- Email: [email protected]
Test-retest of automated segmentation with different motion correction strategies: A comparison of prospective versus retrospective methods
NeuroImage 209 (2020) 116494
Contents lists available at ScienceDirect
NeuroImage journal homepage: www.elsevier.com/locate/neuroimage
Test-retest of automated segmentation with different motion correction strategies: A comparison of prospective versus retrospective methods Steven R. Kecskemeti *, Andrew L. Alexander University of Wisconsin-Madison, USA
A R T I C L E I N F O
A B S T R A C T
Keywords: MPnRAGE Motion correction PROMO Segmentation FIRST Freesurfer Volume
Test-retest of automated image segmentation algorithms (FSL FAST, FSL FIRST, and FREESURFER) are computed on magnetic resonance images from 12 unsedated children aged 9.42.6 years ([min,max] ¼ [6.5 years, 13.8 years]) using different approaches to motion correction (prospective versus retrospective). The prospective technique, PROMO MPRAGE, dynamically estimates motion using specially acquired navigator images and adjusts the remaining acquisition accordingly, whereas the retrospective technique, MPnRAGE, uses a selfnavigation property to retrospectively estimate and account for motion during image reconstruction. To increase the likelihood and range of motions, participants heads were not stabilized with padding during repeated scans. When motion was negligible both techniques had similar performance. When motion was not negligible, the automated image segmentation and anatomical labeling software tools showed the most consistent performance with the retrospectively corrected MPnRAGE technique (80% volume overlaps for 15 of 16 regions for FIRST and FREESURFER, with greater than 90% volume overlaps for 12 regions with FIRST and 11 regions with FREESURFER). Prospectively corrected MPRAGE with linear view-ordering also demonstrated lower performance than MPnRAGE without retrospective motion correction.
1. Introduction T1-weighted (T1w) structural images are acquired in nearly all neuroimaging studies. The commonly used magnetization-prepared, rapid gradient echo method, termed MPRAGE (Mugler and Brookeman, 1990), provides three-dimensional (3D), whole-brain images with exquisite anatomical contrast at high spatial resolution in all dimensions. T1w images may be used to characterize the properties of brain morphometry – volume, shape, thickness, length, etc., which have formed the basis of many neuroimaging studies. Several popular techniques are widely used for automated segmentation of T1w brain images including Freesurfer (Dale et al., 1999; Dale and Sereno, 1993; Fischl et al., 2002, 2004) and tools from FSL (FMRIB Software Library) including FAST (Zhang et al., 2001) and FIRST (Patenaude et al., 2011). MPRAGE images are sensitive to even small motions during the approximately four to eight minute acquisition. Motion artifacts in MPRAGE manifest as ghosting, ringing and blurring. In particular, brain imaging studies in young children often suffer from motion artifacts. These motion artifacts may lead to biased estimates of structural brain measurements, possibly with increased variance (Sarlls et al., 2018; Watanabe, Liao, Jara and Sakai, 2013). Consequently, there is a
significant need for more robust structural imaging methods. Prospective motion correction methods dynamically adjust the imaging gradient waveforms based upon head motion estimates. Some strategies utilize optical tracking systems to measure head motion in real time (Forman et al., 2011; Maclaren et al., 2012); however, these methods require additional hardware and are currently under development. Alternatively, motion estimation using navigator images are used for PROMO (PROspective MOtion) correction algorithms which dynamically adjust the imaging gradients in near real-time (White et al., 2010). Several studies have demonstrated the utility of PROMO to reduce MPRAGE motion artifacts in both adult (Sarlls et al., 2018; Watanabe et al., 2016) and pediatric subjects (Brown et al., 2010; Kuperman et al., 2011). Another approach for T1w imaging called MPnRAGE (Kecskemeti et al., 2016) combines inversion RF pulse preparation with 3D radial k-space sampling. This method can generate a large number (“n ~ 300 to 400”) of MPRAGE volumes with a range of inversion-recovery contrasts including a high contrast T1w image in a single scan about the length of a traditional fully sampled MPRAGE acquisition. The radial k-space sampling pattern inherently reduces the sensitivity of MPnRAGE to motion, producing blurring artifacts as opposed to ringing or ghosting artifacts
* Corresponding author. E-mail address: [email protected] (S.R. Kecskemeti). https://doi.org/10.1016/j.neuroimage.2019.116494 Received 11 September 2019; Received in revised form 22 December 2019; Accepted 23 December 2019 Available online 30 December 2019 1053-8119/© 2019 Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
can be found in (White et al., 2010).
(Glover and Pauly, 1992). Moreover, the 3D radial k-space sampling enables retrospective motion correction utilizing self-navigation during the image reconstruction (Kecskemeti et al., 2018). Motion-corrected MPnRAGE was shown to be highly effective for producing high quality T1w images in a study of 32 children (133 years) with autism and 12 age-matched (123 years) controls (Kecskemeti et al., 2018). While that study demonstrated that image quality of motion-corrected MPnRAGE was consistently highly rated by neuroradiologists and automated measures of image sharpness, it did not evaluate the effects on morphometry estimation tools like Freesurfer and FSL. This study evaluates and compares the performance of Freesurfer and the FSL FAST and FIRST tools applied to MPRAGE with PROMO and MPnRAGE both with and without retrospective motion correction. Repeated MPRAGE and MPnRAGE acquisitions were performed on a cohort of 12 children to assess the reliability of structural morphometry measures. The reliability of these automated segmentation methods are assessed both in terms of the volume sizes and the regional overlap of repeated measures.
2.2.3. MPnRAGE acquisition parameters Whole brain coverage with 1.0 mm isotropic resolution was acquired in the axial orientation with 200 mm coverage in the superior/inferior direction. Parameters included delay time TD ¼ 500 ms, TR ¼ 4.9 ms, TE ¼ 1.8 ms, n ¼ 386 views along the recovery curve. The excitation flip angles were 4 /8 for the first 325/remaining 61 radial views. The scan time was 7 min, the same time allocated for MPRAGE-PROMO. The MPnRAGE composite image was reconstructed both with and without motion correction using the summation of all data across the inversion recovery curve. 2.2.4. MPnRAGE motion correction MPnRAGE uses a quasi-random 3D radial k-space trajectory designed so that the spokes acquired within the readout window (n ¼ 386) after each magnetization preparation module are approximately uniformly angularly distributed. The spokes are rotated between readout windows so that the summation of all spokes can be used for the final MPnRAGE image reconstruction, while the spokes within each readout window (~1900 ms) can be used to form a series of 3D navigator images to estimate the amount of motion within each readout window. The original k-space data are adjusted for translations using the Fourier Shift Theorem, while rotations are accounted for by rotating the original k-space trajectory. Motion within the readout window is not corrected and will result in blurring. More details about the retrospective motion correction can be found in (Kecskemeti et al., 2018).
2. Materials and methods 2.1. Study population Imaging experiments were performed with institutional review board approval and informed consent/assent. Twelve children (9.4 2.6 years, min ¼ 6.5 years, max ¼ 13.8 years, 6 male and 6 female) without known neurological health concerns were selected for imaging. Recruitment was not based on likelihood of subjects remaining still during the scan.
