Skip to main content
NCBI home page
Human Brain Mapping logoLink to Human Brain Mapping
. 2024 Jan 30;45(2):e26570. doi: 10.1002/hbm.26570

Diffusion MRI head motion correction methods are highly accurate but impacted by denoising and sampling scheme

Matthew Cieslak 1,2,✉, Philip A Cook 3,4, Golia Shafiei 1,2, Tinashe M Tapera 1,2, Hamsanandini Radhakrishnan 1,2, Mark Elliott 3, David R Roalf 2, Desmond J Oathes 2, Dani S Bassett 2,4,5,6,7,8, M Dylan Tisdall 2, Ariel Rokem 9, Scott T Grafton 10, Theodore D Satterthwaite 1,2,11,✉
PMCID: PMC10826632  PMID: 38339908

Abstract

Head motion correction is particularly challenging in diffusion‐weighted MRI (dMRI) scans due to the dramatic changes in image contrast at different gradient strengths and directions. Head motion correction is typically performed using a Gaussian Process model implemented in FSL's Eddy. Recently, the 3dSHORE‐based SHORELine method was introduced that does not require shell‐based acquisitions, but it has not been previously benchmarked. Here we perform a comprehensive evaluation of both methods on realistic simulations of a software fiber phantom that provides known ground‐truth head motion. We demonstrate that both methods perform remarkably well, but that performance can be impacted by sampling scheme and the extent of head motion and the denoising strategy applied before head motion correction. Furthermore, we find Eddy benefits from denoising the data first with MP‐PCA. In sum, we provide the most extensive known benchmarking of dMRI head motion correction, together with extensive simulation data and a reproducible workflow.

Practitioner points

  • Both Eddy and SHORELine head motion correction methods performed quite well on a large variety of simulated data.

  • Denoising with MP‐PCA can improve head motion correction performance when Eddy is used.

  • SHORELine effectively corrects motion in non‐shelled diffusion spectrum imaging data.

Keywords: artifact, diffusion MRI, diffusion spectrum imaging, ensemble average propagator, head motion, software

Head motion correction is a critical step in processing of diffusion MRI data. Here we compare the performance of dMRI head motion correction methods to ground truth data. We find that although performance is highly accurate across methods, it varies with respect to data acquisition type and preprocessing steps.

graphic file with name HBM-45-e26570-g001.jpg

1. INTRODUCTION

Diffusion‐weighted MRI (dMRI) modulates the MR signal to encode information about the distribution of water diffusion, which is constrained by the orientation and permeability of tissue (Basser, 1995; Callaghan, 1993; Stejskal & Tanner, 1965). This method has become widely used to non‐invasively image the structural properties of white matter in the brain. Over the course of three decades, dMRI sequences have advanced to capture signal in many directions (e.g., higher angular resolution) and diffusion sensitizations (Tuch, 2004; Wedeen et al., 2005) with most modern sequences capturing hundreds of images over the course of 5–30 min of scanning.

Any scanning sequence where multiple images are acquired over time is highly susceptible to artifacts related to head motion during the scan. The effects of head motion during functional MRI (fMRI), another imaging technique that acquires images in a series over time, are well‐known and typically addressed by simply aligning each image to a reference image using a rigid or affine transformation (Jenkinson et al., 2002), followed by further corrections to the time series data in each voxel (Ciric et al., 2018). The use of a single reference image works well for spatially correcting fMRI because the contrast and SNR remain relatively constant over the acquisition.

dMRI sequences acquire images that can have dramatically different spatial contrasts and SNR depending on the diffusion‐encoding gradient moment (i.e., b‐value in s/mm2) and direction. The set of directions and b‐values that define a dMRI sampling scheme are what allow the method to estimate the ensemble average diffusion propagator (EAP) in each voxel (Callaghan, 1993). However, such differences in contrast also preclude the use of a single image as the registration target for head motion correction. Instead, for each b > 0 image in the dMRI series, an image with similar spatial contrast must be generated as if it were aligned with all other images in the dMRI series. Each image can then be registered to the target image, correcting the effect of bulk head motion in each volume.

At present the most widely‐used method for dMRI head motion correction is Eddy (Andersson & Sotiropoulos, 2015), which is included in the fMRIB software library (FSL). Eddy has been widely adopted, including by large imaging consortia such as the Human Connectome Project (Glasser et al., 2013) and the UK BioBank (Alfaro‐Almagro et al., 2018). In addition to estimating bulk head motion, Eddy estimates and corrects spatial warping related to eddy currents, fills in dropped slices (Andersson et al., 2016), estimates intra‐volume motion and optionally incorporates susceptibility distortion correction if a fieldmap is estimated using the TOPUP tool (Andersson & Sotiropoulos, 2016). Many of these features rely on Eddy's algorithm for generating registration targets. Eddy operates on shelled dMRI sequences, which acquire multiple gradient directions at the same b > 0 value. Given that all images on the same shell are sampling the surface of a sphere in q‐space, the differences in their signal can be represented as a Gaussian process (GP) on the azimuth and elevation coordinates on the unit sphere S2. Eddy estimates one GP per shell and uses the GP to produce registration targets. A rigid registration is performed between each b > 0 image and its GP prediction, followed by a linear or quadratic warp in the Phase Encoding Direction of the dMRI acquisition to correct for Eddy current‐related distortion.

Although shelled acquisitions are popular, there are other methods of sampling q‐space that have unique advantages. Diffusion spectrum imaging (DSI) samples a Cartesian grid in q‐space, enabling the direct reconstruction of the EAP with a simple Fourier transform (Wedeen et al., 2005). Shelled schemes require more complex modeling and leave open the issue of which b‐values to acquire shells at: a question with no universal answer. DSI scans typically have required the acquisition of more than 200 images, resulting in long scan times—particularly when multiband imaging is not available. However, sparse, random subsets of the Cartesian grid scheme along with a compressed‐sensing reconstruction approach (CS‐DSI) have been shown to provide comparable EAP reconstructions to the full grid sampling scheme at a fraction of the scanning time (Merlet & Deriche, 2013; Paquette et al., 2015). However, until recently, there has been no widely accepted head motion correction for these non‐shelled schemes, limiting the application of DSI in translational research.

