Abstract
Purpose:
In echo-planar diffusion-weighted imaging, correcting for susceptibility-induced artifacts typically requires acquiring pairs of images, known as blip-up blip-down acquisitions, to create an undistorted volume as a target to correct distortions that are often focal where regions with differences in magnetic susceptibility interface, such as the frontal and temporal areas. However, blip-up blip-down acquisitions are not always available, and distortion effects may not be specifically localized to such areas, with subtle effects potentially extending throughout the brain. Here, we apply a deep learning technique to generate an undistorted volume to correct susceptibility-induced artifacts and demonstrate implications for image fidelity and diffusion-based inference outside of areas where high focal distortion is present.
Methods:
To demonstrate differences due to susceptibility artifact correction, uncorrected baseline images were compared to identical images where correction was performed using an undistorted target volume produced by the deep learning tool “PreQual”. Widespread geometric distortion was assessed visually by referencing diffusion-weighted images to T1-weighted images. Tract-based spatial statistics (TBSS) were utilized to perform whole brain analysis of fractional anisotropy (FA) values to assess differences between subject groups (depressed vs. non-depressed) via permutation-based, voxel-wise testing. Multivariate regression models were then used to contrast TBSS results between corrected and non-corrected diffusion images.
Results:
Susceptibility artifact correction resulted in visible, widespread improvement in image fidelity when referenced to T1-weighted images. TBSS results were dependent on susceptibility artifact correction with correction resulting in widespread structural alterations of the mean FA skeleton, changes in skeletal FA, and additional positive tests of significance of regression coefficients in subsequent regression models.
Conclusion:
Our results indicated that EPI distortion effects are not purely focal, and that reducing distortion can result in significant differences in the interpretation of diffusion data, even in areas remote from high distortion.
Keywords: Diffusion weighted imaging, Preprocessing, Distortion correction, Quality assurance
1. Introduction
Diffusion magnetic resonance imaging (dMRI) is a noninvasive, neuroimaging technique used to infer in vivo white matter integrity and trajectory [1]. Single-shot, echo-planer imaging (EPI) sequences are a popular method for acquiring dMRI because of their rapid scan times. Unfortunately, EPI sequences are prone to geometric and intensity distortions caused by a combination of susceptibility-induced field inhomogeneities and low bandwidth in the phase-encoding direction, producing spatial distortion along the phase-encoding axis [2,3]. EPI distortions are typically maximal in areas where there are large differences in magnetic susceptibility, such as the temporal and frontal regions [2,4]. However, distortion may be widespread and can vary in severity, resulting in inhomogeneities across the imaging plane [5] potentially causing misalignment with high-resolution structural images [6], differences in diffusion tensor metrics [7], and altered interpretations of diffusion-weighted images (DWIs) [8].
The presence of EPI distortion may be particularly relevant for applications involving group comparisons, where registering images to a common space is required to achieve an anatomical correspondence across subjects to make sense of group-level statistics on brain maps [9], and facilitate the comparison of experiments and meta-analysis [10,11]. Bringing EPI images into a common space typically requires rigidly co-registering EPI and structural images first to then normalize EPI images through applying a structural-to-template transformation. This indirect structural-based transformation procedure is often the preferred method for normalizing EPI images since computing a deformation from a structural image and a template is generally easier, due to the high contrast in high-resolution structural images, than computing a deformation between a subject’s EPI image and a template. Unfortunately, EPI distortion introduces geometric deformations between the EPI image and high-resolution structural image, though reducing EPI distortion to recover brain geometry has been shown to raise the accuracy of registration between DWIs and high-resolution structural images [12–20]. The influence of correcting for EPI distortion on inference during diffusion-based applications had been noted in several studies; due to the increased correspondence in in the anatomical overlap between subjects, EPI distortion correction has been shown to enhance statistical power in group studies [12], and improve re-test reliability of group comparisons involving diffusion analysis [21], including regions of the brain relatively remote from source of EPI distortion [22].
A popular correction strategy for EPI distortion involves acquiring pairs of images with reverse phase encoding directions, known as a “blip-up blip-down” acquisition, to generate an undistorted image as a target for unwarping [3,23,24]. This technique, available as the tool “topup” in FSL [25], can significantly improve extensive EPI distortion while allowing for the incorporation of subject motion into the correction. As image fidelity is restored by topup processing, this can provide a more accurate calculation of several diffusion metrics. For example, correcting EPI distortion via topup has been shown to raise the accuracy of estimates of white matter integrity, as measured by diffusion tensor imaging, through recovering signal lost due to compression along the phase encoding axis [26]. Additionally, performing topup has been shown to increase the anatomical accuracy and reproducibility of diffusion fiber tractography through reducing errors in underlying tensor fields [27], with similar benefits in tractography when using distortion correction methods in general [6,19,28–30]. Though topup is becoming more commonplace, not all DWIs are obtained with blip-up blip-down acquisitions. A novel diffusion preprocessing pipeline available to overcome this problem is the PreQual tool, a single integrated pipeline that offers a combination of established software tools for DWI preprocessing, including denoising, susceptibility correction, and quality assurance [31]. When blip-up blip-down acquisitions are not available, PreQual utilizes the Synb0-DisCo approach [32], which applies deep learning to synthesize an undistorted image from a T1-weighted structural image. The synthesized volume is then entered as an image with infinite bandwidth (i.e., without distortion) into topup. Following topup, PreQual uses FSL’s “eddy” tool [33,34] to correct for eddy current-induced distortions, inter-volume motion, and slice-wise signal dropout.
In this work, we evaluated the benefits of employing PreQual for correcting widespread geometric distortion of EPI data by comparing images corrected with PreQual to two baseline correction methods; FSL’s “eddy” tool, and FSL’s “eddy” tool combined with a denoising technique to improve signal-to-noise ratio. We identified differences in geometric fidelity, specific to the removal of susceptibility artifact, by visually comparing images processed using the PreQual technique, where susceptibility correction was performed, against duplicate input images processed identically excluding correcting for susceptibility artifact. Two baseline conditions, i.e., with and without denoising, were included to rule out preprocessing effects from denoising. To perform susceptibility correction, we utilized the Synb0-DisCo approach of PreQual to synthesize an undistorted b0 volume for input into topup, this allowed for susceptibility correction though all images were collected along the same phase-encoding direction with no reverse-phase encoded acquisitions collected. Matching preprocessing procedures between PreQual and the baseline methods, apart from correcting for susceptibility distortion, allowed for a depiction of differences in image fidelity particular to susceptibility artifact correction.