2.3. Image analyses 2.2. Image acquisition One mid-axial, one off-center sagittal, and two coronal slices (anterior and midbrain) from non-corrected and motion-corrected MPnRAGE and motion-corrected MPRAGE PROMO acquisitions were presented for visual image quality scoring. Two reviewers assed the images in consensus using the following 4-point Likert scale: 1 – severe motion artifacts (severe blurring or image ghosting obscuring detection of even large WM structures and/or complete loss of WM/GM borders), 2- moderate motion artifacts (WM/GM boundaries obscured by blurring or image ghosting, but generally still detectable), 3-mild motion artifacts (some localized image blurring or ghosting detectable, but not widespread) 4 – no noticeable image blurring or ghosting artifacts detected. Reviewers were blinded to acquisition and reconstruction type and the order of all images were randomly mixed. Region-of-interest based signal to noise ratio (SNR) and WM/GM contrast measurements were performed on the subject who had highest overlaps from FSL FAST to gauge the image quality when motion artifacts were not expected to significantly degrade image quality. MPnRAGE and MPRAGE images of like modalities were co-registered within subject using flirt from the FMRIB Software Library v5.0. No registrations were performed between modalities or across subjects. Coregistered images were then processed with fsl_anat from the FMRIB Software Library v5.0 using the default settings. This included bias-field correction [FAST], brain-extraction [FNIRT-based], tissue-type segmentation [FAST], and subcortical structure segmentation [FIRST]. This yielded segmentations of cerebral white matter, gray matter and CSF and 15 subcortical structures. Co-registered images were corrected for receiver inhomogeneities using N4BiasFieldCorrection (Tustison et al., 2010) and sent through the default Freesurfer pipeline, recon-all, to segment the same 15 subcortical structures as FIRST. The default biasfield correction in recon-all was still used. Dice (or volume-)-overlap-coefficients (DOCs) for each region of interest for all possible pairs of similar scans were calculated. The intraclass correlation coefficient (ICC) was computed using the volume measurements for each region of interest. A paired-sample t-test and two-sample F-test was used to compare the mean and standard deviations of the volume measurements of each region-of-interest between the MPnRAGE
All exams took place on a 3T MRI scanner (Discovery MR750, GE Healthcare, Waukesha, WI) without the use of sedation. Participants watched a video of their choice and were instructed to remain still. The participants heads were stabilized within a 32 channel phased array head coil (Nova Medical, Wilmington, MA) using the NoMoCo pillow support system (NoMoCo Pillow, Inc., La Jolla, Ca) before receiving an MPnRAGE and MPRAGE-PROMO scan. To assess greater variability of head motions, the padding was then removed and two additional scans of each method were performed in alternating order. The order of all scans was counter-balanced across subjects. 2.2.1. MPRAGE-PROMO acquisition parameters A works in progress version of the MPRAGE-PROMO acquisition version was used that acquired k-space with a linear centric viewordering scheme. Whole brain coverage with 1.0 mm isotropic resolution was acquired using a sagittal acquisition with 192 slices and a 256 256 mm in-plane acquisition matrix. Additional parameters include TI ¼ 900 ms, TR/TE ¼ 6.952 ms/2.92 ms, bandwidth ¼ 31.25 kHz, flip angle 8 , time between magnetization preparation pulses ¼ 2488 ms, and ARC acceleration of 2 and 1.25 along the phase encode and slice directions, respectively. The acquisition time was 6 min with an additional 1 min of data acquisition allowed as needed. 2.2.2. MPRAGE-PROMO motion correction MPRAGE-PROMO utilizes the period of “free recovery” between each SPGR readout block and the next magnetization preparation module to collect a series of orthogonal-2D “navigator” images to estimate the amount of motion between successive magnetization preparation pulses (2488 ms) and then adjust the excitation and imaging gradients if a certain motion criterion is met. Motion within the readout window (Nz*TR ¼ 1335 ms) is not corrected and will result in motion induced ghost artifacts. After a complete data set is acquired, an optional rescan period (1 min in this study) continually reacquires kspace data one kyplane at a time, each time choosing the ky-plane that is most corrupt by motion as determined by the navigator motion estimates. More details 2
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
plots of the motion transformations both show nearly continuous and jittery rotations around the x-axis and translations along the z-axis. Major ghost artifacts in the MPRAGE-PROMO images posed challenges for FIRST, which assigns a small, non-hippocampus-like ROI to a region slightly posterior to the expected location. The reduced resolution of the uncorrected MPnRAGE images do not pose significant challenges for FIRST. The results of the image quality assessment are shown in Table 1. Representative images of each score for each modality are show in Fig. 3. The images show a clear distinction between the blurring artifacts of MPnRAGE and ghosting artifacts of MPRAGE. Supplemental Fig. 1, an animated GIF, illustrates the residual blurring and correction capabilities of MPnRAGE for example cases with minor, moderate, and severe motions. The supplementary materials also contain motion plots and select orthogonal views for each scan. The mean scores for MPnRAGE increased from 2.8 without motion correction to 3.6 with motion correction, while the mean score for prospectively corrected MPRAGE PROMO was 2.7. Similarly, the standard deviation of scores, which represents the fluctuation of image quality that can be expected for each method, decreased from 1.2 to 0.5 when motion correction was applied to MPnRAGE, while it was 1.0 for prospectively corrected MPRAGE PROMO. A Mann-Whitney U test rejected the null hypothesis of equal medians before and after MPnRAGE motion correction (reviewer 1: P ¼ 0.004, z ¼ 2.9, reviewer 2: P ¼ 0.006, z ¼ 2.7, mean across reviewers: P ¼ 0.01,z ¼ 2.6), and between motion corrected MPnRAGE and MPRAGE PROMO (reviewer 1: P < 0.001, z ¼ 4.1, reviewer 2: P < 0.001, z ¼ 3.7, mean across reviewers: P < 0.001,z ¼ 3.9), but could not reject the null hypothesis of equal medians between uncorrected MPnRAGE and MPRAGE PROMO (reviewer 1: P ¼ 0.6, z ¼ 0.5, reviewer 2: P ¼ 0.5, z ¼ 0.6, mean across reviewers: P ¼ 0.6,z ¼ 0.5), The results from the FSL FAST segmentations are presented in Fig. 4, and Supplemental Tables 1–2. MPnRAGE with retrospective motion correction had the highest ICCs (>0.98 all regions), mean DOCs (88% for GM and 93% for WM), and lowest variability in DOCs across subjects ( 4% for GM and 2% for WM), while MPnRAGE without motion correction performed better in all areas than MPRAGE with prospective motion correction. The boxplots of volumes show a larger spread of measurements for prospectively corrected MPRAGE, as well as the presence of at least one measurement considered to be an outlier. The mean volumes and standard deviations are statistically significant (p < 0.001) between prospectively corrected MPRAGE and retrospectively corrected MPnRAGE for CSF and GM, but only the mean standard deviation is significant for WM (p ¼ 0.196 for mean volume comparison). The differences of the average and standard deviations of the DOCs are significant between all modalities. The results from the FSL FIRST and Freesurfer segmentations are presented in Figs. 5 and 6, and Supplemental Tables 3–9. FreeSurfer was unable to complete segmentation on six prospectively corrected MPRAGE images that came from four different subjects. All analysis involving prospectively corrected MPRAGE images and FreeSurfer used only those subjects in which all images were segmented. Comparisons in FreeSurfer determined volumes between MPRAGE and MPnRAGE also used only the subset of participants with segmentations for all scans. FreeSurfer determined volume overlaps for MPnRAGE were computed using all subjects. The ICC values of the volume measurements generally increased from prospectively motion corrected MPRAGE to MPnRAGE without motion correction and to MPnRAGE with retrospective motion correction. MPnRAGE without motion correction has six ROIs from FIRST and two from FreeSurfer with ICC values greater than 0.70, while MPnRAGE with motion correction had 14 ROIs from FIRST (remaining ROI had ICC of 0.697) and 11 ROIs with ICC values greater than 0.70. Prospectively corrected MPRAGE had much lower ICC values (max of 0.51) with 11 ROIs from FIRST and 12 from FreeSurfer at below 0.25. Even with this large difference in ICC values, there were no statistical differences (at p < 0.05) in volumes between measurements made with prospectively corrected MPRAGE and MPnRAGE without correction, although there were four (FIRST determined) and six (FreeSurfer determined) regions
and MPRAGE PROMO scans. Box-and-whisker plots are used to present the distribution of volumes and volume-overlap-coefficients. In the box plots, the red line marks the median, the box-length provides the inter-quartile (25th and 75th percentile ranges), and the whisker end points represent the locations of the last non-outlier measurements. Outlier measurements are points beyond 1.5 times the box length away from the 25th and 75th percentiles. With this convention, a 99.3% coverage of points would be provided for a normal distribution. 3. Results Example images chosen from the participants who had the highest DOC averaged across WM, GM, and CSF masks from FAST are shown in Fig. 1. The DOCs were 91% for both corrected and non-motion corrected MPnRAGE scans and were 89% for prospectively corrected MPRAGE. The SNRs in the putamen and frontal white matter were 27 and 30, respectively, for MPnRAGE, and 21 and 30, respectively, for prospectively corrected MPRAGE, while the contrast between putamen and frontal WM was 1.80 for MPnRAGE and 1.89 for prospectively corrected MPRAGE, suggesting comparable image quality across methods in the absence of large motion artifacts. Neither the uncorrected or motion corrected MPnRAGE images display any obvious motion artifacts, but the second and third time points of the MPRAGE-PROMO scans show mild artifacts typically associated with motion. Example images and hippocampus masks (white ROI) chosen from the participants who had the lowest DOC averaged across both WM, GM, and CSF masks from FAST segmentations are shown in Fig. 2. The DOCs were 61% and 79% for uncorrected and retrospectively motion corrected MPnRAGE scans and was 28% for prospectively corrected MPRAGE. The
Fig. 1. MPnRAGE and prospectively motion corrected MPRAGE images from a subject that provided the highest Dice-overlap-coefficients of whole brain WM, GM, and CSF masks from FSL FAST. Image quality for all MPnRAGE scans is visibly similar to scan one of prospectively corrected MPRAGE. Presence of ringing artifacts in scan 2 and scan 3 of the prospectively corrected MPRAGE images are indicative of a mismatch between the imaging gradients and participant position within the scanner. Since there was no evidence of motion in four of the six scans in this subject and since the order of scans was counterbalanced, the presence of these artifacts could suggest an overzealous detection and correction of minor motions in the prospectively corrected MPRAGE acquisition. 3
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
Fig. 2. Example source images (far left column), including hippocampus masks from FIRST (white overlay in second column) and estimated translations and rotations (right columns) are shown from the case that had the worst Dice-overlap-coefficients of whole brain WM, GM, and CSF masks from FAST. Notice that extreme motion ghosts of MPRAGE cause FIRST to completely fail, whereas even the uncorrected MPnRAGE images can be used to extract the hippocampus with FIRST. Both subjects demonstrated nearly continuous rotations around x and translations in z. There was also a 5 mm drift during the MPnRAGE scan.
with statistical differences (at p < 0.05) between prospectively corrected MPRAGE and retrospectively corrected MPnRAGE. MPnRAGE with retrospective motion correction had 41% less variation in volumetric measurements than prospectively corrected MPRAGE within FIRST regions and 23% less in FreeSurfer regions. Across all regions,
retrospectively corrected MPnRAGE had the highest mean and lowest standard deviation (91% 6% and 90% 4% for FIRST and FreeSurfer determined regions) of the Dice-overlap-coefficients, followed by uncorrected MPnRAGE (86% 12% and 84 17%) and prospectively corrected MPRAGE (79% 28% and 83% 17%). The boxplots (Figs. 5
Fig. 3. Representative example images of each Likert score (1–4) for prospectively motion corrected MPRAGE-PROMO (Fig. 3a) and MPnRAGE (Fig. 3b). MPnRAGE images are shown with retrospective motion correction for Likert Scores 4,3,2 and without motion correction for Likert Score 1. Artifacts from motion are more subtle in MPnRAGE images (slight blurring) as opposed to replicate ghosting in MPRAGE images. Corrected MPnRAGE images may have slight image degradation due to intra-navigator motions and uneven k-space sampling that was not accounted for in the reconstruction.
4
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
and 6) show the presence of outlier DOC measurements for all methods and larger interquartile ranges for most regions in prospectively corrected MPRAGE and uncorrected MPnRAGE compared to retrospectively corrected MPnRAGE. Results from the Freesurfer cortical surface reconstructions are shown in Table 2. Ten MPRAGE PROMO and three non-motion corrected MPnRAGE datasets did not complete processing. The presence of motion artifacts (ghosting for MPRAGE PROMO and blurring for uncorrected MPnRAGE) were confirmed in all failed datasets. All 36 datasets completed processing for motion corrected MPnRAGE. When analyzed using only the largest common subset of successfully completed cortical reconstructions, the only statistical differences in mean (ttest p < 0.05) was between uncorrected and corrected MPnRAGE (p ¼ 0.026).
