Evaluation of EPI distortion correction methods for quantitative MRI of the brain at high magnetic field Xin Hong, Xuan Vinh To, Irvin Teh, Jian Rui Soh, Kai-Hsiang Chuang PII: DOI: Reference:
S0730-725X(15)00155-1 doi: 10.1016/j.mri.2015.06.010 MRI 8375
To appear in:
Magnetic Resonance Imaging
Received date: Accepted date:
12 May 2015 20 June 2015
Please cite this article as: Hong Xin, To Xuan Vinh, Teh Irvin, Soh Jian Rui, Chuang Kai-Hsiang, Evaluation of EPI distortion correction methods for quantitative MRI of the brain at high magnetic field, Magnetic Resonance Imaging (2015), doi: 10.1016/j.mri.2015.06.010
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
T
Evaluation of EPI distortion correction methods for quantitative MRI of the brain at high magnetic field
RI P
Xin Honga, Xuan Vinh Toa, Irvin Tehb,1, Jian Rui Soha, and Kai-Hsiang Chuanga,b,c * a
ED
MA
NU
SC
Magnetic Resonance Imaging Group, Singapore Bioimaging Consortium Agency for Science Technology and Research 11 Biopolis Way, #01-02 Helios Building Singapore 138667 b Clinical Imaging Research Centre, National University of Singapore 14 Medical Drive, #B1-01 Singapore 117599 c Department of Physiology, Yong Loo Lin School of Medicine National University of Singapore Block MD9, 2 Medical Drive #04-01 Singapore 117597
AC
CE
PT
* corresponding author Kai-Hsiang Chuang, PhD Singapore Bioimaging Consortium, 11 Biopolis Way, #02-02, Singapore 138667. Tel +65 64788764; Fax +65 64789957; email: [email protected]
1
Current affiliation: Department of Cardiovascular Medicine, Radcliffe Department of Medicine, University of Oxford. Address: Wellcome Trust Centre for Human Genetics, Roosevelt Drive, Oxford, OX3 7BN, United Kingdom
1
ACCEPTED MANUSCRIPT Abstract
MA
NU
SC
RI P
T
High field MRI has been applied to high-resolution structural and functional imaging of the brain. Echo Planar Imaging (EPI) is an ultrafast acquisition technique widely used in diffusion imaging, functional MRI and perfusion imaging. However, it suffers from geometric and intensity distortions caused by static magnetic field inhomogeneity, which is worse at higher field strengths. Such susceptibility artifacts are particularly severe in relation to the small size of the mouse brain. In this study we compared different distortion correction methods, including nonlinear registration, field map-based, and reversed phase encoding-based approaches, on quantitative imaging of T1 and perfusion in the mouse brain acquired by spin-echo EPI with inversion recovery and pseudo-continuous arterial spin labeling, respectively, at 7T. Our results showed that the 3D reversed phase encoding correction outperformed other methods in terms of geometric fidelity, and that conventional field map-based correction could be improved by combination with affine transformation to reduce the bias in the field map. Both methods improved quantification with smaller fitting error and regional variation. These approaches offer robust correction of EPI distortions at high field strengths and hence could lead to more accurate co-registration and quantification of imaging biomarkers in both clinical and preclinical applications.
ED
Keywords: field inhomogeneity, distortion correction, susceptibility artifact, transgenic mouse,
AC
CE
PT
cerebral blood flow, high magnetic field
2
ACCEPTED MANUSCRIPT 1. Introduction
ED
MA
NU
SC
RI P
T
High field (>= 7T) MRI has become widely adopted for high-resolution structural and functional imaging of the brain [1-4]. However, image artifacts, such as susceptibility artifact, are the limiting factors that compromise the signal-to-noise ratio (SNR) benefit of high fields. Among the image acquisition pulse sequences, Echo Planar Imaging (EPI) is an ultrafast sequence widely applied to functional MRI (fMRI), diffusion imaging and perfusion imaging. It suffers from geometric and intensity distortions as well as signal dropouts that are caused by static magnetic field inhomogeneity [5-6], particularly in the phase encoding direction due to the relatively low pixel bandwidth or in the through plane direction due to the typically lower resolution than inplane. The resultant global and regional distortions are typically seen near the air-tissue or bone-tissue interfaces, where magnetic susceptibility changes rapidly. The issue is especially severe when translating similar methods to transgenic mouse models to understand disease mechanisms and drug effects. Firstly, the mouse brain is much smaller than the human brain. The same spatial extent of spin displacement affects a much larger portion of the mouse brain. Secondly, field inhomogeneity and hence distortions are proportional to the magnetic field strength. With rodent imaging typically performed at higher magnetic field strengths than human imaging, the artifacts worsen. The geometric distortion will result in localization errors, difficulty in image registration and hence group analysis. The intensity distortion in EPI could also bias the quantitative measures such as relaxivity, diffusivity and perfusion, since quantification of these metrics usually involves nonlinear fitting of the EPI series.
AC
CE
PT
One could reduce the signal dropouts commonly seen in gradient echo imaging by incorporating slice shimming, also known as the z-shim technique, into the pulse sequence [7-9]. However this will prolong the image acquisition time depending on how many z-shim steps used. Spin-echo based acquisition, which is less sensitive to field inhomogeneity, could help to reduce the distortion. However, even with spin-echo EPI (SE-EPI), the geometric and intensity distortions are still significant at high field strengths. The artifact is also echo time (TE) dependent; however, ultrashort TE is difficult to achieve with EPI. Another way is to reduce the echo train length by increasing the bandwidth or using partial Fourier, at the cost of lower SNR and/or spatial resolution. Echo train length can also be reduced by multi-shot EPI, but the acquisition time will be longer and the phase error between shots would introduce additional artifacts. Applying localized susceptibility matching materials near the air/bone-tissue interfaces can reduce the regional magnetic inhomogeneity, though it will be difficult for regions away from the skin [10]. Alternatively, postprocessing methods can be applied to reduce the distortions. Several EPI distortion correction algorithms have been proposed and demonstrated successfully in human studies at lower fields. The most common way is based on magnetic field mapping. A field map, measured by the phase differences of two gradient echo images acquired with different TEs, describes the spatial variation of the magnetic field, from which a voxel displacement map can be calculated and used to unwarp the distorted EPI images [11-12]. The main drawback of this technique is the separate acquisition of the multiecho images. Subject movement between the field mapping and EPI acquisition could render the field map inaccurate. In addition, the phase unwrapping process is sensitive to noise; therefore the field map in regions of low SNR is unreliable. Alternatively, one can acquire a phase
3
ACCEPTED MANUSCRIPT encoded multi-reference scan to estimate the magnitude and phase errors due to the field inhomogeneity, and then use that to correct the EPI artifact [13].
SC
RI P
T
Geometric distortions can also be corrected for by nonlinear registration to an anatomically correct image, typically acquired by fast spin echo (FSE). The deformation field could be calculated by minimizing the least squared differences of intensity [14] or log-intensity [15] between the distorted EPI and anatomical images. The undistortion performance highly depends on the registration algorithm, implementation, and parameter optimization. Besides, it only corrects for geometric distortion but not the intensity distortion, which may still result in quantification error.