Compared to the baseline methods, PreQual processed data demonstrated visible widespread improvements in geometric image fidelity, notably extending beyond the typically distorted frontal and temporal regions. To further explore potential correction effects on diffusion analysis, we contrasted corrected vs non-corrected results generated from Tract-Based Spatial Statistics (TBSS), a voxel-wise comparison method used to assess whole brain fractional anisotropy (FA) differences between groups [35]. TBSS results were dependent on which correction method was applied, with PreQual processed DWIs exhibiting altered white matter skeletal structure and higher mean skeletal FA values within significant clusters. Comparing TBSS outcomes in regression analyses revealed that models predicting ROI FA calculated from PreQual processed DWIs resulted in additional significant coefficients over replica models involving data processed only with the baseline correction methods.
2. Materials and methods
2.1. Participants
MR images were obtained from ninety-four participants (mean age:67; range 60–86) who participated in a study examining late-life depression. Informed consent was obtained for each participant according to our Institutional Review Board. Participants were recruited through clinical referrals and advertisements. Fifty-three of the participants met Diagnostic and Statistical Manual of Mental Disorders (DSM-IV-TR) criteria for major depressive disorder (MDD) [36], single episode, recurrent, or chronic. The remaining forty-one subjects had no psychiatric or neurological history.
2.2. Image acquisition
All imaging was collected using a 3 T Philips Achieva system. T1-weighted anatomical images were acquired using an MPRAGE sequence with the following settings: 256 × 256 FOV, 0.88 × 0.88 × 1.2 resolution, 288 × 288 matrix size, 9 degree flip angle, 8.751 s repetition time/4.6 ms echo time, 1.2 mm slice thickness, and 142 slices. The DWI acquisition was performed with a HARDI acquisition protocol in the anterior to posterior phase-encoding direction, using an EPI sequence with diffusion gradients applied in sixty encoding directions. All DWI acquisitions were acquired along a single phase-encoding direction, no pairs of blip-up blip-down acquisitions were obtained. Single b-value scans (b = 2000 s/mm2) were acquired in 60 isotropic diffusion directions using the following settings: 240 × 240 FOV, 2.5 × 2.5 × 2.5 resolution, 96 × 96 matrix size, 90 degree flip angle, 6525 ms repetition time/75 ms echo time, 2.5 mm slice thickness, and 50 slices.
2.3. Denoising
Denoising was achieved through applying the Marchenko-Pastur PCA technique as used in MRTrix3 [37–39]. This method was applied to both the baseline condition with denoising and the PreQual condition.
2.4. PreQual susceptibility distortion correction
Distortion correction was performed retrospectively using deep learning methods implemented by the PreQual pipeline [31]. Briefly, PreQual corrected for susceptibility distortion by using FSL’s topup [23,25] to unwarp images using an undistorted b0 volume estimated from reverse phase encoded images. As reverse phase encoded images were absent, PreQual used Synb0-DisCo [32,40] to apply deep learning to synthesize an undistorted b0 volume of infinite bandwidth from a T1 structural image from the same subject. The undistorted b0 image was then used as an anatomical target for susceptibility-induced distortion correction via topup. To synthesize the undistorted b0 volume, Synb0-DisCo applied a 3D U-Net [41,42] (multi-slice, multi-view, 2 channel input and 1 channel output) generative adversarial network (GAN), based on the original implementation in PyTorch [43], to generate the undistorted b0 volume from a single blip (distorted) b0 and a structural T1 image. The datasets used for training this network consisted of inputs of both structural T1 images and distorted b0 images and a truth of undistorted b0 images. Training datasets were obtained from the Baltimore Longitudinal Study of Aging (BLSA), Human Connectome Project (HCP), and Vanderbilt University. To enable generalization across diverse acquisition parameters, dataset acquisitions were comprised of varying resolutions, signal-to-noise ratios, T1 and diffusion contrasts, magnitudes of distortions, and directions of distortions. For training, the data was partitioned across subjects for test, validation, and training sets. First the data was partitioned into a test set of 100 random subjects and a “learning” set of 850 subjects. The test set was completely withheld. The “learning” set was again partitioned using 5-fold cross validation into training and validation sets randomly shuffled into 680 testing and 170 validation for each fold. This approach was then applied to the withheld test set to show that distortions were successfully corrected. Additionally, the validity of the Synb0-DisCo approach was further demonstrated on several external validation sets that were not included in the original training/validation/testing datasets. Importantly, the external validation datasets consisted of varying acquisition parameters, distortion directions, brain sizes, and brain ages.
After PreQual synthesized the undistorted b0 volume, the synthesized volume was registered to and concatenated with the respective real b0 volume and was then entered into topup. The acquisition parameters for topup processing were set such that the readout time (i.e., time between the center of the first echo and center of last echo) for the synthesized images was set to 0 while the real b0 retained the correct readout time. This informed the topup algorithm that the synthesized volume had infinite bandwidth in the PE direction, with no distortion, thus fixing its geometry when estimating the susceptibility field and allowing warping only of the real b0, forcing the geometry of the real b0 to match the undistorted synthesized image. After topup, the PreQual pipeline used FSL’s “eddy” in conjunction with a brain mask estimated by (bet) on the averaged topup output to perform gradient rotation and correct eddy current distortions, motion, and slice-wise signal dropout [33,34]. Comparison data was ran through the same eddy protocol including a bet estimated brain mask, no other preprocessing was performed.