Table 1 The total number of cases with each Likert score for each method (36 scans x 2 reviewers ¼ 72 scores total).
Likert score 1 Likert score 2 Likert score 3 Likert Score 4
MPnRAGE w/o motion correction
MPnRAGE w/motion correction
MPRAGE w/ PROMO
13
0
11
16
1
17
13
26
25
30
45
19
Fig. 4. Volumes (top) and volume overlaps (bottom) from automated segmentations using FSL FAST. Superscripts 1,2,3 indicate statistical differences in the mean values between prospectively corrected MPRAGE and MPnRAGE without correction (1), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (2), and between uncorrected and retrospectively corrected MPnRAGE (3) at p < 0.05. Subscripts a,b,c indicate statistical differences in the standard deviations between prospectively corrected MPRAGE and MPnRAGE without correction (a), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (b), and between uncorrected and retrospectively corrected MPnRAGE (c) at p < 0.05. Points beyond the whiskers are considered outliers and are shown as red crosses (þ). 5
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
Fig. 5. Volumes (top) and volume overlaps (bottom) from automated segmentations using FSL FIRST. Superscripts 1,2,3 indicate statistical differences in the mean values between prospectively corrected MPRAGE and MPnRAGE without correction (1), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (2), and between uncorrected and retrospectively corrected MPnRAGE (3) at p < 0.05. Subscripts a,b,c indicate statistical differences in the standard deviations between prospectively corrected MPRAGE and MPnRAGE without correction (a), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (b), and between uncorrected and retrospectively corrected MPnRAGE (c) at p < 0.05. Points beyond the whiskers are considered outliers and are shown as red crosses (þ). The large red stars with zero volume and overlaps indicate that data from four participants were excluded from analysis (values of zero were not used to construct the plots) for MPRAGE with prospective motion correction since FreeSurfer would not complete the necessary segmentations.
Similarly, the only statistical differences (fftest < 0.05) in standard deviations was between uncorrected MPnRAGE and the other two methods (p ¼ 0.01 for MPRAGE PROMO and P ¼ 0.002 for corrected MPnRAGE).
scan. Likewise, since both methods also correct for motion at approximately the same rate, it is also unlikely that this was the source of differences when motion was present. Neither method, however, corrects for motion that occurs between the navigator scans. MPRAGE PROMO used a Cartesian k-space acquisition with linear view-ordering, so motion during this time will result in ghost artifacts, whereas MPnRAGE used a radial k-space acquisition that will result in blurring artifacts. It is possible that the segmentation algorithms are more sensitive to ghost artifacts than local blurring. Alternative k-space readout strategies may help to reduce motion sensitivity of MPRAGE PROMO. Incomplete, overcorrection, or unnecessarily modification of the imaging gradients with PROMO and a Cartesian k-space acquisition will also produce ghosting artifacts, but will produce blurring artifacts when a radial kspace acquisition, as in MPnRAGE, is used. Thus, without a ground truth, it is hard to determine whether suboptimal image quality of either method was a result of inaccurate motion estimates or uncorrectable artifacts from intra-navigator motions or a combination of the two. Since the PROMO acquisition in this study utilized a different k-space acquisition strategy than MPnRAGE, it is not possible to conclude if the differences observed in this study were due to the motion correction
4. Discussion When motion was negligible both MPnRAGE and MPRAGE PROMO had similar properties as characterized by SNR, CNR, and DOC of the segmented tissue masks. However, when motion was not negligible, the automated image segmentation and anatomical labeling software tools showed more consistent performance with MPnRAGE than MPRAGEPROMO. In particular, MPnRAGE with motion correction showed very high regional label consistency (80% Dice overlaps for 15 of 16 regions (min ¼ 79% with FIRST, but 84% with Freesurfer) and >90% in 12 of the regions with FIRST segmentations and 11 with Freesurfer). Conversely, prospectively corrected MPRAGE demonstrated lower performance than MPnRAGE without motion correction. The scan order was counter-balanced both within and across subjects, thus the difference in performance in the presence of motion is unlikely due to differences in the degree or type of motions experienced for each 6
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
strategy itself (prospective vs. retrospective), or the k-space acquisition strategy (Cartesian with linear view-ordering vs 3D radial), or a combination of the two. The higher test-retest scores for MPnRAGE without correction compared to MPRAGE PROMO suggests a favorability of 3D radial k-space sampling. Regardless, the two approaches are independent and compatible and the combination of the two could address defeciences in each method. In the retrospectively corrected technique, head rotations will result in uneven k-space sampling that leads to increased streak artifacts that typically manifest as additive noise. The incorporation of PROMO would prospectively rotate the k-space trajectory accordingly to keep the sampling approximately uniform. Since a 3D radial k-space trajectory is used for MPnRAGE, intra-navigator motions and inaccurate PROMO motion estimates will result in blurring as opposed to ghosting artifacts when Cartesian sampling is used. Similar combinations with other k-space trajectories that frequently- or oversample the center of k-space, such as Cones (Gurney et al., 2006), radial Cones (Johnson, 2017), rotated spiral-PR (Irarrazabal and Nishimura, 1995), FLORET (Pipe et al., 2011), and radial fan beam (Madhuranthakam et al., 2009) would likely produce improvements over the
Table 2 Cortical surface reconstructions by motion correction type. The volumes and coefficients of variation use the largest common subset of the successfully completed scans across all motion correction types. Motion correction type
MPRAGE PROMO prospective
none
retrospective
Freesurfer Cortical Reconstructions (# failed/total #) Cortical volume (105 mm3)
10/36
3/36
0/36
6.1 0.6
5.8 1.0 0.17
6.2 0.5
Cortical volume (coefficient of variation)
0.09
MPnRAGE
0.08
traditional linearly view-ordered Cartesian MPRAGE method. With a less motion sensitive k-space trajectory, the need to retake motion corrupt data may be alleviated or eliminated, as there was no data retaken or eliminated with 3D radial MPnRAGE in this study. The evaluation conditions for these methods, young children with
Fig. 6. Volumes (top) and volume overlaps (bottom) from automated segmentations using FreeSurfer. Superscripts 1,2,3 indicate statistical differences in the mean values between prospectively corrected MPRAGE and MPnRAGE without correction (1), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (2), and between uncorrected and retrospectively corrected MPnRAGE (3) at p < 0.05. Subscripts a,b,c indicate statistical differences in the standard deviations between prospectively corrected MPRAGE and MPnRAGE without correction (a), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (b), and between uncorrected and retrospectively corrected MPnRAGE (c) at p < 0.05. Points beyond the whiskers are considered outliers and are shown as red crosses (þ). The large red stars with zero volume and overlaps indicate that data from four participants were excluded from analysis (values of zero were not used to construct the plots) for MPRAGE with prospective motion correction since FreeSurfer would not complete the necessary segmentations.
7
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
motion artifacts. The Freesurfer segmentations of the left thalamus-proper from retrospectively corrected MPnRAGE had ICC ¼ 0.479, but DOCs of 93 2%, which highlights the importance of considering both measures as markers of reliability. The results of the total cortical volume analysis indicate that when datasets with strong motion artifacts are excluded from analysis, no statistical differences in total cortical volumes exist. In summary, we found that MPnRAGE with retrospective motion correction provided highly reliable automated segmentations of regions within the human brain. These techniques are effective for volumetric studies in challenging populations including young children.