CE
PT
ED
MA
NU
Another approach is to acquire additional EPI data with reversal of the phase encoding direction [16-18]. The basic idea is that phase encoding gradients with opposite polarities will produce opposite spatial and intensity distortions in the phase encoding direction. Therefore an anatomically correct image should be the one with minimal difference between the pair of distorted images. Once the displacement for each voxel is determined the intensity can be corrected accordingly. One advantage of this approach is the short imaging time, which is on the scale of seconds for a full EPI volume compared to minutes for a conventional field map acquisition, and hence makes it less sensitive to subject motion. However, since the algorithm is unwarping on a single dimension, the resulted image may show discontinuities in the other two dimensions. An improved implementation is to generate a continuous and smooth 3D displacement field [19]. Further improvement uses computationally more efficient method, achieving subvoxel resolution of the distortion map [20]; however, the nonlinear optimization requires empirically determined regularization parameters. Distortion correction by the reversed phase encoding approach has been demonstrated in diffusion imaging [21-23], fMRI [23] and dynamic susceptibility contrast MRI [24] of the human brain at 3T.
AC
Whether the abovementioned postprocessing approaches can effectively correct for the distortions at high fields and how they affect quantitative imaging like arterial spin labeling are not clear. In this study we compared the effectiveness of different distortion correction methods on SE-EPI of the mouse brain acquired at 7T. Their influence on quantitative T1 and cerebral blood flow (CBF), measured using inversion recovery and pseudo-continuous arterial spin labeling (pCASL) respectively, were assessed.
2. Material and Method Animal preparation The animal study was approved by the Institutional Animal Care and Use Committee (Biomedical Sciences Institutes, Singapore). Eight mice of FVB (n=4, male) and C57BL/6 (n=4, male) backgrounds, aged between 6 and 12 months were used. They were anesthetized and maintained by 1-2% isoflurane in air and O2 (1:1) at 1L/min flow rate via a nose cone, with respiration rate controlled at 100±20 breaths per minute. The head was secured on an MRI-compatible stereotaxic holder. The body temperature and respiration were monitored by an MRI-compatible monitoring system (Small Animal Instruments, Inc, Stony Brook, NY, USA) with temperature maintained at 36±0.5oC using heated air. 2.1. Image acquisition 4
ACCEPTED MANUSCRIPT
ED
MA
NU
SC
RI P
T
MRI was acquired on a 7T scanner (ClinScan, Bruker BioSpin, Ettlingen, Germany) using a 72mm volume transmit coil and a mouse brain 4-channel receive array (Bruker). The brain was carefully positioned at the center of the magnet. Volume selective 2nd-order 3D shimming was carried out to optimize the field homogeneity of the whole brain. T2-weighted (T2w) anatomical images were acquired with a 2D multislice FSE sequence with TR/TE=2760/43ms, 0.1x0.1x0.3mm3 voxel resolution, average=2, number of slices=51, and coil inhomogeneity normalization. T1 mapping was obtained by inversion recovery singleshot SE-EPI with 7 inversion times (TIs) ranging from 10 to 8000 ms, TR/TE=10000/20ms, bandwidth=3552Hz/pixel, average=2, FOV=18x13.5mm, acquisition matrix=64x48, slice thickness=1mm, gap=0mm, and number of slices=12. The images were reconstructed with zero-filling to a matrix of 128x96, resulting in apparent resolution of 0.14x0.14mm2. Perfusion imaging was acquired by pCASL [25-26] with labeling offset=12mm, labeling duration=1600ms, 10 post-labeling delays (PLDs) ranging from 0 to 400ms and SE-EPI with acquisition parameters the same as the T1 mapping sequence except TR=4000ms and 30 measurements, i.e. 15 control and label pairs. Reversed phase encoding SE-EPI was acquired at PLD=0ms for pCASL and each TI of the T1 mapping. The field map was measured by a dual gradient echo sequence with TE=1.8ms and 4.85ms, TR=213ms, flip angle=20 degrees, acquisition matrix=128x96 without interpolation, and the same geometry and reconstructed voxel size as the SE-EPI. The acquisition time, including both the forward and reversed phase EPI and dummy scans, was about 8 min for the T1 mapping and 24 min for the pCASL imaging. 2.2. Distortion correction
CE
PT
The T2w FSE image was rigidly registered to the magnitude image of the field mapping sequence to create a geometrically and anatomically correct gold standard in the EPI space. Five correction methods were applied on the EPI data:
AC
1) nonlinear registration: the EPI image was registered to the gold standard FSE image using FNIRT [27] in FSL (FMRIB Software Library v5.0, University of Oxford, Oxford, UK, http://fsl.fmrib.ox.ac.uk/) with sum of squared difference as the cost function; 2) field map-based correction: the EPI was corrected by FSL FUGUE [12, 27] using a field map created by the phase images of the dual echo sequence followed by unwrapping by FSL PRELUDE; 3) optimized field map-based method: to account for the residual shearing after the field map-based correction, the undistorted EPI image corrected as above was further registered to the gold standard image using 3D affine transformation with correlation ratio as the cost function; 4) 1D reversed phase encoding correction: in-house Matlab codes (MathWorks, Natick, MA, USA) implemented based on the algorithm of Chang et al. [16] was applied on brain-extracted EPI data; and 5) 3D reversed phase encoding correction: FSL TOPUP [19, 28], which generated 3D distortion field on EPI data without brain extraction, was used. Since there were displacements between the field map and the EPI data, the field map-based correction was initially performed on the first and last acquired volumes of the pCASL series, and the one with
5
ACCEPTED MANUSCRIPT
RI P
T
better correction (in terms of higher Jaccard index) was chosen to be the reference volume. Then all the other pCASL and TI series were registered to that reference EPI volume using rigid body transformation. In the reversed phase encoding correction, each TI image was undistorted by its corresponding reversed phase encoding image, and then registered to the first pCASL volume with PLD=0ms. For pCASL images, the distortion field was calculated based on the forward and reversed phase encoding image pair with PLD=0ms, and then applied to all other volumes with different PLDs. 2.3. Data analysis
E S E S i
i
i
i
i
NU
JI E , S
SC
The goodness of geometric correction was assessed quantitatively by an overlap measure, Jaccard index (JI):
i
(1)
PT
ED
MA
where S and E are binary matrices of the brain masks of the gold standard structural image and the EPI image, respectively, and i is the voxel index within each volume. The brain masks of the structural images were created using the automatic 3D-Pulse-coupled Neural Networks (3D-PCNN) algorithm [29] with additional manual editing to correct residual errors. The brain masks of EPI data were drawn manually. In addition to the Jaccard index, the performance of the correction methods was also evaluated by image similarity metrics, including intensity-based cross correlation (CC): 2
Var X Var Y
CE
CC X , Y
Cov X , Y
(2)
and histogram-based mutual information (MI):
AC
MI X , Y H X H Y H X , Y
(3)
where X and Y are the gold standard structural image and the EPI image, respectively, Var(X) is the variance of X, Cov(X,Y) is the covariance of X and Y, H(X) is the Shannon entropy of X, and H(X,Y) is the joint entropy of X and Y. The Fieldman test was performed using Prism (version 5, GraphPad Software, San Diego, CA, USA) to compare the similarity metrics between methods with family-wise error rate controlled. For the T1 mapping and CBF quantification, distortions were corrected for by both the optimized field map-based method and the 3D reversed phase encoding method. The original and undistorted M0 and T1 maps were calculated using nonlinear least square 3-parameter fitting to the inversion recovery signal. The CBF was quantified by nonlinear least square fitting to the multi-PLD data using the following equation [30]:
6
ACCEPTED MANUSCRIPT 0, 0 t t ' CBF M t e t /T1b / 2 M 0 BT1' 1 e t t /T1 , t t t ' M t et /T1b e t t /T1 / 2 M T ' 1 e /T1' , t t 0B 1
RI P
(4)
T
NU
SC
where ∆M(t) is the tissue magnetization difference between the control and label images during the interval t, M0B is the equilibrium magnetization of arterial blood assumed to be the same as the tissue M0 obtained from the T1 mapping, α is the labeling efficiency assumed to be 0.9, λ is the tissue/blood partition coefficient of water assumed to be 0.9ml/g [31], ∆t is arterial transit time to be fitted, τ is the labeling duration (1600ms in our sequence), T'1 is the apparent relaxation time measured by the T1 mapping, T1b is the longitudinal relaxation time of arterial blood assumed to be 2300ms [32]. The goodness of fitting for CBF quantification was evaluated by the coefficient of determination, R2.
