MP2RAGEME: T1, T2 *, and QSM mapping in one sequence at 7 tesla

Abstract Quantitative magnetic resonance imaging generates images of meaningful physical or chemical variables measured in physical units that allow quantitative comparisons between tissue regions and among subjects scanned at the same or different sites. Here, we show that we can acquire quantitative T1, T2 *, and quantitative susceptibility mapping (QSM) information in a single acquisition, using a multi‐echo (ME) extension of the second gradient‐echo image of the MP2RAGE sequence. This combination is called MP2RAGE ME, or MP2RAGEME. The simultaneous acquisition results in large time savings, perfectly coregistered data, and minimal image quality differences compared to separately acquired data. Following a correction for residual transmit B1 +‐sensitivity, quantitative T1, T2 *, and QSM values were in excellent agreement with those obtained from separately acquired, also B1 +‐corrected, MP2RAGE data and ME gradient echo data. The quantitative values from reference regions of interests were also in very good correspondence with literature values. From the MP2RAGEME data, we further derived a multiparametric cortical parcellation, as well as a combined arterial and venous map. In sum, our MP2RAGEME sequence has the benefit in large time savings, perfectly coregistered data and minor image quality differences.

additional B 1 -map may be obtained to correct for transmit field (B 1 + )inhomogeneities, such as by means of a dual refocusing echo acquisition mode in the DREAM-sequence (Nehrke & Börnert, 2012). Other work performed a segmentation-based correction, alleviating the need of a separate B 1 -map . Acquiring multiple multiecho gradient echo (ME-GRE) readouts enabled quantitative multiparameter mapping of R 1 and R 2 * (Weiskopf et al., 2013). These parameters could be robustly estimated over multiple sites. In order to reduce the sensitivity to B 1 + -inhomogeneities and alleviate the need to coregister separately acquired volumes, the MP2RAGE-sequence was proposed. In this inversion-recovery sequence, two GRE-readouts follow after optimized inversion times. This allows for high-resolution imaging at high field (Marques et al., 2010;Marques & Gruetter, 2013). The resulting image is free of T 2 * , M 0 (net magnetization or proton density) and B 1 − effects, but a small and protocol-dependent amount of B 1 + contrast remains. This can be removed by acquiring a B 1 + -map in addition to T 1 -weighted data (Marques & Gruetter, 2013) and use it in the parameter estimation step. Although the MP2RAGE sequence is robust and widely used at 7 T, the acquisition is relatively inefficient because of the long TR required for the magnetization to return to equilibrium. Longer readouts, containing more k-space lines, would lead to shorter scan times, but also result in more T 1 -relaxation during the readout and, hence, incur more T 1 -induced blurring in the images, leading to a poorer PSF. Therefore, most acquisitions opt for relatively short readouts of typically one k-space plane, limiting blurring and accepting the acquisition dead time. Optimally exploiting the dead time, as we will propose in this article, will result in a timeefficient sequence.
Faster imaging at high field was achieved by using a multislice echo planar imaging (EPI) readout (Polders, Leemans, Luijten, & Hoogduin, 2012;van der Zwaag et al., 2018;Wright et al., 2008), but EPI imaging comes at the cost of spatial distortions due to the lower readout bandwidth. Where the MP2RAGE sequence is limited to two readouts, the MPnRAGE performs many more radial readouts with view sharing, sampling the relaxation curve over a wider range (Kecskemeti et al., 2015).
T 2 * relaxometry is commonly performed using a FLASH sequence with a readout comprising of ME. QSM is performed by a dipole deconvolution of the magnetic field, obtained from the phase data of this sequence (Wang & Liu, 2015).
Recent work showed that combined T 1 , T 2 * , and QSM mapping is feasible by extending the MP2RAGE sequence to have ME at both readouts (Metere, Kober, Möller, & Schäfer, 2017). This makes efficient use of the dead time in the MP2RAGE sequence, although the ME on the first image following the inversion necessarily leads to a longer readout block and, subsequently, to a nonoptimal inversion time for T 1 -sensitivity as well as a too-short TE for optimal T 2 * contrast (TEs should at least match the expected T 2 * values of tissues).
T 1 , T 2 * , and QSM are thus widely used contrasts at 7 T and mapping these properties quantitatively is more and more sought after.
While indeed many approaches can be used to measure these in reasonable times, challenges arise in terms of inhomogeneities or bias in the estimated quantities and precision of their alignment. Simultaneous acquisition avoids the need for coregistration and subsequent resampling of separately acquired scans and allows time savings while compromises to both maps are minimal.
Here, we present an extension of the MP2RAGE sequence in which ME are acquired on the second inversion. This allows measurement of T 1 , T 2 * , and QSM simultaneously and efficiently, with optimized inversion and longer echo times, at high resolution in maximally 17 min, reducing the sequence dead time to 6%, and show that the measured quantities agree well with expected measures after correction of B 1 + -inhomogeneities.