The SHORELine algorithm (Cieslak et al., 2021) was introduced to fill this gap and provide a method to generate registration targets for any dMRI sampling scheme with both radial and angular variability in its q‐space sampling scheme. SHORELine is a cross‐validated method where, for each b > 0 image, the 3dSHORE basis set (Özarslan et al., 2013) is fit to all other images using L2‐regularization. A registration target for the left‐out image is estimated based on the 3dSHORE fit and the image is registered with a Rigid (6DOF) or optional Affine (12DOF) transform using ANTs (Figure 1). This process is repeated up to two times based on user specifications. The 3dSHORE basis functions are defined in three‐dimensional space (R3) and are therefore appropriate for multi‐shelled, Cartesian as well as sparse/random sampling schemes. The original SHORELine evaluation showed an overall improvement in the Neighboring DWI Correlation (NDC) quality measure of both shelled and non‐shelled sampling schemes compared to the unprocessed data (Cieslak et al., 2021).

FIGURE 1.

FIGURE 1

Open in a new tab

SHORELine. Images shown here are from an actual scan using the CS‐DSI sampling scheme. The slice in the yellow‐dashed box is the SHORELine‐predicted slice for the left‐out volume. The actual left‐out slice is shown on the left in Step 3. The contrast in the predicted slice is visually very similar to the actual left‐out slice, which enables standard image registration methods to work.

The SHORELine and Eddy methods have some notable differences. No Eddy current correction is explicitly attempted in SHORELine. Like Eddy, SHORELine can incorporate a susceptibility distortion correction along with head motion correction in a single interpolation. Unlike Eddy, shells are not treated separately in SHORELine, which may be beneficial in the presence of sparsely‐sampled shells like ABCD's b = 500 shell. Importantly, SHOREline has a far more permissive license (BSD 3‐Clause) than Eddy, which makes it attractive to clinical and industry users.

Independent of head motion correction, denoising algorithms are often applied prior to head motion correction. One popular denoising algorithm included in the MRtrix package is dwidenoise (Veraart et al., 2016), which implements a non‐aggressive denoising method commonly called Marchenko‐Pasteur PCA (MP‐PCA). MP‐PCA operates on small patches of dMRI signal, estimating a noise level used as an optimal cut‐off for PCA denoising. The application of MP‐PCA denoising is enabled by default in both QSIPrep (Cieslak et al., 2021) and the MRtrix3_connectome (Smith & Connelly, 2020) pipelines. However, denoising is not applied in the HCP (Glasser et al., 2013) or UK Biobank (Alfaro‐Almagro et al., 2018) pipelines. There are reasonable arguments both for and against performing denoising prior to motion correction. The registration targets generated by Eddy and SHOREline may be more realistic if their underlying models are fit with less‐noisy data. However, denoising is not currently recommended by the developers of Eddy because it will affect how noise is distributed in the dMRI signal, possibly violating assumptions of the GP (Andersson, 2019).

Both Eddy and SHORELine are considerably more complex than standard fMRI head motion correction methods. Despite this, it may surprise many investigators that there has not been a systematic evaluation of how well these methods perform on commonly acquired sequences with different levels of head motion, nor has there been a systematic evaluation on the effect of denoising dMRI data prior to head motion correction. Previous evaluations of Eddy on phantom data were performed on a simulated 2‐shell (b = 700,2000, 104 total b > 0) scheme with 2.5 mm voxels (Andersson et al., 2017; Graham et al., 2016) and neither were designed to benchmark motion estimation accuracy. Instead, they demonstrated Eddy's higher accuracy than eddy_correct and demonstrated the importance of slice‐to‐volume correction respectively. Other studies have shown that head motion correction is beneficial in general for calculating anisotropy measures or performing tractography (Brun et al., 2019; Kreilkamp et al., 2016; Sakaie & Lowe, 2010; Yamada et al., 2014). Denoising has likewise been shown to increase the detectability (Moeller et al., 2021) and reliability (Schilling et al., 2021) of dMRI derivatives, but no interaction with head motion correction has been investigated. To address this gap, here we simulated hundreds of thousands of dMRI images from common shelled and non‐shelled schemes incorporating known head motion and realistic MR artifacts to determine how accurate these methods are at estimating true head motion with and without denoising. Critically, all data, operations, and analyses of this are tracked using DataLad in a reproducible workflow. Together, this effort represents the most extensive benchmarking of dMRI head motion correction methods to date, is the first evaluation of SHORELine, and provides extensive simulation data together with a reproducible workflow.

2. MATERIALS AND METHODS

2.1. Simulation of images and motion

MITK FiberFox (Neher et al., 2014) was used to simulate the entire dMRI series for a set of commonly acquired sampling schemes, including ABCD, HCP, a Cartesian DSI half sphere and a Cartesian/random CS‐DSI. The simulation used streamline segments from the ISMRM 2015 fiber phantom (http://www.tractometer.org/ismrm_2015_challenge/data) to define a fiber ODF in each voxel and converted into a diffusion ODF to generate an MRI signal. MRI signal was combined with N/2 aliasing, eddy currents, thermal noise and susceptibility distortion artifacts. Simulation parameters used here are identical to the ISMRM simulated phantom, except that the original study contained only three volumes with simulated head motion, whereas we simulated hundreds of thousands of volumes under different sampling schemes and conditions (see below). A docker image of the exact version of Fiberfox can be downloaded at https://hub.docker.com/r/pennbbl/fiberfox using tag 1.0.

To introduce controlled head motion, a random rigid transform was applied to the streamline data before the MR signal was simulated (Hering et al., 2014). Translation and rotation about each axis were sampled from a random uniform distribution with a maximum absolute displacement of 5 mm and a maximum absolute rotation of 5° (Figure 2). The range of translation and rotation values was selected based on the empirical motion parameters observed in large‐scale public datasets. Specifically, ABCD and HBN were selected because their public releases include scans that include both low and high motion scans (Figure S1). Two simulations were created for each sampling scheme: one with no head motion and one including random motion in every volume. To generate the many individual dMRI series we replaced a random subset of images from the no‐motion simulation with the motion‐included simulated images. Unique dMRI scans were generated to have a specific motion prevalence such that 15%, 30% or 50% of the volumes in the series included head motion. A total of 30 unique scans were generated for each sampling scheme for each of the three motion prevalence values, yielding 90 simulated complete series per scheme. In total this resulted in 360 unique simulated dMRI scans. Other than the empirical estimates of head motion in Figure S1, all subsequent analyses are based on these simulated data.