2.5. TBSS analysis
After preprocessing, FSL’s “dtifit” tool was used to compute the diffusion tensors on a voxel-wise level [44,45] from which FA maps for each participant were created. All FA maps were then registered to a common space using the Tract-Based Spatial Statistics (TBSS) method [35]. To determine a study-specific template in common space, each FA map was non-linearly transformed to every other map and the most “typical” target was selected. The “typical” target FA image was selected by taking each FA image in turn and estimating the amount of warping necessary to align all other images to it; the target was the FA image that required the smallest amount of warping when used as a target. This target was normalized using an affine registration to MNI 152 space. Every FA map was then transformed to MNI 152 space by combining the non-linear transformation to the target FA map with the affine transformation from target to MNI space. For this step, each FA image first had the non-linear transform to the target applied, and then each image had the affine transform to MNI152 space applied. This resulted in each FA image being transformed to MNI 152 space. This transformation method was based on work showing that a non-linear transform to target space, rather than affine-only, is insufficient for proper alignment; additionally, an affine transform to MNI152 space, despite not being non-linear, was found to be accurate [35]. FA maps were then averaged to produce a mean FA map that was thinned to generate a mean white matter skeleton. The skeleton was then thresholded at 0.2 to reduce partial volume effects and between participant misregistration. Each participant’s FA map was then projected onto the white matter skeleton to generate a 4D image containing all participants’ skeletonized FA data.
2.6. ROIs
Two white matter regions of interest were chosen a priori to examine possible downstream inference effects of EPI distortion correction. A white matter region of interest representing the right uncinate fasciculus, and an additional white matter region of interest representing the left anterior cingulum, were determined on the white matter skeleton. These regions were selected based on their relation to the population included in the current sample, i.e., depressed older adults, and because they allowed for investigation of downstream effects in variable locations in diffusion imaging space. Previous research has found that FA in the right uncinate fasciculus is lower in depressed older adults compared to healthy controls [46], and that white matter integrity in the left cingulum varies between depressed older adults and healthy controls [47]. The location of these regions were determined by using the Johns Hopkins University (JHU) white matter tractography atlas [48] as a template to mask regions. Mean FA values from voxels that were simultaneously within the ROIs and on the TBSS mean FA skeleton were extracted for subsequent statistical analyses.
2.7. Statistical analyses
Voxel-wise statistical analysis was performed using FSL’s permutation-based “randomise” tool (using 10,000 permutations). Group differences in FA skeleton voxels between depressed and non-depressed older adults were assessed using a between subject t-test implemented in the randomise tool, controlling for age. Multiple comparison correction was applied using the threshold-free cluster enhancement (TFCE) option in randomise. The significance threshold was set at 0.05 on the corrected P-value data. Mean FA values were obtained from significant clusters using the ‘fslmeants’ command. ROI FA values were obtained through extracting mean FA from voxels that were both on the white matter skeleton and within the JHU regions. Multiple linear regression was then used to test the relation between ROI FA and depression status, age, and their interaction. Results from DWIs preprocessed with the baseline methods were compared against results from DWIs preprocessed with PreQual.
3. Results
3.1. Qualitative results
Fig. 1 displays EPI susceptibility distortion and correction in two example subjects with distortion free high-resolution T1-weighted images for reference. EPI distortion is noticeable as compression of image geometry in the phase encoding direction, coinciding with signal “pile-up” manifesting as areas of high and low intensity most noticeable in the frontal and temporal regions (Fig. 1a, c). The PreQual processed images are corrected to a similar geometry as the distortion free high-resolution T1-weighted images, with improvements in areas with signal pile-up, and distortion correction including and extending beyond the focal distortions present in frontal and temporal regions (Fig. 1b, d).
Fig. 1.
T1-weighted images (top) and b0s (bottom) from two separate subjects, showing sagittal and coronal sections and comparing images from the baseline method on the left side (Fig. 1a, c) to PreQual processed images on the right (Fig. 1b, d). T1-weighted images are provided as a distortion-free reference. For the baseline processing method, b0s display typical susceptibility distortion along the phase encoding direction. After the PreQual processing method, applying distortion correction via a synthesized undistorted image from a T1-weighted image, b0s display noticeable improvements in overall geometric fidelity in areas including occipital regions and near the ventricles. Red circles emphasize major differences in distortion between b0s processed with PreQual vs the baseline method, red lines serve as a guide to help compare b0s with their respective distortion-free T1-weighted reference image, and green lines simply serve to help orient the image position with vertical green lines additionally indicating where slices are being viewed such that vertical lines on sagittal views show the location of the respective coronal view, and vice versa.
3.2. Group differences in TBSS results
A TBSS voxel-wise comparison of FA differences between depressed and non-depressed older adults revealed significant clusters of increased FA following permutation-based TFCE, controlling for age. Fig. 2 shows clusters signifying significantly increased FA for the non-depressed group relative to the depressed group and compares baseline data preprocessed with eddy correction (Fig. 2a) to PreQual processed data (Fig. 2b). Compared to the baseline method, PreQual demonstrated larger FA skeleton size when measured in voxels, larger clusters when measured by total voxels comprising clusters, and greater mean FA values for voxels within significant clusters (see Table 1). Fig. 3 also shows clusters signifying significantly increased FA for the non-depressed group relative to the depressed group, comparing baseline data preprocessed with both eddy correction and denoising (Fig. 3a) to PreQual processed data (Fig. 3b). Compared to the baseline method preprocessed with both eddy correction and denoising, PreQual demonstrated larger FA skeleton size when measured by total voxels, and greater mean FA values for voxels within significant clusters (see Table 1).
Fig. 2.
A comparison of TBSS results between the baseline method on the top (Fig. 2a) and PreQual on the bottom (Fig. 2b). Red regions represent areas of greater FA for healthy compared to depressed older subjects, controlling for effects due to age (P < 0.05, threshold-free cluster enhancement corrected), green tracts represent the mean FA skeleton from entire sample for that particular preprocessing method. Results were superimposed onto the standard MNI152 brain template and displayed in sagittal, coronal, and axial views (from left to right). Compared to the baseline method, TBSS results from DWIs processed with PreQual exhibited altered skeletal structure, a larger number of voxels included in the mean FA skeleton, and greater average FA values for skeletal voxels (see Table 1).
Table 1.
