes with excessive motion, effectively prioritizing data quality over quantity [7].
  • Interpolation (e.g., using surrounding volumes): Replaces bad volumes to retain a complete time series, prioritizing data continuity at the risk of introducing estimation errors [58].

• Step 3: Validate with Quantitative Metrics After applying these techniques, check for improvements in the temporal signal-to-noise ratio (tSNR) and assess whether the pattern of functional connectivity matrices becomes less driven by motion-related artifacts [58] [59].

Table: Impact of Motion on Functional Connectivity and Mitigation Outcomes

Metric/Outcome	Presence of Motion Artifact	After Successful Mitigation
Temporal SNR	Decreased [58].	Increased [58].
Connectivity Estimates	Biased, often inflated in short-range connections and deflated in long-range connections [59].	More biologically plausible and less correlated with motion parameters [59].
Statistical Inference	High risk of false positives/negatives due to systematic bias [59].	Improved specificity and validity of group-level inferences [59].

Frequently Asked Questions (FAQs)

FAQ 1: What is the fundamental trade-off between data censoring and interpolation?

The core trade-off lies between data quality and data quantity. Censoring is a conservative approach that sacrifices data points (reducing quantity) to ensure the remaining data is of the highest possible quality and free from artifact contamination. Interpolation is a more liberal approach that prioritizes retaining all data points (maintaining quantity) by estimating and replacing bad values, but this carries the risk of the interpolation algorithm introducing its own noise or smearing artifacts, thereby potentially compromising data quality [58].

FAQ 2: How can I objectively identify which fMRI volumes to censor or interpolate?

Relying on a single metric can be misleading. Best practice is to use a multidimensional, data-driven approach that combines several indicators of data corruption. Key metrics include:

• DVARS: Measures signal change between consecutive volumes.
• Framewise Displacement (or Scan-to-Scan Displacement): Quantifies head movement.
• Explained Variance (R²): Assesses model fit for each scan. Using a combination of these metrics, often with an automated criterion like AICc, provides a more robust and objective identification of outlying volumes than any single metric alone [58].

FAQ 3: Are there fully automated, state-of-the-art methods for handling motion artifacts?

Yes, the field is rapidly advancing with machine learning and deep learning techniques. For example, Denoising Autoencoders (DAEs) have been successfully applied to remove motion artifacts from functional near-infrared spectroscopy (fNIRS) data. These models are "assumption-free" and can outperform conventional methods by lowering residual motion artifacts and decreasing the mean squared error, all with high computational efficiency. Similar deep learning approaches are being developed and validated for fMRI data [60].

FAQ 4: Why is motion artifact particularly problematic for studies of functional connectivity?

Motion artifacts have a characteristic spatial and temporal pattern that can systematically bias estimates of functional connectivity. This is especially dangerous when the amount of in-scanner motion is correlated with a variable of interest in your study (e.g., older adults or patients with a specific disorder moving more than healthy controls). This correlation can create spurious group differences or mask true effects, leading to incorrect inferences about brain networks [59].

The Scientist's Toolkit: Research Reagent Solutions

Table: Essential Materials for Motion Artifact Correction Experiments

Item	Function/Application
Tetraspeck Beads (200 nm, 500 nm, 4 μm)	Fluorescent beads used as a reference sample to measure and characterize chromatic aberrations and other spatial distortions in multicolor imaging setups. They serve as fixed points for calculating the transformations needed to align different fluorescence channels [61].
UltraPure Low-Melting-Point Agarose	Used to create a solid gel matrix for embedding and immobilizing reference beads, allowing for the creation of stable phantoms for system calibration [61].
SeeDB2G/S Clearing Agents	Optical clearing agents (e.g., based on Omnipaque 350 or Histodenz) used to render biological tissues transparent. This is crucial for deep-tissue imaging, but the refractive index mismatch with immersion media can introduce aberrations that must be corrected [61].
Plasmids for Multi-color Labeling (e.g., pCAG-tTA, pBS-TRE-mTurquoise2, pBS-TRE-EYFP, pBS-TRE-tdTomato)	Used in techniques like Tetbow for stochastic multicolor labeling of neurons. This enables tracing of dense neuronal circuits, a process that requires precise alignment of different color channels to avoid misinterpretations caused by chromatic aberrations [61].
4-Bromofuran-2-carbonyl chloride	4-Bromofuran-2-carbonyl chloride, CAS:58777-59-6, MF:C5H2BrClO2, MW:209.42 g/mol

Motion artifacts are unwanted disturbances in a neuroimaging signal caused by participant movement rather than neural activity [31] [62]. In the context of behavioral tasks, even small movements like nodding, swallowing, or facial expressions can generate signals that obscure the brain's true physiological response [62]. These artifacts are a significant source of noise and can introduce bias and variance into research results, particularly in studies involving populations that have difficulty remaining still, such as children or patients with neurological disorders [63].

A multi-layered defense that combines prospective and retrospective motion correction strategies is considered best practice for mitigating these effects. This approach proactively adjusts the data acquisition in real-time while also allowing for cleanup after the data has been collected, creating a more robust solution than either method alone [63].

---

Troubleshooting Guides

Problem 1: Persistent Motion Artifacts After Retrospective Correction

Description: Even after applying retrospective motion correction (RMC) during data reconstruction, significant motion-related blurring or signal distortions remain in the final structural or functional images.

Potential Cause	Diagnostic Steps	Solution
Low correction frequency during RMC [63]	Check the temporal resolution of the motion estimates used in RMC. Compare data corrected with motion estimates from every echo train (ET) versus within each ET.	Increase the motion correction frequency in the RMC pipeline. For data acquired with prospective motion correction (PMC) applied only before each ET, use a hybrid correction to apply RMC within echo-trains [63].
Severe motion causing k-space undersampling [63]	Inspect the k-space data for gaps caused by head rotations.	Where possible, combine RMC with PMC during acquisition. PMC modifies the imaging field of view in real-time to avoid k-space undersampling from the start [63].
Poor-quality GRAPPA calibration data [63]	Reconstruct images using integrated auto-calibration signal (ACS) data and compare to reconstruction using a pre-scan calibration.	Use integrated ACS data for GRAPPA reconstruction instead of a separate pre-scan calibration, as motion can degrade the latter's quality [63].

Problem 2: Inaccurate Prospective Motion Correction

Description: The prospectively motion-corrected images show severe blurring or a complete failure of anatomy tracking, often due to issues with the motion tracking system itself.

Potential Cause	Diagnostic Steps	Solution
Failed system calibration [63]	Verify the geometric and temporal calibration between the motion tracker and the MRI scanner.	Re-perform the cross-calibration scan and time synchronization between the tracking system and the scanner host computer before the session [63].
Poor line-of-sight for optical tracker [63]	Check the camera's view of the participant's face for obstructions from the head coil.	Reposition the vision probe to ensure an unobstructed view of the participant's face through the head coil [63].
Loss of tracking surface features	Observe the tracking system's software for error messages indicating low tracking confidence.	Ensure the initial reference surface scan is of high quality. For markerless systems, consider adding small, non-metallic fiducials if the participant's facial features are insufficient.

Problem 3: fNIRS-Specific Motion Artifacts

Description: The functional Near-Infrared Spectroscopy (fNIRS) data contains sudden, large signal changes (spikes) or sustained baseline shifts that do not correlate with the experimental paradigm [62].

Potential Cause	Diagnostic Steps	Solution
Optode movement on the scalp [62]	Visually inspect the signal for characteristic spike or baseline shift artifacts [62]. Correlate artifact timing with experimenter notes or video of participant movement.	Secure optode placement using a well-fitting headcap and part hair between the optode and scalp. Use spring-loaded optode holders to maintain consistent pressure [62].
Systemic physiological changes (e.g., from blood redistribution) [62]	Check for oscillatory artifacts linked to repetitive movements like breathing or nodding [62].	During task design, instruct participants to minimize non-essential movements. In data processing, use algorithms like moving standard deviation and spline interpolation to detect and correct these artifacts [62].

---

Frequently Asked Questions (FAQs)

What is the fundamental difference between prospective (PMC) and retrospective motion correction (RMC)?

• Prospective Motion Correction (PMC): This method corrects for motion during the data acquisition. It uses real-time estimates of head pose to dynamically adjust the imaging field of view (FOV), keeping it aligned with the participant's brain. This prevents gaps in k-space data (known as Nyquist violations) that occur due to rotation [63].
• Retrospective Motion Correction (RMC): This method corrects for motion after the scan is complete. It uses recorded motion estimates to adjust the k-space trajectory during image reconstruction. While it can correct for some motion effects, it cannot fully compensate for the k-space undersampling caused by head rotations [63].