and without head stabilization padding, may be more challenging than many clinical or research scans. Thus, the degrees of motion are likely larger than encountered in many neuroimaging studies. Further, the scanner manufacturer strongly recommends the use of stabilizing padding for applications with MPRAGE PROMO. It is highly likely that the results for both methods would improve with the addition of stabilization pads and/or for more compliant subjects where motion is expected to be less severe, but that was not evaluated in this study. This study set the total allotted scan time to be 7 min for each method, which meant that MPRAGE-PROMO had a maximum rescan period of 1 min (6 min basescan þ 1 min rescan). A similar study of 9 pediatric subjects (Brown et al., 2010) aged 10.73 yrs 0.54 yrs used an unlimited rescan period and found that 8 of the subjects required less than a 1 min rescan (14.3s 18.7s). The remaining subject needed nearly 5 min of rescanning. It is possible that lengthening the rescan period for PROMO will produce better results. However Supplemental Fig. 2 shows that participant behavior during the rescan period of this study was consistent with behavior during the acquisition, so rescanning becomes less efficient the more often somebody moves, which is precisely when it is needed the most. Based on each participant’s motion, a 5 min rescan period would be sufficient for 30 of 36 scans, assuming the additional scan length does not produce less compliance. The remaining six scans (consisting of scans from 5 unique participants) are estimated to need at least 10 min (10 min, 13 min, 19 min, 32 min, 37 min, and 41 min respectively) to satisfy a maximum motion metric of 0.7 mm used in PROMO. Increasing the scan time for MPnRAGE could result in higher SNR, allow retrospective rejection of data (currently all data is used irregardless of motion metric), and increased spatial resolution. Traditionally, gradient non-linearity correction is performed as a single post-processing inverse-warping procedure where the amount of warping depends on the position of each voxel with respect to iso-center (Janke et al., 2004). When motion occurs, the final images represent the superposition of multiple positions so straightforward correction is no longer possible. However, the amount of warping generally varies slowly with position. We have estimated that the maximal change in the degree of warping resulting from a 10 rotation around the S/I axis is less than 0.1 mm for a head scan on our system. Thus, localization errors when determining the gradient warps are not expected to cause significant degradation of image quality. On systems where gradient nonlinearity is more problematic, non-linear gradient correction could be performed on small subsets of the data using positional estimates from the navigators (PROMO or MPnRAGE navigators) to help determine more accurate positional locations. A similar procedure was presented in for motion correction scans at 7T (Yarach et al., 2015). Previous studies of test-retest reliability, interscanner reliability, and comparison to manual segmentations using FIRST with MPRAGE without PROMO were previously performed (Nugent et al., 2013) in adult subjects and demonstrated reliable segmentations (high intraclass correlation coefficients (ICC) > 0.83, consistent volumes, and high Dice-overlap-coefficients (DOC)) of the thalamus, caudate, putamen, hippocampus, and pallidum (marginal, but acceptable) for repeat scans from the same scanner platform, but unreliable segmentations of the accumbens and amygdala. The less reliable segmentations of the accumbens and amygdala were also demonstrated in an earlier test-retest study of FIRST and Freesurfer (Morey et al., 2010), which also suggested that the reduced reliability of the accumbens and amygdala may be due to its smaller size, which makes it more susceptible to measurement error. In this study, retrospectively motion corrected MPnRAGE demonstrated acceptable segmentations in terms of volumes (ICCs > 0.7) and shapes (DOCs) for all regions except the right accumbens regions (FIRST, ICC ¼ 0.697), the left thalamus-proper (Freesurfer, ICC ¼ 0.479), left and right pallidum (Freesurfer, ICC ¼ 0.554 and 0.6917), left amygdala (Freesurfer, ICC ¼ 0.606). The amygdala and accumbens regions have also posed difficult in previous test-retest studies (Morey et al., 2010; Nugent et al., 2013) using adult subjects and images without obvious
Author contributions section Steven Kecskemeti: Conceptualization, Methodology, Software, Validation, Formal Analysis, Investigation, Data Curation, Writing -Original Draft & Editing, Visualization, Project administration, Funding acquisition. Andrew Alexander: Conceptualization, Methodology, Formal Analysis, Resources, Writing- Original Draft and Editing, Visualization, Supervision Acknowledgements Funding from the Department of Radiology, University of Wisconsin – Madison was received. The data collection from this study was supported by GE Healthcare and the UW-Madison Department of Radiology. The methods used in the study were supported in part by a core grant to the Waisman Center from the National Institute of Child Health and Human Development (IDDRC U54 HD090256). Additional support was provided by National Institutes of Health grants DA050258, MH097464, MH100031, MH101504, HD094715, AG051216, and AG15001. Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi. org/10.1016/j.neuroimage.2019.116494. References Brown, T.T., Kuperman, J.M., Erhart, M., White, N.S., Roddey, J.C., Shankaranarayanan, A., et al., 2010. Prospective motion correction of high-resolution magnetic resonance imaging data in children. Neuroimage 53 (1), 139–145. https:// doi.org/10.1016/j.neuroimage.2010.06.017. Dale, A.M., Fischl, B., Sereno, M.I., 1999. Cortical surface-based analysis. I. Segmentation and surface reconstruction. Neuroimage 9 (2), 179–194. https://doi.org/10.1006/ nimg.1998.0395. Dale, A.M., Sereno, M.I., 1993. Improved localizadon of cortical activity by combining EEG and MEG with MRI cortical surface reconstruction: a linear approach. J. Cogn. Neurosci. 5 (2), 162–176. https://doi.org/10.1162/jocn.1993.5.2.162. Fischl, B., Salat, D.H., Busa, E., Albert, M., Dieterich, M., Haselgrove, C., et al., 2002. Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain. Neuron 33 (3), 341–355. Fischl, B., Salat, D.H., van der Kouwe, A.J.W., Makris, N., Segonne, F., Quinn, B.T., Dale, A.M., 2004. Sequence-independent segmentation of magnetic resonance images. Neuroimage 23 (Suppl. 1), S69–S84. https://doi.org/10.1016/ j.neuroimage.2004.07.016. Forman, C., Aksoy, M., Hornegger, J., Bammer, R., 2011. Self-encoded marker for optical prospective head motion correction in MRI. Med. Image Anal. 15 (5), 708–719. https://doi.org/10.1016/j.media.2011.05.018. Glover, G.H., Pauly, J.M., 1992. Projection reconstruction techniques for reduction of motion effects in MRI. Magn. Reson. Med. 28 (2), 275–289. https://doi.org/10.1002/ mrm.1910280209. Gurney, P.T., Hargreaves, B.A., Nishimura, D.G., 2006. Design and analysis of a practical 3D cones trajectory. Magn. Reson. Med. 55 (3), 575–582. https://doi.org/10.1002/ mrm.20796. Irarrazabal, P., Nishimura, D.G., 1995. Fast three dimensional magnetic resonance imaging. Magn. Reson. Med. 33 (5), 656–662. https://doi.org/10.1002/ mrm.1910330510. Janke, A., Zhao, H., Cowin, G.J., Galloway, G.J., Doddrell, D.M., 2004. Use of spherical harmonic deconvolution methods to compensate for nonlinear gradient effects on MRI images. Magn. Reson. Med. 52 (1), 115–122. https://doi.org/10.1002/ mrm.20122. 8
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494 interscanner reliability, and comparison to manual segmentation. Hum. Brain Mapp. 34 (9), 2313–2329. https://doi.org/10.1002/hbm.22068. Patenaude, B., Smith, S.M., Kennedy, D.N., Jenkinson, M., 2011. A Bayesian model of shape and appearance for subcortical brain segmentation. Neuroimage 56 (3), 907–922. https://doi.org/10.1016/j.neuroimage.2011.02.046. Pipe, J.G., Zwart, N.R., Aboussouan, E.A., Robison, R.K., Devaraj, A., Johnson, K.O., 2011. A new design and rationale for 3D orthogonally oversampled k-space trajectories. Magn. Reson. Med. 66 (5), 1303–1311. https://doi.org/10.1002/mrm.22918. Sarlls, J.E., Lalonde, F., Rettmann, D., Shankaranarayanan, A., Roopchansingh, V., Talagala, S.L., 2018. Effectiveness of navigator-based prospective motion correction in MPRAGE data acquired at 3T. PLoS One 13 (6), 1–12. https://doi.org/10.1371/ journal.pone.0199372. Tustison, N.J., Avants, B.B., Cook, P.A., Zheng, Y., Egan, A., Yushkevich, P.A., Gee, J.C., 2010. N4ITK: improved N3 bias correction. IEEE Trans. Med. Imaging 29 (6), 1310–1320. https://doi.org/10.1109/TMI.2010.2046908. Watanabe, K., Kakeda, S., Igata, N., Watanabe, R., Narimatsu, H., Nozaki, A., et al., 2016. Utility of real-time prospective motion correction (PROMO) on 3D T1-weighted imaging in automated brain structure measurements. Sci. Rep. 6, 38366. https:// doi.org/10.1038/srep38366. Retrieved from. Watanabe, M., Liao, J.H., Jara, H., Sakai, O., 2013. Multispectral quantitative MR imaging of the human brain: lifetime age-related effects. RadioGraphics: Rev. Publ. Radiol. Soc. North Am. Inc 33 (5), 1305–1319. https://doi.org/10.1148/rg.335125212. White, N., Roddey, C., Shankaranarayanan, A., Han, E., Rettmann, D., Santos, J., et al., 2010. PROMO: real-time prospective motion correction in MRI using image-based tracking. Magn. Reson. Med. 63 (1), 91–105. Yarach, U., Luengviriya, C., Danishad, A., Stucht, D., Godenschweger, F., Schulze, P., Speck, O., 2015. Correction of gradient nonlinearity artifacts in prospective motion correction for 7T MRI. Magn. Reson. Med. 73 (4), 1562–1569. https://doi.org/ 10.1002/mrm.25283. Zhang, Y., Brady, M., Smith, S., 2001. Segmentation of brain MR images through a hidden Markov random field model and the expectation-maximization algorithm. IEEE Trans. Med. Imaging 20 (1), 45–57. https://doi.org/10.1109/42.906424.
Johnson, K.M., 2017. Hybrid radial-cones trajectory for accelerated MRI. Magn. Reson. Med. 77 (3), 1068–1081. https://doi.org/10.1002/mrm.26188. Kecskemeti, S., Samsonov, A., Hurley, S.A., Dean, D.C., Field, A., Alexander, A.L., 2016. MPnRAGE: a technique to simultaneously acquire hundreds of differently contrasted MPRAGE images with applications to quantitative T1mapping. Magn. Reson. Med. 75 (3) https://doi.org/10.1002/mrm.25674. Kecskemeti, Steven, Samsonov, A., Velikina, J., Field, A.S., Turski, P., Rowley, H., et al., 2018. Robust motion correction strategy for structural MRI in unsedated children demonstrated with three-dimensional radial MPnRAGE. Radiology 289 (2), 509–516. https://doi.org/10.1148/radiol.2018180180. Kuperman, J.M., Brown, T.T., Ahmadi, M.E., Erhart, M.J., White, N.S., Roddey, J.C., et al., 2011. Prospective motion correction improves diagnostic utility of pediatric MRI scans. Pediatr. Radiol. 41 (12), 1578–1582. https://doi.org/10.1007/s00247-0112205-1. Maclaren, J., Armstrong, B.S.R., Barrows, R.T., Danishad, K.A., Ernst, T., Foster, C.L., et al., 2012. Measurement and correction of microscopic head motion during magnetic resonance imaging of the brain. PLoS One 7 (11), e48088. https://doi.org/ 10.1371/journal.pone.0048088. Madhuranthakam, M.S., B, E., M, A., 2009. Azimuthal Sorting in Tandem with Elliptical Reordering (ASTER) : a new k-space reordering scheme for reduced motion sensitivity. In: International society for magnetic resonance in medicine annual meeting. Morey, R.A., Selgrade, E.S., Wagner, H. R. 2nd, Huettel, S.A., Wang, L., McCarthy, G., 2010. Scan-rescan reliability of subcortical brain volumes derived from automated segmentation. Hum. Brain Mapp. 31 (11), 1751–1762. https://doi.org/10.1002/ hbm.20973. Mugler, J.P., Brookeman, J.R., 1990. Three-dimensional magnetization-prepared rapid gradient-echo imaging (3D MP RAGE). Magn. Reson. Med. 15 (1), 152–157. https:// doi.org/10.1002/mrm.1910150117. Nugent, A.C., Luckenbaugh, D.A., Wood, S.E., Bogers, W., Zarate, C.A.J., Drevets, W.C., 2013. Automated subcortical segmentation using FIRST: test-retest reliability,
9
Contents lists available at ScienceDirect
NeuroImage journal homepage: www.elsevier.com/locate/neuroimage
Test-retest of automated segmentation with different motion correction strategies: A comparison of prospective versus retrospective methods Steven R. Kecskemeti *, Andrew L. Alexander University of Wisconsin-Madison, USA
A R T I C L E I N F O
A B S T R A C T
Keywords: MPnRAGE Motion correction PROMO Segmentation FIRST Freesurfer Volume
Test-retest of automated image segmentation algorithms (FSL FAST, FSL FIRST, and FREESURFER) are computed on magnetic resonance images from 12 unsedated children aged 9.42.6 years ([min,max] ¼ [6.5 years, 13.8 years]) using different approaches to motion correction (prospective versus retrospective). The prospective technique, PROMO MPRAGE, dynamically estimates motion using specially acquired navigator images and adjusts the remaining acquisition accordingly, whereas the retrospective technique, MPnRAGE, uses a selfnavigation property to retrospectively estimate and account for motion during image reconstruction. To increase the likelihood and range of motions, participants heads were not stabilized with padding during repeated scans. When motion was negligible both techniques had similar performance. When motion was not negligible, the automated image segmentation and anatomical labeling software tools showed the most consistent performance with the retrospectively corrected MPnRAGE technique (80% volume overlaps for 15 of 16 regions for FIRST and FREESURFER, with greater than 90% volume overlaps for 12 regions with FIRST and 11 regions with FREESURFER). Prospectively corrected MPRAGE with linear view-ordering also demonstrated lower performance than MPnRAGE without retrospective motion correction.