Differences between the baseline methods and PreQual by number of total skeletal voxels, number of voxels included in TFCE clusters determined by randomise to have significantly greater FA for healthy vs depressed older adults, and mean FA for each.
| Region | Voxel Count: Baseline/Baseline with Denoise/PreQual | Mean FA: Baseline/Baseline with Denoise/PreQual |
|---|---|---|
| • Mean FA skeleton (threshold set at 0.2) | 122,609/124,329/126,748 | 0.359/0.373/0.371 |
| • Clusters with significantly higher FA for healthy vs patients (P < 0.05, threshold-free cluster enhancement corrected, controlling for age) | 7497/9256/8714 | 0.427/0.440/0.444 |
Fig. 3.
A comparison of TBSS results between the baseline method with denoising on the top (Fig. 3a) and PreQual on the bottom (Fig. 3b). Red regions represent areas of greater FA for healthy compared to depressed older subjects, controlling for effects due to age (P < 0.05, threshold-free cluster enhancement corrected), green tracts represent the mean FA skeleton from entire sample for that particular preprocessing method. Results were superimposed onto the standard MNI152 brain template and displayed in sagittal, coronal, and axial views (from left to right). Compared to the baseline method, TBSS results from DWIs processed with PreQual exhibited altered skeletal structure, a larger number of voxels included in the mean FA skeleton, and greater average FA values for skeletal voxels within significant clusters (see Table 1).
3.3. Linear regression
Age and depression status interacted to predict uncinate fasciculus ROI FA for both the baseline preprocessing method (ΔR2 = 0.048, β = −0.283, P = 0.023) (Table 2) and the PreQual preprocessing method (ΔR2 = 0.044, β = −0.272, P = 0.027) (Table 4), but not for the baseline preprocessing method that applied denoising (ΔR2 = 0.033, β = −0.237, P = 0.059) (Table 3). Notably the effect of age was significant only in the model testing PreQual data (β = −0.247, P = 0.047), not for the baseline processed data (β = −0.210, P = 0.094), or the baseline with denoising data (β = −0.226, P = 0.075). For models predicting cingulum ROI FA, age and depression status interacted to predict FA only for the PreQual preprocessing method (ΔR2 = 0.038, β = −0.251, P = 0.044) (Table 7), not for the baseline preprocessing method (ΔR2 = 0.019, β = −0.176, P = 0.152) (Table 5), or the baseline preprocessing method with denoising (ΔR2 = 0.019, β = −0.176, P = 0.154) (Table 6). Individual level ROI FA values, from PreQual and both baseline preprocessing methods, are included to provide information regarding individual level improvements, before and after distortion correction (Supplementary Table S1).
Table 2.
Regression model predicting Uncinate Fasciculus FA (baseline method) from Age and Depression Status and the interaction between Age and Depression Status.
| Coefficients | ||||||
|---|---|---|---|---|---|---|
| Model | Unstandardized | Standard Error | Standardized | t | p | |
| 1 | (Intercept) | 0.354 | 0.003 | 122.018 | <0.001 | |
| Centered Age | −0.001 | 3.583e – 4 | −0.388 | −3.907 | <0.001 | |
| Depression Status | −0.005 | 0.004 | −0.128 | −1.290 | 0.200 | |
| R2 = 0.145 | ||||||
| 2 | (Intercept) | 0.353 | 0.003 | 123.453 | <0.001 | |
| Centered Age | −7.574e – 4 | 4.476e – 4 | −0.210 | −1.692 | 0.094 | |
| Depression Status | −0.005 | 0.004 | −0.133 | −1.370 | 0.174 | |
| Age X Status | −0.002 | 7.184e – 4 | −0.283 | −2.304 | 0.023 | |
| R2 = 0.193 |
Table 4.
Regression model predicting Uncinate Fasciculus FA (PreQual method) from Age and Depression Status and the interaction between Age and Depression Status.
| Coefficients | ||||||
|---|---|---|---|---|---|---|
| Model | Unstandardized | Standard Error | Standardized | t | p | |
| 1 | (Intercept) | 0.367 | 0.003 | 121.443 | <0.001 | |
| Centered Age | −0.002 | 3.735e – 4 | −0.419 | −4.273 | <0.001 | |
| Depression Status | −0.006 | 0.004 | −0.154 | −1.573 | 0.119 | |
| R2 = 0.170 | ||||||
| 2 | (Intercept) | 0.366 | 0.003 | 122.714 | <0.001 | |
| Centered Age | −9.413e – 4 | 4.671e – 4 | −0.247 | −2.015 | 0.047 | |
| Depression Status | −0.007 | 0.004 | −0.159 | −1.657 | 0.101 | |
| Age X Status | −0.002 | 7.498e – 4 | −0.272 | −2.249 | 0.027 | |
| R2 = 0.214 |
Table 3.
Regression model predicting Uncinate Fasciculus FA (baseline with denoise method) from Age and Depression Status and the interaction between Age and Depression Status.
| Coefficients | ||||||
|---|---|---|---|---|---|---|
| Model | Unstandardized | Standard Error | Standardized | t | p | |
| 1 | (Intercept) | 0.369 | 0.003 | 117.961 | <0.001 | |
| Centered Age | −0.001 | 3.870e – 4 | −0.376 | −3.772 | <0.001 | |
| Depression Status | −0.007 | 0.004 | −0.163 | −1.634 | 0.106 | |
| R2 = 0.141 | ||||||
| 2 | (Intercept) | 0.369 | 0.003 | 118.343 | <0.001 | |
| Centered Age | −8.793e – 4 | 4.877e – 4 | −0.226 | −1.803 | 0.075 | |
| Depression Status | −0.007 | 0.004 | −0.167 | −1.699 | 0.093 | |
| Age X Status | −0.001 | 7.828e – 4 | −0.237 | −1.911 | 0.059 | |
| R2 = 0.174 |
Table 7.
Regression model predicting Cingulum FA (PreQual method) from Age and Depression Status and the interaction between Age and Depression Status.
| Model | Unstandardized | Standard Error | Standardized | t | p | |
|---|---|---|---|---|---|---|
| 1 | (Intercept) | 0.472 | 0.004 | 133.897 | <0.001 | |
| Centered Age | −0.002 | 4.359e – 4 | −0.383 | −3.874 | <0.001 | |
| Depression Status | −0.010 | 0.005 | −0.201 | −2.036 | 0.045 | |
| R2 = 0.153 | ||||||
| 2 | (Intercept) | 0.471 | 0.003 | 134.729 | <0.001 | |
| Centered Age | −9.912e – 4 | 5.478e – 4 | −0.225 | −1.809 | 0.074 | |
| Depression Status | −0.010 | 0.005 | −0.206 | −2.115 | 0.037 | |
| Age X Status | −0.002 | 8.793e – 4 | −0.251 | −2.045 | 0.044 | |
| R2 = 0.191 |
Table 5.