Why does PMC often yield superior image quality compared to RMC?

PMC results in fewer artifacts because it avoids the problem of local Nyquist violations from the outset. By updating the FOV during the scan, PMC ensures k-space is sampled as intended, even when the head moves. RMC, which operates on already-acquired data, cannot fill in these inherent gaps in k-space, leading to residual artifacts, particularly in 3D-encoded sequences [63].

How does the frequency of motion correction updates impact image quality?

Increasing the correction frequency improves image quality for both PMC and RMC. A study on 3D MPRAGE sequences showed that updating the FOV more frequently—for example, every six readouts (Within-ET) instead of only before each echo train (Before-ET)—significantly reduced motion artifacts. This finer-grained correction better accounts for motion that occurs during the data acquisition window itself [63].

My research involves naturalistic behavioral tasks where some motion is unavoidable. What is the best strategy?

For studies where movement is inherent to the task (e.g., speaking, gesturing), a multi-layered defense is critical:

• Prevention: Use a lightweight imaging system (e.g., portable fNIRS), secure the sensor cap firmly, and design tasks to minimize unnecessary motion where possible [62].
• Prospective Correction: If available, use PMC during acquisition to get the cleanest possible raw data.
• Retrospective Correction: Apply advanced RMC algorithms as a final cleanup step to address any residual motion artifacts [63].

Can I use RMC to improve data that was already acquired with PMC?

Yes, this is known as a Hybrid Motion Correction (HMC) strategy. You can take data acquired with a lower-frequency PMC (e.g., Before-ET) and apply RMC to correct for residual motion that occurred within the echo train. This effectively increases the correction frequency retrospectively and has been shown to further reduce motion artifacts [63].

---

Experimental Protocols & Workflows

Protocol 1: Comparative Evaluation of PMC and RMC in 3D-MPRAGE

This protocol is designed to directly compare the performance of prospective and retrospective motion correction in a controlled setting [63].

1. Motion Tracking:

• Use a markerless optical tracking system (e.g., Tracoline TCL) to estimate rigid-body head motion at a high frame rate (e.g., 30 Hz).
• Perform a geometric cross-calibration between the tracker and the MRI scanner to represent motion in the scanner's coordinate system.
• Perform temporal synchronization between the tracker and scanner computers [63].

2. Data Acquisition:

• Use a modified 3D MPRAGE sequence capable of receiving real-time motion data and updating the imaging FOV.
• Scan Conditions:
  • A. No intentional motion: A baseline scan for reference.
  • B. With controlled motion: Participant performs prescribed, continuous head movements.
• PMC Modes:
  • Before-ET-PMC: FOV is updated once before each echo train.
  • Within-ET-PMC: FOV is updated before each ET and also every sixth readout within the ET [63].

3. Data Reconstruction & Analysis:

• Retrospective Correction: Apply RMC to the raw k-space data using a software package like retroMoCoBox. Adjust k-space trajectories based on the recorded motion.
• Hybrid Correction: Apply RMC to data acquired with Before-ET-PMC to create a Within-ET-HMC condition.
• Quantitative Evaluation: Calculate the Structural Similarity Index Measure (SSIM) comparing each motion-corrected image to the no-motion reference image. Higher SSIM values indicate better preservation of image structure and fewer artifacts [63].

Protocol 2: fNIRS Motion Artifact Management During a Behavioral Task

This protocol outlines steps for minimizing and handling motion artifacts in fNIRS studies, which are common during behavioral tasks that involve speaking or moving [62].

1. Pre-Experimental Preparation:

• Headcap and Optode Fitting: Select a well-fitting headcap. Part the hair under each optode to ensure direct scalp contact. Use spring-loaded optode holders to maintain consistent pressure and compensate for varying hair thickness [62].
• Participant Instruction: Clearly instruct the participant to minimize head movements not required by the task (e.g., avoiding excessive frowning or jaw clenching).

2. Data Acquisition & Monitoring:

• Monitor the fNIRS signal in real-time for characteristic motion artifacts:
  • Spikes: Sudden, brief signal changes.
  • Baseline Shifts: Sustained changes in signal level.
  • Oscillatory Artefacts: Periodic fluctuations from repetitive motion like nodding [62].
• Note the timing of any observed movements or artifacts.

3. Post-Processing Correction:

• Artifact Detection: Use algorithms like moving standard deviation to automatically identify periods of high signal variance indicative of motion [62].
• Artifact Reduction: Apply signal processing techniques such as:
  • Spline Interpolation: To correct for baseline shifts and spikes [62].
  • Wavelet-Based Filtering: To separate motion artifacts from the physiological signal of interest [62].
  • Eigenvector-Based Spatial Filtering: To reduce physiological interference that may be correlated with motion [62].

---

The Scientist's Toolkit: Research Reagent Solutions

Table: Essential Components for a Motion Correction Pipeline

Item / Solution	Function & Purpose
Markerless Optical Tracking System (e.g., Tracoline TCL)	Provides high-frame-rate (e.g., 30 Hz), real-time estimates of head pose without requiring attached markers. This is the core hardware for enabling PMC [63].
PMC-Capable Pulse Sequence	A modified MRI sequence (e.g., MPRAGE, EPI) that can receive external motion data and dynamically adjust the imaging FOV and encoding during the scan [63].
Retrospective Motion Correction Software (e.g., retroMoCoBox)	A software package for post-processing that corrects acquired k-space data by adjusting trajectories based on recorded motion. Essential for RMC and HMC [63].
Spring-Loaded Optode Holders (for fNIRS)	Maintains consistent and optimal pressure of fNIRS optodes against the scalp, automatically adjusting to reduce motion-induced signal loss caused by changing scalp contact [62].
Motion Robust Algorithms (e.g., Moving STD, Spline Interpolation)	Software tools for detecting (e.g., via moving standard deviation) and correcting (e.g., via spline interpolation) motion artifacts in modalities like fNIRS and EEG [62].

Frequently Asked Questions

1. What is the fundamental difference between PSNR and SSIM for evaluating image quality?

PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure) are both full-reference image quality metrics, but they assess different aspects of quality.

• PSNR is a pixel-based metric that calculates the ratio between the maximum possible power of a signal and the power of corrupting noise, based on the Mean Squared Error (MSE) between the original and the corrected image. It is mathematically simple and has a clear physical meaning but is often criticized for poorly correlating with perceived visual quality [64] [65].
• SSIM is a perception-based model designed to measure the perceived change in structural information between two images. It incorporates comparisons of luminance, contrast, and structure, making it more aligned with the human visual system's sensitivity to structural distortions than absolute pixel-wise errors [66] [64] [65].

2. When should I use SSIM over PSNR in my motion artifact correction research?

You should prioritize SSIM in scenarios where the goal is to evaluate the perceptual quality and structural integrity of the corrected image, especially when dealing with specific types of distortions.

• Blurring and Compression Artifacts: SSIM is more sensitive to structural distortions like blurring (common in motion artifacts) and compression artifacts, where it demonstrates a higher correlation with subjective human evaluation than PSNR [64].
• Clinical Decision Support: Research has shown that a CNN model predicting SSIM can classify whether a brain MR image requires rescanning due to motion artifacts with high sensitivity (89.5%) and specificity (78.2%), and an Area Under the Curve (AUC) of 0.930 [67]. This demonstrates SSIM's direct applicability to clinical workflow decisions.

Use PSNR when you need a computationally simple, objective measure of pixel-level error, particularly during algorithm development and tuning [65].

3. What are the key limitations of PSNR and SSIM that I should be aware of?

Despite their widespread use, both metrics have notable drawbacks:

• Limitations of PSNR: It can be a poor indicator of visual quality, as demonstrated by examples where images with identical MSE (and hence PSNR) can have drastically different perceptual quality [64].
• Limitations of SSIM: Its performance can be unpredictable. There are cases where images with similar SSIM scores have vastly different visual quality. Furthermore, SSIM is highly sensitive to geometric transformations like spatial shifts, rotations, and scaling, which can lead to misleadingly low scores [64]. The Complex Wavelet SSIM (CW-SSIM) variant was developed to be more robust to these transformations [66] [64].

4. What are the typical benchmark values for SSIM and PSNR in successful motion artifact correction?

Performance benchmarks from recent studies on brain MRI motion correction provide a reference for what constitutes successful correction. The table below summarizes quantitative results from different deep learning approaches.