1. Introduction T1-weighted (T1w) structural images are acquired in nearly all neuroimaging studies. The commonly used magnetization-prepared, rapid gradient echo method, termed MPRAGE (Mugler and Brookeman, 1990), provides three-dimensional (3D), whole-brain images with exquisite anatomical contrast at high spatial resolution in all dimensions. T1w images may be used to characterize the properties of brain morphometry – volume, shape, thickness, length, etc., which have formed the basis of many neuroimaging studies. Several popular techniques are widely used for automated segmentation of T1w brain images including Freesurfer (Dale et al., 1999; Dale and Sereno, 1993; Fischl et al., 2002, 2004) and tools from FSL (FMRIB Software Library) including FAST (Zhang et al., 2001) and FIRST (Patenaude et al., 2011). MPRAGE images are sensitive to even small motions during the approximately four to eight minute acquisition. Motion artifacts in MPRAGE manifest as ghosting, ringing and blurring. In particular, brain imaging studies in young children often suffer from motion artifacts. These motion artifacts may lead to biased estimates of structural brain measurements, possibly with increased variance (Sarlls et al., 2018; Watanabe, Liao, Jara and Sakai, 2013). Consequently, there is a
significant need for more robust structural imaging methods. Prospective motion correction methods dynamically adjust the imaging gradient waveforms based upon head motion estimates. Some strategies utilize optical tracking systems to measure head motion in real time (Forman et al., 2011; Maclaren et al., 2012); however, these methods require additional hardware and are currently under development. Alternatively, motion estimation using navigator images are used for PROMO (PROspective MOtion) correction algorithms which dynamically adjust the imaging gradients in near real-time (White et al., 2010). Several studies have demonstrated the utility of PROMO to reduce MPRAGE motion artifacts in both adult (Sarlls et al., 2018; Watanabe et al., 2016) and pediatric subjects (Brown et al., 2010; Kuperman et al., 2011). Another approach for T1w imaging called MPnRAGE (Kecskemeti et al., 2016) combines inversion RF pulse preparation with 3D radial k-space sampling. This method can generate a large number (“n ~ 300 to 400”) of MPRAGE volumes with a range of inversion-recovery contrasts including a high contrast T1w image in a single scan about the length of a traditional fully sampled MPRAGE acquisition. The radial k-space sampling pattern inherently reduces the sensitivity of MPnRAGE to motion, producing blurring artifacts as opposed to ringing or ghosting artifacts
* Corresponding author. E-mail address: [email protected] (S.R. Kecskemeti). https://doi.org/10.1016/j.neuroimage.2019.116494 Received 11 September 2019; Received in revised form 22 December 2019; Accepted 23 December 2019 Available online 30 December 2019 1053-8119/© 2019 Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
can be found in (White et al., 2010).
(Glover and Pauly, 1992). Moreover, the 3D radial k-space sampling enables retrospective motion correction utilizing self-navigation during the image reconstruction (Kecskemeti et al., 2018). Motion-corrected MPnRAGE was shown to be highly effective for producing high quality T1w images in a study of 32 children (133 years) with autism and 12 age-matched (123 years) controls (Kecskemeti et al., 2018). While that study demonstrated that image quality of motion-corrected MPnRAGE was consistently highly rated by neuroradiologists and automated measures of image sharpness, it did not evaluate the effects on morphometry estimation tools like Freesurfer and FSL. This study evaluates and compares the performance of Freesurfer and the FSL FAST and FIRST tools applied to MPRAGE with PROMO and MPnRAGE both with and without retrospective motion correction. Repeated MPRAGE and MPnRAGE acquisitions were performed on a cohort of 12 children to assess the reliability of structural morphometry measures. The reliability of these automated segmentation methods are assessed both in terms of the volume sizes and the regional overlap of repeated measures.
2.2.3. MPnRAGE acquisition parameters Whole brain coverage with 1.0 mm isotropic resolution was acquired in the axial orientation with 200 mm coverage in the superior/inferior direction. Parameters included delay time TD ¼ 500 ms, TR ¼ 4.9 ms, TE ¼ 1.8 ms, n ¼ 386 views along the recovery curve. The excitation flip angles were 4 /8 for the first 325/remaining 61 radial views. The scan time was 7 min, the same time allocated for MPRAGE-PROMO. The MPnRAGE composite image was reconstructed both with and without motion correction using the summation of all data across the inversion recovery curve. 2.2.4. MPnRAGE motion correction MPnRAGE uses a quasi-random 3D radial k-space trajectory designed so that the spokes acquired within the readout window (n ¼ 386) after each magnetization preparation module are approximately uniformly angularly distributed. The spokes are rotated between readout windows so that the summation of all spokes can be used for the final MPnRAGE image reconstruction, while the spokes within each readout window (~1900 ms) can be used to form a series of 3D navigator images to estimate the amount of motion within each readout window. The original k-space data are adjusted for translations using the Fourier Shift Theorem, while rotations are accounted for by rotating the original k-space trajectory. Motion within the readout window is not corrected and will result in blurring. More details about the retrospective motion correction can be found in (Kecskemeti et al., 2018).
2. Materials and methods 2.1. Study population Imaging experiments were performed with institutional review board approval and informed consent/assent. Twelve children (9.4 2.6 years, min ¼ 6.5 years, max ¼ 13.8 years, 6 male and 6 female) without known neurological health concerns were selected for imaging. Recruitment was not based on likelihood of subjects remaining still during the scan.
2.3. Image analyses 2.2. Image acquisition One mid-axial, one off-center sagittal, and two coronal slices (anterior and midbrain) from non-corrected and motion-corrected MPnRAGE and motion-corrected MPRAGE PROMO acquisitions were presented for visual image quality scoring. Two reviewers assed the images in consensus using the following 4-point Likert scale: 1 – severe motion artifacts (severe blurring or image ghosting obscuring detection of even large WM structures and/or complete loss of WM/GM borders), 2- moderate motion artifacts (WM/GM boundaries obscured by blurring or image ghosting, but generally still detectable), 3-mild motion artifacts (some localized image blurring or ghosting detectable, but not widespread) 4 – no noticeable image blurring or ghosting artifacts detected. Reviewers were blinded to acquisition and reconstruction type and the order of all images were randomly mixed. Region-of-interest based signal to noise ratio (SNR) and WM/GM contrast measurements were performed on the subject who had highest overlaps from FSL FAST to gauge the image quality when motion artifacts were not expected to significantly degrade image quality. MPnRAGE and MPRAGE images of like modalities were co-registered within subject using flirt from the FMRIB Software Library v5.0. No registrations were performed between modalities or across subjects. Coregistered images were then processed with fsl_anat from the FMRIB Software Library v5.0 using the default settings. This included bias-field correction [FAST], brain-extraction [FNIRT-based], tissue-type segmentation [FAST], and subcortical structure segmentation [FIRST]. This yielded segmentations of cerebral white matter, gray matter and CSF and 15 subcortical structures. Co-registered images were corrected for receiver inhomogeneities using N4BiasFieldCorrection (Tustison et al., 2010) and sent through the default Freesurfer pipeline, recon-all, to segment the same 15 subcortical structures as FIRST. The default biasfield correction in recon-all was still used. Dice (or volume-)-overlap-coefficients (DOCs) for each region of interest for all possible pairs of similar scans were calculated. The intraclass correlation coefficient (ICC) was computed using the volume measurements for each region of interest. A paired-sample t-test and two-sample F-test was used to compare the mean and standard deviations of the volume measurements of each region-of-interest between the MPnRAGE
All exams took place on a 3T MRI scanner (Discovery MR750, GE Healthcare, Waukesha, WI) without the use of sedation. Participants watched a video of their choice and were instructed to remain still. The participants heads were stabilized within a 32 channel phased array head coil (Nova Medical, Wilmington, MA) using the NoMoCo pillow support system (NoMoCo Pillow, Inc., La Jolla, Ca) before receiving an MPnRAGE and MPRAGE-PROMO scan. To assess greater variability of head motions, the padding was then removed and two additional scans of each method were performed in alternating order. The order of all scans was counter-balanced across subjects. 2.2.1. MPRAGE-PROMO acquisition parameters A works in progress version of the MPRAGE-PROMO acquisition version was used that acquired k-space with a linear centric viewordering scheme. Whole brain coverage with 1.0 mm isotropic resolution was acquired using a sagittal acquisition with 192 slices and a 256 256 mm in-plane acquisition matrix. Additional parameters include TI ¼ 900 ms, TR/TE ¼ 6.952 ms/2.92 ms, bandwidth ¼ 31.25 kHz, flip angle 8 , time between magnetization preparation pulses ¼ 2488 ms, and ARC acceleration of 2 and 1.25 along the phase encode and slice directions, respectively. The acquisition time was 6 min with an additional 1 min of data acquisition allowed as needed. 2.2.2. MPRAGE-PROMO motion correction MPRAGE-PROMO utilizes the period of “free recovery” between each SPGR readout block and the next magnetization preparation module to collect a series of orthogonal-2D “navigator” images to estimate the amount of motion between successive magnetization preparation pulses (2488 ms) and then adjust the excitation and imaging gradients if a certain motion criterion is met. Motion within the readout window (Nz*TR ¼ 1335 ms) is not corrected and will result in motion induced ghost artifacts. After a complete data set is acquired, an optional rescan period (1 min in this study) continually reacquires kspace data one kyplane at a time, each time choosing the ky-plane that is most corrupt by motion as determined by the navigator motion estimates. More details 2