Regression model predicting Cingulum FA (baseline method) from Age and Depression Status and the interaction between Age and Depression Status.
| Coefficients | ||||||
|---|---|---|---|---|---|---|
| Model | Unstandardized | Standard Error | Standardized | t | p | |
| 1 | (Intercept) | 0.438 | 0.003 | 130.137 | <0.001 | |
| Centered Age | −0.002 | 4.154e – 4 | −0.417 | −4.287 | <0.001 | |
| Depression Status | −0.010 | 0.005 | −0.219 | −2.248 | 0.027 | |
| R2 = 0.181 | ||||||
| 2 | (Intercept) | 0.437 | 0.003 | 129.552 | <0.001 | |
| Centered Age | −0.001 | 5.279e – 4 | −0.306 | −2.473 | 0.015 | |
| Depression Status | −0.010 | 0.005 | −0.222 | −2.292 | 0.024 | |
| Age X Status | −0.001 | 8.474e – 4 | −0.176 | −1.445 | 0.152 |
Table 6.
Regression model predicting Cingulum FA (baseline with denoise method) from Age and Depression Status and the interaction between Age and Depression Status.
| Coefficients | ||||||
|---|---|---|---|---|---|---|
| Model | Unstandardized | Standard Error | Standardized | t | p | |
| 1 | (Intercept) | 0.458 | 0.004 | 130.476 | <0.001 | |
| Centered Age | −0.002 | 4.339e – 4 | −0.392 | −4.002 | <0.001 | |
| Depression Status | −0.012 | 0.005 | −0.247 | −2.519 | 0.014 | |
| R2 = 0.171 | ||||||
| 2 | (Intercept) | 0.458 | 0.004 | 129.872 | <0.001 | |
| Centered Age | −0.001 | 5.515e – 4 | −0.280 | −2.254 | 0.027 | |
| Depression Status | −0.012 | 0.005 | −0.250 | −2.564 | 0.012 | |
| Age X Status | −0.001 | 8.852e – 4 | −0.176 | −1.436 | 0.154 | |
| R2 = 0.190 |
4. Discussion
Here, we assessed a method for EPI susceptibility distortion correction and found that without correction it is possible for susceptibility artifact to produce widespread influences on image fidelity that can extend beyond the frontal and temporal areas where distortion is typically observed. Specifically, we applied a retrospective correction of susceptibility artifact using deep learning (the PreQual software tool) and referenced susceptibility artifact corrected diffusion images to high resolution T1-weighted images and confirmed improvements in image fidelity in frontal and temporal regions and more broadly in additional areas including occipital regions and regions around the ventricles.
To further explore possible downstream effects of susceptibility artifact correction, TBSS and regression analyses targeting fractional anisotropy were performed and found that utilizing susceptibility artifact correction altered analysis outcomes in areas removed from where regions of different magnetic susceptibility interface. For PreQual processed diffusion data, compared to data processed with the baseline methods, differences in TBSS analyses included: widespread changes in the structure and intensity of the mean FA skeleton including regions not near focal artifact, a greater number of total voxels comprising the mean FA skeleton, changes in the size and location of the clusters denoting significant difference between groups, and greater mean FA within significant clusters. Subsequent regression analyses revealed significant outcomes that were insignificant prior to susceptibility artifact correction. Importantly, the differentiation of diffusion-based analyses found here existed above applying eddy correction and denoising preprocessing procedures, suggesting that non-local effects from susceptibility artifact can persist despite removing noise components to reduce signal fluctuations and enhance signal-to-noise ratios.
The PreQual tool utilized a synthesized, undistorted EPI image to enable topup processing to use a Gaussian predictor to correct for susceptibility distortion. In contrast, the baseline methods could not correct for susceptibility-induced distortion, which could explain why images processed with PreQual resulted in different FA values. FA estimates, thought to index white matter integrity [4], are important for a wide range diffusion-based applications ranging from investigations on schizophrenia [49], and music processing [50] to research focusing on measuring white pathology in depression [51], and aging [52]. When FA maps are produced from distorted DWIs, there is potential for a break-down of coherent signal structure across subjects where signal compression has misplaced FA information and reduced alignment between subjects [8]. In line with the current research, correcting for EPI distortion has been shown refine alignment between FA images and high-resolution structural images [53,54], and increase the accuracy of FA estimates and reduce FA variability between subjects [16,19,22]. Considering these associations between FA and EPI distortion correction, and the improvements in geometric fidelity due to PreQual processing, FA estimates computed from PreQual processed images are likely higher in accuracy compared to FA estimates computed from images processed without EPI distortion correction.
The observed differences in FA results between preprocessing methods were spatially variable, rather than global shifts across the TBSS skeleton. These spatially specific shifts may be explained by the regionally variable nature of susceptibility-induced artifacts, which in addition to disrupting field homogeneity and warping DWIs, could impact co-registrations during the TBSS pipeline. The TBSS skeletonization process involves projecting each subject’s aligned FA image onto the mean FA skeleton. However, this process is constrained such that a particular voxel may only be assigned to a single section of skeleton. If there is substantial residual misalignment between subjects, as can occur when susceptibility artifact is present, then the mean white matter skeleton may represent anatomically disparate sections of each subject’s individual white matter structure, potentially weaking the sensitivity and specificity of any subsequent statistical analyses. This concern, combined with complications created by susceptibility artifact in and of itself, may explain the dissimilarities in white matter skeleton structure and subsequent voxel-wise statistics found between DWIs processed with PreQual vs the baseline method. The observation that TBSS results depended on preprocessing method suggests the importance of EPI distortion correction when looking for differences in white matter. In this work, results from TBSS, a method for aligning DWIs across subjects to localize group differences, depended on EPI distortion correction, with significant differences in FA only being detected after distortion correction. This suggests that significant findings from TBSS may be obscured when EPI distortion is present, and that a key feature of TBSS, namely minimizing anatomical misalignment between subjects’ white matter structures, still benefits from EPI distortion correction, an outcome in agreement with research arguing that reliability of TBSS analyses can hinge on EPI distortion correction [21].