CV / .
(5)
AC
CE
PT
ED
MA
The influence of undistortion on quantification was evaluated in five regions of interest (ROIs) with minimal to severe distortions. The ROIs were manually delineated on the original EPI image and in the corresponding locations on the undistorted EPI images. The ROIs from minimally distorted regions included the right ectorhinal cortex (Ect) and the right thalamus (Thal); from substantially compressed regions in the cortex like the right primary somatosensory cortex barrel field (S1BF); from moderately stretched regions like the left amygdala (Amyg); and from severely compressed regions like the right entorhinal cortex (Ent). The mean M0, T1, CBF and R2 values within each ROI were compared between the original and undistorted maps. The R2 of the corrected images were subtracted by and then normalized to the R2 of the original image. To evaluate the variation within each ROI, coefficient of variance (CV) was calculated as the ratio of the standard deviation over the mean:
Pair-wise comparisons were performed by two sample t-tests with the significance level of 0.05 adjusted by the Bonferroni method.
Results Fig.1 shows a typical single-shot spin-echo EPI of a mouse brain acquired at 7T and the undistorted images by the five correction methods. The original SE-EPI showed mild compression in cortical areas (purple arrows), moderate stretching in the amygdala (yellow arrows), and severe compression in the entorhinal cortex (blue and orange arrows). Nonlinear registration and both 1D and 3D reversed phase encoding corrections were able to correct for the cortical distortion. However, nonlinear registration failed to correct for intensity distortion due to voxel compression (red arrows). Field map-based method compensated distortions partly, while in-plane shearing in the corrected images was seen consistently (green arrows). An additional affine registration to the reference image improved the overall correction of the field map-based method. Based on our preliminary results, 1D reversed phase correction performs better in brain extracted EPI data than non-extracted ones. Discontinuity at the brain
7
ACCEPTED MANUSCRIPT
RI P
T
boundary in the phase-encoding (dorsal-ventral) direction in the images corrected by the 1D reversed phase encoding method could be noticed (white arrows), which was caused by optimizing displacement field within the brain by removing the non-brain tissues, but that did not affect the following quantitative evaluation. The 3D reversed phase encoding method, without brain extraction, was able to correct for both geometric and intensity distortions even in regions near the ear canals where other methods failed.
MA
NU
SC
Since the distortion is region-dependent, the distortion and effectiveness of each correction method were evaluated by the Jaccard index at each slice position (Fig. 2). Large distortion could be seen in posterior (caudal) brain, which may be due to susceptibility changes near ear canals and poorer shimming away from the iso-center. Less distortion was seen in the middle of cerebrum but worsened towards the nose. Even though distortion correction methods could improve the quality, the improvement depended on the original image quality so that the Jaccard indexes after correction still followed similar slice-dependent trend. Both reversed phase encoding methods provided the most effective and robust performance across most of the slice locations, followed by the optimized field map-based method.
PT
ED
The mean Jaccard index, cross correlation and mutual information of the whole brain were compared to evaluate the performance of the correction methods (Fig. 3). Nonlinear registration achieved significant undistortion only in terms of MI. Conventional field map-based method showed lowest JI, CC, MI among all, with some improvement when combined with affine registration. Both reversed phase encoding methods were effective with significant increase in all similarity metrics.
AC
CE
Since the optimized field map-based and 3D reversed phase encoding methods were the more effective approaches, their influence on T1 and CBF quantification was compared. Fig.4 shows examples of T1 and CBF maps from one mouse estimated before and after undistortion. It was clear that the 3D reversed phase encoding correction generated higher geometric fidelity and tissue uniformity in the quantitative maps. The improvements in T1, M0 and CBF quantification were compared in various ROIs (Fig. 5). The locations of the five ROIs defined on the original and the undistorted EPI were shown in Fig. 5a and 5b, respectively. Distortions were typically seen in the amygdala and Ent, less in S1BF and minimal in Ect, and Thalamus. Therefore the latter 2 ROIs were used as reference areas not affected by distortion. In the Ect, where no distortion was seen, no differences in T1, M0, CBF, and R2 were detected after correction, nor between the two correction methods (Fig. 5c, e, g, i). In the S1BF, where small distortion was observed, the only difference seen was in the R2 where the 3D reversed phase encoding correction achieved better fitting than the optimized field map correction. In the amygdala, where large distortion was easily visible, the 3D reversed phase encoding correction resulted in higher M0 and R2, while the optimized field map correction produced lower T1 and higher CBF than the original. In the Ent, where the distortion was most significant, the 3D reversed phase correction showed improved T1 and CBF toward the typical gray matter values seen in Ect and S1BF, while both methods had better model fitting with more than 100% increase in R2. In the thalamus, T1 and M0 were higher after the field map correction, while no difference was found in R2 after either correction. Although the T1 and CBF may not
8
ACCEPTED MANUSCRIPT
RI P
T
necessarily change after correction (except in highly distorted regions like the Ent), the increased R2 in most regions indicated that the quality of curve fitting was improved. The significantly reduced coefficient of variance in all four quantification values in the Ent further confirmed improved quantification in severely distorted regions.
4. Discussion
AC
CE
PT
ED
MA
NU
SC
Geometric and intensity distortions as well as signal dropouts have hampered the application of EPI for structural and functional imaging at high fields. Compared to other EPI distortion correction methods tested, the 3D reversed phase encoding method provides the most significant improvements in both geometry and intensity accuracy. Its performance at the boundary is superior to that of the 1D reversed phase encoding method due to the better estimate of the 3D displacement profile. Undistortion by nonlinear registration does not require additional data other than the anatomical image, but the results highly depend on the fine-tuning of the registration parameters, which is subjective and time-consuming. In this study the registration parameters, including the cost function, masking, sub-sampling levels, iteration number, regularization, etc, were optimized based on our pilot study. Besides, nonlinear registration does not correct for intensity distortion, which is obvious in areas of voxel compression near the ear canals. We observed that the widely used field map-based correction did not perform well. This may be due to errors in phase unwrapping, which is especially challenging in areas with large susceptibility differences and hence rapid field changes, e.g. at the boundary of brain tissues and the airfilled ear canals. Also, residual phase error due to asymmetric bipolar readout in the dual echo sequence may contribute to errors in the field map. Zero and first order phase shift in the dual echo sequence may be modeled, estimated, and hence corrected for, but it requires collection of additional phantom data [33]. Our results suggest that a simple additional step of affine registration could improve the field mapbased correction. The T1 values estimated based on the 3D reversed phase encoding corrected images are in good agreement to those reported in the previous study at 7T [34]. The regional CBF values estimated here are comparable to those reported by Kober et al. [35], while higher than other studies [36-38]. Absolute CBF quantification is highly affected by various physiological and methodological factors, including the animal's strain, age, physiological state, anesthesia, labeling technique, readout method, and assumptions in the quantification. Direct comparison of the absolute CBF value between studies may not be conclusive without careful controlling of these factors. In this study we used areas with minimal distortion as the reference. The quantified CBF after undistortion is consistent with these reference areas. It should be noted that there is no absolute gold standard against which the performance of undistortion is evaluated. However, given the high anatomical fidelity of the T2-weighted image and high geometric fidelity of the magnitude image of the multi-echo gradient echo data, the reference generated by a rigid registration of the former to the later gives a reasonable gold standard with the same geometric settings as the EPI images.