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
plots of the motion transformations both show nearly continuous and jittery rotations around the x-axis and translations along the z-axis. Major ghost artifacts in the MPRAGE-PROMO images posed challenges for FIRST, which assigns a small, non-hippocampus-like ROI to a region slightly posterior to the expected location. The reduced resolution of the uncorrected MPnRAGE images do not pose significant challenges for FIRST. The results of the image quality assessment are shown in Table 1. Representative images of each score for each modality are show in Fig. 3. The images show a clear distinction between the blurring artifacts of MPnRAGE and ghosting artifacts of MPRAGE. Supplemental Fig. 1, an animated GIF, illustrates the residual blurring and correction capabilities of MPnRAGE for example cases with minor, moderate, and severe motions. The supplementary materials also contain motion plots and select orthogonal views for each scan. The mean scores for MPnRAGE increased from 2.8 without motion correction to 3.6 with motion correction, while the mean score for prospectively corrected MPRAGE PROMO was 2.7. Similarly, the standard deviation of scores, which represents the fluctuation of image quality that can be expected for each method, decreased from 1.2 to 0.5 when motion correction was applied to MPnRAGE, while it was 1.0 for prospectively corrected MPRAGE PROMO. A Mann-Whitney U test rejected the null hypothesis of equal medians before and after MPnRAGE motion correction (reviewer 1: P ¼ 0.004, z ¼ 2.9, reviewer 2: P ¼ 0.006, z ¼ 2.7, mean across reviewers: P ¼ 0.01,z ¼ 2.6), and between motion corrected MPnRAGE and MPRAGE PROMO (reviewer 1: P < 0.001, z ¼ 4.1, reviewer 2: P < 0.001, z ¼ 3.7, mean across reviewers: P < 0.001,z ¼ 3.9), but could not reject the null hypothesis of equal medians between uncorrected MPnRAGE and MPRAGE PROMO (reviewer 1: P ¼ 0.6, z ¼ 0.5, reviewer 2: P ¼ 0.5, z ¼ 0.6, mean across reviewers: P ¼ 0.6,z ¼ 0.5), The results from the FSL FAST segmentations are presented in Fig. 4, and Supplemental Tables 1–2. MPnRAGE with retrospective motion correction had the highest ICCs (>0.98 all regions), mean DOCs (88% for GM and 93% for WM), and lowest variability in DOCs across subjects ( 4% for GM and 2% for WM), while MPnRAGE without motion correction performed better in all areas than MPRAGE with prospective motion correction. The boxplots of volumes show a larger spread of measurements for prospectively corrected MPRAGE, as well as the presence of at least one measurement considered to be an outlier. The mean volumes and standard deviations are statistically significant (p < 0.001) between prospectively corrected MPRAGE and retrospectively corrected MPnRAGE for CSF and GM, but only the mean standard deviation is significant for WM (p ¼ 0.196 for mean volume comparison). The differences of the average and standard deviations of the DOCs are significant between all modalities. The results from the FSL FIRST and Freesurfer segmentations are presented in Figs. 5 and 6, and Supplemental Tables 3–9. FreeSurfer was unable to complete segmentation on six prospectively corrected MPRAGE images that came from four different subjects. All analysis involving prospectively corrected MPRAGE images and FreeSurfer used only those subjects in which all images were segmented. Comparisons in FreeSurfer determined volumes between MPRAGE and MPnRAGE also used only the subset of participants with segmentations for all scans. FreeSurfer determined volume overlaps for MPnRAGE were computed using all subjects. The ICC values of the volume measurements generally increased from prospectively motion corrected MPRAGE to MPnRAGE without motion correction and to MPnRAGE with retrospective motion correction. MPnRAGE without motion correction has six ROIs from FIRST and two from FreeSurfer with ICC values greater than 0.70, while MPnRAGE with motion correction had 14 ROIs from FIRST (remaining ROI had ICC of 0.697) and 11 ROIs with ICC values greater than 0.70. Prospectively corrected MPRAGE had much lower ICC values (max of 0.51) with 11 ROIs from FIRST and 12 from FreeSurfer at below 0.25. Even with this large difference in ICC values, there were no statistical differences (at p < 0.05) in volumes between measurements made with prospectively corrected MPRAGE and MPnRAGE without correction, although there were four (FIRST determined) and six (FreeSurfer determined) regions
and MPRAGE PROMO scans. Box-and-whisker plots are used to present the distribution of volumes and volume-overlap-coefficients. In the box plots, the red line marks the median, the box-length provides the inter-quartile (25th and 75th percentile ranges), and the whisker end points represent the locations of the last non-outlier measurements. Outlier measurements are points beyond 1.5 times the box length away from the 25th and 75th percentiles. With this convention, a 99.3% coverage of points would be provided for a normal distribution. 3. Results Example images chosen from the participants who had the highest DOC averaged across WM, GM, and CSF masks from FAST are shown in Fig. 1. The DOCs were 91% for both corrected and non-motion corrected MPnRAGE scans and were 89% for prospectively corrected MPRAGE. The SNRs in the putamen and frontal white matter were 27 and 30, respectively, for MPnRAGE, and 21 and 30, respectively, for prospectively corrected MPRAGE, while the contrast between putamen and frontal WM was 1.80 for MPnRAGE and 1.89 for prospectively corrected MPRAGE, suggesting comparable image quality across methods in the absence of large motion artifacts. Neither the uncorrected or motion corrected MPnRAGE images display any obvious motion artifacts, but the second and third time points of the MPRAGE-PROMO scans show mild artifacts typically associated with motion. Example images and hippocampus masks (white ROI) chosen from the participants who had the lowest DOC averaged across both WM, GM, and CSF masks from FAST segmentations are shown in Fig. 2. The DOCs were 61% and 79% for uncorrected and retrospectively motion corrected MPnRAGE scans and was 28% for prospectively corrected MPRAGE. The
Fig. 1. MPnRAGE and prospectively motion corrected MPRAGE images from a subject that provided the highest Dice-overlap-coefficients of whole brain WM, GM, and CSF masks from FSL FAST. Image quality for all MPnRAGE scans is visibly similar to scan one of prospectively corrected MPRAGE. Presence of ringing artifacts in scan 2 and scan 3 of the prospectively corrected MPRAGE images are indicative of a mismatch between the imaging gradients and participant position within the scanner. Since there was no evidence of motion in four of the six scans in this subject and since the order of scans was counterbalanced, the presence of these artifacts could suggest an overzealous detection and correction of minor motions in the prospectively corrected MPRAGE acquisition. 3
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
Fig. 2. Example source images (far left column), including hippocampus masks from FIRST (white overlay in second column) and estimated translations and rotations (right columns) are shown from the case that had the worst Dice-overlap-coefficients of whole brain WM, GM, and CSF masks from FAST. Notice that extreme motion ghosts of MPRAGE cause FIRST to completely fail, whereas even the uncorrected MPnRAGE images can be used to extract the hippocampus with FIRST. Both subjects demonstrated nearly continuous rotations around x and translations in z. There was also a 5 mm drift during the MPnRAGE scan.
with statistical differences (at p < 0.05) between prospectively corrected MPRAGE and retrospectively corrected MPnRAGE. MPnRAGE with retrospective motion correction had 41% less variation in volumetric measurements than prospectively corrected MPRAGE within FIRST regions and 23% less in FreeSurfer regions. Across all regions,
retrospectively corrected MPnRAGE had the highest mean and lowest standard deviation (91% 6% and 90% 4% for FIRST and FreeSurfer determined regions) of the Dice-overlap-coefficients, followed by uncorrected MPnRAGE (86% 12% and 84 17%) and prospectively corrected MPRAGE (79% 28% and 83% 17%). The boxplots (Figs. 5
Fig. 3. Representative example images of each Likert score (1–4) for prospectively motion corrected MPRAGE-PROMO (Fig. 3a) and MPnRAGE (Fig. 3b). MPnRAGE images are shown with retrospective motion correction for Likert Scores 4,3,2 and without motion correction for Likert Score 1. Artifacts from motion are more subtle in MPnRAGE images (slight blurring) as opposed to replicate ghosting in MPRAGE images. Corrected MPnRAGE images may have slight image degradation due to intra-navigator motions and uneven k-space sampling that was not accounted for in the reconstruction.
4
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
and 6) show the presence of outlier DOC measurements for all methods and larger interquartile ranges for most regions in prospectively corrected MPRAGE and uncorrected MPnRAGE compared to retrospectively corrected MPnRAGE. Results from the Freesurfer cortical surface reconstructions are shown in Table 2. Ten MPRAGE PROMO and three non-motion corrected MPnRAGE datasets did not complete processing. The presence of motion artifacts (ghosting for MPRAGE PROMO and blurring for uncorrected MPnRAGE) were confirmed in all failed datasets. All 36 datasets completed processing for motion corrected MPnRAGE. When analyzed using only the largest common subset of successfully completed cortical reconstructions, the only statistical differences in mean (ttest p < 0.05) was between uncorrected and corrected MPnRAGE (p ¼ 0.026).
Table 1 The total number of cases with each Likert score for each method (36 scans x 2 reviewers ¼ 72 scores total).