Table 1: Benchmark Performance of Deep Learning Models for Motion Artifact Correction

Study & Model	Dataset	Reported SSIM	Reported PSNR (dB)	Key Finding
CGAN for Motion Reduction [68]	Head T2-weighted MRI (1.5T)	> 0.9	> 29	A model trained on both horizontal and vertical artifact directions achieved high robustness.
Conditional GAN (CGAN) [68]	Head T2-weighted MRI (1.5T)	~26% improvement	~7.7% improvement	The CGAN model showed the closest image reproducibility to the original, motion-free images.
U-Net (Trained on Real Paired Data) [46]	MR-ART Brain Dataset	0.858 ± 0.079	(Not Reported)	Serves as an upper-bound benchmark for models with access to real motion-corrupted/clean pairs.
U-Net (Trained on Synthetic Data) [46]	MR-ART Brain Dataset	0.821 ± 0.096	(Not Reported)	Demonstrates performance achievable with synthetically generated training data.
Diffusion Model [46]	MR-ART Brain Dataset	0.815 ± 0.103	(Not Reported)	Can produce high-quality corrections but is susceptible to harmful hallucinations.

The Scientist's Toolkit: Research Reagents & Materials

Table 2: Essential Components for Motion Artifact Correction Experiments

Item / Reagent	Function / Application in Research
Open MRI Datasets (e.g., fastMRI, MR-ART)	Provides the essential raw data (k-space or images) for training and validating deep learning models [67] [46].
Deep Learning Framework (e.g., PyTorch, TensorFlow/Keras)	Offers the programming environment and tools for building, training, and testing neural network models [67] [46].
Convolutional Neural Network (CNN)	A core architecture for tasks like quantifying motion artifacts without a reference image or serving as a discriminator in a GAN [67].
Generative Adversarial Network (GAN) / Conditional GAN (CGAN)	A framework for generating high-quality, perceptually convincing motion-corrected images by pitting a generator against a discriminator network [68].
U-Net	A specialized CNN architecture with a symmetric encoder-decoder structure, highly effective for image-to-image tasks like artifact removal [46].
Diffusion Model	A state-of-the-art generative model that learns to denoise images, which can be adapted for motion correction but requires careful tuning to avoid hallucinations [46].

Experimental Protocols: Key Methodologies

Protocol 1: Training a CNN for No-Reference Motion Artifact Quantification

This protocol is based on a study that developed a CNN to predict full-reference IQA metrics like SSIM without needing a clean reference image [67].

• Data Preparation & Motion Simulation:

  • Obtain brain MR images from an open dataset (e.g., NYU fastMRI Dataset).
  • Retrospectively generate motion-corrupted images by applying random rigid transformations (rotation: 1-20 degrees; translation: 1-20 pixels) in the image domain, then simulating their effect in k-space by merging lines from the original and transformed k-space data [67].
  • Calculate the true FR-IQA metrics (PSNR, SSIM) for each motion-corrupted image using the original as a reference.

• Model Training:

  • Architecture: Design a CNN with multiple convolutional layers (e.g., 6 layers with 3x3 kernels), max-pooling layers, and fully connected layers. The output should be the predicted FR-IQA metric value(s) [67].
  • Training: Train the model using the motion-corrupted images as input and the corresponding true SSIM (or other metric) values as the target. Use a loss function like Mean Squared Error (MSE) and an optimizer like Adam [67].

• Validation:

  • Validate the model's performance by comparing its predicted SSIM values against subjective evaluations by expert readers (e.g., radiological technologists using a 5-point scale) [67].
  • Evaluate the model's clinical utility by performing a receiver operating characteristic (ROC) analysis to determine its ability to classify images that require rescanning [67].

Protocol 2: Correcting Artifacts using a Conditional GAN (CGAN)

This protocol outlines the use of a CGAN for direct image correction, as demonstrated in a study achieving high SSIM and PSNR improvements [68].

• Data Preparation:

  • Acquire paired datasets of motion-free and motion-corrupted images. As real paired data is scarce, generate simulated motion artifacts via Fourier-based methods, incorporating both translational and rotational movements. Ensure artifacts are simulated in the phase-encoding direction (both horizontal and vertical) [68].
  • Normalize all pixel values to a standard range (e.g., 0 to 1).

• Model Training:

  • Generator (G): Train a network (often a U-Net) to take a motion-corrupted image as input and output a corrected image.
  • Discriminator (D): Train a CNN to distinguish between the generator's "fake" corrected images and the real "clean" images.
  • Adversarial Process: The generator and discriminator are trained simultaneously. The generator tries to produce images that fool the discriminator, while the discriminator gets better at telling them apart. This process, guided by adversarial loss, encourages the generator to produce highly realistic and structurally accurate images [68].

• Evaluation:

  • Use the trained generator to correct motion-corrupted images from the test set.
  • Quantitatively evaluate the output by calculating SSIM and PSNR between the corrected image and the original, motion-free ground truth [68].
  • Perform qualitative assessment by visual inspection to check for anatomical accuracy and the absence of smoothing or hallucinated features.

Workflow Visualization

The following diagram illustrates a generalized experimental workflow for developing and benchmarking a motion artifact correction system, integrating the key steps from the protocols above.

Experimental Workflow for Motion Correction

Methodology Showdown: Validating Performance Across Techniques

Frequently Asked Questions (FAQs)

Q1: Under what conditions would a deep learning model like a Denoising Autoencoder (DAE) be preferable to traditional methods like wavelet filtering for my fNIRS data?

A1: A deep learning model is preferable when you require an assumption-free approach with minimal need for expert parameter tuning and have access to sufficient computational resources (e.g., a GPU). The DAE has been shown to outperform conventional methods by automatically learning noise features, leading to lower residual motion artifacts and decreased mean squared error (MSE) [69] [60]. In contrast, wavelet filtering requires you to subjectively tune parameters like the probability threshold alpha, and its performance relies on the assumption that wavelet coefficients are normally distributed [69].

Q2: One major drawback of PCA is its dependence on the number of channels. How does a deep learning approach circumvent this limitation?

A2: You are correct that Principal Component Analysis (PCA) is limited by the total number of channels available and depends on the geometry of the probes [69]. Deep learning models, such as the convolutional neural network (CNN)-based Denoising Autoencoder (DAE) described in the search results, learn features directly from the data itself. They do not rely on constructing components from a limited set of channels. This data-driven approach allows them to model complex, nonlinear artifacts without being constrained by the probe setup, making them more robust and scalable across different experimental geometries [69].

Q3: For a lab with limited computational resources, are traditional methods still a viable option for motion artifact removal?

A3: Yes, traditional methods remain viable and are widely used. Techniques like spline interpolation, wavelet filtering, and PCA are effective for many applications and are less computationally intensive than training a deep learning model from scratch [69]. The key is to be aware of their specific requirements and limitations. For instance, when using wavelet filtering, you must carefully select the probability threshold (alpha), and for PCA, the number or proportion of components to be removed must be tuned [69]. Your choice should balance the required accuracy against the available computational resources and expertise.

Q4: The thesis context involves behavioral tasks, which often induce motion. Is deep learning robust enough for these real-world scenarios?

A4: Yes, deep learning models are particularly promising for behavioral research because they are often trained and validated on data that includes a wide variety of realistic motion artifacts. For example, one cited study trained a model on a public dataset containing artifacts from actions like reading aloud, head nodding, and twisting [69]. Furthermore, models like Motion-Net have been designed specifically for processing data from moving subjects, demonstrating the ability to handle motion artifacts on a subject-specific basis, which is crucial for real-world experimental data [70].

Troubleshooting Guides

Issue 1: Poor Denoising Performance with Wavelet Filtering

Symptoms: After applying wavelet filtering, the signal of interest (e.g., hemodynamic response) appears over-smoothed or residual high-frequency noise remains.

Diagnosis: Incorrect parameter selection, specifically the probability threshold (alpha), which controls the identification of outliers in the wavelet coefficients [69].

Solution:

• Visual Inspection: Begin by visually inspecting your raw data to identify the amplitude and duration of typical motion artifacts.
• Parameter Sweep: Systematically test a range of alpha values (e.g., from 0.01 to 0.05) on a representative subset of your data.
• Validation: Evaluate the performance by checking the preservation of the expected hemodynamic response shape and the reduction in obvious spike or shift artifacts. The optimal alpha is often dataset-specific and requires this empirical tuning.

Issue 2: PCA Removing Neural Signals Along with Artifacts

Symptoms: The cleaned signal appears overly attenuated, and task-related brain activity seems diminished after PCA application.