9
ACCEPTED MANUSCRIPT
MA
NU
SC
RI P
T
Most of the previous works on EPI distortion correction only evaluated the performance by visual inspection. In this study, quantitative assessment of the correction was determined regionally in each slice location and globally in the whole brain. However, the evaluation metrics used are global measures and have their limitations. The Jaccard index measures only the degree of overlap between the outlines of the two images, while omitting the rich structural information within the outlines, and hence is not sensitive to shifts of internal structures and intensity distortions. In addition, the manual editing of brain mask required for the Jaccard index calculation might be another source of errors. The cross correlation incorporates intensities of all the voxels within the brain, but its value is very sensitive to outliers due to its strong dependence on the intensity distribution of the images, and hence may not be a robust metric, especially for EPI images with intensity distortions, artifacts and/or low SNR. Mutual information, a histogram-based measure of the statistical dependence between images, could be more robust than intensity-based correlation metric against image degradation, but it provides no direct information about structural correspondence between images. Despite the differences in sensitivity and robustness, all evaluation metrics used in this study suggest the same best method, the 3D reversed phase encoding correction, which is in agreement with its superior performance in terms of both geometry and intensity compensation as seen by qualitative visual inspection.
AC
CE
PT
ED
Image displacement by motion and field drifting during the scan is another factor that may affect the performance of correction. Although the mouse brain was secured in the stereotaxic holder under anesthesia, slight movement and vibration could not be completely eliminated. In addition, field drifting due to the heating of the gradient and shim coils could lead to EPI image displacement in the phase encoding direction. The translational displacement in our EPI data was 0.45±0.07mm, which was largely due to field drifting over the long scanning period. In our processing it was assumed that these two effects were independent of regional field inhomogeneity, so the image displacement was corrected using a motion correction method. Since the field inhomogeneity is caused by the tissue susceptibility differences of the subject within the scanner, subject movements will add a dynamic component to the magnetic field variation, resulting in interaction between movements and susceptibility distortions [39]. The field drifting, while mostly global, may also introduce time-dependent regional inhomogeneity. As each TI image was corrected by its own reversed phase image acquired immediately after the forward phase encoding, such displacement was only seen between different TIs and can be accounted for by registration. Although the forward-reversed phase encoding pair was acquired at PLD=0ms and hence only one deformation field was applied to the entire pCASL dataset, the undistortion was comparable to that of the T1 mapping data. Since brain EPI, such as fMRI, ASL, and diffusion imaging, includes multiple repetitions lasting for minutes, inserting multiple reversed phase encoding EPI volumes throughout the entire acquisition would be desirable to minimize movement dependent field change and field drifting. Another variant implementation is to split the required EPI repetitions into multiple blocks with alternating phase-encoding directions. Both ways could achieve not only dynamic distortion correction, but also higher time efficiency and higher SNR by combining both corrected datasets. Furthermore, field drifting could be measured and compensated for by multiple field mapping sequences throughout the scan [40] or by navigator echo [41-42], which would further improve the accuracy of undistortion.
10
ACCEPTED MANUSCRIPT
SC
RI P
T
In this study, both T1 and CBF mapping were acquired by SE-EPI. While the distortion correction methods performed well, whether it will work for gradient echo EPI (GE-EPI) is still not clear. GE-EPI is the major method used for fMRI and resting-state fMRI. The static field inhomogeneity and susceptibility induced distortions are even worse due to additional through-plane signal dropouts in GE-EPI. Furthermore, shifts of effective TE in GE-EPI may lead to additional intensity distortions on top of the intensity changes due to voxel compression/stretching. It has been demonstrated in the human brain at 3T that the correction field created by reversed phase encoding SE-EPI pairs could be applied to GE-EPI images [20]. In small animal imaging at high fields, however, there are large regional distortions and signal dropouts. Considering the emerging application of resting-state fMRI in mouse models [43], further investigation will be needed to determine the optimal approach for GE-EPI.
AC
CE
PT
ED
MA
NU
Not all distortion correction methods available were evaluated in this study. Point spread function (PSF) of each voxel, which is measured by incorporating a varying gradient pre-winder in the phase encoding in repeated EPI acquisitions, can provide information of the spatial and intensity shift. PSF based correction has been demonstrated successful in human images of up to 7T field strength [44-47]. Compared to field mapping, this approach is computationally more efficient, and provides more reliable and robust correction, especially in regions with high field inhomogeneity. Furthermore, since the PSF is measured in the same EPI readout timing, it can also correct for distortions induced by eddy currents and concomitant gradients. A recent study of extended PSF method by combining with reverse phase correction was demonstrated successful in human data [48]. It would be interesting to evaluate the PSF method on mouse imaging at high fields. Recently, a hybrid approach combining geometric distortion correction by conventional field mapping and intensity correction by reversed phase encoding method was proposed [49]. This method is straightforward and demonstrated to be more effective than the field map-based method in human SE-EPI data. However, given the unsatisfactory performance in geometric undistortion using the field map-based method in our mouse data, it is unlikely that the approach would be as effective.
5. Conclusions
In summary, we evaluated several distortion correction techniques for quantitative T1 and CBF imaging of the mouse brain using SE-EPI at 7T. Our results suggest that the 3D reversed phase encoding method provides the most robust performance, and improves the estimation accuracy of relaxation time and perfusion. The optimized field map-based correction, which combines affine registration with the conventional field map-based correction, also improves the image quality and quantification to a lesser extent. These methods would be helpful in SE-EPI based acquisition of perfusion, diffusion, fMRI and in translational studies at high field strengths.
Acknowledgement This study was supported by the Intramural Research program of the Singapore Bioimaging Consortium, Biomedical Sciences Institutes, Agency for Science, Technology and Research (A*STAR), Singapore. We thank Dr. Danny JJ Wang, UCLA, for providing the pCASL pulse sequence.
11
ACCEPTED MANUSCRIPT Figure legends
NU
SC
RI P
T
Figure 1. The original SE-EPI images (left-most column) and the images corrected by five methods (from left to right) at different slice positions of a mouse, overlaid with the brain contour of the gold standard FSE image (depicted in red). Geometric and intensity distortions are highlighted in color arrows: stretching in yellow, compression with signal loss in blue, compression with elevated intensity in orange and purple. Red arrows highlight the intensity distortion uncorrected by the field map-based method. Green arrows depict the shearing artifacts introduced by the field map correction. White arrows show the discontinuity outside of the brain in 1D Reversed phase encoding-based correction. Nonlinear Reg.: nonlinear registration; Fieldmap: conventional field map-based correction; Opt. Fieldmap: optimized field map-based correction; 1D R. Phase: 1D reversed phase encoding correction; 3D R. Phase: 3D reversed phase encoding correction.
ED
MA
Figure 2. The extent of geometric distortion and performance of each correction method evaluated by the Jaccard index at each slice position of all mice. The first and the last slice were not included due to partial data cutoff by registration.