Likert score 1 Likert score 2 Likert score 3 Likert Score 4
MPnRAGE w/o motion correction
MPnRAGE w/motion correction
MPRAGE w/ PROMO
13
0
11
16
1
17
13
26
25
30
45
19
Fig. 4. Volumes (top) and volume overlaps (bottom) from automated segmentations using FSL FAST. Superscripts 1,2,3 indicate statistical differences in the mean values between prospectively corrected MPRAGE and MPnRAGE without correction (1), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (2), and between uncorrected and retrospectively corrected MPnRAGE (3) at p < 0.05. Subscripts a,b,c indicate statistical differences in the standard deviations between prospectively corrected MPRAGE and MPnRAGE without correction (a), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (b), and between uncorrected and retrospectively corrected MPnRAGE (c) at p < 0.05. Points beyond the whiskers are considered outliers and are shown as red crosses (þ). 5
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
Fig. 5. Volumes (top) and volume overlaps (bottom) from automated segmentations using FSL FIRST. Superscripts 1,2,3 indicate statistical differences in the mean values between prospectively corrected MPRAGE and MPnRAGE without correction (1), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (2), and between uncorrected and retrospectively corrected MPnRAGE (3) at p < 0.05. Subscripts a,b,c indicate statistical differences in the standard deviations between prospectively corrected MPRAGE and MPnRAGE without correction (a), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (b), and between uncorrected and retrospectively corrected MPnRAGE (c) at p < 0.05. Points beyond the whiskers are considered outliers and are shown as red crosses (þ). The large red stars with zero volume and overlaps indicate that data from four participants were excluded from analysis (values of zero were not used to construct the plots) for MPRAGE with prospective motion correction since FreeSurfer would not complete the necessary segmentations.
Similarly, the only statistical differences (fftest < 0.05) in standard deviations was between uncorrected MPnRAGE and the other two methods (p ¼ 0.01 for MPRAGE PROMO and P ¼ 0.002 for corrected MPnRAGE).
scan. Likewise, since both methods also correct for motion at approximately the same rate, it is also unlikely that this was the source of differences when motion was present. Neither method, however, corrects for motion that occurs between the navigator scans. MPRAGE PROMO used a Cartesian k-space acquisition with linear view-ordering, so motion during this time will result in ghost artifacts, whereas MPnRAGE used a radial k-space acquisition that will result in blurring artifacts. It is possible that the segmentation algorithms are more sensitive to ghost artifacts than local blurring. Alternative k-space readout strategies may help to reduce motion sensitivity of MPRAGE PROMO. Incomplete, overcorrection, or unnecessarily modification of the imaging gradients with PROMO and a Cartesian k-space acquisition will also produce ghosting artifacts, but will produce blurring artifacts when a radial kspace acquisition, as in MPnRAGE, is used. Thus, without a ground truth, it is hard to determine whether suboptimal image quality of either method was a result of inaccurate motion estimates or uncorrectable artifacts from intra-navigator motions or a combination of the two. Since the PROMO acquisition in this study utilized a different k-space acquisition strategy than MPnRAGE, it is not possible to conclude if the differences observed in this study were due to the motion correction
4. Discussion When motion was negligible both MPnRAGE and MPRAGE PROMO had similar properties as characterized by SNR, CNR, and DOC of the segmented tissue masks. However, when motion was not negligible, the automated image segmentation and anatomical labeling software tools showed more consistent performance with MPnRAGE than MPRAGEPROMO. In particular, MPnRAGE with motion correction showed very high regional label consistency (80% Dice overlaps for 15 of 16 regions (min ¼ 79% with FIRST, but 84% with Freesurfer) and >90% in 12 of the regions with FIRST segmentations and 11 with Freesurfer). Conversely, prospectively corrected MPRAGE demonstrated lower performance than MPnRAGE without motion correction. The scan order was counter-balanced both within and across subjects, thus the difference in performance in the presence of motion is unlikely due to differences in the degree or type of motions experienced for each 6
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
strategy itself (prospective vs. retrospective), or the k-space acquisition strategy (Cartesian with linear view-ordering vs 3D radial), or a combination of the two. The higher test-retest scores for MPnRAGE without correction compared to MPRAGE PROMO suggests a favorability of 3D radial k-space sampling. Regardless, the two approaches are independent and compatible and the combination of the two could address defeciences in each method. In the retrospectively corrected technique, head rotations will result in uneven k-space sampling that leads to increased streak artifacts that typically manifest as additive noise. The incorporation of PROMO would prospectively rotate the k-space trajectory accordingly to keep the sampling approximately uniform. Since a 3D radial k-space trajectory is used for MPnRAGE, intra-navigator motions and inaccurate PROMO motion estimates will result in blurring as opposed to ghosting artifacts when Cartesian sampling is used. Similar combinations with other k-space trajectories that frequently- or oversample the center of k-space, such as Cones (Gurney et al., 2006), radial Cones (Johnson, 2017), rotated spiral-PR (Irarrazabal and Nishimura, 1995), FLORET (Pipe et al., 2011), and radial fan beam (Madhuranthakam et al., 2009) would likely produce improvements over the
Table 2 Cortical surface reconstructions by motion correction type. The volumes and coefficients of variation use the largest common subset of the successfully completed scans across all motion correction types. Motion correction type
MPRAGE PROMO prospective
none
retrospective
Freesurfer Cortical Reconstructions (# failed/total #) Cortical volume (105 mm3)
10/36
3/36
0/36
6.1 0.6
5.8 1.0 0.17
6.2 0.5
Cortical volume (coefficient of variation)
0.09
MPnRAGE
0.08
traditional linearly view-ordered Cartesian MPRAGE method. With a less motion sensitive k-space trajectory, the need to retake motion corrupt data may be alleviated or eliminated, as there was no data retaken or eliminated with 3D radial MPnRAGE in this study. The evaluation conditions for these methods, young children with
Fig. 6. Volumes (top) and volume overlaps (bottom) from automated segmentations using FreeSurfer. Superscripts 1,2,3 indicate statistical differences in the mean values between prospectively corrected MPRAGE and MPnRAGE without correction (1), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (2), and between uncorrected and retrospectively corrected MPnRAGE (3) at p < 0.05. Subscripts a,b,c indicate statistical differences in the standard deviations between prospectively corrected MPRAGE and MPnRAGE without correction (a), prospectively corrected MPRAGE and retrospectively corrected MPnRAGE (b), and between uncorrected and retrospectively corrected MPnRAGE (c) at p < 0.05. Points beyond the whiskers are considered outliers and are shown as red crosses (þ). The large red stars with zero volume and overlaps indicate that data from four participants were excluded from analysis (values of zero were not used to construct the plots) for MPRAGE with prospective motion correction since FreeSurfer would not complete the necessary segmentations.
7
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494
motion artifacts. The Freesurfer segmentations of the left thalamus-proper from retrospectively corrected MPnRAGE had ICC ¼ 0.479, but DOCs of 93 2%, which highlights the importance of considering both measures as markers of reliability. The results of the total cortical volume analysis indicate that when datasets with strong motion artifacts are excluded from analysis, no statistical differences in total cortical volumes exist. In summary, we found that MPnRAGE with retrospective motion correction provided highly reliable automated segmentations of regions within the human brain. These techniques are effective for volumetric studies in challenging populations including young children.