The current research adds to work on EPI distortion through characterizing an example where spatial distortion alters the interpretation of diffusion data outside of temporal and frontal regions. We argue that the remote distortions we are exhibiting in this work can have negative consequences for the interpretation of diffusion data, through problems such as signal compression and disrupted co-registration between diffusion and anatomical images. Furthermore, as previous research has argued that EPI distortion can alter common diffusion metrics, such as fraction anisotropy, in frontal and temporal regions where distortion is commonly identified, we believe it should be noted that similar effects could occur any place where distortion is significantly present, as in the remote regions we exhibited.
The key takeaways from this work are that (1) EPI distortions are not focal, (2) EPI distortions are not random, and (3) interpreting diffusion MRI in distorted space is not equivalent to corrected space, even in areas remote from focal distortions. Historically, these findings would cast yet another Achilles’ heel toward the diffusion MRI community and use of historical data. Yet, recent innovations with deep learning allow geometric correction using model tools (FSL’s topup). We prospectively applied these tools and found high qualitative agreement with structural anatomy and quantitatively increased effects for a condition with known white matter involvement.
4.1. Limitations
Synthesizing an undistorted image for use as an anatomical target for distortion correction does pose several potential limitations, outlined in detail by Schilling & colleagues [32]. Briefly, image contrast may be affected by filtering, learned network structure may be affected by large differences in acquisitions settings, and appropriate image contrast may not be predicted in certain abnormalities (e.g., tumors). However, in the current sample this distortion correction process still resulted in improved matching with the geometry of undistorted anatomical images, suggesting this method as a source for improvements in susceptibility distortion with potential benefits for DWI registration and subsequent statistical interpretations.
Another potential limitation of the current work is the resolution of subjects’ DWIs. The current work was conducted using diffusion subject data with an isotropic resolution (2.5 mm × 2.5 mm × 2.5 mm). Though investigations at higher resolutions could be beneficial for refining between subject registration, especially where smaller anatomical structures are located, analyzing this patient data permitted us to study a diverse population of older adults, including patients and controls, which permitted us to apply our results to the type of MRI data that might be obtained in a clinical research setting.
5. Conclusion
In this work, we assessed a method for retrospective distortion correction and found widespread improvement in geometric fidelity and significant differences in subsequent statistical analyses. The primary result confirmed here is that EPI susceptibility distortion is not limited to localized areas, and that analyses involving distorted space may not be equivalent with those conducted in corrected space, even in regions removed from areas of high focal distortion.
Supplementary Material
Grant support
This research was supported by the National Institutes of Health under award number R01MH10224. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Footnotes
CRediT authorship contribution statement
John P. Begnoche: Conceptualization, Writing – original draft, Writing – review & editing, Methodology, Formal analysis, Visualization. Kurt G. Schilling: Conceptualization, Software. Brian D. Boyd: Software, Resources. Leon Y. Cai: Methodology, Software. Warren D. Taylor: Resources, Supervision, Funding acquisition. Bennett A. Landman: Conceptualization, Supervision, Writing – review & editing.
Appendix A. Supplementary data
Supplementary data to this article can be found online at https://doi.org/10.1016/j.mri.2022.05.016.
References
- [1].Le Bihan D, Breton E, Lallemand D, Grenier P, Cabanis E, Laval-Jeantet M. MR imaging of intravoxel incoherent motions: application to diffusion and perfusion in neurologic disorders. Radiology 1986;161:401–7. 10.1148/radiology.161.2.3763909. [DOI] [PubMed] [Google Scholar]
- [2].Jezzard P, Balaban RS. Correction for geometric distortion in echo planar images from B0 field variations. Magn Reson Med 1995;34:65–73. 10.1002/mrm.1910340111. [DOI] [PubMed] [Google Scholar]
- [3].Andersson JLR, Skare S. Image distortion and its correction in diffusion MRI. In: Jones DK, editor. Diffus. MRI Oxford University Press; 2010. p. 285–302. 10.1093/med/9780195369779.003.0017. [DOI] [Google Scholar]
- [4].Jones DK, editor. Diffusion MRI: theory, methods, and application. Oxford; New York: Oxford University Press; 2010. [Google Scholar]
- [5].Farzaneh F, Riederer SJ, Pelc NJ. Analysis of T2 limitations and off-resonance effects on spatial resolution and artifacts in echo-planar imaging. Magn Reson Med 1990;14:123–39. 10.1002/mrm.1910140112. [DOI] [PubMed] [Google Scholar]
- [6].Irfanoglu MO, Walker L, Sarlls J, Marenco S, Pierpaoli C. Effects of image distortions originating from susceptibility variations and concomitant fields on diffusion MRI tractography results. NeuroImage 2012;61:275–88. 10.1016/j.neuroimage.2012.02.054. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [7].Brun L, Pron A, Sein J, Deruelle C, Coulon O. Diffusion MRI: assessment of the impact of acquisition and preprocessing methods using the BrainVISA-diffuse toolbox. Front Neurosci 2019;13:536. 10.3389/fnins.2019.00536. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [8].Embleton KV, Haroon HA, Morris DM, Ralph MAL, Parker GJM. Distortion correction for diffusion-weighted MRI tractography and fMRI in the temporal lobes. Hum Brain Mapp 2010;31:1570–87. 10.1002/hbm.20959. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [9].Friston KarlJ, Ashburner J, Frith CD, Poline J-B, Heather JD Frackowiak RSJ. Spatial registration and normalization of images. Hum Brain Mapp 1995;3:165–89. 10.1002/hbm.460030303. [DOI] [Google Scholar]