AC
CE
PT
Figure 3. The box & whiskers plot comparing the similarity, as evaluated by a) Jaccard index, b) cross correlation, and c) mutual information, between the gold standard and the EPI images before and after correction. The box represents the 25th to 75th percentiles and the line in the box represents the median. The whiskers and outliers are plotted according to the Turkey method. Significance was determined by the Fieldman test with correction of family wise error rate. *: p
S0730-725X(15)00155-1 doi: 10.1016/j.mri.2015.06.010 MRI 8375
To appear in:
Magnetic Resonance Imaging
Received date: Accepted date:
12 May 2015 20 June 2015
Please cite this article as: Hong Xin, To Xuan Vinh, Teh Irvin, Soh Jian Rui, Chuang Kai-Hsiang, Evaluation of EPI distortion correction methods for quantitative MRI of the brain at high magnetic field, Magnetic Resonance Imaging (2015), doi: 10.1016/j.mri.2015.06.010
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
T
Evaluation of EPI distortion correction methods for quantitative MRI of the brain at high magnetic field
RI P
Xin Honga, Xuan Vinh Toa, Irvin Tehb,1, Jian Rui Soha, and Kai-Hsiang Chuanga,b,c * a
ED
MA
NU
SC
Magnetic Resonance Imaging Group, Singapore Bioimaging Consortium Agency for Science Technology and Research 11 Biopolis Way, #01-02 Helios Building Singapore 138667 b Clinical Imaging Research Centre, National University of Singapore 14 Medical Drive, #B1-01 Singapore 117599 c Department of Physiology, Yong Loo Lin School of Medicine National University of Singapore Block MD9, 2 Medical Drive #04-01 Singapore 117597
AC
CE
PT
* corresponding author Kai-Hsiang Chuang, PhD Singapore Bioimaging Consortium, 11 Biopolis Way, #02-02, Singapore 138667. Tel +65 64788764; Fax +65 64789957; email: [email protected]
1
Current affiliation: Department of Cardiovascular Medicine, Radcliffe Department of Medicine, University of Oxford. Address: Wellcome Trust Centre for Human Genetics, Roosevelt Drive, Oxford, OX3 7BN, United Kingdom
1
ACCEPTED MANUSCRIPT Abstract
MA
NU
SC
RI P
T
High field MRI has been applied to high-resolution structural and functional imaging of the brain. Echo Planar Imaging (EPI) is an ultrafast acquisition technique widely used in diffusion imaging, functional MRI and perfusion imaging. However, it suffers from geometric and intensity distortions caused by static magnetic field inhomogeneity, which is worse at higher field strengths. Such susceptibility artifacts are particularly severe in relation to the small size of the mouse brain. In this study we compared different distortion correction methods, including nonlinear registration, field map-based, and reversed phase encoding-based approaches, on quantitative imaging of T1 and perfusion in the mouse brain acquired by spin-echo EPI with inversion recovery and pseudo-continuous arterial spin labeling, respectively, at 7T. Our results showed that the 3D reversed phase encoding correction outperformed other methods in terms of geometric fidelity, and that conventional field map-based correction could be improved by combination with affine transformation to reduce the bias in the field map. Both methods improved quantification with smaller fitting error and regional variation. These approaches offer robust correction of EPI distortions at high field strengths and hence could lead to more accurate co-registration and quantification of imaging biomarkers in both clinical and preclinical applications.
ED
Keywords: field inhomogeneity, distortion correction, susceptibility artifact, transgenic mouse,
AC
CE
PT
cerebral blood flow, high magnetic field
2
ACCEPTED MANUSCRIPT 1. Introduction
ED
MA
NU
SC
RI P
T
High field (>= 7T) MRI has become widely adopted for high-resolution structural and functional imaging of the brain [1-4]. However, image artifacts, such as susceptibility artifact, are the limiting factors that compromise the signal-to-noise ratio (SNR) benefit of high fields. Among the image acquisition pulse sequences, Echo Planar Imaging (EPI) is an ultrafast sequence widely applied to functional MRI (fMRI), diffusion imaging and perfusion imaging. It suffers from geometric and intensity distortions as well as signal dropouts that are caused by static magnetic field inhomogeneity [5-6], particularly in the phase encoding direction due to the relatively low pixel bandwidth or in the through plane direction due to the typically lower resolution than inplane. The resultant global and regional distortions are typically seen near the air-tissue or bone-tissue interfaces, where magnetic susceptibility changes rapidly. The issue is especially severe when translating similar methods to transgenic mouse models to understand disease mechanisms and drug effects. Firstly, the mouse brain is much smaller than the human brain. The same spatial extent of spin displacement affects a much larger portion of the mouse brain. Secondly, field inhomogeneity and hence distortions are proportional to the magnetic field strength. With rodent imaging typically performed at higher magnetic field strengths than human imaging, the artifacts worsen. The geometric distortion will result in localization errors, difficulty in image registration and hence group analysis. The intensity distortion in EPI could also bias the quantitative measures such as relaxivity, diffusivity and perfusion, since quantification of these metrics usually involves nonlinear fitting of the EPI series.
AC
CE
PT
One could reduce the signal dropouts commonly seen in gradient echo imaging by incorporating slice shimming, also known as the z-shim technique, into the pulse sequence [7-9]. However this will prolong the image acquisition time depending on how many z-shim steps used. Spin-echo based acquisition, which is less sensitive to field inhomogeneity, could help to reduce the distortion. However, even with spin-echo EPI (SE-EPI), the geometric and intensity distortions are still significant at high field strengths. The artifact is also echo time (TE) dependent; however, ultrashort TE is difficult to achieve with EPI. Another way is to reduce the echo train length by increasing the bandwidth or using partial Fourier, at the cost of lower SNR and/or spatial resolution. Echo train length can also be reduced by multi-shot EPI, but the acquisition time will be longer and the phase error between shots would introduce additional artifacts. Applying localized susceptibility matching materials near the air/bone-tissue interfaces can reduce the regional magnetic inhomogeneity, though it will be difficult for regions away from the skin [10]. Alternatively, postprocessing methods can be applied to reduce the distortions. Several EPI distortion correction algorithms have been proposed and demonstrated successfully in human studies at lower fields. The most common way is based on magnetic field mapping. A field map, measured by the phase differences of two gradient echo images acquired with different TEs, describes the spatial variation of the magnetic field, from which a voxel displacement map can be calculated and used to unwarp the distorted EPI images [11-12]. The main drawback of this technique is the separate acquisition of the multiecho images. Subject movement between the field mapping and EPI acquisition could render the field map inaccurate. In addition, the phase unwrapping process is sensitive to noise; therefore the field map in regions of low SNR is unreliable. Alternatively, one can acquire a phase
3
ACCEPTED MANUSCRIPT encoded multi-reference scan to estimate the magnitude and phase errors due to the field inhomogeneity, and then use that to correct the EPI artifact [13].
SC
RI P
T
Geometric distortions can also be corrected for by nonlinear registration to an anatomically correct image, typically acquired by fast spin echo (FSE). The deformation field could be calculated by minimizing the least squared differences of intensity [14] or log-intensity [15] between the distorted EPI and anatomical images. The undistortion performance highly depends on the registration algorithm, implementation, and parameter optimization. Besides, it only corrects for geometric distortion but not the intensity distortion, which may still result in quantification error.