| METHODS
The MP2RAGE sequence (Marques et al., 2010) was modified so as to acquire ME in the second inversion while maintaining the single-echo acquisition in the first inversion ( Figure 1). Specifically, the longest TE of the second inversion will be designed to be in the range of reported T 2 * values of GM and WM in the brain.
The signal model for the MP2RAGEME sequence for given inver-

| Subjects
Eight healthy volunteers were scanned at 7 T (Philips, Best, NL). The

Multiple Interleaved Scanning Sequences environment provided by
Philips was used to execute the alternation between the two gradient echo blocks. All subjects provided written informed consent and the study was approved by the local Medical Ethical Committee. From two subjects, the data acquisition was not completed due to hardware failure. For the other six subjects (age range 18-38, four males, and two females), the data were further analyzed.

| Sequence
The MP2RAGEME data were compared with separately acquired MP2RAGE and ME-GRE data in the same session. Parameters in common between all three sequences were: field of view: • ME-GRE: TR = 31.4 ms, α = 12 , TE A-D = 3/11.5/20/28.5 ms.

| Image processing and quantification
T 2 * maps were obtained from the ME-GRE and MP2RAGEME data using a single-exponential fit. QSMs were obtained and averaged from TE 2-4 of the ME-GRE and MP2RAGEME data with respect to cerebrospinal fluid (CSF) using STI Suite (Liu, Li, Tong, Yeom, & Kuzminski, 2015). Here, Laplacian-based phase unwrapping was performed (Li, Avram, Wu, Xiao, & Liu, 2014). A brain mask was obtained from the first echo TE 1 magnitude image using SPM8 (Ashburner & Friston, 2005) and eroded by five voxels to remove veins, CSF and regions with low SNR due to susceptibility-induced intravoxel dephasing. The resulting susceptibility maps were normalized to zero with respect to the whole brain average.
To explore the cortical parcellations that can be derived from MP2RAGEME-data, for a single subject, midcortical maps of T 1 , T 2 * , and QSM were z-scored parallel to the cortex, allowing for visual inspection of the relative contrast. For quantitative comparison, average distances between the MP2RAGE and MP2RAGEME-based cortical surfaces over all cortical regions (left and right cerebrum, cerebellum) were computed per subject.
From the MGDM segmentation, median T 1 , T 2 * , and QSM values were obtained from the following regions of interest (ROIs): WM (all), nucleus caudate, putamen, thalamus, cortical gray matter (all). For the red nucleus, substantia nigra and subthalamic nucleus (STN), ROIs were defined by coregistration of an atlas, using the maximum probability labels (Keuken et al., 2017), that were thresholded at 10%. The image intensity distributions were subsequently visualized using histograms averaged over all subjects in MATLAB (The MathWorks, Inc., Natick, MA). To assess reproducibility, correlation plots were made and the Pearson correlation coefficient r 2 was calculated. Bland-Altman plots were generated (Bland & Altman, 1999), and reproducibility coefficients and coefficients of variation were computed, with a Kolmogorov-Smirnov (KS) test on non-Gaussianity of the difference data.

| Simulations
The sensitivity to B 1 + inhomogeneity was simulated by numerically calculating signal intensities using the Bloch equations for both The simulations of the amount of T 1 -induced blurring for both protocols are shown in Figure 3. As the TI 1 of the MP2RAGEME is shorter than the TI 1 of the MP2RAGE sequence, here is more T 1 -evolution expected during the (equally long) readout, and, hence, increased blurring in the slice direction. The simulated slice profiles of the synthetic images show that there is indeed some blurring at the hard boundaries. The blurring is not visible in the complete profiles ( Figure 3a,b), but in the zoomed panels in Figure 3b some effects are seen. The MPRAGE protocol has minimal blurring at the WM/gray matter boundary, and shows moderate effects at the larger T 1 -difference boundary between gray matter and CSF. In the MP2RA-GEME differences in signal intensity of the voxels immediately neighboring the boundary can be observed at both frontiers. For both boundaries and acquisition protocols, the contrast-to-noise ratio (CNR) necessary to observe such blurring is higher than what is typically achieved in 0.6 mm acquisitions. The CNR for GM/WM was measured to be 10 in one data set, while the blurring induced was less than 10% of the difference between these two tissues.
Regarding the contrast between tissue types, for MP2RAGE-ME, the GM/WM contrast is maintained. The GM/CSF contrast, that was relatively high in the MP2RAGE sequence, is reduced by 40% to approximately the level of the GM/WM contrast.
The image quality overall was excellent for the derived T 1 -weighted images, as well as for the quantitative T 1 , T 2 * , and QSM maps. Figure 4 shows example slices of an MP2RAGEME data set at each orientation for all four derived images, including enlarged sections of specific ROIs in the axial plane. Because of the differences in the acquisition parameters, the T 1 -weighted image intensity distributions differ significantly between the MP2RAGE and MP2RAGEME.
Despite the different signal intensity distributions, the T 1 -maps showed good agreement (see also Figure 5).   No non-Gaussian data distributions were observed, since KS p-values were .149, .261, and .090 for T 1 , T 2 * , and QSM, respectively. Repeatability coefficients reflecting the absolute difference's 95% confidence interval were 99 ms, 3.0 ms, and 23 ppb for T 1 , T 2 * , and QSM, the latter being heightened by variable findings in deep brain nuclei, possibly caused by partial voluming or coregistration errors. Also, susceptibility maps are less robust and sensitive to streaking artifacts that can be due to the ill-posed nature of the problem or originating from phase errors in vessels. The CV was low in T 1 (3.3%) and T 2 * (6.6%), and higher in QSM (33%), again because of variable results in the deep nuclei. For T 1 ,  The average cortical boundary distances generated from the MP2RAGE and MP2RAGEME data differed over all subjects on average by 0.1 AE 1.1 mm (WM/GM) and 0.1 AE 1.2 mm (GM/CSF). There is thus negligible bias, that is, much smaller than the voxel size and intersubject variation.  are attributed to small motion leading to displacement between subsequently scanned series. Due to this, certain arteries or veins may or may not be included in the ROI used for intensity projection. Also, locally varying shading patterns may be caused by subject motion.