FIGURE 2.

FIGURE 2

Open in a new tab

Simulation of head motion in DWI images with Fiberfox. Fiberfox simulates MRI data based on a set of streamlines. Here we show the streamlines used for benchmarking (top row). We introduced motion to our test data by applying rigid (6‐DOF) transformations to the streamlines before the volumes were simulated. The diffusion restriction introduced by the fibers represented by the streamlines is included in the signal attenuation in each voxel, which is then “acquired” in k‐space where artifacts can be introduced (middle row). Finally, the k‐space data is reconstructed into realistic 3d volumes that are used for benchmarking.

2.2. Image processing

Eddy and SHORELine were run on each simulated scan using QSIPrep (v0.14.3). This version of QSIPrep included FSL version 6.0.3. Eddy was run twice, once with Linear and once with Quadratic models. However, a comparison of these approaches revealed that these options do not appear to have a substantial impact on head motion estimation. Accordingly, we report the results from the Eddy with the quadratic setting because this is the way it is typically used as part of the HCP Pipelines. SHORELine was also run twice, once with a Rigid (6DOF) and once with an Affine (12DOF) transformation model. The rigid model had minor benefits and also a shorter run time; as such, the results for the rigid model are used to characterize SHORELine's performance. Each configuration was run with and without MP‐PCA denoising (from MRTrix 3.0.3) prior to head motion correction. In all, a total of 2160 QSIPrep preprocessing runs were executed, encompassing the processing of 375,840 b > 0 images (Table 1).

TABLE 1.

Properties of simulated datasets and how they were processed.

Algorithm Model Denoising Type ABCD HCP DSIQ5 HASC55
Shelled Shelled Cartesian Random
Max. b 3000 3000 5000 5000
Num. b > 0 96 270 257 55
Motion prevalence # Pipeline runs
Eddy Linear MP‐PCA 15% 30 30 0 0
30% 30 30 0 0
50% 30 30 0 0
None 15% 30 30 0 0
30% 30 30 0 0
50% 30 30 0 0
Quadratic MP‐PCA 15% 30 30 0 0
30% 30 30 0 0
50% 30 30 0 0
None 15% 30 30 0 0
30% 30 30 0 0
50% 30 30 0 0
SHORELine Rigid MP‐PCA 15% 30 30 30 30
30% 30 30 30 30
50% 30 30 30 30
None 15% 30 30 30 30
30% 30 30 30 30
50% 30 30 30 30
Affine MP‐PCA 15% 30 30 30 30
30% 30 30 30 30
50% 30 30 30 30
None 15% 30 30 30 30
30% 30 30 30 30
50% 30 30 30 30
Total scans 360 360 180 180
Total b > 0 volumes 69,120 194,400 92,520 19,800
Open in a new tab

Note: Each sequence was simulated such that 15%, 30% and 50% of volumes contained head motion. Each level of motion was processed both with and without denoising.

Importantly, to ensure complete reproducibility, the entire benchmarking experiment was run using the FAIRly big workflow (Wagner et al., 2022). This approach uses DataLad to track and distribute the data and code used during data analysis. The entire FiberFox phantom dataset along with a singularity image of the software used to run the processing are publicly available. Each run of QSIPrep was recorded as a git commit and can be reproduced locally by anyone with DataLad who clones from the repository (https://github.com/PennLINC/dMRI_HMC_Benchmark).

2.3. Outcome measures

Performance was evaluated according to multiple metrics. First, the mean error in head motion parameter estimation was calculated to see whether the algorithm is an unbiased estimator. Second, to characterize the expected error of the estimators, we calculated the RMSE of the estimated head motion. Low RMSE reflects accurate head motion estimation and higher RMSE indicates greater estimation error. To understand the factors that affect motion parameter estimation, linear models were fit with root mean squared error (RMSE) as the dependent variable. As rotation and translation are in different units, we fit two linear models of RMSE for translation and rotation separately. Relative performance between SHORELine and Eddy was calculated by subtracting SHORELine's RMSE from Eddy's, resulting in positive values when SHORELine's was more accurate at estimating motion parameters than Eddy.

Third, we compared the interpolation‐related smoothness of the corrected images. Image smoothness is a measure of blurring during preprocessing, which reduces anatomical detail; preprocessing should seek to minimize the introduction of additional image smoothness. We estimated the full‐width at half‐maximum (FWHM) of the mean b = 0 image in the preprocessed data using AFNI's 3dFWMx program. Fourth, we evaluated a summary measure of data quality – the neighboring DWI Correlation (NDC) (Yeh et al., 2019). NDC summarizes the pairwise spatial correlation between each pair of dMRI volumes that sample the closest points in q‐space; lower values reflect reduced data quality, driven by noise and misalignment between dMRI volumes.

3. RESULTS

3.1. Both Eddy and SHORELine accurately correct simulated head motion

Both Eddy and SHORELine demonstrated excellent performance in correcting head motion. While collapsing across all experimental conditions, the mean error (calculated as the mean difference between the estimated motion parameter and the ground truth motion parameter) in estimated head rotation was very small: only 0.194°. Similarly, the mean error in estimated head translation was only ~1/100th of a voxel: 0.012 mm. Such miniscule mean errors suggest that both Eddy and SHORELine are accurate and unbiased estimators of head motion parameters. Mean errors and RMSE (the first and second moments of the error distribution) are provided in Table 2.

TABLE 2.

Mean error of head motion correction methods.

Sampling scheme Method Translation (mm) Rotation (°)
Mean error RMSE Mean error RMSE
ABCD Eddy + MP‐PCA 0.016 0.71 0.096 0.70
Eddy 0.071 1.1 0.27 2.3
SHORELine −0.086 0.80 −0.037 1.6
HCP Eddy + MP‐PCA −0.044 0.73 0.0093 0.56
Eddy 0.063 0.71 −0.063 1.0
SHORELine −0.051 0.56 0.48 0.86
DSIQ5 SHORELine 0.10 0.61 0.24 0.96
HASC55 SHORELine 0.059 0.99 0.33 1.5
Open in a new tab

3.2. Head motion correction accuracy varies by sampling scheme and denoising

Next, we evaluated what factors impacted error (hereafter referring to RMSE) following head motion correction (Figure 3). As described below, results indicate that error following motion correction is primarily due to uncontrolled factors such as the amount of head motion present in the data but also preprocessing choices (use of denoising) and factors related to experimental design (sampling scheme); see Table S1 for complete statistical results.