CE
PT
ED
MA
NU
Another approach is to acquire additional EPI data with reversal of the phase encoding direction [16-18]. The basic idea is that phase encoding gradients with opposite polarities will produce opposite spatial and intensity distortions in the phase encoding direction. Therefore an anatomically correct image should be the one with minimal difference between the pair of distorted images. Once the displacement for each voxel is determined the intensity can be corrected accordingly. One advantage of this approach is the short imaging time, which is on the scale of seconds for a full EPI volume compared to minutes for a conventional field map acquisition, and hence makes it less sensitive to subject motion. However, since the algorithm is unwarping on a single dimension, the resulted image may show discontinuities in the other two dimensions. An improved implementation is to generate a continuous and smooth 3D displacement field [19]. Further improvement uses computationally more efficient method, achieving subvoxel resolution of the distortion map [20]; however, the nonlinear optimization requires empirically determined regularization parameters. Distortion correction by the reversed phase encoding approach has been demonstrated in diffusion imaging [21-23], fMRI [23] and dynamic susceptibility contrast MRI [24] of the human brain at 3T.
AC
Whether the abovementioned postprocessing approaches can effectively correct for the distortions at high fields and how they affect quantitative imaging like arterial spin labeling are not clear. In this study we compared the effectiveness of different distortion correction methods on SE-EPI of the mouse brain acquired at 7T. Their influence on quantitative T1 and cerebral blood flow (CBF), measured using inversion recovery and pseudo-continuous arterial spin labeling (pCASL) respectively, were assessed.
2. Material and Method Animal preparation The animal study was approved by the Institutional Animal Care and Use Committee (Biomedical Sciences Institutes, Singapore). Eight mice of FVB (n=4, male) and C57BL/6 (n=4, male) backgrounds, aged between 6 and 12 months were used. They were anesthetized and maintained by 1-2% isoflurane in air and O2 (1:1) at 1L/min flow rate via a nose cone, with respiration rate controlled at 100±20 breaths per minute. The head was secured on an MRI-compatible stereotaxic holder. The body temperature and respiration were monitored by an MRI-compatible monitoring system (Small Animal Instruments, Inc, Stony Brook, NY, USA) with temperature maintained at 36±0.5oC using heated air. 2.1. Image acquisition 4
ACCEPTED MANUSCRIPT
ED
MA
NU
SC
RI P
T
MRI was acquired on a 7T scanner (ClinScan, Bruker BioSpin, Ettlingen, Germany) using a 72mm volume transmit coil and a mouse brain 4-channel receive array (Bruker). The brain was carefully positioned at the center of the magnet. Volume selective 2nd-order 3D shimming was carried out to optimize the field homogeneity of the whole brain. T2-weighted (T2w) anatomical images were acquired with a 2D multislice FSE sequence with TR/TE=2760/43ms, 0.1x0.1x0.3mm3 voxel resolution, average=2, number of slices=51, and coil inhomogeneity normalization. T1 mapping was obtained by inversion recovery singleshot SE-EPI with 7 inversion times (TIs) ranging from 10 to 8000 ms, TR/TE=10000/20ms, bandwidth=3552Hz/pixel, average=2, FOV=18x13.5mm, acquisition matrix=64x48, slice thickness=1mm, gap=0mm, and number of slices=12. The images were reconstructed with zero-filling to a matrix of 128x96, resulting in apparent resolution of 0.14x0.14mm2. Perfusion imaging was acquired by pCASL [25-26] with labeling offset=12mm, labeling duration=1600ms, 10 post-labeling delays (PLDs) ranging from 0 to 400ms and SE-EPI with acquisition parameters the same as the T1 mapping sequence except TR=4000ms and 30 measurements, i.e. 15 control and label pairs. Reversed phase encoding SE-EPI was acquired at PLD=0ms for pCASL and each TI of the T1 mapping. The field map was measured by a dual gradient echo sequence with TE=1.8ms and 4.85ms, TR=213ms, flip angle=20 degrees, acquisition matrix=128x96 without interpolation, and the same geometry and reconstructed voxel size as the SE-EPI. The acquisition time, including both the forward and reversed phase EPI and dummy scans, was about 8 min for the T1 mapping and 24 min for the pCASL imaging. 2.2. Distortion correction
CE
PT
The T2w FSE image was rigidly registered to the magnitude image of the field mapping sequence to create a geometrically and anatomically correct gold standard in the EPI space. Five correction methods were applied on the EPI data:
AC
1) nonlinear registration: the EPI image was registered to the gold standard FSE image using FNIRT [27] in FSL (FMRIB Software Library v5.0, University of Oxford, Oxford, UK, http://fsl.fmrib.ox.ac.uk/) with sum of squared difference as the cost function; 2) field map-based correction: the EPI was corrected by FSL FUGUE [12, 27] using a field map created by the phase images of the dual echo sequence followed by unwrapping by FSL PRELUDE; 3) optimized field map-based method: to account for the residual shearing after the field map-based correction, the undistorted EPI image corrected as above was further registered to the gold standard image using 3D affine transformation with correlation ratio as the cost function; 4) 1D reversed phase encoding correction: in-house Matlab codes (MathWorks, Natick, MA, USA) implemented based on the algorithm of Chang et al. [16] was applied on brain-extracted EPI data; and 5) 3D reversed phase encoding correction: FSL TOPUP [19, 28], which generated 3D distortion field on EPI data without brain extraction, was used. Since there were displacements between the field map and the EPI data, the field map-based correction was initially performed on the first and last acquired volumes of the pCASL series, and the one with
5
ACCEPTED MANUSCRIPT
RI P
T
better correction (in terms of higher Jaccard index) was chosen to be the reference volume. Then all the other pCASL and TI series were registered to that reference EPI volume using rigid body transformation. In the reversed phase encoding correction, each TI image was undistorted by its corresponding reversed phase encoding image, and then registered to the first pCASL volume with PLD=0ms. For pCASL images, the distortion field was calculated based on the forward and reversed phase encoding image pair with PLD=0ms, and then applied to all other volumes with different PLDs. 2.3. Data analysis
E S E S i
i
i
i
i
NU
JI E , S
SC
The goodness of geometric correction was assessed quantitatively by an overlap measure, Jaccard index (JI):
i
(1)
PT
ED
MA
where S and E are binary matrices of the brain masks of the gold standard structural image and the EPI image, respectively, and i is the voxel index within each volume. The brain masks of the structural images were created using the automatic 3D-Pulse-coupled Neural Networks (3D-PCNN) algorithm [29] with additional manual editing to correct residual errors. The brain masks of EPI data were drawn manually. In addition to the Jaccard index, the performance of the correction methods was also evaluated by image similarity metrics, including intensity-based cross correlation (CC): 2
Var X Var Y
CE
CC X , Y
Cov X , Y
(2)
and histogram-based mutual information (MI):
AC
MI X , Y H X H Y H X , Y
(3)
where X and Y are the gold standard structural image and the EPI image, respectively, Var(X) is the variance of X, Cov(X,Y) is the covariance of X and Y, H(X) is the Shannon entropy of X, and H(X,Y) is the joint entropy of X and Y. The Fieldman test was performed using Prism (version 5, GraphPad Software, San Diego, CA, USA) to compare the similarity metrics between methods with family-wise error rate controlled. For the T1 mapping and CBF quantification, distortions were corrected for by both the optimized field map-based method and the 3D reversed phase encoding method. The original and undistorted M0 and T1 maps were calculated using nonlinear least square 3-parameter fitting to the inversion recovery signal. The CBF was quantified by nonlinear least square fitting to the multi-PLD data using the following equation [30]:
6
ACCEPTED MANUSCRIPT 0, 0 t t ' CBF M t e t /T1b / 2 M 0 BT1' 1 e t t /T1 , t t t ' M t et /T1b e t t /T1 / 2 M T ' 1 e /T1' , t t 0B 1
RI P
(4)
T
NU
SC
where ∆M(t) is the tissue magnetization difference between the control and label images during the interval t, M0B is the equilibrium magnetization of arterial blood assumed to be the same as the tissue M0 obtained from the T1 mapping, α is the labeling efficiency assumed to be 0.9, λ is the tissue/blood partition coefficient of water assumed to be 0.9ml/g [31], ∆t is arterial transit time to be fitted, τ is the labeling duration (1600ms in our sequence), T'1 is the apparent relaxation time measured by the T1 mapping, T1b is the longitudinal relaxation time of arterial blood assumed to be 2300ms [32]. The goodness of fitting for CBF quantification was evaluated by the coefficient of determination, R2.