- [10].Gorgolewski KJ, Varoquaux G, Rivera G, Schwarz Y, Ghosh SS, Maumet C, et al. NeuroVault.org: a web-based repository for collecting and sharing unthresholded statistical maps of the human brain. Front Neuroinform 2015:9. 10.3389/fninf.2015.00008. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [11].Wager TD, Lindquist M, Kaplan L. Meta-analysis of functional neuroimaging data: current and future directions. Soc Cogn Affect Neurosci 2007;2:150–8. 10.1093/scan/nsm015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [12].Cusack R, Brett M, Osswald K. An evaluation of the use of magnetic field maps to undistort Echo-planar images. NeuroImage 2003;18:127–42. 10.1006/nimg.2002.1281. [DOI] [PubMed] [Google Scholar]
- [13].Tao R, Fletcher PT, Gerber S, Whitaker RT. In: Prince JL, Pham DL, Myers KJ, editors. A variational image-based approach to the correction of susceptibility artifacts in the alignment of diffusion weighted and structural MRI. Inf. Process. Med. Imaging, vol. 5636. Berlin, Heidelberg: Springer Berlin Heidelberg; 2009. p. 664–75. 10.1007/978-3-642-02498-6_55. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [14].Hutton C, Bork A, Josephs O, Deichmann R, Ashburner J, Turner R. Image distortion correction in fMRI: a quantitative evaluation. NeuroImage 2002;16: 217–40. 10.1006/nimg.2001.1054. [DOI] [PubMed] [Google Scholar]
- [15].Dohmatob E, Varoquaux G, Thirion B. Inter-subject registration of functional images: do we need anatomical images? Front Neurosci 2018;12:64. 10.3389/fnins.2018.00064. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [16].Irfanoglu MO, Modi P, Nayak A, Hutchinson EB, Sarlls J, Pierpaoli C. DR-BUDDI (diffeomorphic registration for blip-up blip-down diffusion imaging) method for correcting echo planar imaging distortions. NeuroImage 2015;106:284–99. 10.1016/j.neuroimage.2014.11.042. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [17].Duong STM, Phung SL, Bouzerdoum A, Schira MM. An unsupervised deep learning technique for susceptibility artifact correction in reversed phase-encoding EPI images. Magn Reson Imaging 2020;71:1–10. 10.1016/j.mri.2020.04.004. [DOI] [PubMed] [Google Scholar]
- [18].Zahneisen B, Baeumler K, Zaharchuk G, Fleischmann D, Zeineh M. Deep flow-net for EPI distortion estimation. NeuroImage 2020;217:116886. 10.1016/j.neuroimage.2020.116886. [DOI] [PubMed] [Google Scholar]
- [19].Qiao Y, Shi Y. Unsupervised deep learning for FOD-based susceptibility distortion correction in diffusion MRI. IEEE Trans Med Imaging 2021:1165–75. 10.1109/TMI.2021.3134496. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [20].Duong STM, Phung SL, Bouzerdoum A, Ang SP, Schira MM. Correcting susceptibility artifacts of MRI sensors in brain scanning: a 3D anatomy-guided deep learning approach. Sensors 2021;21:2314. 10.3390/s21072314. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [21].Wang S, Peterson DJ, Gatenby JC, Li W, Grabowski TJ, Madhyastha TM. Evaluation of field map and nonlinear registration methods for correction of susceptibility artifacts in diffusion MRI. Front Neuroinform 2017;11. 10.3389/fninf.2017.00017. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [22].Wu M, Chang L-C, Walker L, Lemaitre H, Barnett AS, Marenco S, et al. In: Metaxas D, Axel L, Fichtinger G, Sźekely G, editors. Comparison of EPI distortion correction methods in diffusion tensor MRI using a novel framework. Med. Image Comput. Comput.-Assist. Interv. – MICCAI 2008, vol. 5242. Berlin, Heidelberg: Springer Berlin Heidelberg; 2008. p. 321–9. 10.1007/978-3-540-85990-1_39. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [23].Andersson JLR, Skare S, Ashburner J. How to correct susceptibility distortions in spin-echo echo-planar images: application to diffusion tensor imaging. NeuroImage 2003;20:870–88. 10.1016/S1053-8119(03)00336-7. [DOI] [PubMed] [Google Scholar]
- [24].Morgan PS, Bowtell RW, McIntyre DJO, Worthington BS. Correction of spatial distortion in EPI due to inhomogeneous static magnetic fields using the reversed gradient method. J Magn Reson Imaging 2004;19:499–507. 10.1002/jmri.20032. [DOI] [PubMed] [Google Scholar]
- [25].Smith SM, Jenkinson M, Woolrich MW, Beckmann CF, Behrens TEJ, Johansen-Berg H, et al. Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage 2004;23:S208–19. 10.1016/j.neuroimage.2004.07.051. [DOI] [PubMed] [Google Scholar]
- [26].Gu X, Eklund A. Evaluation of six phase encoding based susceptibility distortion correction methods for diffusion MRI. Front Neuroinform 2019;13:76. 10.3389/fninf.2019.00076. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [27].Lahti K, Parkkola R, Jääsaari P, Haataja L, Saunavaara V, On behalf of the PIPARI Study Group, et al. The impact of susceptibility correction on diffusion metrics in adolescents. Pediatr Radiol 2021;51:1471–80. 10.1007/s00247-021-05000-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [28].Pintjens W, Poot DHJ, Verhoye M, Van Der Linden A, Sijbers J. In: Reinhardt JM, Pluim JPW, editors. Susceptibility correction for improved tractography using high field DT-EPI; 2008. 69142I. 10.1117/12.771096. San Diego, CA. [DOI] [Google Scholar]
- [29].Techavipoo U, Okai AF, Lackey J, Shi J, Dresner MA, Leist TP, et al. Toward a practical protocol for human optic nerve DTI with EPI geometric distortion correction. J Magn Reson Imaging 2009;30:699–707. 10.1002/jmri.21836. [DOI] [PubMed] [Google Scholar]