FIGURE 3.

FIGURE 3

Open in a new tab

Head motion correction error varies by motion present in input data, sampling scheme, and de‐noising. Means and standard deviations of the motion parameter estimate RMSE are plotted for both Eddy and SHORELine. Eddy does not support non‐shelled schemes and therefore could not be evaluated for DSIQ5 and HASC55. Error bars reflect the standard deviation of the RMSE for the sample, each of which consist of 30 simulated scans. Error varied by sampling scheme, with lower errors present in simulated HCP data than ABCD data. Notably, greater motion in the input data was associated with greater error. Additionally, greater error was observed when Eddy was used without MP‐PCA denoising; error was lower in Eddy than SHORELine when denoising was used, but higher when no denoising was performed.

In most cases, higher motion in the input data was associated with greater error following head motion correction. However, this effect was sometimes impacted by an interaction with the denoising and motion correction methods chosen. While denoising with MP‐PCA had almost no impact on the error present in SHORELine output, it had a major impact on Eddy: error was systematically lower across both shelled schemes when the data was denoised first. Somewhat surprisingly, and in contrast to nearly all other parameter combinations evaluated, the amount of rotation in the input data was not associated with greater error when Eddy was used in conjunction with MP‐PCA (this was not true for translations).

Error also varied substantially across acquisition schemes. For example, among shelled schemes, error was systematically lower in the HCP than ABCD sampling scheme. In general, acquisition schemes that acquired a greater number of directions tended to have less error following head motion correction (Figure 4). The interaction in Table S1 is driven by the exception to this trend: CS‐DSI has the fewest number of directions but retained a low RMSE. As SHORELine is the only existing algorithm that can process non‐shelled schemes, data from these acquisition schemes could not be evaluated using Eddy.

FIGURE 4.

FIGURE 4

Open in a new tab

Head motion correction error varies by number of directions in sample scheme. RMSE as a function of the number of sampled directions. Points near x = 103 are from the ABCD sequence, near 270 are HCP, x = 258 is DSIQ5 and x = 55 is HASC55. SHORELine's performance was equivalent with and without denoising, so results without denoising are shown. In general, better performance was seen with more directions and when MP‐PCA was used in conjunction with Eddy.

Next, we directly compared Eddy and SHORELine (Figure 5) using simulated data from the shelled ABCD and HCP sampling schemes where both methods were applicable. Overall, differences between the methods were quite small and depended in part on use of denoising, sampling scheme, and whether rotations or translations were evaluated (see full statistical results in Table S2). For ABCD, SHORELine had less error than Eddy in all scenarios when no denoising was applied first. However, when the data was first denoised with MP‐PCA, Eddy showed a slight superiority that scaled with the prevalence of motion in the input data.

FIGURE 5.

FIGURE 5

Open in a new tab

De‐noising impacts relative performance of motion correction methods. Violin plots of the difference between Eddy's and SHORELine's RMSE for shelled schemes. Positive values indicate that SHORELine had lower RMSE than Eddy, while negative values mean Eddy had lower RMSE than SHORELine. In general, Eddy performed slightly better than SHORELine when MP‐PCA denoising was performed first, whereas SHORELine was slightly superior when Eddy was used without denoising. However, overall absolute differences were quite small.

In contrast, results from the data simulated with the HCP scheme were more heterogeneous, and related to both the measure evaluated (rotation vs. translation) and the amount of motion present in the input data. For rotations, Eddy following MP‐PCA outperformed SHORELine as more motion is present in the simulation. However, without MP‐PCA denoising, SHORELine modestly outperformed Eddy overall for HCP data. For translations, SHORELine error was lower across all conditions, but differences were quite small.

3.3. Output image smoothness is impacted primarily by motion present in data

Next, we evaluated the image smoothness (quantified as FWHM) of the preprocessed data following head motion correction (Figure 6). Image smoothness is a measure of blurring during preprocessing, which reduces anatomical detail. Ideally, preprocessing minimizes the introduction of additional image smoothness. As expected, we found that the largest driver of output image smoothness was the motion prevalence in the simulated data: across all sampling schemes and motion correction methods, more motion in the input data was associated with greater smoothness in the output images. Somewhat surprisingly, denoising did not significantly impact output image smoothness (see Table S3). Additionally, although differences were small (i.e., <0.5 mm FWHM), SHORELine produced significantly sharper output than Eddy.

FIGURE 6.

FIGURE 6

Open in a new tab

Spatial smoothness following motion correction is impacted by the amount of motion present in the input data. Both methods produce blurrier images as the amount of motion present increases. Denoising with MP‐PCA did not significantly impact smoothness, so variation in de‐noising is not shown. Eddy cannot process non‐shelled schemes, so only data from SHORELine is shown. In general, Eddy produced blurrier images than SHORELine, although differences were small.

3.4. Output image quality is improved by head motion correction and denoising

As a final step, we quantified the quality of the output images using the neighboring DWI correlation (NDC; Figure 7). We calculated NDC for both the unprocessed input data and the output from both SHORELine and Eddy; this allowed us to examine how much pre‐processing improved data quality compared to the raw input data. Notably, head motion correction yielded substantial improvements in NDC across levels of input data motion and sampling schemes. Indeed, the improvement in NDC with preprocessing in general scaled with the prevalence of motion in the input data (see bottom row, Figure 7). However, in nearly all cases, greater prevalence of motion were associated with reduced output data quality even following preprocessing. One important exception to this was the use of Eddy with HCP data, where higher prevalence of motion was not associated with reduced NDC.

FIGURE 7.

FIGURE 7

Open in a new tab

Motion correction and denoising improve data quality. Data quality was quantified as the neighboring DWI correlation (NDC). The top row displays the mean and standard deviation of the NDC values after preprocessing. The bottom row shows the change in NDC from the NDC calculated on the unprocessed scans. In general, NDC was improved by motion correction, especially following MP‐PCA denoising. Eddy yielded improvements in NDC that were significantly higher than SHORELine, although differences were small.