CV / .
(5)
AC
CE
PT
ED
MA
The influence of undistortion on quantification was evaluated in five regions of interest (ROIs) with minimal to severe distortions. The ROIs were manually delineated on the original EPI image and in the corresponding locations on the undistorted EPI images. The ROIs from minimally distorted regions included the right ectorhinal cortex (Ect) and the right thalamus (Thal); from substantially compressed regions in the cortex like the right primary somatosensory cortex barrel field (S1BF); from moderately stretched regions like the left amygdala (Amyg); and from severely compressed regions like the right entorhinal cortex (Ent). The mean M0, T1, CBF and R2 values within each ROI were compared between the original and undistorted maps. The R2 of the corrected images were subtracted by and then normalized to the R2 of the original image. To evaluate the variation within each ROI, coefficient of variance (CV) was calculated as the ratio of the standard deviation over the mean:
Pair-wise comparisons were performed by two sample t-tests with the significance level of 0.05 adjusted by the Bonferroni method.
Results Fig.1 shows a typical single-shot spin-echo EPI of a mouse brain acquired at 7T and the undistorted images by the five correction methods. The original SE-EPI showed mild compression in cortical areas (purple arrows), moderate stretching in the amygdala (yellow arrows), and severe compression in the entorhinal cortex (blue and orange arrows). Nonlinear registration and both 1D and 3D reversed phase encoding corrections were able to correct for the cortical distortion. However, nonlinear registration failed to correct for intensity distortion due to voxel compression (red arrows). Field map-based method compensated distortions partly, while in-plane shearing in the corrected images was seen consistently (green arrows). An additional affine registration to the reference image improved the overall correction of the field map-based method. Based on our preliminary results, 1D reversed phase correction performs better in brain extracted EPI data than non-extracted ones. Discontinuity at the brain
7
ACCEPTED MANUSCRIPT
RI P
T
boundary in the phase-encoding (dorsal-ventral) direction in the images corrected by the 1D reversed phase encoding method could be noticed (white arrows), which was caused by optimizing displacement field within the brain by removing the non-brain tissues, but that did not affect the following quantitative evaluation. The 3D reversed phase encoding method, without brain extraction, was able to correct for both geometric and intensity distortions even in regions near the ear canals where other methods failed.
MA
NU
SC
Since the distortion is region-dependent, the distortion and effectiveness of each correction method were evaluated by the Jaccard index at each slice position (Fig. 2). Large distortion could be seen in posterior (caudal) brain, which may be due to susceptibility changes near ear canals and poorer shimming away from the iso-center. Less distortion was seen in the middle of cerebrum but worsened towards the nose. Even though distortion correction methods could improve the quality, the improvement depended on the original image quality so that the Jaccard indexes after correction still followed similar slice-dependent trend. Both reversed phase encoding methods provided the most effective and robust performance across most of the slice locations, followed by the optimized field map-based method.
PT
ED
The mean Jaccard index, cross correlation and mutual information of the whole brain were compared to evaluate the performance of the correction methods (Fig. 3). Nonlinear registration achieved significant undistortion only in terms of MI. Conventional field map-based method showed lowest JI, CC, MI among all, with some improvement when combined with affine registration. Both reversed phase encoding methods were effective with significant increase in all similarity metrics.
AC
CE
Since the optimized field map-based and 3D reversed phase encoding methods were the more effective approaches, their influence on T1 and CBF quantification was compared. Fig.4 shows examples of T1 and CBF maps from one mouse estimated before and after undistortion. It was clear that the 3D reversed phase encoding correction generated higher geometric fidelity and tissue uniformity in the quantitative maps. The improvements in T1, M0 and CBF quantification were compared in various ROIs (Fig. 5). The locations of the five ROIs defined on the original and the undistorted EPI were shown in Fig. 5a and 5b, respectively. Distortions were typically seen in the amygdala and Ent, less in S1BF and minimal in Ect, and Thalamus. Therefore the latter 2 ROIs were used as reference areas not affected by distortion. In the Ect, where no distortion was seen, no differences in T1, M0, CBF, and R2 were detected after correction, nor between the two correction methods (Fig. 5c, e, g, i). In the S1BF, where small distortion was observed, the only difference seen was in the R2 where the 3D reversed phase encoding correction achieved better fitting than the optimized field map correction. In the amygdala, where large distortion was easily visible, the 3D reversed phase encoding correction resulted in higher M0 and R2, while the optimized field map correction produced lower T1 and higher CBF than the original. In the Ent, where the distortion was most significant, the 3D reversed phase correction showed improved T1 and CBF toward the typical gray matter values seen in Ect and S1BF, while both methods had better model fitting with more than 100% increase in R2. In the thalamus, T1 and M0 were higher after the field map correction, while no difference was found in R2 after either correction. Although the T1 and CBF may not
8
ACCEPTED MANUSCRIPT
RI P
T
necessarily change after correction (except in highly distorted regions like the Ent), the increased R2 in most regions indicated that the quality of curve fitting was improved. The significantly reduced coefficient of variance in all four quantification values in the Ent further confirmed improved quantification in severely distorted regions.
4. Discussion
AC
CE
PT
ED
MA
NU
SC
Geometric and intensity distortions as well as signal dropouts have hampered the application of EPI for structural and functional imaging at high fields. Compared to other EPI distortion correction methods tested, the 3D reversed phase encoding method provides the most significant improvements in both geometry and intensity accuracy. Its performance at the boundary is superior to that of the 1D reversed phase encoding method due to the better estimate of the 3D displacement profile. Undistortion by nonlinear registration does not require additional data other than the anatomical image, but the results highly depend on the fine-tuning of the registration parameters, which is subjective and time-consuming. In this study the registration parameters, including the cost function, masking, sub-sampling levels, iteration number, regularization, etc, were optimized based on our pilot study. Besides, nonlinear registration does not correct for intensity distortion, which is obvious in areas of voxel compression near the ear canals. We observed that the widely used field map-based correction did not perform well. This may be due to errors in phase unwrapping, which is especially challenging in areas with large susceptibility differences and hence rapid field changes, e.g. at the boundary of brain tissues and the airfilled ear canals. Also, residual phase error due to asymmetric bipolar readout in the dual echo sequence may contribute to errors in the field map. Zero and first order phase shift in the dual echo sequence may be modeled, estimated, and hence corrected for, but it requires collection of additional phantom data [33]. Our results suggest that a simple additional step of affine registration could improve the field mapbased correction. The T1 values estimated based on the 3D reversed phase encoding corrected images are in good agreement to those reported in the previous study at 7T [34]. The regional CBF values estimated here are comparable to those reported by Kober et al. [35], while higher than other studies [36-38]. Absolute CBF quantification is highly affected by various physiological and methodological factors, including the animal's strain, age, physiological state, anesthesia, labeling technique, readout method, and assumptions in the quantification. Direct comparison of the absolute CBF value between studies may not be conclusive without careful controlling of these factors. In this study we used areas with minimal distortion as the reference. The quantified CBF after undistortion is consistent with these reference areas. It should be noted that there is no absolute gold standard against which the performance of undistortion is evaluated. However, given the high anatomical fidelity of the T2-weighted image and high geometric fidelity of the magnitude image of the multi-echo gradient echo data, the reference generated by a rigid registration of the former to the later gives a reasonable gold standard with the same geometric settings as the EPI images.