| DISCUSSION
We presented the MP2RAGEME sequence, which allows simultaneous measurement of T 1 , T 2 * and QSM at high resolution. The sequence's two inversion readouts were flexibly designed to have a single echo and ME, respectively, leading to a time-efficient sequence of 16 min, maintaining a relatively long TE for good T 2 * and QSM contrasts. The reconstructed parameter maps are naturally aligned, allowing for cortical reconstructions and providing parameter values that are in range with reported literature values.
While the optimal parameters for maximum GM/WM and CSF/GM contrast in MP2RAGE can be simulated using the Bloch equations (Marques et al., 2010), the number of parameters to be set is higher in the MP2RAGEME sequence, complicating these simulations. Moreover, as there are multiple contrasts-of-interest, the target of such simulations is also not clear (Metere et al., 2017). Hence, we chose sequence parameters for the MP2RAGEME to deliver: (1) optimal contrast in the T 2 * map and QSM data, (2) optimal GM/WM contrast in the first inversion image, and (3)  Our work builds upon previous work where two identical ME readouts were employed for both inversions (Metere et al., 2017).
Here, we could shorten the first readout to a single echo, and lengthen the longest TE of the second readout to 28.5 ms, as compared to 18.91 ms in (Metere et al., 2017), leading to an improved contrast. Our longest TE is thus in line with GRE sequences used elsewhere, reporting 28.4 ms (Deistung et al., 2013b) and 29.6 ms , respectively. Our larger flip angle of the first inversion (7 instead of 4 ) resulted in a higher SNR, whereas the higher B 1 + -variation was corrected for by including a B 1 + -map (Nehrke, Versluis, Webb, & Börnert, 2014). Also note that our overall scanning time, although not directly comparable, was slightly shorter, that is, 17 compared to the reported 19 min in Metere et al. (2017), highlighting the time efficiency of the proposed MP2RAGE-ME sequence.
TI 1 was chosen to match the TI of a brain-stem specific protocol (670 ms, (Tourdias, Saranathan, Levesque, Su, & Rutt, 2014)), to facilitate midbrain segmentation using multiple-contrast data . Because the second inversion image cannot overlap in time with the first, the minimum possible value for TI 2 became nearly 4 ss. The longer TI 2 reduces somewhat the T 1 -sensitivity of the T 1 -weighted images.
The flip angles in the MP2RAGEME of 7 and 6 were set for optimal T 1 contrast and minimal B 1 + -sensitivity of the T 1 -weighted images but are lower than the Ernst angle of 12 that was used in the ME-GRE. This implies that the MP2RAGEME was acquired with a lower SNR than the ME-GRE, calculated to be 25% lower in WM and 12% lower in GM, respectively. The impact on, for example, manual delineation of deep brain nuclei is thus small.
Both readouts of the MP2RAGE and MP2RAGEME sequences are long relative to the T 1 -relaxation, meaning that there is significant relaxation during the acquisition of especially the first inversion image.
This T 1 -decay leads to a broadening of the point spread function in the slice encoding direction (Deichmann, Hahn, & Haase, 1999), though this is only a small effect for the protocols compared here ( Figure 3).
For the iron-rich nuclei of the subcortex and cerebellum, the exact colocalization of the three contrasts can help delineate more precise boundaries across the different contrasts, for example, for the globus pallidus or the STN. For the vasculature, a precise  Table 1 with the literature values, the only noticeable differences are found for the values reported in the Caudate nucleus, with somewhat longer T 1 values and lower QSM values found here than in previous work.
Other work also noted that the Caudate nucleus and Putamen are difficult to distinguish . This might be due to partial inclusion of the CSF in the neighboring ventricles. The relatively large SD in QSM values in Table 1, both for the MP2RAGEME and ME-GRE data, are also observed in the literature references. Still, the susceptibility values obtained from the MP2RAGEME data were similar to those obtained from the separately acquired ME-GRE data.
Other ROIs were also possibly affected by imperfect segmentation: The red nucleus T 2 * values were shorter than expected from literature (Table 1), possibly due to segmentation errors. Since the WM mask contained subcortical structures, reported QSM values in WM are higher than in the literature. This is also visible in Figure 5f as a small "shoulder" of low-T 2 * values in the histogram.