Additionally, we found that denoising with MP‐PCA improved NDC values in all scenarios that we evaluated (Table S4). Denoising related improvements were particularly marked for non‐shelled schemes, which could only be processed with SHORELine. Furthermore, we found that Eddy had slightly higher NDC scores than SHORELine; however, the difference was reduced when MP‐PCA was not performed. The small relative advantage for Eddy over SHORELine scaled with greater motion prevalence in the input data.

4. DISCUSSION

In this study, we conducted the largest known benchmarking of dMRI head motion methods to date, and the first quantitative evaluation of SHORELine. While prior studies have considered hundreds of b > 0 images (or even selected voxels), here we evaluated over 350,000 simulated b > 0 images across multiple sampling schemes and denoising conditions. We found that both Eddy and SHORELine work quite well. As expected, both data quantity and quality were the primary determinants of performance. However, benchmarking yielded the unexpected result that Eddy performance is improved when the data is first denoised with MP‐PCA. Moving forward, the publicly‐shared simulated data and our use of a completely reproducible workflow with DataLad provide a data resource that will accelerate future advances in dMRI motion correction.

Our primary finding is that across all methods, denoising, motion prevalence, and sampling schemes, both Eddy and SHORELine estimated head motion correctly, with observed errors being unbiased and quite small. This analysis is a prerequisite to show that head motion correction is accurate on real dMRI data: there would be little hope that these methods would work on real world data if these analyses showed biased estimation of motion, or errors of more than a fraction of a voxel translation or degree rotation were present. This was clearly not the case, which should reassure users of both methods, whether they are already in wide use (Eddy) or recently introduced (SHORELine).

While observed errors were on average quite small, our evaluation did identify several factors that influenced error magnitude. These included input data quality (specifically motion prevalence), denoising, and acquisition scheme. Across all scenarios examined, the single biggest determinant of the amount of error observed was the prevalence of motion in the simulated input data. Similarly, more prevalent motion in the input data resulted in greater image blurring (e.g., FWHM) and reduced data quality (as quantified by the NDC). These results emphasize that acquiring high quality data remains of critical importance for all studies, as no amount of image processing can fully compensate for extensive in‐scanner motion.

Nonetheless, one preprocessing step that is not always used in standard pipelines—denoising—had remarkably salutary effects on the outcomes we evaluated. Across all scenarios, data denoised with MP‐PCA improved in quality without impacting smoothness, meaning that a simple increase in smoothness did not drive the increased performance (Woods, Grafton, Holmes, et al., 1998; Woods, Grafton, Watson, et al., 1998). However, use of denoising did interact with choice of head motion correction method in an unexpected manner: while denoising did not impact the error estimated by SHORELine, estimation error was markedly reduced when Eddy consumed data that had been first denoised by MP‐PCA. This result was unanticipated: Eddy developers do not recommend denoising prior to head motion correction because of theoretical concerns regarding the way denoising might change the noise distribution in the data. When paired with MP‐PCA, Eddy performed uniquely well in some contexts, with rotation error failing to scale with severity of motion. While some field‐standard pipelines—such as QSIPrep—do by default apply MP‐PCA denoising prior to head motion correction, other widely used pipelines (such as HCP pipelines) do not. These empirical results suggest that this recommendation may require re‐evaluation, in particular for the many large‐scale data resources that rely on processing pipelines that apply Eddy without denoising. These results may not generalize to other denoising methods, particularly those that denoise by spatial smoothing. Future work should consider other methods such as Patch2Self (Fadnavis et al., 2020) and complex denoising (Moeller et al., 2021).

Our study considered four different acquisition schemes—including two commonly used shelled schemes and two non‐shelled schemes—that allowed us to examine how outcomes were impacted by this important experimental design choice. We found that sequences with more directions tended to have lower error, likely because increased data volume facilitated model training and fitting. However, perhaps the most important result of comparing these four schemes was the finding that non‐shelled schemes could be successfully corrected as well as shelled schemes. Notably, processing of non‐shelled schemes was only possible with SHORELine. As DSI methods have important advantages for modeling the average ensemble propagator, this represents a milestone for the preprocessing of non‐shelled schemes and may accelerate their adoption by the neuroimaging community. Furthermore, the particularly impressive performance on the very‐brief 55‐direction CS‐DSI scheme emphasizes the promise of compressed sensing methods for the many translational applications where scan time is limited and motion may be prominent (i.e., children and clinical populations).

In contrast to the impact of input data quality, denoising, and acquisition scheme, direct comparisons of Eddy and SHORELine were notable mainly for the small effects that were observed. There were small but significant differences in estimation error between the two methods, but the direction of effect largely depended on whether MP‐PCA was also used. As noted above, adding denoising resulted in a notable reduction in observed error for Eddy. Images processed by SHORELine were slightly sharper than Eddy, but the magnitude of difference was of uncertain practical significance (i.e., 0.06 mm FWHM). Conversely, we found that Eddy had a higher NDC than SHORELine. However, this difference was also of unclear practical impact. To put the NDC difference in context, Eddy on average had an NDC that was 0.02 units higher than SHORELine‐processed images; the difference in NDC between unprocessed data and processed images was approximately 20 times larger. Overall, our results emphasize that both methods perform quite well.

Several limitations of this study should be noted. First, we compared Eddy and SHORELine, but other dMRI head motion estimation methods exist. While not evaluated, we expect the MAPMRI‐based method implemented in TORTOISE (Irfanoglu et al., 2017) is likely to perform similarly to SHORELine, as the MAPMRI and 3dSHORE basis sets are very similar and should yield similar image predictions for registration. Both methods also use ITK‐based (McCormick et al., 2014) registration. Second, SHORELine requires sampling schemes with at least two unique non‐zero b‐values, preventing a comparison of performance on common single shell schemes. Third, real world data may differ in both the types of artifacts and the types of movement observed. For example, the FiberFox simulations do not simulate within‐volume motion and the corresponding slice dropout. This complicates the comparison between Eddy and SHORELine, as Eddy may be particularly useful in the presence of high motion/dropped slices in real data. Our simulated data also only included linear Eddy current distortion. Despite these limitations, our use of simulated data allowed us to have a known ground truth by which to benchmark these methods and systematically manipulate multiple distinct parameters in a factorial design. A limitation of the NDC metric is that it is influenced by the spatial smoothness of the images (Cieslak et al., 2021). In the digital phantom data evaluated here the Eddy results are generally slightly smoother (Figure 6) due to a second interpolation, artificially raising their NDC.