and without head stabilization padding, may be more challenging than many clinical or research scans. Thus, the degrees of motion are likely larger than encountered in many neuroimaging studies. Further, the scanner manufacturer strongly recommends the use of stabilizing padding for applications with MPRAGE PROMO. It is highly likely that the results for both methods would improve with the addition of stabilization pads and/or for more compliant subjects where motion is expected to be less severe, but that was not evaluated in this study. This study set the total allotted scan time to be 7 min for each method, which meant that MPRAGE-PROMO had a maximum rescan period of 1 min (6 min basescan þ 1 min rescan). A similar study of 9 pediatric subjects (Brown et al., 2010) aged 10.73 yrs 0.54 yrs used an unlimited rescan period and found that 8 of the subjects required less than a 1 min rescan (14.3s 18.7s). The remaining subject needed nearly 5 min of rescanning. It is possible that lengthening the rescan period for PROMO will produce better results. However Supplemental Fig. 2 shows that participant behavior during the rescan period of this study was consistent with behavior during the acquisition, so rescanning becomes less efficient the more often somebody moves, which is precisely when it is needed the most. Based on each participant’s motion, a 5 min rescan period would be sufficient for 30 of 36 scans, assuming the additional scan length does not produce less compliance. The remaining six scans (consisting of scans from 5 unique participants) are estimated to need at least 10 min (10 min, 13 min, 19 min, 32 min, 37 min, and 41 min respectively) to satisfy a maximum motion metric of 0.7 mm used in PROMO. Increasing the scan time for MPnRAGE could result in higher SNR, allow retrospective rejection of data (currently all data is used irregardless of motion metric), and increased spatial resolution. Traditionally, gradient non-linearity correction is performed as a single post-processing inverse-warping procedure where the amount of warping depends on the position of each voxel with respect to iso-center (Janke et al., 2004). When motion occurs, the final images represent the superposition of multiple positions so straightforward correction is no longer possible. However, the amount of warping generally varies slowly with position. We have estimated that the maximal change in the degree of warping resulting from a 10 rotation around the S/I axis is less than 0.1 mm for a head scan on our system. Thus, localization errors when determining the gradient warps are not expected to cause significant degradation of image quality. On systems where gradient nonlinearity is more problematic, non-linear gradient correction could be performed on small subsets of the data using positional estimates from the navigators (PROMO or MPnRAGE navigators) to help determine more accurate positional locations. A similar procedure was presented in for motion correction scans at 7T (Yarach et al., 2015). Previous studies of test-retest reliability, interscanner reliability, and comparison to manual segmentations using FIRST with MPRAGE without PROMO were previously performed (Nugent et al., 2013) in adult subjects and demonstrated reliable segmentations (high intraclass correlation coefficients (ICC) > 0.83, consistent volumes, and high Dice-overlap-coefficients (DOC)) of the thalamus, caudate, putamen, hippocampus, and pallidum (marginal, but acceptable) for repeat scans from the same scanner platform, but unreliable segmentations of the accumbens and amygdala. The less reliable segmentations of the accumbens and amygdala were also demonstrated in an earlier test-retest study of FIRST and Freesurfer (Morey et al., 2010), which also suggested that the reduced reliability of the accumbens and amygdala may be due to its smaller size, which makes it more susceptible to measurement error. In this study, retrospectively motion corrected MPnRAGE demonstrated acceptable segmentations in terms of volumes (ICCs > 0.7) and shapes (DOCs) for all regions except the right accumbens regions (FIRST, ICC ¼ 0.697), the left thalamus-proper (Freesurfer, ICC ¼ 0.479), left and right pallidum (Freesurfer, ICC ¼ 0.554 and 0.6917), left amygdala (Freesurfer, ICC ¼ 0.606). The amygdala and accumbens regions have also posed difficult in previous test-retest studies (Morey et al., 2010; Nugent et al., 2013) using adult subjects and images without obvious
Author contributions section Steven Kecskemeti: Conceptualization, Methodology, Software, Validation, Formal Analysis, Investigation, Data Curation, Writing -Original Draft & Editing, Visualization, Project administration, Funding acquisition. Andrew Alexander: Conceptualization, Methodology, Formal Analysis, Resources, Writing- Original Draft and Editing, Visualization, Supervision Acknowledgements Funding from the Department of Radiology, University of Wisconsin – Madison was received. The data collection from this study was supported by GE Healthcare and the UW-Madison Department of Radiology. The methods used in the study were supported in part by a core grant to the Waisman Center from the National Institute of Child Health and Human Development (IDDRC U54 HD090256). Additional support was provided by National Institutes of Health grants DA050258, MH097464, MH100031, MH101504, HD094715, AG051216, and AG15001. Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi. org/10.1016/j.neuroimage.2019.116494. References Brown, T.T., Kuperman, J.M., Erhart, M., White, N.S., Roddey, J.C., Shankaranarayanan, A., et al., 2010. Prospective motion correction of high-resolution magnetic resonance imaging data in children. Neuroimage 53 (1), 139–145. https:// doi.org/10.1016/j.neuroimage.2010.06.017. Dale, A.M., Fischl, B., Sereno, M.I., 1999. Cortical surface-based analysis. I. Segmentation and surface reconstruction. Neuroimage 9 (2), 179–194. https://doi.org/10.1006/ nimg.1998.0395. Dale, A.M., Sereno, M.I., 1993. Improved localizadon of cortical activity by combining EEG and MEG with MRI cortical surface reconstruction: a linear approach. J. Cogn. Neurosci. 5 (2), 162–176. https://doi.org/10.1162/jocn.1993.5.2.162. Fischl, B., Salat, D.H., Busa, E., Albert, M., Dieterich, M., Haselgrove, C., et al., 2002. Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain. Neuron 33 (3), 341–355. Fischl, B., Salat, D.H., van der Kouwe, A.J.W., Makris, N., Segonne, F., Quinn, B.T., Dale, A.M., 2004. Sequence-independent segmentation of magnetic resonance images. Neuroimage 23 (Suppl. 1), S69–S84. https://doi.org/10.1016/ j.neuroimage.2004.07.016. Forman, C., Aksoy, M., Hornegger, J., Bammer, R., 2011. Self-encoded marker for optical prospective head motion correction in MRI. Med. Image Anal. 15 (5), 708–719. https://doi.org/10.1016/j.media.2011.05.018. Glover, G.H., Pauly, J.M., 1992. Projection reconstruction techniques for reduction of motion effects in MRI. Magn. Reson. Med. 28 (2), 275–289. https://doi.org/10.1002/ mrm.1910280209. Gurney, P.T., Hargreaves, B.A., Nishimura, D.G., 2006. Design and analysis of a practical 3D cones trajectory. Magn. Reson. Med. 55 (3), 575–582. https://doi.org/10.1002/ mrm.20796. Irarrazabal, P., Nishimura, D.G., 1995. Fast three dimensional magnetic resonance imaging. Magn. Reson. Med. 33 (5), 656–662. https://doi.org/10.1002/ mrm.1910330510. Janke, A., Zhao, H., Cowin, G.J., Galloway, G.J., Doddrell, D.M., 2004. Use of spherical harmonic deconvolution methods to compensate for nonlinear gradient effects on MRI images. Magn. Reson. Med. 52 (1), 115–122. https://doi.org/10.1002/ mrm.20122. 8
S.R. Kecskemeti, A.L. Alexander
NeuroImage 209 (2020) 116494 interscanner reliability, and comparison to manual segmentation. Hum. Brain Mapp. 34 (9), 2313–2329. https://doi.org/10.1002/hbm.22068. Patenaude, B., Smith, S.M., Kennedy, D.N., Jenkinson, M., 2011. A Bayesian model of shape and appearance for subcortical brain segmentation. Neuroimage 56 (3), 907–922. https://doi.org/10.1016/j.neuroimage.2011.02.046. Pipe, J.G., Zwart, N.R., Aboussouan, E.A., Robison, R.K., Devaraj, A., Johnson, K.O., 2011. A new design and rationale for 3D orthogonally oversampled k-space trajectories. Magn. Reson. Med. 66 (5), 1303–1311. https://doi.org/10.1002/mrm.22918. Sarlls, J.E., Lalonde, F., Rettmann, D., Shankaranarayanan, A., Roopchansingh, V., Talagala, S.L., 2018. Effectiveness of navigator-based prospective motion correction in MPRAGE data acquired at 3T. PLoS One 13 (6), 1–12. https://doi.org/10.1371/ journal.pone.0199372. Tustison, N.J., Avants, B.B., Cook, P.A., Zheng, Y., Egan, A., Yushkevich, P.A., Gee, J.C., 2010. N4ITK: improved N3 bias correction. IEEE Trans. Med. Imaging 29 (6), 1310–1320. https://doi.org/10.1109/TMI.2010.2046908. Watanabe, K., Kakeda, S., Igata, N., Watanabe, R., Narimatsu, H., Nozaki, A., et al., 2016. Utility of real-time prospective motion correction (PROMO) on 3D T1-weighted imaging in automated brain structure measurements. Sci. Rep. 6, 38366. https:// doi.org/10.1038/srep38366. Retrieved from. Watanabe, M., Liao, J.H., Jara, H., Sakai, O., 2013. Multispectral quantitative MR imaging of the human brain: lifetime age-related effects. RadioGraphics: Rev. Publ. Radiol. Soc. North Am. Inc 33 (5), 1305–1319. https://doi.org/10.1148/rg.335125212. White, N., Roddey, C., Shankaranarayanan, A., Han, E., Rettmann, D., Santos, J., et al., 2010. PROMO: real-time prospective motion correction in MRI using image-based tracking. Magn. Reson. Med. 63 (1), 91–105. Yarach, U., Luengviriya, C., Danishad, A., Stucht, D., Godenschweger, F., Schulze, P., Speck, O., 2015. Correction of gradient nonlinearity artifacts in prospective motion correction for 7T MRI. Magn. Reson. Med. 73 (4), 1562–1569. https://doi.org/ 10.1002/mrm.25283. Zhang, Y., Brady, M., Smith, S., 2001. Segmentation of brain MR images through a hidden Markov random field model and the expectation-maximization algorithm. IEEE Trans. Med. Imaging 20 (1), 45–57. https://doi.org/10.1109/42.906424.
Johnson, K.M., 2017. Hybrid radial-cones trajectory for accelerated MRI. Magn. Reson. Med. 77 (3), 1068–1081. https://doi.org/10.1002/mrm.26188. Kecskemeti, S., Samsonov, A., Hurley, S.A., Dean, D.C., Field, A., Alexander, A.L., 2016. MPnRAGE: a technique to simultaneously acquire hundreds of differently contrasted MPRAGE images with applications to quantitative T1mapping. Magn. Reson. Med. 75 (3) https://doi.org/10.1002/mrm.25674. Kecskemeti, Steven, Samsonov, A., Velikina, J., Field, A.S., Turski, P., Rowley, H., et al., 2018. Robust motion correction strategy for structural MRI in unsedated children demonstrated with three-dimensional radial MPnRAGE. Radiology 289 (2), 509–516. https://doi.org/10.1148/radiol.2018180180. Kuperman, J.M., Brown, T.T., Ahmadi, M.E., Erhart, M.J., White, N.S., Roddey, J.C., et al., 2011. Prospective motion correction improves diagnostic utility of pediatric MRI scans. Pediatr. Radiol. 41 (12), 1578–1582. https://doi.org/10.1007/s00247-0112205-1. Maclaren, J., Armstrong, B.S.R., Barrows, R.T., Danishad, K.A., Ernst, T., Foster, C.L., et al., 2012. Measurement and correction of microscopic head motion during magnetic resonance imaging of the brain. PLoS One 7 (11), e48088. https://doi.org/ 10.1371/journal.pone.0048088. Madhuranthakam, M.S., B, E., M, A., 2009. Azimuthal Sorting in Tandem with Elliptical Reordering (ASTER) : a new k-space reordering scheme for reduced motion sensitivity. In: International society for magnetic resonance in medicine annual meeting. Morey, R.A., Selgrade, E.S., Wagner, H. R. 2nd, Huettel, S.A., Wang, L., McCarthy, G., 2010. Scan-rescan reliability of subcortical brain volumes derived from automated segmentation. Hum. Brain Mapp. 31 (11), 1751–1762. https://doi.org/10.1002/ hbm.20973. Mugler, J.P., Brookeman, J.R., 1990. Three-dimensional magnetization-prepared rapid gradient-echo imaging (3D MP RAGE). Magn. Reson. Med. 15 (1), 152–157. https:// doi.org/10.1002/mrm.1910150117. Nugent, A.C., Luckenbaugh, D.A., Wood, S.E., Bogers, W., Zarate, C.A.J., Drevets, W.C., 2013. Automated subcortical segmentation using FIRST: test-retest reliability,
9