The MP2RAGEME sequence is relatively time efficient, with a dead time of only 6%. Since only the second inversion was extended to a ME readout, the first inversion time, which defines the T 1 -contrast in the MP2RAGE image, can be chosen optimally and total scanning time is shortened compared to an ME readout of both inversions (Metere et al., 2017).
The use of less B 1 + inhomogeneity sensitive radio frequency pulses for the inversion, such as the FOCI pulse, would improve the extent of brain area with homogeneous contrast . The use of SPINS pulses for the echo trains would also yield more homogeneous images, limiting the variation in FIGURE 7 Comparison of the cortical contrasts. Maps of mid-cortical T 1 , T 2 * , and QSM variations obtained for a single subject (as in Figure 1) with the separate MP2RAGE and GREME or the combined MP2RAGEME approach. Cortical depth was estimated with volumetric layering and values were smoothed along the mid-cortical depth with a 0.64 mm FWHM Gaussian kernel. Surface reconstructions were obtained from the underlying T 1 maps. To compare local patterns all maps were z-scored parallel to the cortex based on median and interquartile range. Locations of primary motor, auditory and visual cortices are indicated in T 1 maps, as well as lower QSM values in the anterior part of the cortex [Color figure can be viewed at wileyonlinelibrary.com] FIGURE 8 Arterial and venous vasculature. Vasculature reconstructed from the separate MP2RAGE and GREME or the MP2RAGEME-sequence data (maximum intensity projections over 20 slices in axial and sagittal directions, colored in red for structures extracted from T 1 maps (arteries) and blue for structures extracted from 1/T 2 * maps (veins). Note the tight interaction of arteries and veins locations, making precise coregistration of the contrast particularly important. Images are for the same single subject as in Figure 1 [Color figure can be viewed at wileyonlinelibrary.com] the ME-GRE readout due to imperfect flip angles (Malik, Keihaninejad, Hammers, & Hajnal, 2012 Subject motion was limited in the experiments performed in this article and did not lead to noticeable image degradation. Only in minimum and maximum intensity projections (Figure 7), locally varying shading patterns can be seen that may be attributed to motion. However, the long, high resolution acquisitions are susceptible to motion artifacts, especially in possible applications to study disease or large populations. Interleaving the acquisition with fat image navigators (Gallichan & Marques, 2017) or using real-time field control (Özbay, Duerst, Wilm, Pruessmann, & Nanz, 2017) would be candidate approaches to correct for motion artifacts either retrospectively or prospectively.
The proposed algorithm is limited by a longer TR compared to the MP2RAGE sequence, which reduced the CNR between CSF and GM ( Figure 3), however, to an acceptable level comparable to the GM/WM contrast. The reduced sequence dead time was accounted for in the signal model, that is, Tc in Equation (A1), and thus does not affect T 1 -quantification. Because of more and longer gradient switching, more heating of the system and consequential resonance frequency drift may occur.
The specific absorption rate (SAR) and a possible raise thereof is of limited concern in MP2RAGE and MP2RAGEME sequences.
Because of the low flip angle readout trains, and the inversion pulses being far apart, SAR levels are low for both sequences. If anything the SAR per unit of time is reduced in the MP2RAGEME due to the increased repetition time. This is in strong contrast to, for example, TSE readouts, where much higher flip angles (>100 ) are repetitively applied or 2D imaging with simultaneous multi-slice excitation.

| CONCLUSION
We show that quantitative T 1 , T 2 * , and QSM information can be acquired in a single acquisition. Furthermore, we show that the resulting quantitative values are comparable to current separately acquired sequences. Our MP2RAGEME sequence has the benefit in large time savings, perfectly coregistered data and minor image quality differences.
ACKNOWLEDGMENT M.W.A.C. is the stock owner of Nico-lab International Ltd.