Moving forward, we anticipate that preprocessing methods for DWI will continue to advance. In particular, several important features that are included in Eddy—such as eddy current correction—could be included in SHORELine in future releases. Furthermore, the advent of cutting‐edge denoising techniques—such as Patch2Self (Fadnavis et al., 2020)—that leverage self‐supervised learning may have important implications for head motion correction, including providing slice dropout correction when applied after SHORELine. Finally, we have released all simulated images, processing software, analytic code, and results associated with this work; these may prove useful for future benchmarking efforts and facilitate comparisons to existing methods. This open and fully reproducible workflow both bolsters confidence in the current results and are likely to accelerate evaluation studies moving forward.

CONFLICT OF INTEREST STATEMENT

The authors declare that they have no competing or conflicting interests.

Supporting information

TABLE S1. Statistical comparisons of denoising, scheme type, number of directions, denoising method, and motion prevalence on RMSE of estimated motion.

TABLE S2. Statistical comparison of Eddy and SHORELine in shelled schemes. The outcome measure was the difference in error (RMSE) between Eddy and SHORELine (Eddy–SHORELine); the model included main effects of denoising, motion prevalence in the input data, and shell scheme; interactions of these main effects were modeled as well.

TABLE S3. Statistical comparisons of image smoothness between Eddy and SHORELine. The outcome measure was the difference in image smoothness (FWHM) between Eddy and SHORELine (Eddy–SHORELine); the model included main effects of denoising, motion prevalence in the input data, and sampling scheme.

TABLE S4. Statistical comparisons of image quality between Eddy and SHORELine. The outcome measure was the difference in image quality (NDC) between Eddy and SHORELine (Eddy–SHORELine); the model included main effects of denoising, motion prevalence in the input data, and shell scheme.

FIGURE S1. Empirical motion parameters in large‐scale datasets. QSIPrep outputs from the ABCD data processing and data from the public release of Healthy Brain Network (HBN; Richie‐Halford et al., 2022) study were used to obtain automated quality control (QC) metrics for ABCD and HBN datasets, respectively. The distribution of maximum translation (in mm) and rotation (in degrees) values are depicted for both datasets. The empirical motion parameters range widely, such that maximum translation = [0.22 9.09] and maximum rotation = [0.10 6.60] for ABCD (top row), and maximum translation = [0.09 20.78] and maximum rotation = [0.01 10.61] for HBN (bottom row).

Click here for additional data file. (550.8KB, docx)

ACKNOWLEDGMENTS

This study was supported by grants from the National Institutes of Health: R01MH112847, R01MH120482, R37MH125829, R01EB022573, R01MH113550, R01MH123550, RF1MH116920, U01AG052943, RF1MH121867, R01MH119185, R01MH120174. Additional support was provided by the AE Foundation and the Penn/CHOP Lifespan Brain Institute. Scott T. Grafton was supported by the Institute for Collaborative Biotechnologies under Cooperative Agreement W911NF‐19‐2‐0026 from the Army Research Office.

Cieslak, M. , Cook, P. A. , Shafiei, G. , Tapera, T. M. , Radhakrishnan, H. , Elliott, M. , Roalf, D. R. , Oathes, D. J. , Bassett, D. S. , Tisdall, M. D. , Rokem, A. , Grafton, S. T. , & Satterthwaite, T. D. (2024). Diffusion MRI head motion correction methods are highly accurate but impacted by denoising and sampling scheme. Human Brain Mapping, 45(2), e26570. 10.1002/hbm.26570

Contributor Information

Matthew Cieslak, Email: matthew.cieslak@pennmedicine.upenn.edu.

Theodore D. Satterthwaite, Email: sattertt@pennmedicine.upenn.edu.

DATA AVAILABILITY STATEMENT

All data, figures and processing code are available to the public under a BSDv3 license. dMRI phantom data and preprocessing: The fiberfox simulation data is available at https://osf.io/38vce/, but is most conveniently accessed via DataLad from https://github.com/PennLINC/dMRI_HMC_Benchmark. Instructions for cloning this dataset are in the README of this repository, and cloning will also provide the singularity image containing all the software required to re‐run processing along with the input data. Figures and statistical tests: The source code for all statistical tests and figures is publicly available at https://github.com/PennLINC/SHORELINE_PaperDemo, with a visual walkthrough available here https://pennlinc.github.io/SHORELINE_PaperDemo/articles/dmri_hmc_benchmark.html.