9
ACCEPTED MANUSCRIPT
MA
NU
SC
RI P
T
Most of the previous works on EPI distortion correction only evaluated the performance by visual inspection. In this study, quantitative assessment of the correction was determined regionally in each slice location and globally in the whole brain. However, the evaluation metrics used are global measures and have their limitations. The Jaccard index measures only the degree of overlap between the outlines of the two images, while omitting the rich structural information within the outlines, and hence is not sensitive to shifts of internal structures and intensity distortions. In addition, the manual editing of brain mask required for the Jaccard index calculation might be another source of errors. The cross correlation incorporates intensities of all the voxels within the brain, but its value is very sensitive to outliers due to its strong dependence on the intensity distribution of the images, and hence may not be a robust metric, especially for EPI images with intensity distortions, artifacts and/or low SNR. Mutual information, a histogram-based measure of the statistical dependence between images, could be more robust than intensity-based correlation metric against image degradation, but it provides no direct information about structural correspondence between images. Despite the differences in sensitivity and robustness, all evaluation metrics used in this study suggest the same best method, the 3D reversed phase encoding correction, which is in agreement with its superior performance in terms of both geometry and intensity compensation as seen by qualitative visual inspection.
AC
CE
PT
ED
Image displacement by motion and field drifting during the scan is another factor that may affect the performance of correction. Although the mouse brain was secured in the stereotaxic holder under anesthesia, slight movement and vibration could not be completely eliminated. In addition, field drifting due to the heating of the gradient and shim coils could lead to EPI image displacement in the phase encoding direction. The translational displacement in our EPI data was 0.45±0.07mm, which was largely due to field drifting over the long scanning period. In our processing it was assumed that these two effects were independent of regional field inhomogeneity, so the image displacement was corrected using a motion correction method. Since the field inhomogeneity is caused by the tissue susceptibility differences of the subject within the scanner, subject movements will add a dynamic component to the magnetic field variation, resulting in interaction between movements and susceptibility distortions [39]. The field drifting, while mostly global, may also introduce time-dependent regional inhomogeneity. As each TI image was corrected by its own reversed phase image acquired immediately after the forward phase encoding, such displacement was only seen between different TIs and can be accounted for by registration. Although the forward-reversed phase encoding pair was acquired at PLD=0ms and hence only one deformation field was applied to the entire pCASL dataset, the undistortion was comparable to that of the T1 mapping data. Since brain EPI, such as fMRI, ASL, and diffusion imaging, includes multiple repetitions lasting for minutes, inserting multiple reversed phase encoding EPI volumes throughout the entire acquisition would be desirable to minimize movement dependent field change and field drifting. Another variant implementation is to split the required EPI repetitions into multiple blocks with alternating phase-encoding directions. Both ways could achieve not only dynamic distortion correction, but also higher time efficiency and higher SNR by combining both corrected datasets. Furthermore, field drifting could be measured and compensated for by multiple field mapping sequences throughout the scan [40] or by navigator echo [41-42], which would further improve the accuracy of undistortion.
10
ACCEPTED MANUSCRIPT
SC
RI P
T
In this study, both T1 and CBF mapping were acquired by SE-EPI. While the distortion correction methods performed well, whether it will work for gradient echo EPI (GE-EPI) is still not clear. GE-EPI is the major method used for fMRI and resting-state fMRI. The static field inhomogeneity and susceptibility induced distortions are even worse due to additional through-plane signal dropouts in GE-EPI. Furthermore, shifts of effective TE in GE-EPI may lead to additional intensity distortions on top of the intensity changes due to voxel compression/stretching. It has been demonstrated in the human brain at 3T that the correction field created by reversed phase encoding SE-EPI pairs could be applied to GE-EPI images [20]. In small animal imaging at high fields, however, there are large regional distortions and signal dropouts. Considering the emerging application of resting-state fMRI in mouse models [43], further investigation will be needed to determine the optimal approach for GE-EPI.
AC
CE
PT
ED
MA
NU
Not all distortion correction methods available were evaluated in this study. Point spread function (PSF) of each voxel, which is measured by incorporating a varying gradient pre-winder in the phase encoding in repeated EPI acquisitions, can provide information of the spatial and intensity shift. PSF based correction has been demonstrated successful in human images of up to 7T field strength [44-47]. Compared to field mapping, this approach is computationally more efficient, and provides more reliable and robust correction, especially in regions with high field inhomogeneity. Furthermore, since the PSF is measured in the same EPI readout timing, it can also correct for distortions induced by eddy currents and concomitant gradients. A recent study of extended PSF method by combining with reverse phase correction was demonstrated successful in human data [48]. It would be interesting to evaluate the PSF method on mouse imaging at high fields. Recently, a hybrid approach combining geometric distortion correction by conventional field mapping and intensity correction by reversed phase encoding method was proposed [49]. This method is straightforward and demonstrated to be more effective than the field map-based method in human SE-EPI data. However, given the unsatisfactory performance in geometric undistortion using the field map-based method in our mouse data, it is unlikely that the approach would be as effective.
5. Conclusions
In summary, we evaluated several distortion correction techniques for quantitative T1 and CBF imaging of the mouse brain using SE-EPI at 7T. Our results suggest that the 3D reversed phase encoding method provides the most robust performance, and improves the estimation accuracy of relaxation time and perfusion. The optimized field map-based correction, which combines affine registration with the conventional field map-based correction, also improves the image quality and quantification to a lesser extent. These methods would be helpful in SE-EPI based acquisition of perfusion, diffusion, fMRI and in translational studies at high field strengths.
Acknowledgement This study was supported by the Intramural Research program of the Singapore Bioimaging Consortium, Biomedical Sciences Institutes, Agency for Science, Technology and Research (A*STAR), Singapore. We thank Dr. Danny JJ Wang, UCLA, for providing the pCASL pulse sequence.
11
ACCEPTED MANUSCRIPT Figure legends
NU
SC
RI P
T
Figure 1. The original SE-EPI images (left-most column) and the images corrected by five methods (from left to right) at different slice positions of a mouse, overlaid with the brain contour of the gold standard FSE image (depicted in red). Geometric and intensity distortions are highlighted in color arrows: stretching in yellow, compression with signal loss in blue, compression with elevated intensity in orange and purple. Red arrows highlight the intensity distortion uncorrected by the field map-based method. Green arrows depict the shearing artifacts introduced by the field map correction. White arrows show the discontinuity outside of the brain in 1D Reversed phase encoding-based correction. Nonlinear Reg.: nonlinear registration; Fieldmap: conventional field map-based correction; Opt. Fieldmap: optimized field map-based correction; 1D R. Phase: 1D reversed phase encoding correction; 3D R. Phase: 3D reversed phase encoding correction.
ED
MA
Figure 2. The extent of geometric distortion and performance of each correction method evaluated by the Jaccard index at each slice position of all mice. The first and the last slice were not included due to partial data cutoff by registration.
AC
CE
PT
Figure 3. The box & whiskers plot comparing the similarity, as evaluated by a) Jaccard index, b) cross correlation, and c) mutual information, between the gold standard and the EPI images before and after correction. The box represents the 25th to 75th percentiles and the line in the box represents the median. The whiskers and outliers are plotted according to the Turkey method. Significance was determined by the Fieldman test with correction of family wise error rate. *: p