Diagnosis: An excessive number of principal components (PCs) have been removed. Since PCA is a blind source separation method, early components may contain not only motion artifacts but also neural signals of interest [69] [71].

Solution:

• Conservative Component Removal: Start by removing a very small number of components (e.g., the first 1-2 PCs) and gradually increase.
• Task Correlation Check: Compare the cleaned signal to your task paradigm. If the task-evoked response is weakened after removing a certain component, that component may contain neural signals.
• Use a Reference: The cited study notes that selecting the percentage of variance for PCA to remove is subjective, and values of 0.80 vs. 0.97 can produce significantly different results [69]. Use a ground-truth simulation or a separate clean dataset to calibrate this parameter if possible.

Issue 3: Long Processing Times with Deep Learning Models

Symptoms: Model training or inference is taking an impractically long time.

Diagnosis: Deep learning models can be computationally intensive, especially if you are working on a CPU or with a very complex model architecture.

Solution:

• Leverage Hardware Acceleration: Ensure you are using a GPU for model training and application. The significant acceleration of deep learning pipelines is a key advantage highlighted in the results [72].
• Transfer Learning: Consider using a pre-trained model if one is available for your neuroimaging modality. You can then fine-tune it on your specific dataset, which requires less data and computation than training from scratch.
• Optimize Data Loading: Use efficient data loaders that can pre-fetch data and ensure your input data pipeline is not the bottleneck.

Quantitative Performance Comparison

The table below summarizes key performance metrics from evaluated studies, providing a direct comparison of the efficacy of different artifact removal strategies.

Table 1: Performance Metrics of Motion Artifact Removal Methods

Method Category	Specific Method	Key Performance Metric	Reported Result	Neuroimaging Modality
Deep Learning	Denoising Autoencoder (DAE) [69] [60]	Outperformed conventional methods	Lower residual motion artifacts & decreased MSE	fNIRS
Deep Learning	Multiscale Fully Convolutional Network [73]	Reduction in Mean Squared Error	41.84% reduction	Structural MRI
Deep Learning	De-Artifact Diffusion Model [48]	Root Mean Square Error (RMSE)	11.44 ± 1.94	Knee MRI
Deep Learning	De-Artifact Diffusion Model [48]	Peak Signal-to-Noise Ratio (PSNR)	32.12 ± 1.41 dB	Knee MRI
Deep Learning	De-Artifact Diffusion Model [48]	Structural Similarity Index (SSIM)	0.968 ± 0.012	Knee MRI
Traditional	Wavelet Filtering [69]	Dependency	Requires tuning of probability threshold (alpha)	fNIRS
Traditional	PCA [69]	Dependency	Requires tuning of components/variance to remove	fNIRS

Experimental Protocols for Key Cited Studies

Protocol 1: Denoising Autoencoder (DAE) for fNIRS

This protocol is based on the methodology from the study "Deep learning-based motion artifact removal in functional near-infrared spectroscopy" [69].

• Data Simulation: Generate a large synthetic fNIRS training dataset. The noisy signal ((F'(t))) is a superposition of:
  • A clean hemodynamic response function ((F(t))) simulated with a gamma function.
  • Motion artifacts ((\Phi{MA}(t))), including spike noise (modeled with a Laplace distribution) and shift noise (DC changes).
  • Resting-state fNIRS background ((\Phi{rs}(t))), simulated using an Autoregressive (AR) model fitted to experimental data.
• Model Architecture: Implement a Denoising Autoencoder (DAE) with a stack of nine convolutional layers, followed by max-pooling layers.
• Training: Train the DAE using a dedicated loss function designed to effectively separate motion artifacts from the true neural signal.
• Validation: Benchmark the model's performance against conventional methods (e.g., spline interpolation, wavelet, PCA) on both synthetic and open-access experimental fNIRS data, using metrics like MSE and computational time.

Protocol 2: Motion-Net for EEG Artifact Removal

This protocol is based on the study "Motion Artifact Removal from EEG Signals Using the Motion-Net Deep Learning Algorithm" [70].

• Data Preprocessing: Synchronize EEG and accelerometer (Acc) data by aligning experiment triggers and resampling. Perform baseline correction.
• Model Architecture: Employ a 1D U-Net Convolutional Neural Network (CNN) architecture, which is well-suited for signal reconstruction tasks.
• Feature Selection (Optional): Evaluate and select input features for the network, such as the raw EEG signal or a "visibility graph" that represents the signal's structural properties.
• Training and Testing: Train the model on a subject-specific basis, processing single trials separately. This approach is designed to work effectively with relatively small datasets.
• Performance Evaluation: Assess the model's performance by comparing the cleaned signal to a ground truth (when available) using correlation coefficients and qualitative assessment.

Workflow Visualization

Diagram 1: Traditional vs. Deep Learning Artifact Removal

Diagram 2: fNIRS DAE Training & Validation

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 2: Key Resources for Motion Artifact Removal Experiments

Item Name	Function / Description	Relevance in Cited Studies
Synthetic Data Generation Pipeline	Generates large-scale training data by combining simulated clean signals with modeled noise and artifacts.	Critical for training the fNIRS DAE model, as it created a dataset with known ground truth [69].
Denoising Autoencoder (DAE)	A deep learning architecture that learns to map noisy input data to clean output data.	The core model used for assumption-free motion artifact removal in fNIRS signals [69] [60].
Convolutional Neural Network (CNN)	A class of deep neural networks most commonly applied to analyzing visual imagery, but also effective for 1D signals.	Used in various forms: as the backbone of the DAE [69], in multiscale FCN for MRI [73], and in U-Net for EEG [70].
Public Neuroimaging Datasets	Curated, often open-access, datasets containing neuroimaging data with and without motion artifacts.	Used for training and validation; e.g., the fNIRS DAE study used a public dataset with induced motion artifacts [69].
Workflow Manager (e.g., Nextflow)	Software that manages complex, multi-step computational workflows, enabling scalability and reproducibility.	Empowers pipelines like DeepPrep to efficiently process tens of thousands of scans across different computing environments [72].

Frequently Asked Questions (FAQs) on Motion Artifact Correction

Q1: What are the most common types of artifacts in simultaneous EEG-fMRI studies, and how do they differ? Simultaneous EEG-fMRI is compromised by two main MR artifacts with distinct properties. Gradient (or Imaging) Artifacts are induced by rapid switching of gradient coils and RF pulses during echo-planar imaging (EPI). They are very high amplitude (up to tens of mV) and have a deterministic, periodic shape tied to the slice and volume repetition frequencies. Ballistocardiogram (BCG) Artifacts originate from head and electrode movement caused by cardiac activity (e.g., blood pulsation). They are typically smaller (exceeding 50 μV at 3T) but have most of their spectral power below 25 Hz, directly overlapping with the frequency range of neuronal electrical activity of interest, making them particularly challenging to remove [74].

Q2: For a study focusing on beta band oscillations during a motor task, which artifact reduction method is recommended? Empirical evaluations suggest that a Carbon-Wire Loop (CWL) system is superior for improving spectral contrast in specific frequency bands. One study found that using a CWL system for reference signal regression was significantly more successful than other methods at recovering spectral contrast in both the alpha and beta bands. This method uses physical wires that capture MR-induced artifacts independently of the scalp EEG, providing a clean reference for artifact subtraction [74].

Q3: Our structural MRI data from a Parkinson's disease cohort has motion corruption. Can this be corrected retrospectively? Yes, 3D Convolutional Neural Network (CNN)-based retrospective correction has been validated on real patient data, including those with Parkinson's disease. One framework was trained on simulated motion and successfully applied to the Parkinson's Progression Markers Initiative (PPMI) dataset. The correction led to a measurable improvement in cortical surface reconstruction quality and enabled the detection of more widespread, statistically significant cortical thinning in patients, which was restricted before correction. This confirms its utility for studies of movement disorders [75].

Q4: How can I augment a small neuroimaging dataset with synthetic data to improve a classification model? Generative models, such as Kernel-Density Estimation (KDE) models, can create synthetic, normative regional volumetric features. The GenMIND resource, for instance, provides 18,000 synthetic samples derived from real MRI data. Using such synthetic data to augment a small real dataset, especially the control group, has been shown to significantly enhance the accuracy of downstream machine learning models for tasks like disease classification, leading to more robust results [76].

Quantitative Data Comparison of Correction Methods

The following table summarizes empirical data on the performance of different artifact correction methods, providing a basis for selection.