REFERENCES

  1. Alfaro‐Almagro, F. , Jenkinson, M. , Bangerter, N. K. , Andersson, J. L. R. , Griffanti, L. , Douaud, G. , Sotiropoulos, S. N. , Jbabdi, S. , Hernandez‐Fernandez, M. , Vallee, E. , Vidaurre, D. , Webster, M. , McCarthy, P. , Rorden, C. , Daducci, A. , Alexander, D. C. , Zhang, H. , Dragonu, I. , Matthews, P. M. , … Smith, S. M. (2018). Image processing and quality control for the first 10,000 brain imaging datasets from UK Biobank. NeuroImage, 166, 400–424. 10.1016/j.neuroimage.2017.10.034 [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Andersson, J. (2019, January 19). FSL Archives [FSL Discussion Board]. JISCMail. https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=fsl;590ecadd.1901
  3. Andersson, J. L. R. , & Sotiropoulos, S. N. (2015). Non‐parametric representation and prediction of single‐ and multi‐shell diffusion‐weighted MRI data using Gaussian processes. NeuroImage, 122, 166–176. 10.1016/j.neuroimage.2015.07.067 [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Andersson, J. L. R. , & Sotiropoulos, S. N. (2016). An integrated approach to correction for off‐resonance effects and subject movement in diffusion MR imaging. NeuroImage, 125, 1063–1078. 10.1016/j.neuroimage.2015.10.019 [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Andersson, J. L. R. , Graham, M. S. , Drobnjak, I. , Zhang, H. , Filippini, N. , & Bastiani, M. (2017). Towards a comprehensive framework for movement and distortion correction of diffusion MR images: Within volume movement. NeuroImage, 152, 450–466. 10.1016/j.neuroimage.2017.02.085 [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Andersson, J. L. R. , Graham, M. S. , Zsoldos, E. , & Sotiropoulos, S. N. (2016). Incorporating outlier detection and replacement into a non‐parametric framework for movement and distortion correction of diffusion MR images. NeuroImage, 141, 556–572. 10.1016/j.neuroimage.2016.06.058 [DOI] [PubMed] [Google Scholar]
  7. Basser, P. J. (1995). Inferring microstructural features and the physiological state of tissues from diffusion‐weighted images. NMR in Biomedicine, 8, 333–344. [DOI] [PubMed] [Google Scholar]
  8. Brun, L. , Pron, A. , Sein, J. , Deruelle, C. , & Coulon, O. (2019). Diffusion MRI: Assessment of the impact of acquisition and preprocessing methods using the brainVISA‐diffuse toolbox. Frontiers in Neuroscience, 13, 536. 10.3389/fnins.2019.00536 [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Callaghan, P. T. (1993). Principles of nuclear magnetic resonance microscopy. Oxford University Press. [Google Scholar]
  10. Cieslak, M. , Cook, P. A. , He, X. , Yeh, F.‐C. , Dhollander, T. , Adebimpe, A. , Aguirre, G. K. , Bassett, D. S. , Betzel, R. F. , Bourque, J. , Cabral, L. M. , Davatzikos, C. , Detre, J. A. , Earl, E. , Elliott, M. A. , Fadnavis, S. , Fair, D. A. , Foran, W. , Fotiadis, P. , … Satterthwaite, T. D. (2021). QSIPrep: An integrative platform for preprocessing and reconstructing diffusion MRI data. Nature Methods, 18(7), 775–778. 10.1038/s41592-021-01185-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Ciric, R., Rosen, A. F. G., Erus, G., Cieslak, M., Adebimpe, A., Cook, P. A., Bassett, D. S., Davatzikos, C., Wolf, D. H., & Satterthwaite, T. D . (2018). Mitigating head motion artifact in functional connectivity MRI. Nature Protocols, 13(12), 12. 10.1038/s41596-018-0065-y [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Fadnavis, S., Batson, J., & Garyfallidis, E . (2020). Patch2Self: Denoising Diffusion MRI with Self‐Supervised Learning. NeurIPS, 1–11. [Google Scholar]
  13. Glasser, M. F. , Sotiropoulos, S. N. , Wilson, J. A. , Coalson, T. S. , Fischl, B. , Andersson, J. L. , Xu, J. , Jbabdi, S. , Webster, M. , Polimeni, J. R. , Van Essen, D. C. , & Jenkinson, M. (2013). The minimal preprocessing pipelines for the human connectome project. NeuroImage, 80, 105–124. 10.1016/j.neuroimage.2013.04.127 [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Graham, M. S. , Drobnjak, I. , & Zhang, H. (2016). Realistic simulation of artefacts in diffusion MRI for validating post‐processing correction techniques. NeuroImage, 125, 1079–1094. 10.1016/j.neuroimage.2015.11.006 [DOI] [PubMed] [Google Scholar]
  15. Hering, J. , Neher, P. , Meinzer, H.‐P. , & Maier‐Hein, K. (2014). Construction of ground‐truth data for head motion correction in diffusion MRI. ISMRM annual meeting 2014 .
  16. Irfanoglu, M. , Nayak, A. , Jenkins, J. , & Pierpaoli, C. (2017). Improvements and new features of the NIH diffusion MRI processing pipeline. Proceedings of the 25th annual meeting of ISMRM. ISMRM, Hawaii, HI.
  17. Jenkinson, M. , Bannister, P. , Brady, M. , & Smith, S. (2002). Improved optimization for the robust and accurate linear registration and motion correction of brain images. NeuroImage, 17(2), 825–841. 10.1006/nimg.2002.1132 [DOI] [PubMed] [Google Scholar]
  18. Kreilkamp, B. A. K. , Zacà, D. , Papinutto, N. , & Jovicich, J. (2016). Retrospective head motion correction approaches for diffusion tensor imaging: Effects of preprocessing choices on biases and reproducibility of scalar diffusion metrics. Journal of Magnetic Resonance Imaging, 43(1), 99–106. 10.1002/jmri.24965 [DOI] [PubMed] [Google Scholar]
  19. McCormick, M., Liu, X., Ibanez, L., Jomier, J., & Marion, C . (2014). ITK: Enabling reproducible research and open science. Frontiers in Neuroinformatics, 8, 13. 10.3389/fninf.2014.00013 [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Merlet, S. L. , & Deriche, R. (2013). Continuous diffusion signal, EAP and ODF estimation via compressive sensing in diffusion MRI. Medical Image Analysis, 17(5), 556–572. 10.1016/j.media.2013.02.010 [DOI] [PubMed] [Google Scholar]
  21. Moeller, S. , Pisharady, P. K. , Ramanna, S. , Lenglet, C. , Wu, X. , Dowdle, L. , Yacoub, E. , Uğurbil, K. , & Akçakaya, M. (2021). Noise reduction with distribution corrected (NORDIC) PCA in dMRI with complex‐valued parameter‐free locally low‐rank processing. NeuroImage, 226, 117539. 10.1016/j.neuroimage.2020.117539 [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Neher, P. F. , Laun, F. B. , Stieltjes, B. , & Maier‐Hein, K. H. (2014). Fiberfox: Facilitating the creation of realistic white matter software phantoms. Magnetic Resonance in Medicine, 72(5), 1460–1470. 10.1002/mrm.25045 [DOI] [PubMed] [Google Scholar]
  23. Özarslan, E., Koay, C. G., & Basser, P. J . (2013). Simple harmonic oscillator based reconstruction and estimation for one‐dimensional Q‐space magnetic resonance (1D‐SHORE). In Applied and Numerical Harmonic Analysis (pp. 373–399). Springer International Publishing. 10.1007/978-0-8176-8379-5_19 [DOI] [Google Scholar]
  24. Paquette, M. , Merlet, S. , Gilbert, G. , Deriche, R. , & Descoteaux, M. (2015). Comparison of sampling strategies and sparsifying transforms to improve compressed sensing diffusion spectrum imaging. Magnetic Resonance in Medicine, 73(1), 401–416. 10.1002/mrm.25093 [DOI] [PubMed] [Google Scholar]
  25. Richie‐Halford, A. , Cieslak, M. , Ai, L. , Caffarra, S. , Covitz, S. , Franco, A. R. , Karipidis, I. I. , Kruper, J. , Milham, M. , Avelar‐Pereira, B. , Roy, E. , Sydnor, V. J. , Yeatman, J. , Consortium, T. F. C. S. , Satterthwaite, T. D. , & Rokem, A. (2022). An analysis‐ready and quality controlled resource for pediatric brain white‐matter research. Scientific Data, 9(1), 616. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Sakaie, K. E. , & Lowe, M. J. (2010). Quantitative assessment of motion correction for high angular resolution diffusion imaging. Magnetic Resonance Imaging, 28(2), 290–296. 10.1016/j.mri.2009.07.004 [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Schilling, K. G. , Fadnavis, S. , Batson, J. , Visagie, M. , Combes, A. J. E. , McKnight, C. D. , Bagnato, F. , Garyfallidis, E. , Landman, B. A. , Smith, S. A. , & O'Grady, K. P. (2021). Patch2Self denoising of diffusion MRI in the cervical spinal cord improves intra‐cord contrast, signal modelling, repeatability, and feature conspicuity. medRxiv (2021.10.04.21264389). 10.1101/2021.10.04.21264389 [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Smith, R. E. , & Connelly, A. (2020). MRtrix3_connectome: A BIDS application for quantitative structural connectome construction. OHBM. [Google Scholar]
  29. Stejskal, E. O. , & Tanner, J. E. (1965). Spin diffusion measurements: Spin echoes in the presence of a time‐dependent field gradient. The Journal of Chemical Physics, 42(1), 288–292. 10.1063/1.1695690 [DOI] [Google Scholar]
  30. Tuch, D. S. (2004). Q‐ball imaging. Magnetic Resonance in Medicine, 52(6), 1358–1372. [DOI] [PubMed] [Google Scholar]
  31. Veraart, J. , Novikov, D. S. , Christiaens, D. , Ades‐aron, B. , Sijbers, J. , & Fieremans, E. (2016). Denoising of diffusion MRI using random matrix theory. NeuroImage, 142, 394–406. 10.1016/j.neuroimage.2016.08.016 [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Wagner, A. S. , Waite, L. K. , Wierzba, M. , Hoffstaedter, F. , Waite, A. Q. , Poldrack, B. , Eickhoff, S. B. , & Hanke, M. (2022). FAIRly big: A framework for computationally reproducible processing of large‐scale data. Scientific Data, 9(1), 80. 10.1038/s41597-022-01163-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Wedeen, V. J. , Hagmann, P. , Tseng, W.‐Y. I. , Reese, T. G. , & Weisskoff, R. M. (2005). Mapping complex tissue architecture with diffusion spectrum magnetic resonance imaging. Magnetic Resonance in Medicine, 54(6), 1377–1386. [DOI] [PubMed] [Google Scholar]
  34. Yamada, H. , Abe, O. , Shizukuishi, T. , Kikuta, J. , Shinozaki, T. , Dezawa, K. , Nagano, A. , Matsuda, M. , Haradome, H. , & Imamura, Y. (2014). Efficacy of distortion correction on diffusion imaging: Comparison of FSL Eddy and Eddy_correct using 30 and 60 directions diffusion encoding. PLoS One, 9(11), e112411. 10.1371/journal.pone.0112411 [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Yeh, F. C. , Zaydan, I. M. , Suski, V. R. , Lacomis, D. , Richardson, R. M. , Maroon, J. C. , & Barrios‐Martinez, J. (2019). Differential tractography as a track‐based biomarker for neuronal injury. NeuroImage, 202, 116131. 10.1016/j.neuroimage.2019.116131 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