- [30].Qiao Y, Sun W, Shi Y. FOD-based registration for susceptibility distortion correction in brainstem connectome imaging. NeuroImage 2019;202:116164. 10.1016/j.neuroimage.2019.116164. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [31].Cai LY, Yang Q, Hansen CB, Nath V, Ramadass K, Johnson GW, et al. PreQual: an automated pipeline for integrated preprocessing and quality assurance of diffusion weighted MRI images. Magn Reson Med 2021:mrm.28678. 10.1002/mrm.28678. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [32].Schilling KG, Blaber J, Huo Y, Newton A, Hansen C, Nath V, et al. Synthesized b0 for diffusion distortion correction (Synb0-DisCo). Magn Reson Imaging 2019;64: 62–70. 10.1016/j.mri.2019.05.008. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [33].Andersson JLR, Sotiropoulos SN. An integrated approach to correction for off-resonance effects and subject movement in diffusion MR imaging. NeuroImage 2016;125:1063–78. 10.1016/j.neuroimage.2015.10.019. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [34].Andersson JLR, Graham MS, Zsoldos E, Sotiropoulos SN. Incorporating outlier detection and replacement into a non-parametric framework for movement and distortion correction of diffusion MR images. NeuroImage 2016;141:556–72. 10.1016/j.neuroimage.2016.06.058. [DOI] [PubMed] [Google Scholar]
- [35].Smith SM, Jenkinson M, Johansen-Berg H, Rueckert D, Nichols TE, Mackay CE, et al. Tract-based spatial statistics: Voxelwise analysis of multi-subject diffusion data. NeuroImage 2006;31:1487–505. 10.1016/j.neuroimage.2006.02.024. [DOI] [PubMed] [Google Scholar]
- [36].American Psychiatric Association. Diagnostic and statistical manual of mental disorders, fourth edition, text revision (DSM-IV-TR). 4th ed.vol. 1. Arlington, VA: American Psychiatric Association; 2000. 10.1176/appi.books.9780890423349. [DOI] [Google Scholar]
- [37].Cordero-Grande L, Christiaens D, Hutter J, Price AN, Hajnal JV. Complex diffusion-weighted image estimation via matrix recovery under general noise models. NeuroImage 2019;200:391–404. 10.1016/j.neuroimage.2019.06.039. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [38].Veraart J, Novikov DS, Christiaens D, Ades-aron B, Sijbers J, Fieremans E. Denoising of diffusion MRI using random matrix theory. NeuroImage 2016;142: 394–406. 10.1016/j.neuroimage.2016.08.016. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [39].Veraart J, Fieremans E, Novikov DS. Diffusion MRI noise mapping using random matrix theory: diffusion MRI Noise Mapping. Magn Reson Med 2016;76:1582–93. 10.1002/mrm.26059. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [40].Schilling KG, Blaber J, Hansen C, Cai L, Rogers B, Anderson AW, et al. Distortion correction of diffusion weighted MRI without reverse phase-encoding scans or field-maps. PLoS One 2020;15:e0236418. 10.1371/journal.pone.0236418. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [41].Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. ArXiv150504597 Cs. 2015. [Google Scholar]
- [42].Çiçek ö, Abdulkadir A, Lienkamp SS, Brox T, Ronneberger O. 3D U-Net: learning dense volumetric segmentation from sparse annotation. ArXiv160606650 Cs. 2016. [Google Scholar]
- [43].Ketkar N, Moolayil J. Automatic differentiation in deep learning. Deep learn. Python, Berkeley, CA: Apress; 2021. p. 133–45. 10.1007/978-1-4842-5364-9_4. [DOI] [Google Scholar]
- [44].Basser PJ, Mattiello J, LeBihan D. MR diffusion tensor spectroscopy and imaging. Biophys J 1994;66:259–67. 10.1016/S0006-3495(94)80775-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [45].Pierpaoli C, Basser PJ. Toward a quantitative assessment of diffusion anisotropy. Magn Reson Med 1996;36:893–906. 10.1002/mrm.1910360612. [DOI] [PubMed] [Google Scholar]
- [46].Charlton RA, Lamar M, Zhang A, Yang S, Ajilore O, Kumar A. White-matter tract integrity in late-life depression: associations with severity and cognition. Psychol Med 2014;44:1427–37. 10.1017/S0033291713001980. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [47].Taylor WD, Zhao Z, Ashley-Koch A, Payne ME, Steffens DC, Krishnan RR, et al. Fiber tract-specific white matter lesion severity Findings in late-life depression and by AGTR1 A1166C genotype. Hum Brain Mapp 2013;34:295–303. 10.1002/hbm.21445. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [48].Mori S, Oishi K, Jiang H, Jiang L, Li X, Akhter K, et al. Stereotaxic white matter atlas based on diffusion tensor imaging in an ICBM template. NeuroImage 2008;40: 570–82. 10.1016/j.neuroimage.2007.12.035. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [49].Alba-Ferrara LM, de Erausquin GA. What does anisotropy measure? Insights from increased and decreased anisotropy in selective fiber tracts in schizophrenia. Front Integr Neurosci 2013;7. 10.3389/fnint.2013.00009. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [50].Moore E, Schaefer RS, Bastin ME, Roberts N, Overy K. Diffusion tensor MRI tractography reveals increased fractional anisotropy (FA) in arcuate fasciculus following music-cued motor training. Brain Cogn 2017;116:40–6. 10.1016/j.bandc.2017.05.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [51].Cole J, Chaddock CA, Farmer AE, Aitchison KJ, Simmons A, McGuffin P, et al. White matter abnormalities and illness severity in major depressive disorder. Br J Psychiatry 2012;201:33–9. 10.1192/bjp.bp.111.100594. [DOI] [PubMed] [Google Scholar]
- [52].Grieve SM, Williams LM, Paul RH, Clark CR, Gordon E. Cognitive aging, executive function, and fractional anisotropy: a diffusion tensor MR imaging study. AJNR Am J Neuroradiol 2007;28:226–35. [PMC free article] [PubMed] [Google Scholar]
- [53].Bhushan C, Haldar JP, Joshi AA, Leahy RM. Correcting susceptibility-induced distortion in diffusion-weighted MRI using constrained nonrigid registration. In: Signal Inf Process Assoc Annu Summit Conf APSIPA Asia-Pac Asia-Pac Signal Inf Process Assoc Annu Summit Conf 2012; 2012. http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6412009. [PMC free article] [PubMed] [Google Scholar]
- [54].Lee J, Lazar M, Lee J, Holden J, Terasawa-Grilley E, Alexander AL. Correction of B0 EPI distortions in diffusion tensor imaging and white matter tractography. Proc Int Soc Magn Reson Med 2004;11:2172. [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.