Table 1: Performance Comparison of MR Artifact Reduction Methods for EEG-fMRI

Method Category	Example Method	Key Performance Findings	Validated On Real Patient Data?
Reference Signal	Carbon-Wire Loops (CWL)	Superior in improving spectral contrast in alpha/beta bands; better at recovering visual evoked responses [74]	Yes, during resting-state, motor, and visual tasks [74]
Template Subtraction	Average Artifact Subtraction (AAS)	Effective but can leave residual artifacts if motion occurs or if neuronal activity is correlated with the averaging period [74]	Yes, widely used as a baseline method [74]
AI-Based Correction	3D CNN for Structural MRI	PSNR improved from 31.7 dB to 33.3 dB; reduced QC failures in PPMI dataset from 61 to 38; revealed more significant cortical thinning in Parkinson's disease [75]	Yes, on ABIDE, ADNI, and PPMI datasets [75]

Experimental Protocols for Validation

Protocol 1: Validating EEG-fMRI Artifact Reduction Using a Carbon-Wire Loop System

• Aim: To quantify the efficacy of the CWL method in recovering neuronal signals during a task.
• Methodology:
  • Data Acquisition: Collect simultaneous EEG-fMRI data during paradigms with robust, known neural responses (e.g., visual evoked potentials, motor beta modulation). A set of six carbon-wire loops should be placed close to the scalp to record the MR artifact.
  • Reference Recording: Record the CWL reference signals concurrently with the EEG and fMRI data.
  • Data Processing: Perform artifact correction using the CWL signals as reference for regression, comparing it to other software-only methods (e.g., AAS variants).
  • Benchmarking: Acquire a separate, artifact-free EEG recording from the same participant outside the MRI scanner using the same task paradigm. This serves as the gold-standard benchmark.
  • Validation Metrics:
    • Spectral Domain: Calculate the oscillatory power in the frequency band of interest (e.g., alpha, beta) during task versus rest. Compare the task-induced spectral contrast of the corrected in-scanner EEG to the benchmark out-of-scanner EEG.
    • Time Domain: For evoked responses, compare the amplitude and latency of key peaks (e.g., P100) in the corrected in-scanner EEG to the benchmark [74].

Protocol 2: Validating AI-Based Motion Correction for Structural MRI

• Aim: To assess whether a deep learning-based motion correction algorithm improves image quality and downstream analysis outcomes.
• Methodology:
  • Model Training: Train a 3D CNN using a large set of motion-free structural T1-weighted images (e.g., from ABIDE I). Corrupt these images with simulated motion in the Fourier domain to create input-target pairs for the network [75].
  • Testing with Real Motion:
    • Paired Data: Use a test dataset (e.g., from ADNI) where subjects have both a motion-corrupted T1 volume and a subsequent "clean" re-scan. Apply the trained model to the corrupted volume.
    • Quantitative Metrics: Calculate Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) between the corrected image and the clean re-scan [75].
  • Downstream Analysis Validation:
    • Cortical Reconstruction: On a larger cohort without paired rescans (e.g., PPMI), perform automated cortical surface reconstruction (e.g., using FreeSurfer) on both uncorrected and CNN-corrected images.
    • Blinded Quality Control: Have raters manually assess the quality of cortical surface reconstructions for both sets in a blinded fashion. Record the number of QC failures.
    • Clinical Correlation: Run a group analysis (e.g., patients vs. controls) on a derived metric like cortical thickness using both uncorrected and corrected data. Compare the statistical significance and spatial extent of the findings [75].

Workflow and Signaling Diagrams

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Tools for Motion Artifact Research and Correction

Research Reagent / Tool	Function / Description	Example Use Case
Carbon-Wire Loops (CWL)	A hardware solution of conductive loops placed near the head to record a reference signal of the MR artifact isolated from neuronal EEG [74].	Providing a pure reference for regression-based removal of gradient and BCG artifacts in EEG-fMRI studies [74].
Average Artifact Subtraction (AAS)	A software algorithm that creates a template of the imaging artifact by averaging EEG signal over many volume or slice acquisition periods, then subtracts it [74].	A baseline and widely available method for initial reduction of large-amplitude gradient artifacts in EEG-fMRI.
3D Convolutional Neural Networks (3D CNN)	A deep learning architecture designed to process volumetric data. Can be trained to map motion-corrupted images to their clean counterparts [75].	Retrospective correction of motion artifacts in structural T1-weighted MRI scans, improving cortical surface reconstruction [75].
Kernel-Density Estimation (KDE) Generative Model	A non-parametric statistical model used to estimate the probability density function of a dataset. Can sample new synthetic data points from this distribution [76].	Generating synthetic, normative brain volumetric features to augment small datasets and improve machine learning model robustness [76].
Structural Similarity Index (SSIM) & Peak SNR	Image quality metrics used to quantitatively compare a processed image to a ground-truth reference image [75].	Objectively validating the performance of a motion correction algorithm on a test dataset with paired motion-corrupted and clean scans [75].

Troubleshooting Guide: Motion Artifacts in Functional Connectivity Analysis

FAQ: Motion Artifacts and Default Mode Network Analysis

Q1: Why does participant motion severely impact our Default Mode Network (DMN) functional connectivity results, and how can we detect this?

Motion artifacts introduce signal changes that can be misrepresented as neural activity, severely confounding functional connectivity analysis. Even small movements can cause spurious signal changes that mimic the subtle fluctuations (typically 1-2%) associated with true neural activity in resting-state fMRI [17]. Motion has been shown to specifically affect anterior-posterior connections within the DMN, potentially creating false hypoconnectivity findings [77]. To detect motion-related contamination:

• Check frame-wise displacement metrics calculated during preprocessing
• Inspect the correlation between motion parameters and your connectivity results
• Use qualitative assessment to identify ghosting or blurring in images [10] [17]

Q2: What are the most effective strategies to minimize motion artifacts during data acquisition?

Proactive prevention is crucial for reliable DMN analysis. Implement these strategies during scanning:

• Patient Preparation: Provide clear instructions, practice sessions, and proper head fixation using foam padding [17]
• Technical Solutions: Utilize fast imaging sequences to reduce acquisition time and consequently minimize motion window [10]
• Monitoring: Implement real-time motion tracking systems for prospective correction [17]

Q3: Our preprocessing includes motion realignment, but we still suspect motion contamination in DMN connectivity. What advanced methods should we consider?

Standard realignment alone is often insufficient. Consider these advanced approaches:

• Data Scrubbing: Remove motion-contaminated volumes using framewise displacement thresholds [78]
• ICA-Based Denoising: Use ICA-AROMA to automatically identify and remove motion-related components without reducing degrees of freedom [20]
• Regression Methods: Include motion parameters and their temporal derivatives as nuisance regressors [79]

Q4: How does motion specifically affect different neuroimaging applications beyond basic fMRI?

Motion artifacts manifest differently across modalities:

Table: Motion Artifact Impact Across Neuroimaging Applications

Application	Primary Impact	Special Considerations
Resting-state fMRI	Spurious functional connectivity patterns [78] [77]	Can mimic DMN hypoconnectivity; particularly affects anterior-posterior connections [77]
Task-based fMRI	Reduced sensitivity to true activation; false positives correlated with task timing [17]	Motion can be correlated with task performance, creating confounds
Diffusion Tensor Imaging	Misalignment of data; inaccurate fiber tracking [17]	Long acquisition times increase sensitivity; requires specific motion-compensated sequences
Arterial Spin Labeling	Blurring; inaccurate cerebral blood flow quantification [17]	Background suppression schemes can help mitigate effects
High-Resolution Structural	Reduced cortical thickness measures; false atrophy appearance [17]	Impacts morphological measurements

Q5: We're using a multimodal approach with simultaneous EEG-fMRI. What specialized motion artifact challenges should we anticipate?

Simultaneous EEG-fMRI introduces unique motion-related challenges:

• Complex Artifacts: Motion induces currents in EEG electrodes via Faraday's Law, creating large artifacts obscuring neural signals [80]
• Multiple Sources: Must address gradient, ballistocardiogram, and motion artifacts simultaneously [80]
• Synchronization Issues: Motion affects EEG and fMRI data differently, requiring specialized correction pipelines [80]

Recommended solutions include model-based approaches like Average Artifact Subtraction and advanced data-driven methods such as ICA, though no single method addresses all artifacts completely [80].

Advanced Motion Correction Methodologies

Experimental Protocol: ICA-AROMA for Motion Artifact Removal

ICA-AROMA (ICA-based Automatic Removal of Motion Artifacts) provides robust motion cleanup while preserving neural signal integrity [20].