TABLE S1. Statistical comparisons of denoising, scheme type, number of directions, denoising method, and motion prevalence on RMSE of estimated motion.

TABLE S2. Statistical comparison of Eddy and SHORELine in shelled schemes. The outcome measure was the difference in error (RMSE) between Eddy and SHORELine (Eddy–SHORELine); the model included main effects of denoising, motion prevalence in the input data, and shell scheme; interactions of these main effects were modeled as well.

TABLE S3. Statistical comparisons of image smoothness between Eddy and SHORELine. The outcome measure was the difference in image smoothness (FWHM) between Eddy and SHORELine (Eddy–SHORELine); the model included main effects of denoising, motion prevalence in the input data, and sampling scheme.

TABLE S4. Statistical comparisons of image quality between Eddy and SHORELine. The outcome measure was the difference in image quality (NDC) between Eddy and SHORELine (Eddy–SHORELine); the model included main effects of denoising, motion prevalence in the input data, and shell scheme.

FIGURE S1. Empirical motion parameters in large‐scale datasets. QSIPrep outputs from the ABCD data processing and data from the public release of Healthy Brain Network (HBN; Richie‐Halford et al., 2022) study were used to obtain automated quality control (QC) metrics for ABCD and HBN datasets, respectively. The distribution of maximum translation (in mm) and rotation (in degrees) values are depicted for both datasets. The empirical motion parameters range widely, such that maximum translation = [0.22 9.09] and maximum rotation = [0.10 6.60] for ABCD (top row), and maximum translation = [0.09 20.78] and maximum rotation = [0.01 10.61] for HBN (bottom row).

Click here for additional data file. (550.8KB, docx)

Data Availability Statement

All data, figures and processing code are available to the public under a BSDv3 license. dMRI phantom data and preprocessing: The fiberfox simulation data is available at https://osf.io/38vce/, but is most conveniently accessed via DataLad from https://github.com/PennLINC/dMRI_HMC_Benchmark. Instructions for cloning this dataset are in the README of this repository, and cloning will also provide the singularity image containing all the software required to re‐run processing along with the input data. Figures and statistical tests: The source code for all statistical tests and figures is publicly available at https://github.com/PennLINC/SHORELINE_PaperDemo, with a visual walkthrough available here https://pennlinc.github.io/SHORELINE_PaperDemo/articles/dmri_hmc_benchmark.html.

Articles from Human Brain Mapping are provided here courtesy of Wiley

RESOURCES