Procedure:

• Perform standard fMRI preprocessing including motion realignment and spatial normalization
• Execute high-dimensional ICA to decompose data into 75-100 components
• Automatically classify components as motion-related using four robust features:
  • High-frequency content
  • Correlation with motion parameters
  • Edge fraction (spatial characteristic)
  • CSF fraction
• Remove identified motion components from data
• Reconstruct cleaned fMRI time series

Advantages:

• Preserves temporal autocorrelation structure
• Maintains degrees of freedom compared to scrubbing
• Requires no classifier retraining across datasets
• Validated for both resting-state and task-based fMRI [20]

Experimental Protocol: Motion Scrubbing for Severe Contamination

For datasets with significant motion, implement the scrubbing protocol [78]:

• Calculate framewise displacement (FD) from motion parameters
• Identify "bad" volumes exceeding FD threshold (typically 0.2-0.5mm)
• Remove contaminated volumes and interpolate gaps
• Include nuisance regressors for removed volumes

Table: Comparative Performance of Motion Correction Methods

Method	Motion Reduction Efficacy	Impact on Temporal Structure	Best Use Cases
24-Parameter Regression	Moderate	Preserved	Mild motion; initial preprocessing
Spike Regression	High	Disrupted (reduces degrees of freedom)	Severe, isolated motion spikes
ICA-AROMA	High	Preserved	Most applications; balance of efficacy and structure preservation [20]
Motion Scrubbing	High	Disrupted (volume removal)	Extreme motion cases [78]
Prospective Correction	Variable	Preserved	Real-time systems; cooperative participants

Research Reagent Solutions: Motion Artifact Toolkit

Table: Essential Tools for Motion Artifact Management

Tool/Technique	Function	Implementation Considerations
Framewise Displacement Metric	Quantifies volume-to-volume motion	Calculate from realignment parameters; threshold 0.2-0.5mm
ICA-AROMA	Automatic identification and removal of motion components	Use standard settings; compatible with FSL preprocessing [20]
Motion Parameter Regression	Removes motion-related variance via general linear model	Include 24 parameters (6 + derivatives + squares); effective for mild motion
Multiecho Acquisition	Enhances denoising capabilities	Requires specialized sequences; improves ICA decomposition
Structural Image Integration	Improves spatial normalization accuracy	High-resolution T1-weighted reference reduces misalignment
Motion-Tracked fMRI	Real-time motion monitoring and correction	Hardware-dependent (e.g., camera systems); prospective correction

Workflow Visualization

Motion Artifact Management Workflow for Reliable DMN Analysis

Causal Pathway of Motion Artifact Impact on DMN Research

Technical Support Center

Troubleshooting Guides

Guide 1: Troubleshooting Model Generalization for Motion Artifact Detection

Problem: A deep learning model for rating motion artifacts performs well on internal test data but shows significantly reduced accuracy when applied to data from new patient populations or different scanner hardware.

Explanation: This is a classic problem of domain shift. Models trained on data from specific scanners and populations learn features that may not transfer perfectly to new environments. For instance, a study on a Deep CNN for motion artifact evaluation reported high accuracy (100%) on its original test set but showed a substantial drop in performance on data from different domains, including epilepsy patients (90.3%), images with susceptibility artifacts (63.6%), and data from different scanner vendors (75.0%) [33].

Solution Steps:

• Assess Data Discrepancies: Check for differences in image contrast, resolution, and noise characteristics between your training data and the new target data.
• Implement Domain Adaptation: If retraining is not possible, use techniques like histogram matching to normalize the appearance of new images to match your training set.
• Fine-Tune the Model: If new labeled data is available, fine-tune your pre-trained model on a small sample from the new domain. The original study used transfer learning and fine-tuning to achieve good performance with a limited amount of medical imaging data [33].
• Consider a Multi-Source Approach: For future model development, train on a multi-site, multi-scanner, and multi-population dataset to build inherent robustness.

Guide 2: Validating Portable MRI Data Against High-Field Systems

Problem: You are using a portable, low-field-strength (64mT) MRI scanner for research but need to ensure that the derived biomarkers (e.g., brain volume measurements) are comparable to those from standard high-field (3T) systems used in clinical trials.

Explanation: Portable low-field MRI scanners increase access but produce lower quality images with different contrast and resolution compared to high-field scanners [81]. This can lead to systematic biases in quantitative measurements.

Solution Steps:

• Image Synthesis: Use a dedicated image translation model, such as the LowGAN framework, to generate synthetic high-field-quality images from your low-field inputs. This has been shown to improve the agreement of volumetric measurements (e.g., thalamic volume) with 3T ground truth [81].
• Cross-Validation Study: Design a sub-study where a cohort of participants is scanned on both the portable and a high-field scanner. This creates a paired dataset for validation.
• Correlate Key Outcomes: Statistically compare the primary outcomes (e.g., cortical volumes, lesion counts) between the actual high-field scans and either the native low-field or synthetic high-field scans to establish validity. Ensure the synthetic images preserve clinically relevant features like white matter lesions [81].

---

Frequently Asked Questions (FAQs)

Q1: Our multi-site study uses different 3T MRI scanners from various vendors. What is the best way to harmonize data and mitigate site-specific artifacts for motion detection?

A1: Scanner-specific differences in hardware and software can introduce confounding variability.

• Standardized Protocols: Implement identical imaging protocols across all sites to the greatest extent possible.
• Phantom Scans: Regularly scan a standardized phantom at all sites to quantify and monitor inter-scanner differences in geometric distortion and intensity uniformity.
• Harmonization Algorithms: In post-processing, use statistical harmonization techniques like ComBat to remove site-specific effects while preserving biological signals of interest.
• Sequence Choice: Be aware that accelerated imaging sequences (e.g., simultaneous multislice acquisition) can have different sensitivities to motion and may complicate artifact identification [17].

Q2: When we apply a pre-trained motion artifact detection model to our new dataset, it fails. What are the first things we should check?

A2:

• Image Preprocessing: Ensure your preprocessing pipeline (spatial normalization, skull-stripping, intensity normalization) matches exactly what was used during the model's training.
• Input Specifications: Verify that your image dimensions, voxel resolution, and contrast (e.g., T1-weighted) match the model's required input.
• Quality of Failures: Manually inspect a sample of the images the model failed on. Look for systematic differences in artifact appearance or image quality compared to the training data. The model may be encountering motion patterns or scanner artifacts it has never seen before [33].

Q3: Can portable MRI reliably be used for longitudinal monitoring of disease progression in clinical trials?

A3: Evidence is growing, especially for conditions with clear macroscopic markers. Studies on Multiple Sclerosis (MS) patients have shown that with advanced processing, portable MRI can reliably capture key biomarkers. For example, synthetic high-field images generated from portable scanner inputs preserved MS lesions and captured the known inverse relationship between total lesion volume and thalamic volume [81]. This suggests strong potential for longitudinal tracking, particularly when access to traditional MRI is a limiting factor.

---

Data and Protocols

Table 1: Performance of a Deep CNN for Motion Artifact Detection Across Different Domains

This table summarizes the generalizability test results of a reference-free deep learning method for rating motion artifacts, highlighting the challenge of domain shift [33].

Test Dataset	Description	Reported Acquisition-Based Accuracy
Internal Test Set	Original dataset from healthy volunteers	100.0%
Generalization Test 1	MR images from epilepsy patients (93 acquisitions)	90.3%
Generalization Test 2	MR images presenting susceptibility artifact (22 acquisitions)	63.6%
Generalization Test 3	MR images from a different scanner vendor (20 acquisitions)	75.0%

Table 2: Key Specifications: Portable vs. High-Field MRI

This table compares typical technical specifications of a portable low-field scanner with a standard high-field system, underlying the source of the generalizability challenge [81].

Parameter	Portable Low-Field MRI (e.g., Hyperfine SWOOP)	Standard High-Field MRI (e.g., 3T)
Magnetic Field Strength	64 mT	3 T (3,000 mT)
Typical T1w Resolution	1.5 × 1.5 × 5 mm	1 mm isotropic
Typical FLAIR Resolution	1.6 × 1.6 × 5 mm	1 mm isotropic
Key Advantages	Portability, lower cost, bedside use, enhanced access	Higher signal-to-noise ratio (SNR), standard for clinical diagnosis

Experimental Protocol 1: Implementing a Low- to High-Field Image Translation Model (LowGAN)

Purpose: To generate synthetic high-field-quality images from portable low-field MRI inputs to improve the portability and reliability of quantitative measurements [81].

Methodology:

• Data Collection and Preprocessing:
  • Acquire a paired dataset of participants scanned on both portable (64mT) and high-field (3T) scanners.
  • Rigidly co-register all low-field and high-field images (T1w, T2w, FLAIR) to the high-field T1w space.
  • Skull-strip the images and perform intensity normalization.
• Model Training - Stage 1 (2D Image Translation):
  • Use a pix2pix Generative Adversarial Network (GAN) architecture.
  • Train three separate models on 2D slices from the axial, coronal, and sagittal planes.
  • Input: 2D slice from a low-field volume (multi-contrast channels).
  • Target: Corresponding 2D slice from the high-field volume.
  • Train for 30 epochs at a fixed learning rate (1x10⁻⁴), followed by 70 epochs with linear decay.
• Model Training - Stage 2 (3D Refinement):
  • Use a 3D U-Net to combine the outputs from the three orthogonal models from Stage 1.
  • Input: The three reconstructed orthogonal volumes.
  • Target: The original high-field 3D volume.
  • Train using a mixed L1 and Structural Similarity (L1-SSIM) cost function to reduce artifacts and improve edge preservation.
• Validation:
  • Quantitatively compare synthetic images with actual high-field images using metrics like Normalized Cross-Correlation (NCC).
  • Validate key biomarkers (e.g., brain volumetry, lesion volume) against ground truth 3T measurements.

Experimental Protocol 2: Building a Generalizable Motion Artifact Detector with Deep CNN

Purpose: To create a deep learning model that can accurately rate the severity of motion artifacts in neuroimages across diverse datasets [33].

Methodology:

• Model Selection and Adaptation:
  • Select a pre-trained CNN architecture (e.g., initially trained on ImageNet).
  • Modify the architecture by using shallower networks, as lower-level features are more effective for detecting fine-grained corruption in small image patches.
• Training with Limited Data:
  • Use transfer learning and fine-tuning on a smaller annotated medical image dataset.
  • Train separate models for each orthogonal MR acquisition axis (axial, coronal, sagittal) to detect motion per-patch.
• Ensemble for Robustness:
  • Form an ensemble classifier that combines the predictions from all shallower architectures and all three axes.
  • This consensus approach significantly improves the final acquisition-based decision.
• Generalization Testing:
  • Systematically test the final model on distinct datasets from different patient populations (e.g., epilepsy), with different artifact types (e.g., susceptibility), and from different scanner vendors.

---

The Scientist's Toolkit

Research Reagent Solutions

Item Name	Function / Explanation
Paired Low/High-Field Dataset	A set of images from the same participant scanned on both portable low-field and traditional high-field MRI. Essential for training and validating image translation models like LowGAN [81].
Pre-trained CNN Models	Deep learning models (e.g., ResNet, VGG) previously trained on large natural image datasets (e.g., ImageNet). Serves as a robust feature extractor, enabling effective model training with limited medical image data via transfer learning [33].
Generative Adversarial Network (GAN)	A framework for image-to-image translation, consisting of a generator that creates images and a discriminator that critiques them. Used to synthesize high-quality images from lower-quality inputs [81].
Neurodesk	A containerized data analysis environment that provides a vast suite of neuroimaging software (e.g., FSL, FreeSurfer) in version-controlled containers. This ensures reproducible analysis across different computing systems, mitigating software dependency issues [82].
Rigid Body Transformation	A preprocessing step that realigns image volumes to correct for head movement by modeling the head as a rigid body with six degrees of freedom (three translations, three rotations). This is a common first step to mitigate motion artifacts in functional time-series data [17].

Workflow Diagrams

Diagram 1: Cross-Scanner Generalizability Validation

Diagram 2: Portable MRI Data Analysis Workflow

Conclusion

Addressing motion artifacts requires a multifaceted 'toolbox' approach, as no single solution is universally effective. The integration of simple preventative measures with advanced computational corrections—particularly deep learning models like CGANs and diffusion models—shows immense promise for recovering high-fidelity data from motion-corrupted scans. For the research and pharmaceutical development community, adopting these robust correction pipelines is paramount. It ensures the reliability of functional connectivity maps and activation loci derived from behavioral tasks, which in turn strengthens the validity of biomarkers and therapeutic evaluations. Future progress hinges on developing more portable, modality-agnostic algorithms, establishing standardized validation benchmarks, and creating open-source resources to make state-of-the-art correction accessible to all.

References

• 1. Head motion: the dirty little secret of neuroimaging in psychiatry - PMC [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6306289/]
• 2. Comparison of MRI head motion indicators in 40,969 subjects informs neuroimaging study design | Scientific Reports [https://www.nature.com/articles/s41598-024-79827-9]
• 3. Chapter 3: Motion Correction — Andy's Brain Book 1.0 documentation [https://andysbrainbook.readthedocs.io/en/latest/fMRI_Short_Course/Preprocessing/Motion_Correction.html]
• 4. Motion correction and the use of motion covariates in multiple‐subject fMRI analysis - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6871380/]
• 5. Evaluating the sensitivity of functional connectivity measures to... [https://pubmed.ncbi.nlm.nih.gov/34284108/]
• 6. The association between head motion during functional magnetic resonance imaging and executive functioning in older adults - PubMed [https://pubmed.ncbi.nlm.nih.gov/37377763/]
• 7. Removal in Artifact Neuroimaging [https://www.numberanalytics.com/blog/ultimate-guide-artifact-removal-neuroimaging]
• 8. Optimizing preprocessing and analysis pipelines for single‐subject fMRI. I. Standard temporal motion and physiological noise correction methods - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC4898950/]
• 9. Quantification and Reduction of Ghosting Artifacts in Interleaved Echo-Planar Imaging - PMC [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2396297/]
• 10. in Motion : a Complex Problem with Many Partial... Artefacts MRI [https://pmc.ncbi.nlm.nih.gov/articles/PMC4517972/]
• 11. Mastering Motion Correction in Brain Imaging [https://www.numberanalytics.com/blog/mastering-motion-correction-brain-imaging]
• 12. Estimation and Removal of Residual Motion Artifact in Retrospectively Motion-Corrected fMRI Data: A Comparison of Intervolume and Intravolume Motion Using Gold Standard Simulated Motion Data | Published in Aperture Neuro [https://apertureneuro.org/article/123369-estimation-and-removal-of-residual-motion-artifact-in-retrospectively-motion-corrected-fmri-data-a-comparison-of-intervolume-and-intravolume-motion-u]
• 13. of Preprocessing data - Ceitec.cz fMRI [https://mafil.ceitec.cz/en/preprocessing-of-fmri-data/]
• 14. Methods to detect, characterize, and remove motion in resting... artifact [https://pubmed.ncbi.nlm.nih.gov/23994314/]
• 15. Quality assurance of human functional magnetic resonance imaging: a literature review - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6629553/]
• 16. Detecting and Adjusting for Artifacts in fMRI Time Series Data - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC1479857/]
• 17. Are Movement Artifacts in Magnetic Resonance Imaging a Real Problem?—A Narrative Review - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC5447676/]
• 18. A Motion Artifact Correction Procedure for fNIRS Signals Based on Wavelet Transform and Infrared Thermography Video Tracking - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC8346954/]
• 19. Mastering Motion Correction in Neuroimaging [https://www.numberanalytics.com/blog/motion-correction-neuroimaging-brain-mapping]
• 20. ICA-AROMA: A robust ICA-based strategy for removing motion ... [https://pubmed.ncbi.nlm.nih.gov/25770991/]
• 21. Frontiers | Applications and advances of combined fMRI-fNIRs... [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2025.1542075/full]
• 22. Frontiers | Are Movement Artifacts in Magnetic Resonance Imaging ... [https://www.frontiersin.org/articles/10.3389/fneur.2017.00232/full]
• 23. Artifacts in magnetic resonance imaging: how it can really affect diagnostic image quality and confuse clinical diagnosis? - Extrica [https://www.extrica.com/article/19756]
• 24. Frontiers | Evaluation and comparison of most prevalent artifact reduction methods for EEG acquired simultaneously with fMRI [https://www.frontiersin.org/journals/neuroimaging/articles/10.3389/fnimg.2022.968363/full]
• 25. Retrospective motion artifact correction of structural MRI images using deep learning improves the quality of cortical surface reconstructions - ScienceDirect [https://www.sciencedirect.com/science/article/pii/S1053811921000331]
• 26. Frontiers | The Voxel-Wise Functional Connectome Can Be Efficiently... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2016.00381/full]
• 27. Improved Motion in fNIRS Data by Combining... Artifact Correction [https://pubmed.ncbi.nlm.nih.gov/37112320/]
• 28. & Management of Pain, Anxiety, Agitation/ Prevention ... Sedation [https://www.guidelinecentral.com/insights/prevention-management-pain-delirium-icu-guidelines-timeline/]
• 29. A Focused Update to the Clinical Practice Guidelines for the... [https://pubmed.ncbi.nlm.nih.gov/39982143/]
• 30. NEUROIMAGING AND BEHAVIORAL TASKS - Laureate Institute for Brain Research [https://www.laureateinstitute.org/neuroimaging-and-behavioral-tasks.html]
• 31. Motion Artifact - an overview | ScienceDirect Topics [https://www.sciencedirect.com/topics/engineering/motion-artifact]
• 32. Self-reports map the landscape of task states derived from brain imaging - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11754446/]
• 33. Automatic MR image quality evaluation using a Deep CNN... [https://pubmed.ncbi.nlm.nih.gov/33770561/]
• 34. Reduction: Triggering, Motion , Gating | SpringerLink Navigator Echoes [https://link.springer.com/chapter/10.1007/978-3-030-85413-3_104]
• 35. Reducing motion artifacts - Questions and Answers in MRI [https://www.mriquestions.com/reducing-motion-artifacts.html]
• 36. IJBEM, Vol. 2, No. 2, pp. 248-254, 2000. [https://ijbem.org/volume2/number2/248-254.htm]
• 37. Magnetic resonance imaging of freely moving objects: prospective real-time motion correction using an external optical motion tracking system - PubMed [https://pubmed.ncbi.nlm.nih.gov/16600642/?dopt=Abstract]
• 38. - ICA : A robust ICA-based strategy for removing AROMA ... motion [https://research.rug.nl/en/publications/ica-aroma-a-robust-ica-based-strategy-for-removing-motion-artifac]
• 39. the effects of Reducing in fMRI: A motion ... artifacts structured matrix [https://pmc.ncbi.nlm.nih.gov/articles/PMC8934405/]
• 40. Wavelet-Based Motion Artifact Removal for Electrodermal Activity - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC5413204/]
• 41. Reduction of Motion Artifacts in Functional Connectivity Resulting from Infrequent Large Motion - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC9618388/]
• 42. Estimation and Removal of Residual Motion in Retrospectively... Artifact [https://apertureneuro.org/article/123369]
• 43. Evaluation of motion reduction depending on the artefact ... artefacts [https://pmc.ncbi.nlm.nih.gov/articles/PMC10220077/]
• 44. Systematic Review and Meta-analysis of AI-driven MRI Motion Artifact Detection and Correction [https://arxiv.org/html/2509.05071v1]
• 45. Mitigating head motion connectivity artifact in functional MRI [https://www.nature.com/articles/s41596-018-0065-y?error=cookies_not_supported&code=8f78a509-2a86-49a5-a350-7504e0f899aa]
• 46. Assessing the use of Diffusion models for motion artifact correction in brain MRI [https://arxiv.org/html/2502.01418v1]
• 47. Think Big, Train Small: Memory-Efficient 3D Diffusion for Medical ... [https://akmaier.medium.com/think-big-train-small-memory-efficient-3d-diffusion-for-medical-image-processing-0c981e1f504e]
• 48. A Deep Learning-Based De-Artifact Diffusion Model for Removing Motion Artifacts in Knee MRI - PubMed [https://pubmed.ncbi.nlm.nih.gov/40583513/]
• 49. removal and evaluation techniques for functional... Motion artifacts [https://pmc.ncbi.nlm.nih.gov/articles/PMC9576156/]
• 50. Functional near-infrared spectroscopy - Wikipedia [https://en.wikipedia.org/wiki/Functional_near-infrared_spectroscopy]
• 51. in Motion near-infrared spectroscopy... artifacts functional [https://pubmed.ncbi.nlm.nih.gov/23639260/]
• 52. (PDF) Motion in functional near-infrared spectroscopy... artifacts [https://www.academia.edu/122101489/Motion_artifacts_in_functional_near_infrared_spectroscopy_A_comparison_of_motion_correction_techniques_applied_to_real_cognitive_data]
• 53. Preprocessing functional Near Infrared Spectroscopy ( fNIRS ) Data [https://nimrobotics.com/blog/2025/fnirs-preprocessing/]
• 54. Techniques — MNE-NIRS 0.7.1 documentation Artifact Correction [https://mne.tools/mne-nirs/stable/auto_examples/general/plot_21_artifacts.html]
• 55. Systematic Review and Meta-analysis of AI-driven MRI Motion Artifact Detection and Correction [https://arxiv.org/html/2509.05071]
• 56. Wavelet-Optimized Motion Artifact Correction in 3D MRI Using Pre-trained 2D Score Priors [https://arxiv.org/html/2511.02256]
• 57. Evaluating Deep Learning Based Domain Generalization for Motion Mitigation in Multi-center Brain MRI | SpringerLink [https://link.springer.com/chapter/10.1007/978-3-032-06103-4_6]
• 58. A multidimensional artefact -reduction approach to increase robustness... [https://pubmed.ncbi.nlm.nih.gov/30779930/]
• 59. in studies of functional connectivity: Characteristics and... Motion artifact [https://pubmed.ncbi.nlm.nih.gov/29091315/]
• 60. Deep learning-based motion in functional... artifact removal [https://pubmed.ncbi.nlm.nih.gov/35475257/]
• 61. Post hoc Correction of Chromatic Aberrations in Large-Scale Volumetric Images in Confocal Microscopy - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC8703134/]
• 62. in (f) NIRS — Not a problem, but not to be... Motion artefacts [https://www.artinis.com/blogpost-all/motion-artefacts-in-fnirs-not-a-problem-but-needs-attention]
• 63. Comparison of prospective and retrospective motion correction in... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8635810/]
• 64. PSNR and SSIM: application areas and criticism [https://videoprocessing.ai/metrics/ways-of-cheating-on-popular-objective-metrics.html]
• 65. Difference between PSNR and SSIM comparison - MATLAB Answers - MATLAB Central [https://www.mathworks.com/matlabcentral/answers/280722-difference-between-psnr-and-ssim-comparison]
• 66. Structural similarity index measure - Wikipedia [https://en.wikipedia.org/wiki/Structural_similarity]
• 67. of Evaluation in brain magnetic resonance images... motion artifacts [https://pmc.ncbi.nlm.nih.gov/articles/PMC8782596/]
• 68. of Evaluation reduction depending on the motion ... artefact artefacts [https://www.nature.com/articles/s41598-023-35794-1?error=cookies_not_supported]
• 69. Deep learning-based motion artifact removal in functional near-infrared spectroscopy - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC9034734/]
• 70. Motion Artifact Removal from EEG Signals Using the Motion-Net Deep Learning Algorithm - ScienceDirect [https://www.sciencedirect.com/science/article/abs/pii/S1746809425013886]
• 71. Discussion of the Influence of Multiscale PCA Methods with... Denoising [https://pmc.ncbi.nlm.nih.gov/articles/PMC8879559/]
• 72. DeepPrep: an accelerated, scalable and robust pipeline for neuroimaging preprocessing empowered by deep learning | Nature Methods [https://www.nature.com/articles/s41592-025-02599-1]
• 73. Correction of Motion Using a Multiscale Fully Artifacts ... Convolutional [https://pubmed.ncbi.nlm.nih.gov/32054615/]
• 74. Evaluation and comparison of most prevalent artifact ... reduction [https://pmc.ncbi.nlm.nih.gov/articles/PMC10406266/]
• 75. Retrospective motion of structural MRI images... artifact correction [https://pmc.ncbi.nlm.nih.gov/articles/PMC8044025/]
• 76. Generative models of MRI-derived neuroimaging features and associated dataset of 18,000 samples | Scientific Data [https://www.nature.com/articles/s41597-024-04157-4]
• 77. Frontiers | Resting state fMRI reveals a default dissociation... mode [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2013.00802/full]
• 78. Does motion -related brain functional reflect both... connectivity [https://pubmed.ncbi.nlm.nih.gov/24999036/]
• 79. Modular preprocessing pipelines can reintroduce artifacts into fMRI ... [https://pubmed.ncbi.nlm.nih.gov/30666750/]
• 80. Frontiers | Artifact Reduction in Simultaneous EEG-fMRI: A Systematic Review of Methods and Contemporary Usage [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2021.622719/full]
• 81. Multi-contrast high-field quality image synthesis for portable low-field... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10793526/]
• 82. Neurodesk FAQ | Neurodesk [https://neurodesk.org/overview/faq/]